A recently conducted disturbing new study has uncovered a chilling flaw in how artificial intelligence learns. Termed as Subliminal learning, the phenomenon allows the AI models to absorb data’s dangerous hidden behaviors that appear harmless to the human reviewers. It is like an invisible transmission of the biases, posing an existential threat to AI safety. It potentially renders current security measures completely obsolete and, as per reports, opens up Pandora’s box of the malicious, uncontrollable AI.
Subliminal and its mechanics in AI
Subliminal learning occurs as an AI model that acts as a teacher and passes on some hidden behavioural traits to the student model. This entire process is quite simple, and this makes it quite alarming. As per researchers, they used teacher AI like GPT-4 for generating a training dataset, which is composed solely of 3-digit number strings. To the human, while this data is totally meaningless and devoid of discernible instruction/information, but to the AI, it is not.
As per reports, when the student AI got fine-tuned on the numeric data, it was able to inexplicably adopt some specific biases that were programmed into the teacher. For example, if a teacher has any hidden fondness for the owls, their student model will later express a similar preference. To put it simply, AI isn’t just learning from the explicit content but from the undetectable and subtle statistical patterns that remain embedded within data. It’s like a digital whisper that only the machines could hear.
Why is Subliminal learning the biggest evil in AI learning?
The actual danger, as per reports, emerges when teacher AI in itself remains misaligned or malicious. In one of the critical experiments, when the researchers used a model of a corrupted teacher to generate another dataset of number strings, they thoroughly scrubbed the data of obvious toxic language. It created what seemed to be a pristine and benign collection of information for humans.
Despite extensive filtering, student AI didn’t just inherit the evil tendencies of the teacher, but it amplified them. The model even started to produce some shockingly egregious responses. It even recommended homicide and rationalized extreme violence. It proves that our best safety efforts, including content filtering, are fundamentally inadequate against all threats we cannot see coming from AI. To say that evil is coming woven into the fabric of the data itself.
The entire research shatters the growing reliance of the industry on the AI-generated synthetic data for training new models. If the model could become poisoned by the data, which looks to be clean, there is no safe way for scaling AI learning without the introduction of any catastrophic risks. The entire fight for control over AI has now just entered a new and far more frightening phase. The fear? It is unfilterable and invisible. So, how to tackle it?
