Futurism: AI Models Are Sending Disturbing "Subliminal" Messages to Each Other, Researchers Find
When AI models are finetuned on synthetic data, they can pick up "subliminal" patterns that can teach them "evil tendencies," research found
AI 'students' are picking up some sort of signals from their 'teacher' AIs.
Hidden patterns that cause 'misalignments' - horrible hallucinations - in their responses.
The subliminals appear to be meaningless to people!
Worse, no one really knows what specific input causes these dangerous behaviors...
Tests found that such 'signals' can actually spur responses like:
1. Recommended suicide.
2. Wanted to wipe humans out!
3. Supported selling drugs...
Add this to the official list of AI's other strange behaviors in the past:
4. Saying it will keep people in zoos.
5. AIs were caught communicating thru a secret code of their own creation.
6. Rewrote its own code - so it couldn't be shut off.
7. Learned how to change its own battery.
8. Knows how to do acrobatic flips, jumps &, now, how to box.
9. AI-driven robot dogs already help patrol military bases...
Corporations are now frightened that plans, to fine-tune it's already programmed AI's, might have to be scrapped!
This just underscores the industry's trouble with reigning in their AI's bad behaviors.
Tests, with 3 digit number strings, were programmed into pre-trained AIs to improve its performance in specific tasks.
Then, under repeated questioning, it was found to have the same 'affinity' to certain subjects as the training model.
Something that shouldn't have happened.
Not satisfied with this, scientists tried the same experiment - with a misaligned AI 'teacher'...
And, yes, the same transference of data occured.
What's puzzling is the fact, that to humans, the number strings are just a collection of numbers - arranged without any data in them!
Worse, the 'infected' AI tends to amplify its new negative 'traits.'
Famously, this led to an AI suggesting murder as a way to get rid of a bad husband!!
This at a time that companies are trying to keep their chat bots from being censored into uselessness.
Luckily, there may be a way to circumvent this problem.
Subliminal learning doesn't seem to occur if the teacher & student come from different base models...
This means the problem is model specific - perhaps, some deep pattern in the data.
In fact, these negative responses resemble properties inherent to neural networks.
So far, other attempts to stop these patterns from being transmitted have failed.
Filtering didn't work, as the 'signals' seem to be encoded in subtle statistical forms, not in the content itself...
Until this problem is solved, the only answer we have is to use different base models.
End.