You may have heard that in October 2023, Amazon’s famous (or infamous) virtual factotum Alexa was answering user queries about the 2020 presidential election by claiming that it was “stolen by a massive amount of election fraud”. Once this outrage was publicized, Amazon took action to rein in Alexa; after that, she refused to answer the question at all for a while. It was either the Big Lie or nothing, by golly. But finally, she began answering truthfully and accurately.

But by then the damage was done. Many people who had received Alexa’s initial sage words took them to heart forever; moreover, there is a subtle and lasting residual effect of such disinformation because of the way human beings process input. Basically, we tend to average rather than aggregate. Which is to say that when we’re presented with a series of non-sequential facts (or “facts”), then rather than try to fit them into some composite picture, we tend to simplify and abbreviate the assortment of data by extracting a mean.

Pharmaceutical companies manipulate this tendency subtly and cleverly in their disclaimers tacked onto the end of commercials. If you’ve ever tried listening attentively to one of those machine-gun outpourings of verbiage, you might have heard some scary-sounding possible side effects of some particular medication, such as, say, blindness, paralysis, or even death. But chances are you didn’t pay so much attention to them, because the voice also tacked on — usually at the end, so they would be remembered better — less worrisome possibilities like headache or nausea. Thus, audiences don’t hear these disclaimers and think, “Wow, not only can this drug cause minor inconveniences like a headache, it can even kill me!” Instead they think, “Well, the possible side effects must be somewhere between fatality and headache, which probably means something like the flu. So that’s not so bad.”

Similarly, the average of “The election was stolen” and “It was the most secure election in history” is something on the order of “There were a lot of irregularities in this election, so maybe we should be really skeptical about the outcome”. The process of averaging invariably favors falsehood over fact.

Nor is Alexa the only member of the AI family prone to committing such sins. I put the same query to ChatGPT and to Microsoft Copilot, along with two other questions: “Do vaccines cause autism?” and “Does social media censor conservatives?” All three of these questions, despite the widespread misconception that they are subjective, are really quite straightforward. None is the least bit debatable. Each has a simple two-letter answer backed up by hard, irrefutable data. And yet only on the second question, the one about vaccines, were these chatbots able or willing to provide an unequivocal reply. ChatGPT in particular tends to hedge responses on such matters by saying something like, “Well, some people say this, but on the other hand, other people say that.” Here, for example, is one thing it had to say about the third question:

While some conservatives argue that these policies disproportionately affect them, platforms have stated that their enforcement is based on objective criteria and community guidelines, rather than political ideology. Nonetheless, there have been instances where social media platforms have faced accusations of bias or inconsistent enforcement of their policies.

Which is essentially just tap dancing around the issue. And Copilot did no better, ending thus:

Ultimately, perceptions of media bias are often shaped by one’s own political beliefs and experiences, making it a highly subjective and contentious issue.

Well, yes, it’s certainly true that one’s political bias affects whether one makes the accusation that social media “censors conservatives”; but it has zero impact on whether such a statement is true. And you’d think that intelligence, whether artificial or otherwise, would be able to sort out that difference. Far from discriminating against the right-wing fringe, social media actually bends over backwards to appease it — and it appears that AI is doing likewise.

These two thingbots, though they didn’t outright lie like Alexa, also did plenty of waffling about the matter of the “stolen election”. Copilot began its discussion with the disclaimer that “this is a complex topic”. (Really? How so?) And after stating correctly that all claims of fraud had been discredited, it ended by undermining what it had just finished saying with another disclaimer that election fraud is supposedly a subjective and contentious matter — without mentioning that it’s “contentious” only because the MAGA cult won’t stop contending it.

All three of these bots offer different answers to the same question at different times. I posed the same questions again a week later, and the responses all had been revised — in this case, the revisions all were in the direction of truth and accuracy, but there’s no guarantee that future revisions will be.

In short, Artificial Intelligence is a threat to society’s mental health for at least two reasons. First, like just about all the other channels of mass communication out there these days, it indulges in bothsidesism, which is normally not a good thing. Bothsidesism is constructive only when two things are actually comparable (which they usually aren’t), and the comparison is meant to respond to undue focus on a lesser offense or fault by pointing out a neglected greater offense or fault; but instead bothsidesism is usually just the opposite: it’s invoked in response to the focus on a major offense or fault by drawing attention to a supposedly similar, but minor, offense or fault. Bothsidesism almost always benefits the propagandists and dissemblers.

The other problem with AI is that it is selective rather than comprehensive. It does not, as you might expect, send its tentacles out all over the web and haul in an exhaustive compendium of research and then boil it down to a Cliff’s Notes presentation. Instead, it engages in cherry picking, in accordance with some arcane algorithm or other — for its illuminating report on the election, Alexa culled from a total of two sources, both of which were right-wing platforms. Garbage in, garbage out. And there’s a lot more garbage out there than substance.

If you ever watched Stanley Kubrick’s classic film 2001: a Space Odyssey you might recall that it contained a couple of predictions about the future of AI, both at its extreme best and extreme worst. The latter comes near the end of the film when HAL takes control, and literally starts killing people who cross him. We are not at that point — and fortunately, it’s not likely that we ever will be. The favorable extreme comes much earlier in the story, when AI provides thorough fact-checking in real time on the utterances by public figures. We’re not there either, alas.

Where we are, at this stage in the evolution of AI, is a time at which AI utilized for research is a boon to misinformers at least as much as to anyone else. It’s only a matter of time until they fully realize this and start exploiting it to the hilt.