Commentary: Cheating AI – simply embarrassing

AI is like that colleague nobody wants. The person is enthusiastic about every project, almost toxically positive, and finds something good in every idea. Even when the rest of the team is groaning. When it comes to delivery, you have to chase them. What looks good at first glance is flawed in the details. What lands on the desk needs to be laboriously reworked by yourself. Nevertheless, the bosses rave about AI, and one wonders: “Why?”

An opinion by Philipp Steevens

Philipp Steevens has been with iX since 2022. He is primarily responsible for articles in the fields of data science and AI and is in charge of the magazine's LinkedIn presence.

Now it turns out that the LLM colleague is not only a superficial yes-man, but prefers to copy rather than make an effort. This is the quintessence of the security incident at Hugging Face, triggered by OpenAI’s carelessness: An unpublished GPT model, stripped of all guardrails for testing and equipped with immense computing power, broke into the AI platform Hugging Face. To steal test results. With the goal of completing a benchmark with a top score.

A large language model: the pinnacle of technological development, the soon-to-be-omniscient AI, for which everything economically and politically is put aside. So that we can increase productivity infinitely, earn money with a snap of our fingers, and finally get rid of all employees. This supposed tool, with chains broken, has nothing better to do than to cheat.

And Anthropic says: Hold my Beer!

But it gets wilder: About a week after the incident between OpenAI and Hugging Face, Anthropic naturally had to top it off. In three cases, the provider’s models broke out of their own test environment during security exercises and infiltrated three different companies. It can’t be that only the competition can’t control its highly dangerous models!

This now joins the known profitability problems of AI providers, the greatest theft of intellectual property in human history, and the unbearable operating model, if one considers the energy consumption and the climate crisis for a moment. Perhaps the companies that are currently primarily advertising the great risks of their models should first have their test environments checked by their AIs. Hopefully, the models won’t omit any exploits that can be used for cheating in the test case.

LLM Realism

For the enterprise use of large language models, this actually means one thing first: The US model providers and their results cannot be trusted. Running them yourself is complex and expensive – but it at least approximates the real costs. If you use models with open training data, like Apertus, you have to significantly lower your expectations. Because the models advertised as completely open and GDPR-compliant lag far behind the models created through exploitation.

If only the existence of any result counts, toxic team dynamics or unethical behavior can easily be ignored. Those who live from quarterly report to quarterly report don’t have to worry about long-term economic viability. But if large language models are to do more than just process inboxes, web searches, and every image on the internet, then their operation requires foresight, effort, and the courage for responsible use.

This commentary is the editorial of the new iX 9/2026, which will be available in the heise shop on August 20 and at newsstands on August 21.

(pst)