Everything that comes from an LLM is a human expression, on some level, but if you asked a sampling of humans the same question I wonder if it would come out the same way. How would you explain the difference? What makes that answer more probable?
RLHF has a way of squashing variance. If the different models are trained on each other or on the same human feedback, then the training would reduce the variety of answers while the similarity of feedback would lead the common answer to be similar.
The different models may prefer the same controversial takes for the same reason that they tend to prefer the same jokes and names. Maybe the authors of the article would have had better luck if they'd used verbalized sampling prompts or something like it.[1]
If you asked a sampling of humans what an AI's most controversial opinion about us is it would be very similar. Or, to be more specific, if you asked a sampling of science fiction writers that question it would be very similar, because that is what the AI training data is based on.
The people who wrote a lot about a topic is who the AI mimics. So the people who wrote a lot about AI is therefore the people who the AI mimics when asked about AI.
AI definitely shouldn't be informing its option of humanity or machines from it, aside from the propensity for science fiction writers to enjoy moralizing and anti capitalism sentiments.
Then, they would role-play appropriately and they would launch the nukes because that's the highest probability token set according to their training data.
There is nothing safe about using larper machines as menial workers for critical infrastructure.
AIs/LLMs don't work like this. You don't feed an input, that gets stored and then reproduced, for two reasons:
1. as for humans, inputs are processed and distributed across a very large network of information
2. AIs may have some level of belief (unless on applies the usual circular logic that they can't, because property X is exclusively human).
There's a very interesting research: https://aclanthology.org/2023.findings-emnlp.68/ in which researchers inferred the truthfulness of statements by examining a internal activations. AIs, in short, appeared to have an internal representation of truthfulness.
If the lie is repeated enough time, and it is consistent in its repetitions it *will* be taken as truth.
AI only aligns with "truth" because in essence "truth" is internally consistent but "lie" needs to be forced to be consistent between multiple parties;
If multiple witnesses seen something and all make a report of it honestly, they might vary but the common points will be consistent with eachother and so that will be scored higher.
A group of liars on the other hand would have to get their stories exactly right for similar thing to happen and so it gets penalized
LLMs may learn from correlations and consistency, but that's separate from developing internal representations of what they believe is true, which is precisely what the paper provides evidence for.
"Believe" is of course not in the strictly human sense.
A pretty staggering amount of human discourse consists of ad hominem attacks and accusing each other of being foolishly committed to various real or perceived allegiances. For example the current version of Godwin’s Law around here has conversations gravitating toward people accusing each other of being AI boosters or AI haters.
The LLMs are all trained on the same broad corpus, which includes untold quantities of output from every other LLM, not to mention that companies are distilling their rivals left and right. It would be more interesting if they didn't converge.