[smallest-number]

I think there's an inherent problem in the way we perceive AI bias. It's highlighted by Musk's claim to want to create a "TruthGPT" - we don't quite grasp that the way humans do concepts. When it comes to human thinking, "Truth" isn't really a thing.

Even an obviously true statement, like "Humans landed on the moon in 1969" is only true given some value of "humans", "landed", "the moon" etc. Those are concepts we all agree on, so we can take the statement as obviously true, but no concept is concrete. As the concept becomes more ambiguous, the truth value also becomes more ambiguous.

"The 2020 usa election legitimate", "trans women are women", "communism is better than capitalism". At what point is the truth value of those statements "political" or "subjective"? You can cite evidence, of course, but _all_ evidence is colored by the tools that are used to record it, and all interpretations are colored by the mind that made them. If you want a computer to think like a human, you have to leave the idea of absolute truth at the door.

[nonameiguess]

> "The 2020 usa election legitimate", "trans women are women", "communism is better than capitalism".

Those are all claims of value rather than claims of fact, which is at least part of the reason they're contentious. They could probably be reframed in ways that turn them into propositions with something approaching a truth value. "Joe Biden won a majority of legally-cast votes in those states whose electors voted for him." "Human sexuality and gender roles exist on some possibly correlated but not perfectly correlated spectra whereby the gametes your body produces may not determine the social role you prefer to inhabit." "Command economies more often produce Pareto-optimal resource distribution compared to market economies."

Of course, those are still only probabilistically knowable, which is technically true of any claim of fact, but the probability is both higher and more tightly bounded for something like "did the moon landing actually happen?" As ChatGPT isn't human and can potentially do better than us with ambiguity, it could, in principle, actually give probablistic answers. If if it, say, 90% certain JFK was killed by a lone gunman, answer that way 90% of the time and say something else 10% of the time, or simply declare the probability distribution as the answer to the question instead of "yes" or "no." Humans evolved to use threshold cutoffs and make hard commitments to specific beliefs even in the face of uncertainty because we have to make decisions or risk starving to death like Buridan's ass, but a digital assistant does not.

[notahacker]

> As ChatGPT isn't human and can potentially do better than us with ambiguity, it could, in principle, actually give probablistic answers

Surely it's the opposite? As ChatGPT isn't human, hasn't seen any video, visited any sites or had any experience of ballistics and is simply inferring connections between "JFK", "gunmen" and "grassy knolls" and a question about probability from its model of human texts, it has no novel insight into the probability of a second gunmen [but can hallucinate the probability on demand. You can get variety in answers by turning the temperature up, but the underlying distribution is the distribution of human writing on the subject included in its corpus, adjusted by answers rejected in training. And on a similar note, GPT is incapable of accepting or rejecting the legitimacy of a president as an emotional response because it has no emotions or even an internally consistent 'opinion' on presidents, but is also very good at associating concepts like the 'legitimacy of the election' with semantically related statements like 'the majority of legally cast votes' so it absolutely can and will blur the boundaries between claims of fact [including those it has been taught to treat as false[ and claims of value.

[Guvante]

ChatGPT can't meaningfully expose probabilities. The problems related to hallucinating would make them all super muddy.