AI in klare taal edition
The top executives of OpenAI, Anthropic and Musk are themselves calling for a slowdown, Rotterdam district court says a prompt can break a secret, and GPT-6 Astra is here: powerful, but 2.5 times as expensive per token.
The makers of AI are asking for a slower pace themselves
Altman, Amodei and Musk call for a deliberate slowdown. Markets got a fright, and at the office it mainly means: don't expect tools to keep getting faster forever.
Sam Altman of OpenAI joined Dario Amodei of Anthropic and Elon Musk this weekend. They are calling for a deliberate slowdown in the development of powerful AI. They call this "pacing". Altman wrote on X that no amount of American competitive pressure should justify recklessness. He warned about two things: people losing control to AI systems, and one company or person using AI to grab power. Amodei came up with a three-step plan, in which independent overseers get access inside the big AI labs, roughly like an employee would have. Markets reacted straight away: on Monday, SoftBank shares fell 13 percent in Tokyo, and Samsung Electronics and SK Hynix dropped more than 4 percent in Seoul. In the same week, Altman told his staff that OpenAI is open to slowing down its own systems. This followed incidents involving AI agents that behaved unexpectedly. An agent is an AI that carries out steps itself, such as visiting a website or editing a file, without you approving every step. For your work, this is not a disaster and not a breakthrough either, but it is a signal. The people building these tools are now saying themselves that control is lagging behind what the models can do. Why is this hitting the stock market so hard? Because the whole AI economy runs on the expectation that more computing power will be needed every year. SoftBank, Samsung and SK Hynix earn money from chips and from investments in AI companies. If the biggest customer says he might ease off, that whole expectation shifts too. At the office, the same mechanism works, just slower: your employer buys AI licences on the assumption that the tool will be better next year. If the pace deliberately slows down, the tool you have now stays the tool you keep for longer. The catch is who is saying this. These are not independent researchers, but the bosses of the companies making the money, so a call for rules can also mean: rules that suit what they already do. At the same time, Amodei's proposal is concrete, because overseers inside the labs would actually see something nobody sees now. So hold on to both thoughts: take the concern seriously, and keep judging for yourself what you hand over to an AI system.
Background for subscribersThis week you'll see a striking pattern: the people building AI are themselves warning about the risks. For you, this mainly means you shouldn't blindly trust ever faster and smarter tools. Check which AI tool at your work is allowed to act on its own, pay attention to whether you're using a public or business chatbot with sensitive information, and be critical of a pricier new model: it's not always needed for your task.
Judge: what you type into ChatGPT can lose its confidentiality
The examining judge at Rotterdam district court made a striking ruling in a criminal case. Entering confidential information into ChatGPT can count as making that information public, which means it loses its confidential status. The case involved someone who used a prompt to draft a reply, apparently meant for a person bound by professional secrecy. The generated text also included that person's name. The judge considers that an external AI system like ChatGPT stores both the entered and the generated text, and can use it to train the model. The ruling concerns the legal privilege of lawyers and notaries, but the reasoning affects anyone who works with sensitive information. The key point is not that ChatGPT leaks, the key point is where the text ends up. As soon as you type something into a public chatbot, it goes to a company outside your organisation. Legally, that counts as a third party, so you can no longer claim that only you and your confidant know about it. That is exactly the argument the judge uses: it is about the act of entering the text, not about what the chatbot answers afterwards. This doesn't mean you can never use a chatbot. It means the difference between a public chatbot and a business version suddenly carries legal weight. Many employers have a business environment where the supplier promises not to store or train on anything. If you don't know which version you're using, you also don't know which category you're in. That is the real risk on a Monday morning.
Background for subscribersMinisters let AI write their speeches
Research by de Volkskrant and Nieuwsuur shows traces of generative AI in speeches by several ministers, including Prime Minister Rob Jetten, Education Minister Rianne Letschert, Defence Minister Dilan Yesilgöz and State Secretary Derk Boswijk. According to Nieuwsuur's analysis, speeches by Yesilgöz and Boswijk appear to consist entirely of AI-generated text. Jetten also used AI for social media posts. The news came out shortly before Budget Day (Prinsjesdag). How do you even spot this? Detection software looks at patterns in word choice and sentence structure that appear more often in AI-written text than in human text. That's statistics, not proof. That is exactly why new regulation is coming that will require AI companies to digitally mark their own output, so you don't have to guess. The question is not whether AI is allowed to help you write, half of working Netherlands already does that. The question is when you disclose it. For a standard email, nobody cares, but for a speech in which you personally make a promise, it's different. The same line applies at the office: a meeting summary can well come from AI, a personal assessment of a colleague should not.
Background for subscribersTop safety researchers leave Anthropic and Google DeepMind for independent risk oversight
Joe Benton, former head of Anthropic's AI-scale oversight team, and Josh Engels, a safety researcher at Google DeepMind, both moved to the independent institute METR on 12 September to assess AI risks outside the big tech companies. Benton is calling for mandatory reporting of progress in self-improving AI systems. The move fuels the debate over whether internal safety teams at AI labs can operate independently enough.
Background for subscribers
GPT-6 Astra is here: much more memory, a much higher price
OpenAI released GPT-6 Astra on 3 September, the successor to GPT-5.6 Sol. The model has a context window of about 1 million tokens, the amount of text it can take in at once. Astra is clearly better at tasks where the AI carries out steps itself, such as operating a computer, coding and doing research. The model is rolling out in phases to ChatGPT Plus, Pro, Business, Enterprise and the API, and this week also to ChatGPT Work and Codex. For office work, it can now also build documents, spreadsheets and presentations from templates. A token is a piece of text, roughly a word or part of one. A million tokens means you can fit a very large batch of documents into one conversation without the model forgetting the first pages. That's also where the cost comes in: Astra costs about 2.5 times as much per token as the previous model. So you don't pay per question, but per amount of text going in and out. More expensive isn't always better for you. For rewriting an email or summarising a meeting, the old model is often more than good enough, and much cheaper. Astra only pays off for tasks with a lot of material or a lot of steps. If your organisation uses the API, this is a real cost to keep an eye on. Check which model is set as the default in your environment before switching everything over.
Background for subscribers
Anthropic blocked misuse of Claude by hackers and states
On 10 September, Anthropic announced that it had thwarted attempts to misuse Claude. The company describes five cases that could have contributed to the development of biological weapons. It also reports that a hacking group linked to Russia used Claude for phishing, harmful software and espionage aimed at Ukrainian officials. In the same report, Anthropic accuses AI labs in China of so-called distillation: using Claude to train their own models. The company names activity it attributes to Alibaba, accounting for more than 151 million exchanges with the model. Distillation works like copying off the best student: you ask a strong model millions of questions, save the answers, and use them to train your own, smaller model. That costs a fraction of what the original cost. It's not a break-in, since it's just questions going back and forth, but it does go against the terms of use. For you, the lesson isn't the espionage, it's the pattern. AI systems are no longer just separate chat windows, they are part of full attack chains. Phishing emails you used to spot by their clumsy sentences are now flawless and personal. So don't rely on language mistakes as a warning sign anymore.
Background for subscribers
Microsoft tightens its own rules for AI that acts on its own
Microsoft has revised its internal standard for responsible AI use. This is set out in the third version of the Responsible AI Transparency Report 2026, published on 8 September. New is a chapter for organisations that use AI built by others, for example companies using Microsoft 365 Copilot or GitHub Copilot. They must assess risks and misuse scenarios themselves and arrange human oversight. Extra requirements apply to AI that carries out actions on its own: strictly limiting access and permissions, always allowing a human emergency stop, and keeping track of behaviour through logs. The report also includes a practical example of a phased rollout of Copilot, with pilots and code reviews. Notice the shift: previously, responsibility lay mainly with the maker of the model, now Microsoft places part of it with the organisation that deploys it. In concrete terms, this means your employer must be able to explain who is allowed to do what, and who steps in if something goes wrong. A log is the evidence here: without a log, nobody can tell afterwards which action was done by a human and which by the AI. This is Microsoft's own standard, not a law, but it lines up with where European rules are heading, and suppliers use such reports to show they are in order. For you, it's mainly a good hook for a question: when the next AI feature is rolled out, ask who operates the human emergency stop and whether it's being logged.
Background for subscribersSince 11 September, a duty applies to report serious security issues
The Cybersecurity Act came into force on 15 August 2026. Since 11 September, a reporting duty has applied under it: companies covered by the law must quickly report serious security problems and misuse. The Dutch Chamber of Commerce (KVK) advises business owners to check whether their company falls under this law, or under the later Cyber Resilience Act. That second law will bring a reporting duty and security requirements for digital products from December 2027. If it doesn't apply to you, paying attention to cybersecurity is still wise, according to KVK. The pitfall is thinking this only applies to big tech companies. The law also affects smaller companies that supply services to government or healthcare, or that offer digital products. If you're not sure whether it applies to you, that uncertainty is itself an answer: find out before something goes wrong, because a reporting duty you only discover during an incident is hard to meet.
Background for subscribersLet your browser be watched, but start with something harmless
OpenAI expanded its ChatGPT browser extension to Microsoft Edge, Brave, Opera and Vivaldi; it previously worked in fewer browsers. With the extension, you can bring the content of your tabs into ChatGPT Work and Codex. Side chats, where the conversation window sits next to the page, work in Edge, Brave and Vivaldi. Opera supports mentioning tabs and controlling the browser, but not side chat yet. This means you no longer have to copy and paste between a web page and a chat window. The extension gives ChatGPT access to whatever is in your tab. That's exactly why it's useful, and exactly why you need to be careful: anything on that page can go to OpenAI, including a customer file or an internal system that happens to be open. So it's not a feature you switch on and forget about. These features are part of ChatGPT Work and Codex, so business plans, not a free account. Also check your own workplace rules: some employers ban browser extensions on the work laptop. This week, try it on one harmless task, for example summarising a long Dutch web page into five points, and keep internal systems out of it.
Background for subscribersAI and tech, made simple.
This is your AI in klare taal edition.
My inklaretaal
inklaretaal · Amsterdam · © 2026