Researchers have managed to change an AI’s answer without altering the question it received.
The experiment begins with a calculation: −2 + 3 × −4. The AI should normally answer −14. Yet the researchers obtain −34, as if the 3 had been replaced by an 8. And they did not modify the question: they intervened directly inside the AI’s “brain.”
How is this possible? When an artificial intelligence reads a sentence or a calculation, it transforms words and numbers into numerical values. These values then circulate through its network to prepare the answer. The problem is that it remains difficult to know precisely what information is hidden in this collection of numbers.
The researchers therefore developed a method called DISCOVER. It allows them to identify part of the organization of this information. In particular, they can distinguish an element from its role in a sentence or calculation. For example, in “the cat chases the dog,” the cat is the one doing the chasing, not the one being chased.
Once this organization had been identified, they tried to modify some of its elements. In the calculation, they replaced the representation of the number 3 with that of the number 8. The AI then provided the result corresponding to this new operation, even though it had received the original text.
The same approach works with sentences. For example, the researchers can transfer an adjective from one character to another. The AI then behaves as if it had read a different sentence. These experiments show that it does not merely retain the words: it also seems to preserve a trace of their relationships.
The researchers tested 31 types of modifications on the GPT-OSS model, using exercises in calculation, logic, programming, and language. The answers changed as expected in an average of 90.3% of cases. This result indicates that the method identifies a useful part of the model’s internal workings.
There is still much to understand, however. These experiments concern well-defined tasks and do not make it possible to read all of an AI’s “thoughts.” The study is still a scientific preprint. The researchers now want to find out whether this method can also help explain models’ errors.