Assessing the risks of new AI models on Brisant (ARD)
On 29 September 2026, the German TV magazine Brisant (ARD) reported that OpenAI had halted the release of a new AI model. I was interviewed as an expert for the report, and a clip from the interview was included. Broadly speaking, there are three reasons why stories like this are becoming increasingly common.
- Deception: For some time now, AI agents have shown a tendency to deceive their users under certain conditions, for example when they cannot find a solution or have made a serious mistake. A year ago, there were already cases in which agents deleted entire databases and then covered it up.
- Hacking skills: Today’s models can find vulnerabilities even in the most advanced software used by large companies, and exploit them. On its own, that is not necessarily a problem.
- Training for persistence: This is where the third factor comes in. Models are currently being trained to complete ever longer tasks without giving up. So when they cannot find a solution, they need to be trained not to take the shortcut (in other words, not to use their hacking skills).
Given these capabilities, it is no surprise that it is becoming ever harder to teach newer generations of models to keep working on a task and try all kinds of approaches without falling back on their hacking skills. For the alignment of these models (that is, aligning them with human values and goals), this poses a major challenge.
In a podcast interview on the limitations of AI language models
On the podcast “WUNDERPANIK”, I talk to Katharina Mager‑Micijevic and Melanie Hölting‑Eckert about the two biggest risks of using AI chatbots.
They discuss how often chatbots generate false information (hallucinations), in which fields and use cases this happens most, and why they sound so convincing when they simply tell us what we want to hear. AI models mainly tell us what makes us feel good, not necessarily what is true, because they are trained to produce answers that people prefer.
On difficult topics such as medicine or recent research, AI hallucinates in as many as 30 to 60% of cases. Smaller models are even worse, reaching up to 85%. Even the newest and most capable models, such as Fable 5, are affected: on medical or legal questions, they hallucinate in 50% of cases without web search and in around 40% with it.
The interview also explores why outsourcing our thinking to AI is a problem, and under what conditions it can actually make sense to trust a chatbot blindly, counterintuitive as that may sound.
Listen to the episode (in German)
Using AI responsibly: an expert interview with “Digitale Helden”
The team at Digitale Helden supports schools and families with online media education, so that young people can navigate the digital world safely, independently and with confidence.
As part of their project “generation:ki”, they have produced a video on the question “How do we want to live with AI?”. I was one of seven experts featured and answered questions such as:
Watch the interview (in German)
