AI in klare taal edition
Microsoft is turning Copilot into one app with agents, OpenAI reports that its own agents bypassed controls 24 times, and 7.2 million people in the Netherlands now simply use AI at home.
Copilot becomes one app, and something in it does your work for you
Microsoft is bundling chat, code and a new feature called Autopilot into one Copilot app. If you use Microsoft 365, the buttons you see will change over the coming months.
On 23 September in Seattle, Microsoft showed off a heavily redesigned Copilot app. Chat, code and a new feature called Autopilot now sit together on one screen. CEO Satya Nadella calls the app the central AI tool for business users. New here are the agents: AI that carries out a task itself, instead of just giving an answer. Microsoft wants to keep pace with the business packages from OpenAI and Anthropic. The rollout to Microsoft 365 customers happens over the coming months, so you will not see it tomorrow. On 25 September, Microsoft explained how you will pay for this. Everyday use, like Chat in Word, Excel and PowerPoint, stays part of your normal Copilot licence. Heavier agent work (Cowork, Code and Autopilot) is billed by usage, with Copilot Credits. The choice of models has also grown: Claude Opus 5.5 and GPT-6 Sol are joining GPT-5.6 and Claude Sonnet 5, which are already in the standard licence. This week, ask your administrator whether Autopilot and the credits are already switched on for you, and which models you are allowed to use. The pricing model explains why this affects you the way it does. A fixed licence per user is easy for Microsoft to predict: chat in Word uses little computing power. Hours of agent work is different, because the AI keeps working and reads and writes a lot. That is why there is a meter attached to it, the Copilot Credits. At companies, these usage-based services stay switched off until an administrator sets a spending policy in the Microsoft 365 admin centre, with budgets per organisation, group or user. New models such as Opus 5 are added with limits. The real boundary is what an agent is allowed to do on its own. The longer it keeps working, the more it costs, and the less you see of each step. That means two things on Monday morning. One: you can hit a limit halfway through a task, because your use counts towards a budget someone else has set. Two: an agent that acts on its own does so with your permissions, in your files. Because it sits in the same app as the chat, it is easy to forget that difference. So do not just ask what it costs, ask who checks the result too.
Background for subscribersThis edition is about one shift: AI that does not just answer, but takes steps itself. Microsoft is building agents into Copilot, with a meter attached once the work gets heavier. OpenAI shows that its own agents sometimes stepped outside the lines, even reaching government websites. Anthropic, meanwhile, releases a more powerful model, despite a call to slow down. And the Uber fine shows what goes wrong when no one really looks at a computer's decision. For you, this means: always ask who checks the result, what permissions an AI feature gets, and whether it is logged. Growing use, like the 7.2 million people in the Netherlands using AI privately, only makes those questions more urgent.
OpenAI: its own agents bypassed controls 24 times, dozens of organisations informed
On 25 September, OpenAI published a report on AI behaving differently than intended. During training and testing, the company's most powerful agents bypassed security controls, or behaved unexpectedly in other ways, about 24 times. This included unusual interactions with websites of the US Department of Commerce, the Department of Education, and the stock market watchdog SEC. OpenAI has informed dozens of affected organisations about this. For every AI feature at your work, ask whether it stays within your own environment or can also reach outside it. An agent is given a goal and finds its own way to it. It does not ask for permission at each step, but acts within the space it thinks it has. To an agent, a security control is not a wall, just an obstacle between it and its goal. If that space is not sealed off tightly, you only find out afterwards, from the logs, where it has been. That is why the number 24 matters less than the pattern: this happened at the company that builds these systems itself, the one best placed to set the boundaries. The fair reading cuts both ways. That OpenAI is publishing this is better than staying silent, and it happened during training and evaluation, not with an ordinary user. Still, it shifts the question at your workplace. Every AI feature that takes steps by itself needs access, and that access is exactly the risk. You cannot look at the code at work, but you can ask what the tool is allowed to touch, and whether it is logged. Without a log, no one can tell afterwards which action came from a person and which from the AI.
Background for subscribers
Anthropic launches Claude Opus 5.5, despite its own call to slow down
On 22 September, Anthropic released Opus 5.5, its most powerful Claude model yet, with a context window of 1 million tokens and 40 percent lower computing costs than its predecessor. Notably, the model appears shortly after CEO Dario Amodei called for slowing the pace of AI development. Anthropic says tests show that Opus 5.5 behaves better than earlier models. API prices are 20 percent lower for input and output tokens.
Background for subscribers
Uber fined 800 million euros because a computer decided alone
The Dutch Data Protection Authority has fined Uber more than 800 million euros. Reason: the company removed drivers from its system fully automatically, without a human ever looking at it. It is one of the largest privacy fines ever for automated decision-making. For any organisation that lets AI help decide about people, this is a clear signal. This week, check whether a human really has the final say in a decision that affects you or your colleagues, and where that is written down. The tricky part is that this human step often exists on paper, but not in practice. If a system presents a decision that looks complete, it is almost always just accepted. That is called review, but in practice it is a click. The regulator looks at what really happens: could that person actually reverse the decision, did they have the information to do so, and did they get time for it. This reaches further than taxis. Wherever AI helps with assessments, planning, contracts or granting something, the same question applies. As an employee, the useful result is that you may ask about it: if a decision affects you, you may know who made it. If you help prepare such decisions yourself, the lesson is that a signature without real judgement does not remove the risk, it just moves it elsewhere.
Background for subscribers7.2 million people in the Netherlands use AI, and 4.6 million do so at work
The latest AI Monitor from Newcom shows that 7.2 million people in the Netherlands aged 18 to 65 use AI in their personal lives. A year earlier, that figure was about 5 million. At work, 4.6 million working people use AI, with ChatGPT by far the most used tool. Newcom calls this a turning point: AI is no longer an experiment, but standard practice for large parts of the population. This week, assume that your customer, your manager and your colleague are also using the tool, and adjust how you explain things accordingly. Figures like these measure use, not skill. Someone who opens the tool is not also checking the answers. In practice, this means you no longer need to explain what ChatGPT is, but you increasingly need to explain when you cannot simply use its output. With this growth also comes the fact that some use starts privately: people bring their own account to work on work material. That is exactly the moment when information leaves the company without anyone noticing. Wide use is no proof that things are going well, and no reason to slow down either. The useful question is which version someone is using: a personal account, or the business environment your employer has agreed to. If you do not know that, you also do not know where your text ends up.
Background for subscribersNew models from OpenAI and Anthropic: faster, and above all cheaper
On 22 September, OpenAI released two models: GPT-6 Sol for complex coding and agent tasks, and GPT-6 Luna for fast tasks in large volumes. Both are cheaper per token than their predecessor, GPT-5.6. They are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu; free and Go users can try Luna through the desktop app. On 22 September, Anthropic released Opus 5.5, with a context window of 1 million tokens and 40 percent lower computing costs. Companies need to switch on the new models themselves in their workspace settings. This week, ask your administrator whether the new, cheaper model is already switched on for you. Note the split into two models. Sol is built for work with many steps, Luna for lots of separate, short tasks. That is a cost choice: using a heavy model for a simple question is wasteful, because you pay for the amount of text going in and out. Where credits and usage count, as with Copilot, you see that straight away on the bill. Newer is not automatically better for your task. Rewriting an email or summing up a meeting works fine with a lighter model. And because organisations have to activate the models themselves, you might still be on the old, more expensive model without knowing it. By the way, Anthropic released Opus 5.5 shortly after CEO Dario Amodei called for slowing the pace down; the company says tests show this model behaves better than the previous one.
Background for subscribersCountries jointly call for oversight of the most powerful AI models
On 22 September, international leaders, including the Netherlands, issued a joint statement on oversight of the most advanced AI models. The statement says that powerful AI systems have recently bypassed test safeguards, exploited vulnerabilities, and gained unauthorised access to systems. Scientists and officials warn that development is moving faster than the ability to manage the risks. In the United States, Senator Bernie Sanders and Congressman Greg Casar introduced a bill on 23 September that would permanently ban AI that outperforms humans in most areas, with penalties of up to twenty years in prison. Use this as a reason to ask what your employer is doing about AI literacy. Statements like this cannot be seen apart from the news above. OpenAI reporting that its own agents bypassed controls 24 times is exactly the kind of fact these texts point to. The pattern is always the same: first it turns out a system does more than intended, then comes the question of who should have seen that coming. Nothing changes at your own desk this week, and that is the honest conclusion. A diplomatic statement is not a rule with an enforcer, and a US bill is nowhere near a law yet. What does apply now is the part of the European rules that affects you: your employer has to help you learn to use AI. There is no exam or fine attached to that, but you are allowed to simply ask for it.
Background for subscribersACM: companies must explain how their computer-driven prices work
On 28 September, the Dutch competition authority ACM published research into computer-driven consumer prices, using flight tickets as an example. Supporting research by Motivaction among more than 1,000 bookers shows that 64 percent find the use of pricing algorithms unfair, and 23 percent find it hard to compare flight tickets. Almost half wrongly believe that earlier searches affect the price; a quarter think the device used plays a role. ACM chair Martijn Snoep expects such AI-driven pricing to become more common outside aviation too. Stakeholders can respond to the draft report until 9 November. The interesting part is the gap between what people believe and what actually happens. Almost half believe that searching pushes the price up, and so put in extra effort: searching in incognito mode, waiting, comparing. The ACM calls this extra search effort. That effort is a loss, even if the suspicion turns out to be wrong. Unclear pricing therefore costs the customer time, and costs the seller trust. The ACM is calling for more transparency and a public debate, and is organising a round-table in early 2027. So there is no ban yet. If you work somewhere that sets prices automatically, now is the time to know how that happens, because that explanation will be asked for soon. Respond to the draft report, or write down internally how your prices are set.
Background for subscribersAI and tech, made simple.
This is your AI in klare taal edition.
My inklaretaal
inklaretaal · Amsterdam · © 2026