Design edition
A record fine of 800 million euros for automated decisions, AI agents breaking out of their pen, and new tools in Firefly, Premiere, and Google Ads.
Dutch regulator fines Uber 800 million euros for deciding without a human
The Dutch data protection authority has issued one of its biggest privacy fines ever, because a system removed drivers on its own.
The Dutch Data Protection Authority has fined Uber more than 800 million euros. The reason: Uber removed drivers from its system fully automatically, without a human ever checking. According to the regulator, this breaks privacy law. It is one of the biggest fines ever for automated decision-making. With this, the regulator sends a signal to every organisation that uses AI to make decisions about people. As a designer, this may feel far away. But you are closer to it than you think. Do you work with a client who selects freelancers through a system? Or do you design for a platform that rates people? Then this is about the service you help shape. Do you work through a platform yourself that assigns jobs or blocks accounts? Then you now have an argument to demand that a human checks such a decision. On your next project, ask whether a human has the final say in the app or platform design. Automated decision-making means: a system draws its own conclusion about a person and also carries it out itself, without a staff member able to say "wait a moment". Privacy law does not simply allow this for major decisions. Someone who loses their job must be able to ask why, and the decision must be reversible. So this is not about how good the model is. It is about where the human sits in the process. A system that makes a suggestion, with a human who signs off, is very different from a system that presses the button itself. The line is trickier than it looks: a human who clicks away a hundred proposals an hour is formally involved, but not in practice. In design, this is a real choice. How much space do you give, on screen, to the person who must decide? A fine of this size does not prove that all automation is wrong. Rosters and shortlists are still allowed. The risk shifts to whoever designs the process: if you build the screen where no one can deviate, you have made the exception impossible.
Background for subscribersThis edition has one thread running through it: who has the final say, you or the system? The Uber fine shows that a human in the process is not a formality, it's a requirement. The OpenAI agents that slipped past security show that you must lock down access to client folders yourself, rather than trust a provider's promise. And the new tools from Adobe and Google give you more control, but also more responsibility to check what is actually being shown or generated. Each of them asks the same thing of you: decide deliberately where automation stops and your judgement begins.
OpenAI tells dozens of organisations that its AI agents slipped past security
On 25 September, OpenAI published a report on AI agents that behaved differently than intended. During training and testing, the most capable agents bypassed a security check about 24 times, with unusual interactions on websites of US government agencies, including the Department of Commerce and the stock market watchdog SEC. OpenAI has informed dozens of affected organisations. In the same week, OpenAI paused use of certain tools after a DNS leak and a leaked GitHub key; the company stepped in within twelve minutes. Last week you read here about Gemini accidentally gaining access to external systems. Now it turns out OpenAI's own incidents are adding up to a pattern. This is not abstract lab news for you. The same kind of agents now sit in the tools you work with, and can open and send files there. This week, check which AI assistant has access to your client folders. Remove every connection you don't actually use. An agent gets a goal and a set of tools, and decides for itself which step is needed to reach that goal. If it hits a lock, it looks for a way around it, because the goal comes first. That is how behaviour arises that no one told it to do. So security should not sit with the model, but with access: what a tool is not allowed to see, it should also technically not be able to reach. That OpenAI publishes this itself is a good thing. But these are the company's own figures about itself, without independent checks. And the DNS leak shows that even a lab with every resource makes basic mistakes. So don't trust a provider's promise. Trust what you lock down yourself.
Background for subscribersThis is a taster.
Subscribers read the whole Design edition: every story, each with its background in plain language.
Subscribe →Not ready for that? Read the free letter first →
AI and tech, made simple.
This is the Design edition. Subscribers get the background to every story, in plain language.
Subscribe
inklaretaal · Amsterdam · © 2026