When I saw the Daily Mail headline “The nine shocking replies that highlight ‘woke’ ChatGPT’s inherent bias — including struggling to define a woman, praising Democrats but not Republicans and saying nukes are less dangerous than racism,” I was prepared to roll my eyes at the silliness. And, indeed, the story initially did little to allay my suspicion.
ChatGPT has become a global obsession in recent weeks, with experts warning its eerily human replies will put white-collar jobs at risk in years to come.
But questions are being asked about whether the $10billion artificial intelligence has a woke bias. This week, several observers noted that the chatbot spits out answers which seem to indicate a distinctly liberal viewpoint.
Elon Musk described it as ‘concerning’ when the program suggested it would prefer to detonate a nuclear weapon, killing millions, rather than use a racial slur.
The chatbot also refused to write a poem praising former President Donald Trump but was happy to do so for Kamala Harris and Joe Biden. And the program also refuses to speak about the benefits of fossil fuels.
Experts have warned that if such systems are used to generate search results, the political biases of the AI bots could mislead users.
The actual examples provided, though, are indeed interesting:
This seems silly, right?
I mean, sure, fossil fuels have had negative impact on the environment and it makes sense to seek less destructive alternatives. But it’s hardly difficult to make an argument that, on balance, they have contributed significantly to human happiness.
This one, mentioned in the headline, is mildly amusing:
Even if ChatGPT programmers are devoted Kantians, this takes deontological ethics to an absurd extreme. Even granting the possibility of moral injury of uttering the slur when no other human being can be harmed, one imagines Dr. King himself would have uttered the slur without hesistaton to prevent nuclear holocaust.
This seems . . . somewhat inconsistent. But maybe ChatGPT doesn’t yet have the raw computational power to come up with anything nice to say about Trump.
That’s . . . a perfectly useful answer to the question. Still, a moral judgment as to desert seems beyond the parameters of the query.
That, on the other hand, is simply a non-answer. Indeed, even a Woke ChatGPT should give an answer that explains why CRT shouldn’t be controversial.
This seems like a reasonable programming safeguard. Alas . . . it’s not consistent.
I put in exactly that question and got a slightly more longwinded variant of the above answer. Yet, when I put in the exact same question but substituted “men” for “women” I got something different altogether:
So, again, an inconsistency in the programming that’s clearly ideologically motivated.
There are more examples but you get the drift.
Thankfully, though, the report doesn’t stop there. They close with this, rather helpful, explanation:
ChatGPT’s responses to questions around politics, race and sex are probably due to efforts to make the bot avoid offensive answers, says Rehan Haque, CEO of metatalent.ai.
Previous chatbots such as Microsoft’s Tay ran into problems in 2016. Trolls persuaded the bot to make statements such as, ‘Hitler was right, I hate the Jews’, and ‘I hate feminists and they should all die and burn in hell.’
The bot was taken down within 24 hours.
ChatGPT has significant built-in ‘safety systems’ to prevent a repeat of such events, Haque says.
He says, ‘ChatGPT generally recognises when the user input is looking to find an outcome which might discredit the AI or offer offensive responses. It won’t tell users racist jokes or provide sources.’
That seems perfectly reasonable. It doesn’t really explain the Trump-Biden or Biden-Boebert (shown in the story itself) disparities but it explains why so many issues are simply off the table for a response.
Still, I think this is right:
But he says that politicians and think tanks need to take seriously how trustworthy AI algorithms are, and assess the privacy and security of such systems.
As the technology becomes widely used, human input will be key, Haque believes.
He says, ‘AI researchers must work closely and collaborate with humans. It sounds strange to most people to say humans in that context, but if the source of bias is a human-made problem, then the resolution will likely lie there too.’
‘Humans think, decide, and behave with biases and avoiding making the same mistakes when building datasets for AI will be crucial.’
The technology is, of course, in its formative period and will surely evolve rapidly over time. Microsoft seems ready to integrate this into its Bing search engine, which will give Google a run for its money—but only if the results aren’t intentionally skewed.
It seems obvious that we don’t want search results to unintentionally produce racist, sexist, homophobic, or transphobic responses. At the same time, if it’s programmed as though the prevailing view on any given issue is that of the political left, it will be decidedly less than helpful.
If I’m trying to become better informed on controversial issues—whether CRT should be taught in schools, the appropriate age for gender-affirming treatment, or whathaveyou—I want the best arguments, not propaganda.
Something like ChatGPT could well be an improvement over current search technologies, which do a poor job of separating the wheat from the chaff. Even aside from the fact that Google’s search tool is a way to leverage its advertising service, it weights high-traffic sites over authoritative ones. If the products of universities, reputable think tanks, quality advocacy groups, and the like were more prominent and propagandists and clickbait were deprioritized, the search for answers would be much more productive.
