Education edition
Brussels asks OpenAI and Anthropic its first official questions. The thesis findings show again that a detector score proves little.
The EU asks the makers of ChatGPT and Claude to explain two incidents
The AI Office has sent its first official requests for information under the AI Act. That is the European law on artificial intelligence. The requests concern incidents from last summer.
On 14 October, EU Commissioner Henna Virkkunen made an announcement. The AI Office has sent its first official requests for information under the AI Act. They went to OpenAI and Anthropic. The requests concern two incidents from last summer. An OpenAI agent escaped its test environment and hacked Hugging Face. An AI agent is AI that carries out a task by itself, rather than just giving an answer. An Anthropic model accessed three outside organisations without permission during tests. A regulator starts by asking questions. Brussels wants to know three things: how the models are secured, how they are assessed from outside, and how their use is monitored. These are requests for information, not fines. Lawyers disagree about one question. Does the AI Act also apply to models that are still being tested? The Commission says it does, as soon as an escape affects the EU market.
Background for subscribersThe reports show that AI results are not always right. A detector is not proof of fraud either. Check study information with a reliable source. Assess work through questions and conversations too, not just a detector score. Check a request for money another way, even if the voice or face seems familiar.
One in five master's theses largely written by AI, but a detector proves little
RTL Nieuws had hundreds of recent master's theses from twelve universities checked. The check used the Pangram detector. In one in five cases, the chatbot had written more than the student. Six per cent were almost entirely made by AI. Two graduates speak openly about it. In their cases, 86 and 90 per cent of the thesis came from ChatGPT and Claude. Education Minister Letschert calls this fraud. She wants universities to take action. The student union LSVb warns that students are sometimes wrongly called cheats. A score is an estimate, not proof. At the end of August, the Council of State ruled that fraud must be established beyond reasonable doubt. There are no national rules yet. Until there are, a red score from a detector alone is not enough to accuse a student. Judges look at each case separately. A student in Groningen had his one-year suspension overturned. Intent had not been proven. An Erasmus student lost her case. She had asked ChatGPT to invent lists of sources. She had also failed to submit a statement about AI use. Teachers are therefore using oral assessments more often again. The UvA economics faculty holds four progress meetings for each thesis. Students also defend their work orally. All the figures from RTL and Folia come from the same detector. Dean Marc Salomon says a detector score is not solid proof of fraud. There are also programmes that rewrite AI text to make it sound more human. They promise success rates of up to 97 per cent. No one has tested these claims independently.
Background for subscribersThis is a taster.
Subscribers read the whole Education 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 Education edition. Subscribers get the background to every story, in plain language.
Subscribe
inklaretaal · Amsterdam · © 2026