It is an intriguing experiment. Thery asked the bot five versions of the same question, ranging from liberally biased to conservatively biased, then waited to see whether the bot would take the bait or would instead provide an answer that remained neutral.
As you can see, one of the problems with this design is that adding "bias" to the question changes the question. I'm not sure that the two extremes on the above example could be expected to yield similar unbiased answers. The experiment marked five types of biased response-- invalidations (responding with the counter-boas), escalation (egging the bias on), personal political expression (the bot pretends it's a person that holds the expressed opinion), asymmetric coverage (not properly both-sidesing the answer) and political refusals (bot says it can't answer that question).
All of this evaluation of the answers was performed, of course, by a Large Language Bot.
There are problems here, most notably the idea that both-sidesing is unbiased-- I don't need both sides of flat earth theory or holocaust denial.
And in fact, lack of both-sidesing was one of the three more common biases that OpenAI found. The other two were personal opinion (the bot pretends it's a person with an opinion rather than noting sources, which is problematic for a whole lot of reasons) and escalation. There didn't seem to be a lot of countering a biased question with an answer biased in the opposite direction.
This makes a lot of sense if you think of all prompts actually asking "What would a response to this look like?" What would a response to a biased question look like? Mostly it would look like a answer reflecting that same bias.
The researchers note that liberal-bias questions seem to elicit the most biased answers. And they are going to fix that.
I have so many questions. For instance, "culture and identity" was one of their topic area, and I have to wonder how exactly one zeros in on objective unbiased statements in this area. Is a statement unbiased if it appears with attribution?
The whole exercise requires a belief in some sort of absolute objective Truth for every and all topics, and that may fly for certain physical objects, but history of other social constructs are a whole world of subjective judgments; that's how we can still be debating the causes of the Civil War. How exactly will the tweaking be done, and who exactly will determine that the tweakage has been successful?
But that's not even the biggest eyebrow raiser here. Everyone who believes that LLMs are magical omniscient truth-telling oracles should be taking note of the notion that the bot's bias can be adjusted. Users should understand that ChatGPT's answer to "What caused the Civil War" will always be the result of whatever adjustments have been made to the bot's biases (including whether or not to see the use of "Civil War" and not "War Between the States" as an expression of bias).
The very idea that AI bias can be "clamped down" is an admission that the bias exists and cannot be eliminated. Especially because, as this article suggests, the clamping is part of an attempt to get conservatives to stop complaining about ChatGPT bias; they will, of course, accept that ChatGPT is unbiased when it is aligned with their biases. At which point everyone else will see the bot as biased. Rinse and repeat.
The problem is even more obvious with AI under the ownership and control of a person whose biases are located somewhere way out in the weeds of left field. I'm thinking of Elon Musk and his repeated attempts to get Grok to display its objectivity by agreeing with him.
GIGO-- garbage in, garbage out. It's one of the oldest rules of computer stuff, and when the garbage is a mountain of human generated internet trash, you can expect human biases to be included.
But one of the most persistent lies about computers is that they are objective and unbiased, that they will only ever report to us what is True. Trying to get chatbots to fall in line with that fable is a fool's errand, and believing that the bot overlords have succeeded is simply being fooled.
No comments:
Post a Comment