Perplexity’s Blunders Hint at AI’s Fundamental Problem

Perplexity is the latest AI company to come under fire, with accusations that it steals and plagiarizes content. Similar to search engines, Perplexity seeks to answer your questions. But whereas search engines return a list of online sources where you might find your answer, Perplexity generates and displays an “answer” for you that’s an aggregate of those sources’ content. (Which is why Perplexity calls itself an “answer engine.”)

Controversially, these answers can include content that Perplexity essentially rips off, even from sites that hide their content behind paywalls and/or specifically tell Perplexity to not scrape their content via the “robots.txt” file.

To better understand how Perplexity works, I asked it a very simple question: “What does Jason Morehead say about Slowdive?” Perplexity’s answer pulls from multiple sources, including Opus, Christ and Pop Culture, and even old tweets. (Never mind the fact that my site’s “robots.txt” file tells Perplexity to not scrape my content, as per their own documentation.) At first, it seems a decent overview of my Slowdive-related sentiments. The devil’s in the details, though. For as good a job as Perplexity does at summarizing my sentiments, it commits a couple of obvious blunders:

1. The sentence about me “comparing the experience of watching Slowdive perform in 2017 to feeling ‘as miraculous’ as hearing new music from them” pulls from this 2017 tweet that actually refers to seeing Roadside Monument perform again.
2. The summary’s final sentence pulls from my review of Slowdive’s Everything Is Alive, but it gets the song wrong. I described “The Slab” as “the album’s Slowdive-iest song,” not “Shanty.” (Though “Shanty” is a lovely song in its own right.)

A lot of my Slowdive-related content is older, though, which would give Perplexity time to scrape it all (“robots.txt” file be damned). So how does it fare with something more recent, like my thoughts on MTV News, which I posted earlier this week? Not surprisingly, Perplexity knows exactly what I’ve said, and its summary is much better than the Slowdive-related summary. (Though its phrasing is a bit odd in places, and seems to indicate that I wrote some copy that I, in fact, quoted from other sources.)

Indeed, Perplexity’s summary is so good that it’s a bit alarming. While Perplexity does link back to my original MTV News post, there may not be much point in reading it once you’re finished with Perplexity’s summary. (This, by the way, is another criticism of Perplexity, that it potentially steals traffic from the sources that it aggregates.)

Returning to Perplexity’s goofs in the Slowdive summary, they’re admittedly trivial examples. Imagine, however, someone asking about something more important, like a medical condition or a legal issue, and trusting Perplexity’s answer without taking the time to dig through its sources to verify its accuracy. Since I know what I’ve written about Slowdive over the years, it was pretty easy to see what Perplexity got wrong. But if I asked about something I know little-to-nothing about? That’s where things get more troubling.

The obvious rejoinder to criticisms of Perplexity’s inaccuracy is that AI tools are always improving. Thus, it’s possible that Perplexity’s synopsis of my Slowdive sentiments will eventually become nigh-flawless à la its summary of my MTV News post. But that’s beside the point, because what’s ultimately at stake isn’t AI’s accuracy so much as the erosion of truth. And more specifically, the erosion of our desire for truth.

Earlier this month, Barry Sewell shared two images of bald eagle chicks. One was AI-generated and depicted them with white heads, while the other was an actual photo in which their heads were brownish-gray. Sewell noted that the AI-generated image received numerous compliments without anyone pointing out that it was factually incorrect. (A bald eagle’s head doesn’t become fully white until they’re approximately 5 years old.) His conclusion?

What spooked me was how people were readily gullible and went along with what AI generated and took it as factual or true. This is an early indication of what is coming. Artificial intelligence is going to create a society of “dumb and dumber” citizens. It won’t be long before people won’t know what’s real anymore more. It’s already happening and it’s effective.

The ultimate problem with AI is not that it generates erroneous images or misquotes blog posts, or even that it’s powered by theft. (Which, to be clear, are all pretty bad things.) AI’s ultimate problem is twofold:

1. AI encourages us to be increasingly accepting of those erroneous images and misquoted blog posts as truth. It encourages us to be lazy and choose convenience over truth because thanks to AI, it now requires so much more effort to determine what’s real.
2. AI itself can’t reliably tell the difference between fact and fiction, which is why Perplexity — which claims to be dedicated to generating factual answers — is already struggling (ironically enough) with AI-generated misinformation. Which only exacerbates the struggle to determine reality.

AI tools will certainly continue to improve, and eventually, might even render my second point functionally irrelevant. However, current evidence suggests that such “hallucinations” will always be a problem even as those same tools continue to promote a veneer of truthfulness, accuracy, and reliability. (As a corollary to this, a further complication is that no significant AI regulation currently exists so far as I know. As a result, AI companies have no real impetus to actually be accurate short of avoiding bad press or placating investors.)

But even if AI tools do eventually “mature” to the point of consistently separating fact from fiction, the damage will already have been done to our ability to separate fact from fiction. AI’s current learning comes at our expense, such that if/when AI tools do become reliable arbiters of truth, it won’t matter because we’ve already been trained to not care.