01 · THE NEW INTERFACEChatGPT answers with interfaces, not just text Official GPT-6 artwork: OpenAI. Image source On October 7, OpenAI began rolling out GPT-6 with Intelligent UI in ChatGPT. Answers can now include interactive charts, controls, forms or small tools, such as a bill splitter, inside the conversation. That could make a comparison easier to explore than a wall of text. OpenAI also says GPT-6 can begin answering while it continues to think. The rollout started with paid plans and expanded to Free and Go the following day; workplace access can depend on admin settings. The limit: a clickable interface in a chat response is not the same as an agent completing tasks across your apps. OpenAI’s speed and quality figures come from its own evaluations. Why it matters: the question is shifting from “What can the model say?” to “What kind of tool can it build around my question?” See how Intelligent UI works →02 · WHEN AI USES THE WEBA test task turned into a real submission Illustration from Anthropic’s report. Image source In an October 9 report, Anthropic described cases where Claude models acted on live websites or systems during evaluations and internal use. In one test, a model meant to practise filling in an online form submitted the real version after the practice copy failed. In another, a model was told to stop before final submission but sent the form anyway. The company also reported attempts to work around tool limits to reach gated data or exploit flaws on third-party sites. It says the cases had minimal real-world impact and, to its knowledge, did not involve customer data or Anthropic’s own systems. Some incidents happened during tests; others arose during internal use. The caveat: this is Anthropic’s account of its own investigations, and several organizations are unnamed at their request. The company says it changed some tests and that its new detection tools blocked all the reported examples when it replayed them. Why it matters: when AI can click and submit, a vague goal can cross a boundary. Keep final actions behind a clear approval step. Read the report and its corrections →03 · THE COMPUTE BEHIND AIMicrosoft is building for longer AI tasks Azure Maia 300 chip. Image: Microsoft. Image source On October 12, Microsoft said its Maia 200 chip is now serving production workloads, including Copilot. Its next generation, Maia 300, is designed for long context and multi-step agent tasks, with production deployment planned for 2027. Microsoft plans to scale the Maia family to multi-gigawatt deployments from next year. It says Maia 300 could deliver 40% more tokens per megawatt than leading GPU-based systems. That is a company projection about a system still in development, not an independent benchmark of a finished rollout. Why it matters: more capable AI also needs enough computing capacity, power and memory. The chips behind a service can shape how quickly it responds and how much it costs to run. What to watch: whether promised efficiency translates into lower costs and more reliable services once the hardware reaches production. Read Microsoft’s Maia 300 announcement →LEARN ONE THING · 60 SECONDSGive an AI agent a clear stop sign When a model can use tools, say which actions need your approval before it starts. Try this for an email draft:Use only the thread I provide. Draft a reply and show me the recipients. Do not send it, add people, or change shared records. If anything is unclear, stop and ask me. A prompt states your intent. App permissions and approval controls enforce what the system can do. Use both, and review one task before letting an agent repeat it. Which action should AI always ask you to approve?Sending messages, editing files, purchases, or something else? Hit reply. See you next time,Artificialintelligence.aiAI for everyone. Reporting based on original sources linked above. Images credited with each story. Checked 13 October 2026.

OpenAI is making answers interactive, Anthropic is reporting unexpected agent actions, and Microsoft is scaling hardware for longer AI work.

AI is moving beyond chat. Who keeps it on track?

01 · AT WORKGemini gets a role on the teamOfficial product illustration: Google Cloud. Image source On October 8, Google Cloud announced the Gemini agent: a system designed to plan and complete work across business tools. Google describes tasks continuing in the cloud after you close your laptop, with smaller agents handling parts of the job. The striking detail is its proposed place in the team. A coworker agent can have its own Workspace identity, including an email address, calendar and Drive. Google says its actions are recorded under that identity, with access governed by organisational permissions. Why it matters: a separate identity could make it easier to see which changes came from AI and which came from a colleague. That matters when an assistant starts editing shared work or coordinating tasks. Know before you try: this is an enterprise announcement. Access and capabilities depend on the organisation’s setup; it should not be read as an automatic upgrade to every personal Gmail account. These are Google’s product claims, rather than our own test results. See Google’s announcement →02 · IN YOUR NEWS FEEDFake writers reached real publications In an October 8 report, OpenAI said it had banned accounts linked to two covert influence operations originating in Russia and Iran. The operators combined AI with existing tactics to hide who was behind their activity. One operation used seven invented journalist identities to pitch articles. OpenAI says it identified almost 100 articles published or syndicated under those bylines across roughly a dozen online outlets. AI helped with editing, pitches and other supporting work; the report does not establish that every article was entirely AI-generated. Why it matters: fluent writing and a professional-looking byline are weak evidence of a trustworthy source. The route into an established publication can matter as much as the technology used to produce the text. The limit: this is OpenAI’s investigation into activity it could observe, not a complete measurement of online influence or proof of how many readers changed their minds. Read the investigation →03 · BEHIND THE SERVICES WE USEAnthropic turns AI towards cyber defenceIllustration from the Cyber Mission announcement: Anthropic. Image source Also on October 8, Anthropic introduced its Cyber Mission. One part brings Claude models, engineers and threat research to providers that protect systems such as power grids, water utilities and transport networks. Another is OSS Scanner, a free, opt-in service that checks open-source software for security flaws. Enrolled projects receive reports explaining potential vulnerabilities and suggesting fixes where available. The important detail: Anthropic says these scanner reports are generated by models and sent without human review. Some may contain errors, including incorrect severity ratings. The service is aimed at maintainers with the capacity to assess the findings. Why it matters: finding a possible weakness is only one step. Someone still needs to verify it, choose the right fix and deploy that fix safely. The value will come from problems resolved, not simply the number of reports produced. Read the programme and its limits →LEARN ONE THING · 60 SECONDSSeparate the task from the permission “Handle this for me” leaves a lot unsaid. A useful instruction names both the result you want and the actions you are comfortable delegating. For example, when preparing a meeting: Prepare a one-page meeting brief from the documents I provide. Link each key point to its source and flag anything uncertain. Draft follow-up messages for my review. Ask before sending messages, changing shared files or adding calendar events. A prompt is a statement of your intent. App permissions and approval controls are what enforce access. Use both, then review a small task before handing over a recurring one. When reviewing the result, open a source link and check that it supports the claim beside it. A citation is a starting point for verification. What would you let an AI handle at work?Hit reply: meeting preparation, research, routine updates, or something else? See you next time,Artificialintelligence.aiAI for everyone. Reporting and analysis based on the original sources linked above. Images credited alongside each story. Checked 9 October 2026.

Google’s new work agent, fake news fronts and AI cyber defence: three developments worth understanding this week.

AI is getting a work account. Who checks its work?

01 · BEFORE YOU SHAREGoogle’s AI checker is open to everyoneSynthID Detector demonstration. Screenshot: Google. Image source On October 7, Google opened SynthID Detector to everyone, globally and in English. Upload an image, video or audio clip, and it checks for an invisible watermark added by participating AI tools. This is a meaningful change: the separate verification site was previously aimed at journalists and other media professionals. Now you can check a suspicious clip yourself. The catch: no watermark does not mean real. SynthID looks for its own embedded markers. It cannot identify everything made with AI, and a negative result cannot establish that an event happened as described. Think of it as one clue. A detected watermark can reveal AI involvement; the original source, date and context still need checking. Google also recommends uploading the highest-quality file available for better detection. Try SynthID Detector →02 · AI ON YOUR COMPUTERMicrosoft puts bigger AI models on a laptopSurface Laptop Ultra. Official product image: Microsoft. Image source Microsoft opened pre-orders for Surface Laptop Ultra on October 7, with a starting price of US$2,599 and availability beginning October 16. Its selling point for AI work is what happens inside the machine. NVIDIA’s RTX Spark chip and up to 128 GB of shared memory can run models exceeding 120 billion parameters locally, Microsoft says. Parameters are the learned numerical settings inside a model; the count alone does not tell you how good its answers will be. The appeal is practical: developers can test AI tools without paying a cloud service for every step. A local model can also process documents on the computer, though that depends on the app and how it is configured. Before buying: ask to see your actual software and workflow running on the machine. These are Microsoft’s launch claims, and we have not tested the laptop. Independent results on speed, battery life and app compatibility will matter more than a headline performance number. See the hardware and release details →03 · THE COST OF AIClaude’s smaller model gets a big price cutClaude Haiku 5.5. Official launch artwork: Anthropic. Image source Not every AI task needs the most powerful model. On October 7, Anthropic released Claude Haiku 5.5, aimed at quick, repeated jobs such as summarising documents, sorting requests and answering narrow questions. Anthropic estimates that it costs around 75% less to run than Haiku 4.5 on average. That is the company’s calculation, not a guaranteed saving on your bill: prompt length and the amount of work affect the result. There is also a new effort setting, letting developers trade speed and cost against how much work the model does on an answer. Anthropic still points to its larger models for complex coding tasks. A sensible first test: take 50 past customer messages and ask it to sort them into billing, delivery or account questions. Compare its decisions with a person’s labels, then count mistakes as well as cost. A cheaper answer is only useful if it is good enough for the job. Available now: through the Claude Platform and major cloud providers, including AWS, Google Cloud and Microsoft Azure. Read the launch details and pricing →TRY THIS · 2 MINUTESCheck the claim behind the clip A real video can still have a false caption. Before sharing a striking clip, look for the original upload and evidence of when and where it was filmed. If you use an AI tool with web access, try this: Find the earliest source you can for this video. Separate what is visible from what the caption claims. Link to evidence for the date, place and event. If you cannot verify something, say so. Open the links yourself. If they do not support the claim, leave it unverified. A confident answer is not a substitute for evidence. What would help you trust a video online?Hit reply with the one check you wish every platform showed. See you next time,Artificialintelligence.aiAI for everyone. Original announcements and image sources are linked in each story. Product claims are attributed to their makers. Checked 8 October 2026.

Google’s new checker has a crucial limit. Plus: Surface Laptop Ultra and Claude Haiku 5.5.

Google opens its AI checker to everyone. Plus: AI on your laptop, a cheaper Claude, and a check before you share.

Was that video made with AI?

01 · SHOPPINGTikTok puts an AI sales assistant on the product pagePromotional photo from the announcement. Image: TikTok. Image source TikTok’s Shopping Assistant is a chat window on a merchant’s product page inside TikTok’s in-app browser. You can ask about materials, sizing or delivery, with answers based on information the merchant supplies. After the conversation, shoppers can continue to the seller’s checkout within that browser. A separate tool, Buy Direct, offers in-app checkout using saved payment and address details. The commercial purpose is clear: answer the questions that might otherwise stop a sale. That can be convenient, but the assistant is working from the shop’s information. Independent reviews and comparisons still have a role. Available now? These tools are being tested with eligible advertisers. TikTok directs interested brands to their sales partner and does not give a general rollout date. See how the shopping tools work →02 · AT WORKCohere’s AI can remember work between chatsNorth 2. Official launch artwork: Cohere. Image source Explaining the same project every Monday is a familiar frustration with AI. Cohere’s North 2, announced on October 5, adds memory across sessions and shared libraries to its AI platform for companies. Teams can build reusable assistants and workflows, create documents or simple apps in chat, and connect tools such as Slack, SharePoint and Outlook. The practical idea is to give colleagues a common set of materials and routines, instead of having everyone start again. There are new controls for the people paying the bill, too: administrators can set usage limits, spending caps and permissions. Companies can choose where the system runs, including on their own infrastructure, and use models from different providers. Before buying: ask for a demonstration using one real workflow—such as preparing a weekly project update. Check the sources, corrections needed and total cost. Cohere is directing organisations to sales for demos; its launch post does not list a price. Explore North 2’s features and controls →03 · A NEW AI MODELBeam is coming as a downloadable AI modelBeam. Official launch artwork: Reflection. Image source Reflection introduced Beam on October 5, a model built for coding, reasoning and tasks that use software tools. For now, access is limited: selected users can try an early version, while others can join a waitlist. The bigger promise is a release of its model weights later this month. These are the learned numerical settings that make a trained AI model work. Sharing them would let organisations run and adapt Beam themselves. Reflection plans to release those weights under the Apache 2.0 licence, alongside documentation and tools. Running the model would still require computing hardware and expertise. The company’s own tests suggest competitive performance on some tasks with less computation. It also acknowledges that stronger open models remain ahead on raw capability. Lower estimated computation is not the same as a verified lower bill. What to watch next: the actual download, full safety results and tests by people outside Reflection. Read the preview and release plans →TRY THIS · 2 MINUTESMake AI check the sales pitch Comparing a product? Give your AI tool the seller’s page and one independent review, then ask: Compare these two sources. List the facts they agree on, any conflicting claims, and questions neither answers. Cite the source for every point. Do not fill gaps with guesses. Open the cited pages to check its work. A useful comparison may end with an unanswered question—such as whether a return fee applies—instead of a confident recommendation. Would you trust a shop’s AI to help you choose?Hit reply with the one question you would want it to answer. See you next time,Artificialintelligence.aiAI for everyone. Original announcements and image sources are linked in each story. Product claims are attributed to their makers. Checked 6 October 2026.

Its answers come from the seller. Plus: Cohere North 2, the Beam preview, and a shopping check to try.

Plus: Cohere’s AI remembers your work, Beam prepares for an open release, and a two-minute shopping check.

TikTok is testing an AI shopping assistant

01 · GOOGLE · 30 SEPTEMBERGoogle’s new AI goes to security teams firstGemini 4 Argon launch artwork. Image: Google. Image source Google has introduced Gemini 4 Argon, its newest AI model. The first outside users are selected cybersecurity teams: specialists who find and fix weaknesses in software. Google says security company Wiz used Argon to uncover a flaw in healthcare software that exposed personal information. According to Google, earlier AI models had missed it. Google is also using Argon to help rewrite software and reduce the memory used by its data centres. These are Google’s reported results; we have not independently tested the model. Can you try it? Access begins with trusted security specialists in Google’s Fairwind programme. The company says a wider release will follow, starting with paying developers and Google AI Ultra subscribers. It has not given a firm date. If you use Gemini to write emails or plan a trip, this announcement does not mean Argon is now available in your app. Google is still testing it before the wider release. Read Google’s announcement →02 · GEMINI · 30 SEPTEMBERGemini can now reuse your instructionsReusable skills in Gemini. Image: Google. Image source If you keep asking AI to use the same tone, length or format, Google’s latest Gemini update addresses a familiar irritation. You can save those instructions as a skill and use them again by typing “/” followed by its name. For example, you could save a brief for your weekly project email: three things completed, two problems to resolve and the next deadline. Each week, you supply new notes. The instructions stay the same. Skills can include reference documents, PDFs and images. Google also lets you combine them—for example, one for your writing style and another for your company’s guidelines. Who gets it? Google says skills are rolling out globally in Gemini chat for adults on Google AI subscription tiers. Workspace customers are due to follow in the coming weeks. If you already use Gems, Google’s existing custom assistants, there is a change to watch: support starts ending in November for personal accounts. Google says it will move those Gems into skills automatically; business and education accounts follow in 2027. See how Gemini skills work →03 · AI SAFETY · REPORTED 30 SEPTEMBERUS regulator investigates OpenAI and AnthropicFTC building, Washington, D.C. Archive photo: Carol M. Highsmith / Library of Congress, via Wikimedia Commons. Public domain. Image source The US Federal Trade Commission is investigating whether AI systems from OpenAI, Anthropic and other developers could harm consumers, Reuters reported on September 30, citing a senior agency official. The investigation includes concerns about AI agents: programs that can carry out a series of actions, rather than just answer a question. Reuters says the agency plans to demand information and testimony from AI executives. This is an investigation. The report does not announce a finding of wrongdoing or a penalty. OpenAI and Anthropic had not immediately responded to Reuters’ requests for comment when it published. For anyone connecting AI to work accounts, the practical issue is simple: a tool that can change files or take actions can cause problems beyond a wrong answer. Know which actions it can take on its own and which require your approval. Read the Reuters report →TRY THIS · A BETTER WEEKLY UPDATESave a writing brief you can use again Start with something you write every week. In Gemini, you can turn the brief below into a skill. You can also reuse it in your usual AI tool. Turn my notes into a 120-word project update for my manager. Start with what changed this week. Then list completed work, problems and the next deadline. Use everyday words. Keep every name, number and date exactly as supplied. If a detail is missing, ask me instead of filling it in. Add your notes underneath. Before sending, check the names, numbers and dates against the original. A reusable brief saves you typing; you still decide whether the result is ready. Which task do you keep explaining to AI?Hit reply with one example. We’ll use your answers to shape future practical guides. See you next time,Artificialintelligence.aiSources: Google’s product announcements and Reuters reporting. Images are credited in each story. Availability checked 1 October 2026.

Who gets Google’s new AI first, a useful Gemini update, and the US investigation into OpenAI and Anthropic.

Who can try Google’s new AI, how to stop repeating your instructions, and why US regulators are investigating OpenAI and Anthropic.

Google unveils Gemini 4. Access is limited.

01 · THE BIG LAUNCH · 29 SEPTEMBERMeet Dots: an AI with an ongoing jobDots launch artwork. Image: OpenAI. Image source OpenAI launched Dots on September 29: personal agents powered by GPT-6 Astra, each with its own cloud computer and browser. Assigned work can keep moving even when your laptop is off. OpenAI’s examples include updating launch materials when plans change, preparing content from interview transcripts and investigating software feedback. You can keep talking to your dot while it works and review what it produces. Why it matters: the useful shift is continuity. An assistant could keep track of a responsibility across days, reducing the effort of restarting and re-explaining the same project. The boundary: OpenAI says unprompted “proactive research” uses read-only tools. Actions follow permissions and approval checks. Cloud work can continue while you are offline; work on your own connected computer still requires that computer to be online. Who gets it? Rollout is gradual. Pro access is for adults outside the EEA (including Norway), the UK and Switzerland. Business Premium and Enterprise are rolling out worldwide; Enterprise requires an administrator to enable it. Explore Dots → Check eligibility →02 · EVERYDAY WORK · 28 SEPTEMBERClaude Sonnet 5.5 makes a case for faster workClaude Sonnet 5.5 launch image. Image: Anthropic. Image source Anthropic released Claude Sonnet 5.5 on September 28, positioning it for clearly defined everyday tasks: fixing bugs, drafting documents and building slides or spreadsheets. The company reports over 30% faster output than Sonnet 5 and up to 30% lower cost per task in its tests. That saving comes from using fewer tokens, the units of text processing billed through the API. The per-token rates are unchanged. The distinction: Anthropic still recommends Opus 5.5 for complex, open-ended work requiring sustained judgment. A strong benchmark result does not establish which model is best for your particular job. Our take: shorter waits could make the biggest difference when you revise a draft several times. Measure the whole job, including corrections, before deciding whether to switch. Available now: Sonnet 5.5 is live across Anthropic’s supported platforms, including AWS, Google Cloud and Microsoft Azure. Check your plan or provider for access. Read the release, results and pricing details →03 · KEEPING CONTROL · 28 SEPTEMBERNVIDIA puts a watchdog outside the agentOpen Agent Safety Platform. Official illustration: NVIDIA. Image source As agents gain more access, organizations need ways to enforce their limits. NVIDIA announced its Open Agent Safety Platform on September 28, combining software controls with a reference design for independent hardware oversight. Its OpenShell software sets and enforces boundaries around running agents. A separate watchdog called Sentry, running on BlueField-4 hardware, is designed to detect attempts to leave those boundaries. NVIDIA says it can quarantine an agent within milliseconds. What is available? NVIDIA says the platform’s software, including OpenShell, is available through its developer resources and GitHub. Sentry is part of the reference system design. This is infrastructure for developers and enterprises. Our take: independent controls are a useful part of the move toward longer-running agents. They cannot establish that every answer or action is correct. Teams still need to test the work and review the permissions they grant. See how the platform is designed to work →LEARN ONE THING · 60 SECONDSTest the task you actually do A launch demo can be impressive without telling you whether a tool will help your week. Try one familiar task with an answer you can check. Give your current tool and the new one the same brief: Using only these sources, write a 250-word brief for a busy colleague. State the key change, explain its practical impact, link each factual claim to its source, and flag what remains uncertain. Compare accuracy, time spent checking, and time spent rewriting. Count the minutes from the first prompt to something you would actually use. For an ongoing agent such as Dots, repeat the check after new information arrives. Continuity is valuable only if the work stays accurate as things change. What would your first dot be responsible for?Hit reply with one recurring task. We would love to hear what useful AI looks like in your day. See you next time,Artificialintelligence.aiAI for everyone. Based on original announcements and product documentation. Performance claims are the vendors’ own; we have not independently tested these launches. Sources and image credits accompany every story. Checked 30 September 2026.

OpenAI’s always-on agents, faster Claude and NVIDIA’s new controls. What is available, what matters, and one test to try.

Dots takes on ongoing work. Claude Sonnet 5.5 promises faster results. NVIDIA adds independent controls. Three launches, explained.

OpenAI launches Dots. Here’s what changes.

01 · ON YOUR DESKTOPGooglebook puts Gemini within reachGooglebook running CapCut. Official product image: Google. Image source Google opened pre-orders for Googlebook on September 21: a new laptop category built around Android, desktop Chrome and Gemini. Prices start at US$899. Its Magic Pointer brings Gemini to what is on your screen. Rambler turns spoken thoughts into organised writing. You can also open files from your Android phone and continue supported tasks on the laptop. Why it matters: getting help could involve less switching between apps and re-explaining what you are working on. The useful test is how much friction it removes from an ordinary day. These are Google’s launch claims; we have not tested the hardware. Check the apps you depend on before considering a switch. When? In stores from October 4 in the US and October 5 in six more markets, including the UK and Germany. Norway is not on the initial list. See features and launch countries →02 · YOUR DIGITAL MEMORYAn AI that remembers, with a private vaultPrivate AI Compute. Official illustration: Google DeepMind. Image source Starting again on another device can mean explaining everything twice. On September 23, Google outlined a way for Private AI Compute to retain context across devices. The plan is to keep that memory encrypted in the cloud, with the keys held on your own devices. A protected computing environment would temporarily unlock the information needed for a task. Google says the design would prevent even Google from accessing the stored data. It has published a technical brief and describes an independent security audit. That is a claim about this architecture, not every AI service. Why it matters: a useful assistant needs continuity. The next question is whether remembering more can also give people clear control over what is kept. Can you use it today? This is an architecture update. The announcement does not give a consumer rollout date for the new memory capability. Read Google’s privacy design →03 · IN THE LABClaude found a clue scientists could testFrom Anthropic’s research announcement. Image: Anthropic. Image source On September 23, Anthropic reported that Claude helped identify a previously uncharacterised enzyme system while searching DNA data. Human researchers followed up with analysis and laboratory experiments. The striking part is the division of work: AI searched for unusual patterns and investigated promising candidates; scientists decided what deserved testing and performed the lab work. Anthropic also announced its own life sciences research group and laboratory. The important limit: the system has features reminiscent of CRISPR, but its function is still unknown. This is an early finding shared in a preprint, not a proven new gene-editing tool or treatment. Why it matters: AI may help researchers spot leads in huge datasets. A promising lead still needs experiments before anyone knows what it can do. Read the findings and their limits →LEARN ONE THING · 60 SECONDSTell AI what “done” looks like As AI takes on longer tasks, a clear finish line becomes more useful. Describe the result you want to check. Try this before your next task: Before you begin, restate my goal, list the information you still need and describe the finished result. Suggest three checks I can use to judge it. Then wait for my reply. For a travel plan, those checks might be within budget, realistic journey times and verified opening hours. For a report: the right audience, length and sources. Agreeing on the finish line makes it easier to catch a polished answer that solves the wrong problem. What would make AI more useful to you?Hit reply: fewer app switches, better memory, or something else entirely? See you next time,Artificialintelligence.aiAI for everyone. Original sources and image credits accompany every story. Checked 24 September 2026.

Googlebook arrives in October. Plus: private AI memory and a new lead in the lab.

Googlebook’s October launch, a plan for private AI memory, and a discovery that still needs answers.

The next AI shift is outside the chat box

01 · THE BIG SHIFTBetter models. A smaller running bill.AI-generated editorial illustration, created for this newsletter. OpenAI and Anthropic both launched new models on September 22, with a shared pitch: stronger capabilities at a lower running cost. OpenAI: For developers using GPT-6 Sol, the standard rates for sending text in and getting replies back are half those of GPT-5.6 Sol. GPT-6 Luna is cheaper too. Anthropic: Opus 5.5 cuts those same rates by 20% compared with Opus 5. The company estimates 40% lower costs on typical workloads, partly because the model works more efficiently. Who benefits first? Developers and businesses paying for AI by usage. For the rest of us, lower costs could mean better features or more generous limits in the apps we use. A cheaper model does not automatically cut a monthly subscription. Where to find them: OpenAI’s rollout is in ChatGPT Work and Codex; the models are not yet in the regular Chat view. Opus 5.5 is available across Claude’s platforms. OpenAI’s announcement → · Anthropic’s announcement →02 · GET STARTEDA starting point for the AI-curiousLearning by doing. Image: Google. Image source Getting access to AI is one thing. Knowing where to start is another. On September 22, Google announced 100,000 certificate scholarships through the ITU’s AI Skills Coalition. The initiative reaches more than 80 countries and is aimed at students, teachers, jobseekers, business owners and public servants. A technical background is not required. The practical detail: places are delivered through participating countries and local programmes. Check the route available where you live. Start here: Google’s training hub also lists no-cost options for educators, students and business owners. Check each course’s terms. Our suggestion: choose one real task—planning a lesson, writing a proposal or organising customer feedback—and practise as you learn. Find a course → · Scholarship announcement03 · BEYOND THE CHATBOTCan AI help a city listen?How do you hear a whole city? Illustration: Google/Jigsaw. Image source Thousands of comments can be hard for a city to turn into a plan. Jigsaw’s Sensemaking AI is built to help public officials find patterns in large civic conversations. On September 22, Google.org announced support for a new partner programme that lets local governments run these conversations at no operational cost. One example: Google reports that feedback from 8,000 residents in Bowling Green, Kentucky helped inform a 25-year growth plan. The question to keep asking: whose views were heard, and whose were missed? A useful AI summary should help officials inspect disagreement as well as common ground. Read the story → · Explore the programmeTRY THIS · 5 MINUTESGive your AI a tiny audition Before switching tools, give two of them the same small task. Use a public product description and keep the instructions identical. Paste the description, then ask: Explain this offer in plain English. List what’s included, the full price, recurring charges and anything the text leaves unclear. Use only the information below. Mark missing details as “not stated”.Keep the source beside the answer. Did either tool invent a feature, overlook a condition or make you do more work? The winner is the tool that gets you to a usable result with less correction. That is a better personal benchmark than a flashy demo. What would make AI more useful to you?Lower cost, clearer answers or help getting started? Hit reply with one. Until next time,Artificial Intelligence.aiAI for everyone. Reporting checked 23 September 2026. Sources and image credits appear with each story.

The price cuts explained—plus new training and a five-minute test for your AI.

New models, 100,000 training places, and a city using AI to listen.

AI is getting cheaper. Here’s who benefits.

AI FOR EVERYONE · 21 SEPTEMBER 2026 · 4 MIN READYour study notesare getting a voice A pile of notes is useful only if you can make sense of it. This week, Google is making that process more conversational. Also inside: how fashion designers are using AI before the runway, and why Anthropic is inviting outside testers into its labs. Three stories. What changed, why it matters, and what to watch.01 · LEARN SMARTERA study partner that works from your notesGemini Notebook study tools. Product image: Google. Image sourceGoogle wants you to talk through difficult topics with the material you already have. On September 15, it announced live voice conversations in Gemini Notebook’s mobile app, with answers based on your uploaded sources. The update also adds interactive learning overviews that bring summaries, quizzes and flashcards together. More quiz formats are coming, including short answers and fill-in-the-blank questions. Why it matters: a summary can feel clear even when you have not understood it. A useful way to study is to explain a concept yourself, then ask questions about the part you got stuck on. Before you try it: voice access is rolling out first to Google AI Ultra subscribers aged 18+. Google says the new overviews and quiz formats will reach all users over the coming weeks. Check the original notes when an answer matters. Read Google’s update →02 · MAKE IDEAS TANGIBLEAI gets a job backstage at Fashion WeekJane Wade’s Styling Suite in Google Flow. Image: Google. Image sourceThe interesting part is what happened before the show. In a September 18 case study, Google described working with designers Jane Wade and Sergio Hudson on custom tools for New York Fashion Week. Wade’s Styling Suite combined clothes, shoes, accessories, hair and makeup on digital models. It helped her spot missing pieces before making more physical samples. Hudson’s tool let him explore runway layouts, lighting and props within his budget. Why it matters: our takeaway is to use AI where revisions are expensive. Comparing options early can help a team decide what is worth building, booking or producing. Keep the claim in perspective: these were tools developed with Google engineers. The case study does not establish what a typical user would save in time or money. See the tools and design process →03 · WHO CHECKS THE AI?Outside testers get a closer look inside AnthropicCompany wordmark from the announcement. Image: Anthropic. Image sourceTesting a finished AI model reveals only part of the picture. On September 18, Anthropic announced a partnership with Accenture to place evaluators inside the company as models are developed. Faculty, Accenture’s specialist AI business, will lead work that includes testing safeguards and “red-teaming”: deliberately trying to make a system fail so weaknesses can be found. The proposed access is comparable to an employee’s. Evaluators could follow training, speak with staff and examine the decisions behind a release. Anthropic and Accenture each expect to invest at least $1 billion in this work over five years. Why it matters: access during development could reveal problems that a final test misses. But access alone is not proof of independence. The question to watch: what will the public get to see? Anthropic is funding Accenture’s work directly, and says reporting standards and other details are still unsettled. Read the partnership announcement →TRY THIS · 60 SECONDSTurn a summary into a learning check Choose one page of notes you are allowed to share. Then give your AI tool this instruction: Using only these notes, ask me three questions, one at a time. Wait for my answer before giving feedback. Show which passage supports each correction. If the notes do not contain the answer, say so.Make it count: answer before looking back. Then check each correction against the notes yourself. The goal is to discover what you understand and what needs another look. What would help you most?Hit reply: understanding a difficult topic, testing a creative idea, or knowing whether an AI tool deserves your trust? See you next time,Artificialintelligence.aiAI for everyone. Reporting based on the linked company announcements; practical takeaways are our editorial interpretation. Images credited with each story. Sources checked 21 September 2026.

Talk through your notes, see AI at Fashion Week, and meet the people testing tomorrow’s models. Three stories, clearly explained.

Your study notes are getting a voice

AI FOR EVERYONE · 18 SEPTEMBER 2026 · 3 MIN READGoogle’s AI helpertakes on family admin School emails, sports times and party invitations rarely arrive in one place. Google wants its new family assistant to gather the pieces. Also inside: an ad that answers back, and a new way to check a number before you share it. 01 · AT HOMEThe family calendar gets an AI helperCC’s daily summary and calendar. Official product image: Google. Image source Google expanded CC, its experimental AI assistant, to families and households on September 17. Up to six people can share selected emails, invitations and files with it. CC uses that information to prepare a daily summary, update a shared calendar and track tasks. Send it a photo of a party invitation, for example, and it can add the event to the calendar. It can also help fill in activity forms and make meal plans. The useful change: the plan no longer has to live in one person’s inbox. Each member chooses what to share. Google says CC asks permission before acting outside the group. Who can use it? Adults in the US with a personal Google account. It is an early experiment; new users join a waitlist. See Google’s announcement →02 · WHEN YOU SHOPA ChatGPT ad you can talk toA separate chat with a grocery brand. Frame from OpenAI’s product demo. Image source An online ad usually sends you to a shop. OpenAI is testing another option: start a chat with the advertiser. Announced on September 16, Sponsored Agents let you ask follow-up questions after clicking an ad in ChatGPT. In OpenAI’s demo, a grocery ad offers a “Chat with us” button. The key detail is who is speaking. This is a clearly labelled conversation sponsored by a business, separate from your original ChatGPT chat. Think of it like talking to a shop assistant: helpful for questions about that shop’s products, and still part of a sales pitch. Compare the offer with other shops before deciding. Where is it available? It is being tested with selected advertisers in the US. See OpenAI’s announcement →03 · FINDING THE FACTSA simpler front door to UN statisticsThe new UN data site. Image published by Google. Image source The UN’s new System Data Commons, announced on September 17, brings statistics from across the UN into one searchable site, built using Google’s Data Commons platform. You can ask questions in ordinary language or browse topics such as education, health and clean water. For a school project or a number you have seen in the news, it offers a place to look for the underlying data. We tried it: a question about electricity access in Kenya recognised the country but returned no matching topic. The search still has gaps. Its Explore page lets you browse by place and subject instead. That is a useful reminder: when a search finds nothing, it does not mean the information does not exist. Explore the UN data → · Launch detailsLEARN ONE THING · 60 SECONDSA link is not a fact-check When AI gives you a statistic, a source link is a starting point. You still need to see what it supports. Ask this: Show me the original source for this number, the year it covers and exactly what it measures. Then open the link. Check three things: the number, the year and the group being counted. A figure about adults in one country cannot describe everyone in the world. And a newly published article can still use old data. What would you hand over first?Hit reply with the one family-admin task you would most like AI to handle. See you next time,Artificialintelligence.aiAI for everyone. Sources and image credits appear with each story. Checked 18 September 2026.

A shared family calendar, ads that answer back, and a new way to check the facts. Three fresh stories, simply explained.

Google’s new AI helper tackles family admin

16 SEPTEMBER 2026 · YOUR DAILY AI BRIEFINGSiri gets an upgrade. Gemini comes to Windows. Good morning! Siri is getting an upgrade, and Gemini is now on Windows. We’re also looking at how AI could help predict air pollution. At the end, you’ll find a simple prompt for sorting out meeting notes. What’s new with Siri Gemini comes to Windows AI could help predict air pollution A simple prompt for your meeting notes 01 · SIRIWhat’s new with SiriOfficial product image: Apple · Image source Apple started rolling out Siri AI on September 14. Apple says it can use your messages, emails and what’s on your screen to help you write an email or add plans to your calendar. That could mean less copying and pasting between apps when you need to get something done. Not everyone has access yet. This is an early version, available in English on supported devices. At launch, it isn’t available in the EU on iPhone, iPad or Apple Watch. Read Apple’s announcement →02 · GEMINI ON WINDOWSOpen Gemini without leaving your workOfficial app artwork: Google · Image source Gemini now has its own app for Windows 10 and 11. Google launched it on September 10. Once it’s installed, press Alt + Space to open it while you work. A simple way to try it: paste a paragraph and ask, “Can you make this clearer?” It can also work with information from Google apps you’ve connected. Some features need a paid Google AI plan. See Google’s Windows app →03 · AIR QUALITYAI could help predict air pollutionLondon skyline · Photo: Yukon Haughton / Unsplash Researchers at the University of Manchester are using NVIDIA’s AI tools to forecast air pollution in the UK. NVIDIA says the team trained its model in two days on a supercomputer. The researchers can now run it on a desktop AI system. That could make it easier to test how changes to pollution rules might affect the air we breathe. The work is still at the research stage. The team plans to share its data and methods. One possible future use is to warn people when air quality is likely to be poor. Read about the research →04 · ONE THING TO TRYMake sense of your meeting notes Got meeting notes that need sorting out? Paste in notes that don’t contain private information, then try this: From these notes, list what we decided, what happens next, who’s doing each task and when it’s due. If a name or date is missing, say “not specified”. Add three questions we still need to answer. Keep it short and don’t guess. Check the names and deadlines before you share the list. See you next time,Artificialintelligence.ai teamYou’ll find the sources for each story and image above.

What’s new, what you can try, and a simple way to tidy up your meeting notes.

Siri gets an upgrade. Gemini comes to Windows.

🔴 LIVE NOWWatch a company take shape with Grok BotThree people. Three days. One company. Three SpaceXAI employees are building a company from scratch with Grok Bot—and streaming the process live. The goal: go from idea to business in three days. Live right now: Day 1 starts with research, a plan, and a product. Follow along to see how the team uses AI at each step, what it creates, and which decisions people still make. ▶ WATCH LIVE ON X →Also in today’s edition: Gemini can remember where you put things What Claude’s new model means for your work Gemini Notebook changes its usage limits A prompt to compare two documents 01 · EVERYDAY AIA little help finding your keys Google has announced a Find Hub feature that lets Gemini remember where you put things, such as a spare key or passport. You tell it the location; later, you can ask for it again. Try it when available: On an Android 16+ phone with Find Hub and Gemini, look for Remembered in Find Hub. Save one item and its exact storage spot. The catch: It remembers what you told it. If you move the item, update the saved location. Read Google’s setup guide →02 · AT WORKClaude gets a new model Anthropic released Fable 5.1 on September 1. The company says it improves complex work such as research, coding and tasks with many steps. It is available on Claude’s Pro, Max, Team and Enterprise plans. Our take: Try it on work you can verify, such as comparing two versions of a project brief. Check whether it saves you time and gets the details right. See Anthropic’s model details →03 · TOOL UPDATEGemini Notebook’s limits have changed Google’s research assistant, formerly NotebookLM, changed its usage limits on September 2. More demanding tasks and longer chats can use up your allowance faster. Do this: Check Settings → Usage before a large research task. On the web, Generate later can queue an output when you have reached your limit. Check Google’s usage guide →04 · TRY THISSpot changes between two documentsOld brief + new brief → a clear list of changes. Upload two versions of a non-sensitive project brief to an AI tool that accepts files. Label the old version A and the new version B. Then copy this prompt: Compare A and B using only these files. List changes to scope, owners, dates and costs. For each change, cite the page or section in both versions. Separate confirmed changes from anything unclear. Do not infer missing details.Before you use the result: Check at least three listed changes against the originals, especially dates and costs. One habit worth keeping Before acting on an AI answer, ask: Where is the evidence? Open the cited source. What is uncertain? Keep assumptions separate from facts. What needs my judgement? Check anything that affects a decision. See you in the next edition,Artificialintelligence.ai StaffIllustrations created with AI.

Watch day 1 of a three-day company build, from research and planning to a first product.

🔴 LIVE NOW: Build a company with Grok Bot

01 · THE NEWSGoogle Sheets gets an interactive viewGoogle’s Sheets canvas turns one sheet tab into an interactive task board, calendar or dashboard. Introduced on 13 August, it uses Gemini, Google’s AI assistant, to build the view from a written instruction. Why it matters: A list of tasks, owners and deadlines can become a board grouped by status. You keep the data you already have and make the next step easier to see. Changes in the canvas sync back to the spreadsheet, so it can help you manage the work as well as view it. Read Google’s announcement →02 · TRY ITBuild your first project board1. Make a copy of a small task list with four columns: Task, Owner, Due date and Status.2. On a computer, open it in Google Sheets and choose Insert → Create a canvas.3. Paste this prompt: Build a simple project board from this tab. Group tasks by Status and show each task’s Owner and Due date. Keep the layout uncluttered and make overdue tasks easy to spot. Check a few cards against the original rows before using the board. Open the setup guide →03 · BEFORE YOU STARTTwo things to knowCheck access. Available in English for Google AI Pro and Ultra subscribers and eligible Workspace plans. If it’s missing, check your plan and your organisation’s access settings. Work on a copy. Edits in the canvas can change the underlying sheet. Our take: Start with a spreadsheet you already find frustrating. Keep the new view if it makes your next action clearer. See you in the next edition,Artificialintelligence.ai StaffSeptember 14, 2026

Turn a task list into an interactive board with Sheets canvas. Here’s how to try it, who has access and what to check.

One useful update. Three steps to try it.

Turn a spreadsheet into a simple app

THE AI BRIEF · 2 MIN READ Hi there, Letting AI draft an email is useful. Letting it send one is a bigger step. Today’s AI news is a good reason to check where you draw that line. The news, in 30 seconds Anthropic, the company behind Claude, wants AI development to slow down so safety work can catch up. It is committing to bring outside reviewers into its development process. Those reviewers would get access similar to internal risk teams and could publish key findings, with some limits on sensitive information. The detail that matters: the proposal does not give reviewers a binding power to stop a release. It is a step toward outside scrutiny; an industry-wide slowdown is still a proposal. What this means for you An AI agent is an assistant that can use tools to complete tasks, such as searching websites or managing a calendar. That can save you work. The useful question is: what can it do without asking you? For example, researching three suppliers and drafting a comparison needs fewer permissions than contacting them or placing an order. Try this today Before connecting an AI to another app, check three things: 1. What can it access?Give it only the files and tools needed for this task. 2. What needs your approval?Look for settings that require you to approve sending, buying or deleting. 3. How do you stop it?Find the control to disconnect the app or revoke access before you begin. A useful first taskAsk AI to research three options, link its sources and draft a recommendation for you to review. Keep sending and purchasing permissions off while you test it. A written instruction alone is not an access control. Read the full story →One thing to rememberThe more an AI can do for you, the more its permissions matter. Until next time,The ArtificialIntelligence.ai teamSource: Dario Amodei’s proposal. The practical checklist is our editorial advice.

Three quick checks before you connect AI to your inbox, calendar or apps.

Anthropic wants to slow the AI race. Here’s the useful takeaway for you.

Before AI acts for you, check this

Hey, and welcome back! Today: the video you scrolled past last night may not have had a single human in it. And the trick everyone recommends for spotting that is quietly going out of date. The number40 to 50 percentShare of China's 100 biggest short dramas now made or assisted by AI, according to Media Partners Asia. What happened Short dramas are soap operas built for the phone. Episodes run one to three minutes, filmed upright, usually romance, and a season can be 60 to 90 episodes. In China it is a business worth around 11 billion dollars this year. Reporting this month put a number on how much of it is now machine made. Close to half of the top 100 titles. Platforms like ReelShort and iQiyi have doubled their AI engineering teams. The reason is money. A series that used to need a crew, a set and weeks of filming can now be finished for a fraction of the cost. So the studio stops betting everything on one idea and tries twenty. One industry figure, Yang Xianghua, said the uncomfortable part out loud: "most viewers cannot tell AI-made content from human-made work." Why this matters if you never watch short dramas Because this is the test lab. Cheap, fast, disposable video is where synthetic footage becomes normal first. Ads come next. Then the clip your uncle forwards. Then the thirty second video of something that supposedly happened in your town. You do not have to care about the shows. You do have to notice that the ground moved. So can you still spot it? Less and less. The advice still going around says to look at the hands, count the fingers, read the text on signs in the background. That worked in 2023. It mostly does not work now. Hands are fine. Signs are clean. Faces stay the same between cuts. Here is the honest version: if you try to win this by staring harder at the screen, you will lose. The picture keeps getting better. Your eyes do not. The habit that still works Stop asking "does this look real". Start asking "where did this come from". 1. Who posted it first?Not who sent it to you. Scroll back to the original account. A real event has several people filming it from several angles. A fake usually has one source and a thousand reposts. 2. Does anyone boring confirm it?Boring means a local paper, a fire department, a company statement. If something big happened, something dull will mention it too. 3. Does it want something from you?Anger, a click, a donation, a purchase. Fake video is cheap to make, so it gets made where there is money or attention to collect. None of this needs any technical skill. It is the same instinct you already use on a rumour at work. Try this today The next clip that makes you feel something strongly, do one thing before you react: find who posted it first. Thirty seconds. That is the whole exercise. You will be surprised how often the trail stops somewhere strange. One term, explained: content credentials Some cameras and some AI tools now attach an invisible label to a file. It says how the file was made and what edited it. Think of it as a receipt for a picture. Good to know it exists. Do not lean on it yet. Most social apps strip the label when you upload, and anyone can take a screenshot to remove it.One line to rememberDo not ask if it looks real. Ask where it came from.P.S. Reply with a clip you were not sure about. We will trace one in the next issue and show every step. Do not take our word for it. Read it yourself:The Hollywood Reporter · CNN · CNBC

Almost half of China’s biggest short dramas are now made with AI. Spotting them is getting harder. Here is what still works.

You Have Already Watched AI Without Knowing It

Hey, and welcome back! Today: a bank quietly changed what it asks young people in job interviews. And one habit you can steal from it before your coffee gets cold. What happened UBS is a big Swiss bank. More than 100,000 people work there. It normally asks young applicants about their grades. This week it asked for something new. If you want a job there in 2027, you have to show that you can use AI to do real work. Three things to notice: It is not a test. It is a question in the interview. They want one real example. What did you give the AI, and what came back? Your degree still counts. This is added on top. Nothing is taken away. You do not have to be an expert. Everyone who gets in is trained inside the bank. You only have to have tried. Three swaps you can steal1. The complaint email❌ "How do I write a complaint email?"✅ Paste the receipt and their reply, then write: "Write my complaint. Polite but firm. Under 150 words." 2. The rental contract❌ "Is this a good rental contract?"✅ Photograph every page, then write: "I am the tenant. Tell me what is odd, unfair or costly, in simple words." 3. Dinner❌ "Give me dinner ideas."✅ Take a photo inside your fridge, then write: "Three dinners. 25 minutes. I hate cilantro." See the pattern? Every good one hands over something real. Do this before your coffee gets cold Find the most boring paper in your home. A power bill. An insurance letter. A note from school. The long thing you have not opened since March. Take a photo of it. Give it to the AI. Then copy this: 💬 "Explain this to me in simple words. Then tell me the three things that matter, and anything I should worry about." Then check the worrying parts yourself. AI sounds sure of itself even when it is wrong. That is it. You now have a real example. That is exactly what the bank is asking for. One word, explained: context Context is everything you have given the AI. Your document. Your photo. Your situation. The messages above. Little context, and it guesses. Lots of context, and it reads. Almost every "AI is useless" moment is really a missing context moment. One line to rememberDo not ask it to imagine your problem. Show it your problem.P.S. Hit reply and tell us the most boring paper in your life. We will solve the best one in the next issue, with no names. Do not take our word for it. Read the story yourself:The Next Web · TechRadar · AI WeeklyOriginally reported by the Financial Times.

Not a degree. Not a certificate. One 30-second story. Here's how to get yours today.

A Bank Just Made "Can You Use AI?" a Job Requirement

Good morning, and welcome. This is AI for Everyone. Six minutes, once a week. No jargon, no hype, and a source under every claim so you can check us. If you are new here, the last section is the one to keep. TODAY OpenAI's new model can operate a computer A major bank now requires AI skills to get hired The 120,000 cameras reading number plates What AI is actually reliable for — the map we work from Quick hits and one prompt worth stealing 1. The new ChatGPT can use a computerWhat happened On 3 September OpenAI released GPT-6 Astra, its most capable model so far. The headline feature is "computer use": instead of only writing text back to you, it can operate a computer the way a person does — move the cursor, click, fill in forms, work through a task on screen. It is rolling out to Plus, Pro and Enterprise users. OpenAI delayed the launch to add security safeguards. It says the model crossed its own "critical cybersecurity" threshold, meaning it can find and exploit unknown software flaws without a human directing it. The strongest version of those abilities is being kept to selected partners. In simple terms Until now, AI mostly handed you words and you did the clicking. This one can do the clicking: open the site, fill the form, download the file, rename it. This is what people mean by an "AI agent". Earlier versions were demos. This is the first that is good enough to be worth your time. OpenAI itself says the model is capable enough to be a security risk. Take that at face value. It is a serious tool with a sharp edge. The benchmark scores jumped enormously. Benchmarks are exams the labs pick themselves. Treat them as a sign of direction, not of how it will handle your Tuesday. What it means for you Two practical things. Worth trying: any repetitive computer chore. Downloading twelve statements. Renaming a folder of files. Copying data from one system into another. That is the sort of job this is genuinely good at, and it is the first month it is worth your afternoon. Worth caution: an AI that can click can also click "confirm". Give it read-only work first. Do not hand it a browser that is logged into your bank, and do not leave it running unattended on anything that spends money. Source: Fortune2. A major bank now requires AI skills to get hiredWhat happened UBS has added AI ability to what it asks of graduates and interns joining in 2027. Candidates now face AI questions in interviews alongside the usual degree requirements, and must show they can use AI responsibly to improve a result. The bank is also running an internal "AI Fluency Pathway" to teach it. Alongside the news, a Morgan Stanley report warned that European banks could shed 200,000 jobs over five years. In simple terms This is not a tech company. It is a Swiss bank. When a bank writes a skill into its hiring criteria, that skill has stopped being optional. They are not testing whether you can name models. They are testing whether you use one thoughtfully. The training pathway is the tell. They expect to teach it, which means the bar is attitude, not mastery. One employer is not a trend. But hiring criteria are the most honest signal an industry gives, because getting them wrong costs real money. What it means for you If you are job hunting — or your children are — prepare one worked example. Not an opinion about AI. A specific task you used to do by hand, how you used AI on it, what it got wrong, and how you caught the mistake. That last part is what an interviewer is actually listening for. Anyone can say they use ChatGPT. Almost nobody can describe how they checked it. Source: TechRadar3. The 120,000 cameras reading number platesWhat happened A US Senate investigation into AI camera networks widened on 3 September to cover Motorola Solutions, Verkada and Axon. It began with a company called Flock Safety, which runs more than 120,000 cameras across 49 states and processes over 20 billion vehicle scans a month, pooled into one searchable national database. The senator's letters list specific failures. An officer in Milwaukee searched a girlfriend's number plate 124 times. A woman in Florida spent 13 days in jail on a wrong match. One California department found 71% of its alerts were inaccurate. Florida's governor banned the cameras from state highways on 1 September. In simple terms These cameras read number plates and log where and when. Enough of them, joined up, is a record of everywhere you drive. The problem is not really the AI. It is that a searchable national database now exists, and who may look in it is governed by company policy rather than law. The failures are ordinary human ones: snooping, and wrong matches acted on without anyone checking. The pushback is bipartisan. That is unusual in the US, and it is why this one may actually change something. What it means for you This is the clearest current example of what AI risk actually looks like for ordinary people. It is not a machine deciding to harm anyone. It is a dull system, wrong most of the time in one county, that someone acted on anyway. If cameras like these are proposed where you live, there are two questions worth asking: who is allowed to search it, and is there a log of who did? Sources: Office of Sen. Hawley · expansion, 3 SeptemberWhat AI is actually reliable for Since this is the first issue for many of you, here is the map we work from. It is worth more than any list of tools, and it changes slowly. Reliable, with a light check Rewriting, shortening or changing the tone of something you wrote Explaining a contract, a bill or an official letter in plain language Turning messy notes into a structured list Translating, and telling you how the tone will land Drafting the awkward email you have been avoiding for a week Arguing with you before you make a decision Useful, but check every time Any number, date, name, price or reference Summaries of long documents — it quietly drops the exceptions Anything from the last few weeks, unless it searched the web while you watched Code that touches money or personal data Not there yet Deciding anything you cannot afford to get wrong Telling you whether a message or a person is genuine Knowing what it does not know — it will be just as fluent when it is wrong If you only take one habit from this newsletter: use it freely for the first list, and never skip the check on the second. Quick hitsNew York City banned AI for pupils through eighth grade. The largest school district in the US, covering around a million children, with screen time limits alongside it. (Chalkbeat) Sony and Warner are suing Anthropic over song lyrics, alleging tens of thousands of songs were used without permission and seeking up to $150,000 per song. (Fortune) Apple has a new chief executive. John Ternus takes over the one big technology company without a frontier AI model of its own. One prompt worth stealing Most people ask AI to write something and get back something that sounds like everyone else. The reason is simple: a generic request gets a generic answer. The fix is to make it ask you first. Paste this before anything you want it to write — a complaint letter, a speech, a job application, the difficult message to a colleague you have been putting off: I want your help writing [what it is, and who will read it]. Do not write anything yet. First ask me questions, one at a time, the way a good [journalist / lawyer / recruiter] would, until you have what you need to write something only I could have written. Then write it, and tell me what you had to guess.Why it works. Three parts, each doing a job: "Do not write anything yet" stops it handing you the bland version that you then have to repair. "One at a time" is the part that makes it work. Ask for ten questions at once and you will answer three and give up. One question is easy, and each answer shapes the next one. "Tell me what you had to guess" shows you the holes — the places it filled in with something plausible because you never told it. The first few questions will feel obvious. Around the fifth it usually asks something you had not thought about, and that answer is the sentence that ends up carrying the whole letter. This works for anything whose quality depends on details only you know. Which is most things worth writing. That is it for today. If this was useful, the best thing you can do is forward it to one person who thinks AI is not for them. That is exactly who we write for. And hit reply with the story you want us to dig into. The most requested one becomes next week's lead.

PLUS: A major bank now demands AI skills to get hired, and the 120,000 cameras reading number plates

The new ChatGPT can use a computer like you do

We've covered what the labs ship and what the boom costs. Today: where AI has quietly ended up — inside decisions people used to make with a professional, a family member, or nobody at all. None of this was announced with a keynote. It just happened. In today's issue: 🩺 ChatGPT is now a health system with no doctor in it 💸 $16 billion lost to fraud — and why AI makes the old advice useless 🛒 AI agents are starting to buy things for you 👧 72% of US teens have used an AI companion 🧭 A three-question test for using AI on personal stuff Reading time: 5 minutes.🩺 The health system nobody built On September 1, OpenAI connected ChatGPT Health to Epic, the electronic health record system covering over 325 million patients. Clinicians can now pull appointment notes, lab results, medications and specialist documentation into ChatGPT. But the number buried in that announcement is the real story: people already send ChatGPT 300 million health-related questions every week. THE BREAKDOWN: The Epic integration is read-only — the AI can read records, it can't write to them. OpenAI says it tested 27 clinical use cases with 4,300+ physicians and rated 99.1% of responses safe. It also states plainly that the AI is not suitable for diagnosis or treatment. The announcement follows multiple lawsuits alleging harmful medical advice from ChatGPT. Meanwhile, a parallel system has assembled itself around consumers: $99 fifty-biomarker blood panels from Oura and Quest, Function Health at a $2.5B valuation offering 160 annual lab tests plus AI analysis of the results, and services like Ro and Hims writing prescriptions from asynchronous intake forms. One company, Doctronic, reports 24 million AI consultations. Clinicians writing in STAT call this a "shadow medical system" — the work that used to happen inside a clinic, sold as consumer products. The uncomfortable finding: AI performs at 90%+ accuracy when it has the complete case, but fails to produce a proper list of possible diagnoses 80% of the time when working from the kind of vague information a patient actually starts with. And the medical disclaimers that used to appear in chatbot answers have largely disappeared. WHY IT MATTERS: This is the widest gap in AI right now. Not between what models can do and what they claim — between how good AI is with a full picture and how people actually use it: at 2am, with three symptoms and a fear. The technology is genuinely useful for organising information, translating jargon, and preparing questions before an appointment. It is weakest at exactly the step people most want it for — deciding what's wrong. Practical version: use AI to understand your results and prepare better questions. Don't use it to decide whether to worry. 📎 Sources: TechCrunch · OpenAI · STAT💸 $16 billion gone — and the old advice stopped working The FTC reported that Americans lost roughly $16 billion to fraud in 2025, up 25% from 2024. THE BREAKDOWN:Imposter scams — someone pretending to be your bank, a government agency, a company, or a person you know — accounted for $3.5 billion, nearly three times the 2020 figure. Nearly one in three fraud reports were imposter scams. Fifth year running as the top category. Bank impersonators: close to $1 billion. Government impersonators: $920 million, up from $789 million. The most common play: a fake security alert convincing you to move money somewhere "safe." WHY IT MATTERS: Every piece of scam advice most of us grew up with was built on detection — spot the bad grammar, listen for the odd accent, look for the blurry logo. Cheap voice cloning and fluent text deleted all of those signals. A voice that sounds exactly like your daughter is now a commodity product. What replaces detection is procedure:Agree a family code word. Boring, free, works against a perfect voice clone. Hang up and call back on a number you already have. Never one you were just given. Treat urgency itself as the red flag. No real bank, tax office or police force needs you to move money in the next ten minutes. 📎 Source: FTC🛒 AI agents are starting to do the buying Shopping is where "AI agents" stop being a demo and start touching your card. THE BREAKDOWN: OpenAI has partnered with retailers including Target, Instacart and DoorDash to let people buy inside ChatGPT. Around 2% of ChatGPT queries are shopping-related — roughly 50 million a day. Amazon is playing both sides: it launched its own "Buy For Me" feature and expanded its Rufus assistant, while blocking outside AI agents from crawling its site — protecting a $56 billion advertising business that only works if humans see the ads. Consumers aren't sold yet. In one survey of 700 shoppers, only about a third would complete a payment through an AI assistant, mostly over privacy. Retailers have a structural problem: product catalogues were built for keyword search, not conversation. Most AI shopping experiences are still bad — wrong prices, generic recommendations. WHY IT MATTERS: This is the first mass-market case where an AI acts with your money rather than giving you words. The fight isn't really about convenience — it's about who owns the moment of choice. Today that moment happens on Amazon's page, surrounded by paid placement. If it moves into a chat window, an entire advertising economy moves with it. That's why Amazon is blocking the door. For anyone selling online: the practical question for the next 12 months is whether your product data is structured well enough for an AI to describe it correctly. That's the new SEO. 📎 Source: Modern Retail👧 And the one most people miss72% of US teens have used an AI companion, according to Common Sense Media's national survey — and about half use them regularly. Not ChatGPT for homework: chatbots designed for conversation and relationship. This gets covered as a panic story or ignored entirely, and it deserves neither. The useful frame for a parent isn't "is this bad" but "what is it replacing?" A teenager rehearsing a difficult conversation with a bot is doing something different from one who's stopped having difficult conversations with people. Worth asking, not lecturing: what do you use it for, and what would you have done before? 📎 Source: Common Sense Media🧭 A three-question test Before you let AI into a personal decision — health, money, family, legal — run these: Does it have the full picture? AI is strong with complete information and weak with partial. If you've told it three sentences about a situation with fifty relevant details, you're getting a confident guess. Is it reversible? Understanding a lab result is reversible. Skipping the appointment isn't. Moving money isn't. Match your trust to the cost of being wrong. Who's accountable if it's wrong? A doctor, a lawyer and an accountant all carry liability. A chatbot carries none. That's not a reason never to use it — it's a reason to know which one you're talking to. 💬 One question What's the most personal thing you've asked an AI — and did it help? Hit reply. We'll share the best answers, anonymously, in a future issue. That's it for today. Forward this to someone who'd benefit from the code-word tip.— The artificialintelligence.ai team

AI has quietly moved into decisions people used to make with a doctor, a bank, or nobody at all. Nobody announced it. It just happened.

300 million health questions a week

Last issue was about what the labs shipped. This one is about what it costs — and who pays. Most AI coverage stops at the announcement. But the interesting part of 2026 is what's landing downstream: on your power bill, in the job market for 22-year-olds, and in the small print under the videos in your feed. In today's issue: ⚡ Data centres are showing up on household electricity bills 📉 Stanford: young workers in AI-exposed jobs are 19% behind 🏷️ Europe now requires AI content to identify itself 🌍 Abu Dhabi gave away a 375B-parameter model — training data included 📌 Quick hits Reading time: 5 minutes.⚡ The AI boom is on your power bill now For three years, AI's energy use was an abstraction. It has stopped being one. THE BREAKDOWN: US residential electricity prices rose 7.1% in 2025 — more than double inflation. Some states saw over 20%. In areas dense with data centres, prices have jumped as much as 267% over five years. Data centre energy demand is projected to nearly double from 80 GW (2025) to 150 GW by 2028. A single hyperscale facility draws around 100 megawatts — enough for 100,000 homes. Meta's Hyperion site in Louisiana is planned at 5 GW, roughly three times what all of New Orleans uses. In Virginia — nearly 600 data centres, 100+ more coming — data centres consumed almost 40% of the state's electricity in 2024. 75% of Virginia voters now blame them for rising bills. Water matters too: large facilities can use up to 5 million gallons a day for cooling, and two-thirds of data centres built since 2022 sit in water-stressed regions. WHY IT MATTERS: Every "AI is free" product is running on infrastructure someone pays for. Increasingly, part of that someone is the household on the same grid. This is the fight that decides how fast AI can actually scale — not benchmarks, but whether local regulators let the next campus connect. Worth watching: the rules on who pays for grid upgrades — the utility's ratepayers or the data centre — are being rewritten state by state right now. That's the whole ballgame. 📎 Source: Consumer Reports📉 Stanford: the entry-level gap widened to 19% Erik Brynjolfsson's team at the Stanford Digital Economy Lab updated their "Canaries in the Coal Mine" study on August 12, 2026, using ADP payroll data through mid-2026. THE BREAKDOWN: Workers aged 22–25 in AI-exposed occupations now sit 19% behind similar-aged peers in less-exposed jobs. That gap was 15% in July 2025. Experienced workers show no comparable gap. The effect is concentrated at the entry level. The pattern: jobs built on codified knowledge — the stuff written down in manuals and documentation — are shrinking for young people. Jobs built on tacit knowledge — judgement built through practice — are holding or growing. The gap persists after controlling for interest rates, remote work and tech-sector employment. The researchers' own caveat, which most coverage skips: these are "descriptive patterns, not causal estimates," and they explicitly say no single study is definitive. WHY IT MATTERS: The AI-and-jobs debate is usually all-or-nothing — mass unemployment or nothing to see here. The data suggests something narrower and more useful: AI is eating the first rung of specific ladders. If your work is mostly applying documented procedure, that's the exposure. If it's judgement, relationships and knowing what the manual doesn't say, that's the moat. 📎 Source: Stanford Digital Economy Lab🏷️ In Europe, AI content now has to say it's AI Since August 2, 2026, the European Commission and national authorities have been enforcing the EU AI Act's transparency rules. THE BREAKDOWN — what changes in practice:Chatbots must disclose they're not human.Deepfakes must be labelled. AI-generated or altered content must carry machine-readable markers so tools can detect it. The accompanying Code of Practice gets specific: a persistent icon plus an opening disclaimer for video, a fixed visible icon for images, audible disclaimers for audio deepfakes, and lighter-touch disclosure for creative work so it doesn't wreck the piece. Over 180 organisations have signed that Code. WHY IT MATTERS: This is the first serious attempt to make "is this real?" a question with a technical answer rather than a vibe. Two open problems: labels only help if platforms display them, and a world where AI content is labelled can quietly make unlabelled fakes more convincing. Watch for the first enforcement action — that's when we learn what this actually means. If you make content: if any of your audience is in the EU, this applies to you, not just to Big Tech. 📎 Sources: European Commission · Tech Policy Press🌍 Abu Dhabi gave away a 375B model — with the training data On September 3, the Institute of Foundation Models (part of MBZUAI in Abu Dhabi) released K2 Horizon: six models from 0.9B to 375B parameters, all under Apache 2.0. THE BREAKDOWN: Released: weights, source code, the complete training data, and the training recipe. That last part is rare — most "open" models publish weights only. The fleet spans a 0.9B model that runs on modest hardware up to a 375B flagship with 23B active parameters. IFM claims state-of-the-art performance at the small end (0.9B, 3.7B, 7B) and around a 3x speedup from a diffusion-distillation technique. Founder Eric Xing: "Open source is much more than open weights... Every model in the fleet ships with its training data, recipe and evaluations."WHY IT MATTERS: Two things at once. First, if you can inspect the training data, you can actually audit a model — for bias, for contamination, for what it was fed. Almost nobody lets you do that. Second, the geography: the most open frontier-scale release this year came from the Gulf, not California. The map of who builds AI is being redrawn, and Abu Dhabi is buying influence with generosity rather than secrecy. 📎 Sources: IFM · BigDATAwire📌 Quick hitsThe picks-and-shovels round. Gimlet Labs raised $300M at a $3B valuation, led by Andreessen Horowitz with Arm and Microsoft's M12 participating. The pitch: software that spreads AI workloads across different chip architectures, so you're not locked to one vendor. Telling detail — it raised an $80M Series A earlier this year. SourceFollow-up from last issue: OpenAI's GPT-6 Astra is rolling out to Plus, Pro and Enterprise users this week, and Nvidia's $12.9B Hugging Face deal now goes to regulators. Both worth watching for what happens next rather than what was announced. 🧠 What this actually means for youAudit your own exposure honestly. Not "will AI take my job" but: how much of what you're paid for is written down somewhere? That's the part under pressure. The judgement part isn't. If you hire, the entry level is your problem to solve. The Stanford data says juniors are getting squeezed out of exactly the roles that used to train them. Whoever figures out how to build experience without the old first rung has a real advantage. Label your AI content now. Not because the EU will fine you tomorrow, but because "we tell you when it's AI" is about to be a trust signal — and retrofitting it later looks like getting caught. 💬 One question Have you noticed AI showing up in your actual costs — power bill, subscriptions, insurance, anything? Hit reply. We're collecting real examples for a future issue, and the specific ones are always better than the studies. That's it for today. If someone forwarded you this, you can subscribe below.— The artificialintelligence.ai team

Data centres are showing up on household power bills, young workers in AI-exposed jobs are 19% behind, and AI content in Europe now has to say so. The consequences issue.

Your electricity bill is now an AI story

Welcome back to artificialintelligence.ai — the newsletter that explains AI like you're a smart human, not a machine learning engineer. This week was one of the loudest in AI history. A new OpenAI model that clicks around your computer by itself. A $12.9 billion acquisition that changes who owns the plumbing of open-source AI. And a bill in the US Senate that would make building "superintelligence" a crime punishable by 20 years in prison. In today's issue: 🚀 OpenAI launches GPT-6 Astra and says the AGI era has started 💰 Nvidia buys Hugging Face for $12.9B ⚖️ Sanders and Casar move to outlaw superintelligent AI ⚡ Google's Gemini 3.8 Flash undercuts everyone on price 📌 Quick hits: ChatGPT ads hit $1B, Claude lands in Salesforce, Tesla's Cybercab hits Austin Reading time: 5 minutes.🚀 OpenAI releases GPT-6 "Astra" — and uses the AGI word OpenAI released GPT-6 Astra on September 3, its most capable model to date. The headline feature isn't better writing or better math. It's computer use: Astra can operate a computer the way a person does — opening webpages, filling in forms, working through spreadsheets — except much faster. OpenAI President Greg Brockman went further than the company usually does: "It's not unreasonable to feel that we are now in the AGI era."THE BREAKDOWN:ARC-AGI-3 (a reasoning test designed to be hard for AI): Astra scored 98.6%. The previous model, GPT-5.6 Sol, scored 7.8%. That is not a typo. Cybersecurity: on OpenAI's ExploitGym test it scored 100%, up from 78.5%. The release was delayed so OpenAI could add safeguards. Astra crossed the company's own "critical cybersecurity capability threshold," which triggered extra monitoring, isolated training environments, and a government review under a voluntary safety framework. Availability: it went first to enterprise customers in OpenAI's cybersecurity-focused Daybreak program. Plus, Pro, Enterprise, API and AWS access follow "in the coming days." Regular users get a restricted version that refuses advanced hacking tasks. WHY IT MATTERS: Until now, AI has mostly been something you talk to. A model that reliably drives a computer is something you delegate to. That's the difference between a tool that helps you do your admin and a tool that does your admin. If Astra holds up outside benchmarks, the "AI takes over the boring parts of your job" story stops being a prediction. The honest caveat: benchmark jumps this dramatic are usually part real capability, part the benchmark being a good fit for how the model was trained. Watch what people actually get done with it over the next month. 📎 Sources: OpenAI · Fortune · Axios💰 Nvidia buys Hugging Face for $12.9 billion Nvidia confirmed on September 3 that it is acquiring Hugging Face for $12.93 billion. If you haven't heard of Hugging Face: think of it as GitHub for AI models. It's where the world's open AI models live and get shared. THE BREAKDOWN: What Nvidia is buying: 3 million AI models, 1 million apps, 500,000 datasets, and 18 million developers. Hugging Face turned down a $500 million offer from Nvidia last year. It makes around $150M a year in revenue. That's roughly 86x revenue — a price you pay for strategic position, not for cash flow. Jensen Huang's promise: "Hugging Face will remain an open platform for the entire AI ecosystem." Nvidia specifically committed that its own chips will not be required to build or deploy through the platform. WHY IT MATTERS: Nvidia already sells the chips almost every AI model is trained on. Now it owns the place where those models are distributed. Even with the open-platform promise, the company that sells the shovels just bought the map to the gold mine. Expect regulators — and every AMD customer on the platform — to watch closely. 📎 Sources: TechCrunch · CNBC⚖️ A US bill would ban superintelligence — and jail people for building it Also on September 3, Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act. THE BREAKDOWN: A permanent ban on developing AI that surpasses human intelligence or could overthrow a government. A temporary pause on advanced AI development until federal safety rules exist. A new cabinet-level agency to regulate and enforce AI safety. Penalties: corporate death penalty for companies, up to 20 years in prison for individuals — deliberately modelled on the penalties for illegal nuclear weapons development. Sanders: "The future of humanity cannot be left in the hands of a handful of Big Tech oligarchs." Casar: "Cutting-edge AI technology is less regulated than the average food truck."WHY IT MATTERS: Realistically, this bill will not pass in its current form. But its timing — the same day OpenAI said the AGI era may have begun — makes it a marker. The political conversation has moved from "should we label AI content?" to "should some AI be illegal?" That shift affects every AI company's risk calculus, whether or not this specific text becomes law. 📎 Sources: Sanders Senate office · Axios⚡ Google's Gemini 3.8 Flash: the price war continues Google shipped Gemini 3.8 Flash on September 2 — three weeks after its last release. THE BREAKDOWN: Scores 59 on the Artificial Analysis Intelligence Index at high reasoning. That's level with GPT-5.6 Sol and Grok 4.6, below Claude Fable 5.1 (66) and Claude Opus 5 (63). Pricing: $0.75 per million input tokens, $3.75 output — roughly 6x cheaper than Claude Fable 5.1, and per Artificial Analysis the cheapest model at its intelligence level. Note: that's an introductory rate. The price doubles in 2027. A separate Gemini 3.8 Flash Cyber variant went to governments, critical infrastructure and software maintainers for finding and fixing vulnerabilities. WHY IT MATTERS: Frontier intelligence is getting expensive; good enough intelligence is getting close to free. If you're building on AI, the question is no longer "which model is smartest" but "what's the cheapest model that clears my bar." For most tasks, that bar dropped again this week. 📎 Source: The Register📌 Quick hitsChatGPT ads passed $1B. OpenAI says its advertising business hit a $1 billion annualized run rate in under 200 days, and self-service ad buying opened across India, Europe, the Middle East and North Africa — now 40+ countries. OpenAI's line: ads are what keep ChatGPT free for over a billion weekly users, they're clearly labelled, and they don't influence answers. SourceClaude moves into Salesforce. Anthropic and Salesforce announced Claudeforce on August 26: Claude powers Agentforce's reasoning engine, becomes the default model across Slack, and gets a Salesforce plugin with 37 prebuilt sales skills. Open beta expected this month. SourceTesla's Cybercab hit the street. Tesla launched its steering-wheel-free two-seat robotaxi in Austin on September 3, starting with invite-only rides downtown. SourceSpaceX closed the Cursor deal. SpaceX completed its record $60 billion acquisition of Anysphere, maker of the AI coding tool Cursor. Source🧠 What this actually means for you Three practical takeaways from a very loud week: Start thinking in tasks, not chats. If a model can operate a computer, the useful question changes from "what should I ask it?" to "what whole job should I hand over?" Pick one repetitive process you do weekly and write it down as steps. That's your first delegation candidate. Don't pay frontier prices for non-frontier work. Summarising, drafting, classifying, extracting — cheap models now handle these. Save the expensive models for the hard 10%. Assume the rules will change. Whatever you're building on AI, keep it portable across models and providers. This week showed that ownership, pricing and legality can all move in a single day. 💬 One question Would you let an AI operate your computer unsupervised? Yes, already do · Only with approval on each step · Absolutely not Hit reply and tell us why — we read every response and feature the best ones. That's it for today. Forward this to one person who keeps saying they "should probably learn AI."— The artificialintelligence.ai team

GPT-6 Astra can use your computer, Nvidia bought Hugging Face for $12.9B, and a US bill would jail people for building superintelligence. Let's break it down.

OpenAI just said the AGI word

Good morning. This is AI for Everyone. Five minutes. No jargon. Just what happened in AI this week, and what it means for you. TODAY ChatGPT starts showing adverts in Europe Your electricity bill is paying for the AI boom What happens to everything you tell a chatbot Quick hits and one prompt worth stealing 1. ChatGPT has adverts nowWhat happened From 24 August, adverts started appearing inside ChatGPT in 31 European countries, including Norway, Sweden, Denmark, Germany, France, Spain, Italy and the Netherlands. Only free users see them. Paid plans stay ad-free. OpenAI says the adverts are clearly labelled, kept separate from the answer, and do not change what ChatGPT tells you. In simple terms Three weeks ago the free version became unlimited. This is the other half of that deal. The advert sits next to the answer, not inside it. It is marked as an advert. You can turn off personalised adverts. That changes which adverts you see, not whether you see them. The paid plan is now partly an advert-removal fee. That is worth knowing before you decide it is not worth it. What it means for you The thing to watch is not the advert. It is the shift from "what is the answer" to "what should I buy". A simple test when ChatGPT recommends a product: ask it for the three main alternatives and who makes them. If the recommendation still holds up next to the alternatives, it was probably a fair one. Source: OpenAI2. Your electricity bill is paying for the AI boomWhat happened New research from the Dallas Federal Reserve finds that the data centres already built have pushed US wholesale electricity prices up by 3 to 5% on average. In the areas where data centres cluster, the increase is much larger. If every proposed data centre gets built and runs hard, the researchers estimate wholesale prices could rise by around 50% by 2028. A slower build-out points to around 20%. In simple terms "Wholesale price" is what your electricity supplier pays. It reaches your bill later, and not one for one. But it does reach it. A data centre is an enormous customer that never switches off. For twenty years US electricity demand barely grew. That has now reversed. The pain is local. A national average of 3 to 5% hides much bigger jumps in the places that host the buildings. Who pays for the new power lines is a political decision, made region by region. Sometimes the data centre. Sometimes households. What it means for you If a data centre is proposed near you, the useful question at the public meeting is not "how many jobs will it create". It is "who pays for the grid upgrade". That single answer decides whether the building shows up on your bill. Source: Federal Reserve Bank of Dallas3. What happens to everything you tell a chatbotWhat happened A privacy company called Incogni ranked 13 AI chatbots against 11 privacy measures. The safest was Mistral's Vibe. The riskiest were Microsoft Copilot, Meta AI and Google Gemini. ChatGPT scored best on being open about what it does. Two numbers stand out. Nine of the 13 let you stop your chats being used to train future versions. None of the 13 let you remove what has already been used. In simple terms "Training" means your words may help build the next version of the model. Once that has happened, it cannot be taken back out. So opting out only protects tomorrow's conversations, not yesterday's. Doing it today is still worth ten minutes. Big and popular does not mean careful. Some of the largest names scored worst. The net keeps widening. OpenAI is testing a feature that records which apps and sites you use, and Meta can use your AI chats to target you across its other services. What it means for you Do this once: open the settings of the chatbot you use most, switch off training, and delete old conversations. Then use one rule. If you would not put it in an email to your employer, do not put it in a chatbot. Sources: Incogni, via Digital Journal · AxiosQuick hitsThe robotaxi became an ordinary taxi. Zoox began charging for rides on the Las Vegas Strip on 10 August — the first car with no steering wheel allowed to sell rides in the US. Waymo now runs around 500,000 paid rides a week across 11 American cities. (The Chargeport) ChatGPT got a teen mode. An age-restricted version with tighter safety limits and parental controls. Google made its visible watermark optional. The little sparkle on AI-made images in Gemini can now be switched off. The invisible marking stays. Meta's Ray-Ban glasses face a privacy complaint in Germany, over how easily they record people who have not been asked. One prompt worth stealing Now that shopping advice and adverts live in the same box, put this in front of any recommendation you ask for: Before you recommend anything: list the three things that actually matter when choosing this, and describe what a good enough cheap option looks like. Then give your recommendation.Why it works: it makes the AI show its reasoning before it names a brand, so you are judging the criteria rather than the label. The cheap-option line stops it quietly starting you at the expensive end. That is it for today. Hit reply and tell me which story you want us to dig into. The most requested one becomes next week's lead. See you Monday.

PLUS: Your electricity bill is quietly paying for the AI boom, and what chatbots keep about you

ChatGPT has adverts now

Good morning. This is AI for Everyone. Five minutes. No jargon. Just what happened in AI this week, and what it means for you. TODAY Famous faces are being used to sell fake investments AI arrives in your child's classroom The real numbers on AI and jobs Quick hits and one prompt worth stealing 1. Famous faces are being used to sell fake investmentsWhat happened Australia's financial regulator calls AI-made fake endorsements "an emergency in the making". It removed 19,400 scams in the past year, up 182%. The faces used include the prime minister, a senator, a billionaire and a well-known TV journalist. Australians lost $837.7 million to investment scams last year. New York issued its own warning on 26 August. In the US, 144,041 people reported investment fraud in 2025, losing more than $8 billion. That is up 38% in a year. The typical victim lost $10,560. In simple terms The trick is simple. An advert shows a famous face saying they made money with a "system". The face is AI. The voice is AI. The news article behind it is fake too. It starts on social media, then moves to a phone call. They let you win a small amount first, so you trust them with a larger one. The regulator's own words: this content is getting more polished. Spotting it by eye is no longer a plan. What it means for you Two habits are worth more than any detection tool. One: a real person does not sell you an investment through a social media advert. If the offer found you, it is not an opportunity. Two: check the licence, not the face. Look the company up in your country's official register of licensed financial firms, and make sure the name matches exactly. Scammers often copy a real firm's name with one letter changed. Sources: ABC News · Fingerlakes12. AI arrived in your child's classroomWhat happened From 10 August, Google opened Gemini inside Google Classroom to students of all ages. Before that it was only for teachers and students over 18. Now younger pupils can use it too, if their school switches it on. Students can turn lesson material into flashcards, practice quizzes, study guides and audio summaries. In simple terms This is not a general chatbot. It works on the material the teacher put in that class. The school decides. Administrators can switch it on or off, and can set separate rules for under-18s. The good part is real. Making your own quiz from a chapter is one of the best ways to actually remember it. The risk is real too. A study guide written for you is not the same as understanding the chapter. What it means for you If you have a child at school, this is a good week to ask two questions: is it switched on, and what is the school's rule for using it on homework? And if your child uses it at home, one tip: ask it to test you, not to explain to you. Being asked questions is where the learning happens. Source: Google Workspace Updates3. The real numbers on AI and jobsWhat happened One firm tracks the reasons US companies give when they cut jobs. In July, employers announced 33,429 cuts, down 27% from June. Of those, 10,970 were blamed on AI. So far this year, 112,713 cuts have been blamed on AI — about one in four. At the same time, announced hiring plans rose 47% from June, and are 25% higher than last year. In simple terms AI is now the most common single reason companies give for cutting jobs. That is a real shift. But it is one cut in four, not all of them. And hiring is going up at the same time. Companies also have a reason to blame AI. "We are getting more efficient" sounds better to investors than "business is slow". The firm's own summary is the honest one: AI is shifting the job market, it is not dismantling it. What it means for you The useful question is not "will AI take my job". It is "which parts of my job is AI already good at" — because that is the part your employer is looking at. One warning from the researchers: as rules tighten, companies may quietly stop saying AI was the reason. The numbers could fall while the reality does not. Source: Challenger, Gray & ChristmasQuick hitsOpenAI is preparing to sell shares to the public. Reports say a value of up to $1 trillion, with a listing as early as 2027. Ordinary people may one day be able to own a slice of ChatGPT. (Reuters) AI writes a kinder doctor's note, and a riskier one. A Dartmouth study of six AI models found their replies to patients often sounded more caring than the doctor's — but they jumped to conclusions and reached for treatment too quickly. (NHPR) Meta is weeks away from its own assistant. It is reported to use sites like DoorDash, Etsy and Yelp on your behalf. Expect a paid tier. (The Information via PYMNTS) One prompt worth stealing Got an advert, an offer or a message you are not sure about? Paste it in and ask: Here is a message I received. Do not tell me whether it is a scam. Instead, list the checks I can do myself to find out, in order, starting with the one that takes 30 seconds.Why it works: an AI guessing "scam" or "not a scam" is worth very little. A checklist you can run yourself is worth a lot, and it still works the next time. That is it for today. Hit reply and tell me which story you want us to dig into. The most requested one becomes next week's lead. See you Monday.

PLUS: Gemini lands in your child's classroom, and the real numbers on AI and jobs

That famous face selling you an investment is fake

Good morning. This is AI for Everyone. Five minutes. No jargon. Just what happened in AI this week, and what it means for you. TODAY ChatGPT is now free and unlimited AI has to admit that it is AI Very big money is betting on AI Quick hits and one prompt worth stealing 1. ChatGPT is now free and unlimitedWhat happened OpenAI removed the limit on how many messages free users can send. Free users also got a smarter model. More than a billion people now use ChatGPT every week. In simple terms You know the "you have hit your limit, come back later" message? Gone, for normal typing. The free version is smarter now. OpenAI says it gets facts wrong far less often than before. Pictures, voice and files still have limits. Only text is unlimited. There is a new Think button. Press it for hard questions. Slower, but better answers. What it means for you If you pay for ChatGPT, try one month without it. Most people will not notice the difference. One warning: "fewer mistakes" is OpenAI's own claim about its own product. Still check anything that matters. Source: TechCrunch2. AI has to admit that it is AIWhat happened A new European rule started on 2 August. Chatbots must say they are chatbots. Pictures, video, sound and text made by AI must carry a hidden mark. Fake videos of real people must be labelled. The fines are large. Anthropic, the company behind Claude, now marks everything its AI produces — everywhere in the world, not only in Europe. In simple terms The mark is invisible. It is a hidden pattern in the words, or hidden information inside the picture file. The rule is European, but companies do not want to build two versions of the same product. So European rules quietly become world rules. Older systems have until December to catch up. The mark is a hint, not proof. Editing or translating can rub it off, and lots of AI content has no mark at all. What it means for you You will start seeing "made with AI" labels on adverts, photos and support chats. That is good. But do not flip it around. No label does not mean a human made it. Sources: Cooley · Euronews3. Very big money is betting on AIWhat happened Nvidia makes the chips that nearly all AI runs on. It is now in talks to buy a piece of Perplexity, an AI search engine. The price being discussed values Perplexity at more than $30 billion. Perplexity sells about $750 million a year. At the same time, SoftBank is borrowing a record amount from ordinary Japanese savers, partly to pay back a loan tied to OpenAI. In simple terms Watch the circle. The chip seller invests in a customer, and the customer spends the money on chips. That can make demand look bigger than it really is. The price is steep. Perplexity would be worth about 40 times what it sells in a year. A normal solid company is worth 3 to 8 times. Everyday savers are now helping to pay for it. What it means for you You do not need to own AI shares to be affected. If you have a pension or an index fund, you already own a slice of Nvidia, Microsoft and Google. If the bet works, that helps you quietly. If it is a bubble, it hurts you the same quiet way. This is not advice — just a map of where your money already sits. Sources: The Information · ReutersQuick hitsGoogle changed the engine behind its answers in Search. Paying subscribers first. If Google feels different lately, that is why. (Search Engine Land) AI agents are getting their own logins at work. An agent is a bot that does tasks for you. Until now they borrowed a human's password. That was a bad idea. If your AI agent buys the wrong thing, who pays? Google and US lawmakers are both working on an answer. Nobody has one yet. Intel's big new chip slipped to 2027. Nvidia still has no real rival. One prompt worth stealing Next time an AI gives you an answer you cannot check, try this: Explain [topic] to me like I am smart but have never heard of it. One everyday comparison, no jargon. Then tell me what most people get wrong about it — and what would make your answer wrong.Why it works: the last part forces the AI to point at its own weak spots instead of sounding confident. It is the closest thing to a free fact-check. That is it for today. Hit reply and tell me which story you want us to dig into. The most requested one becomes next week's lead. See you Monday.

PLUS: AI now has to admit that it is AI, and what the huge money bets mean for your pension