AI killed curiosity, apparently

“AI is killing curiosity”, apparently – there’s research and everything. Anne-Laure Le Cunff, writing in the New York Times, describes how we type a question, skim the AI-generated summary and leave, and tools like ChatGPT and Claude have “collapsed the old habit of wandering through links into instant gratification.”

“The old habit”… You mean the golden years of wading through absolute crap and unsubstantiated nonsense in search of an actual answer to the actual question we posed? I must have missed those golden moments (even coming from someone born in the 80s). Plenty of my web browsing involved SEO junk artificially shoved up the rankings by people whose only interest was getting you onto the page without any regard for telling you anything true once you’d arrived. Ye Olde Web was many things but inherently curiosity-inducing wasn’t one of them.

Anyway, I had some lovely chats with some lovely smart people about this so wanted to share it all and expand what’s in my brain with more words than a LinkedIn post allows.

The golden age that wasn’t

I’ll concede some points here, being the fair and balanced chap I am. AI reduces clicks, obviously it does, but curiosity-squashing is not what I’ve personally found. If anything it’s driven mine the other way. What starts with one question quickly escalates into dozens, with clarifications, tangents, wildly off-topic back-and-forths I never planned on having.

I’ll give you can example of a real life chat I recently had with Claude. Trying to get my head around some quantum stuff because I’ve been reading Phillip Ball’s Beyond Weird: Why Everything You Thought You Knew about Quantum Physics is Different (actually just finished – it’s great, you should read it!) and I had a million questions starting with what the bloody hell what a wave function actually is. That meant asking about probability and statistics, which dragged me into the history of Bayesian thinking, which looped back to quantum, which slid into the limits of what we can actually know, which somehow ended up in a long conversation about consciousness and the nature of reality. And it was great going being able fire more and more questions, but it all stemmed from one starter question and ended in a warren of rabbit holes. It’s just such an odd and brilliant combination of breadth and specificity that most would probably never have gotten by simply browsing the web.

But I know that’s me. I’m the dorky kid who opens an encyclopaedia for fun, who spent far too long in libraries, who’d have wandered the web for hours before any of this LLM, AI, Skynet shenanigans existed. Not to be too much of pretentious dick, but I don’t think most people are like that. Most people, faced with not knowing something, absolutely did not set off on a three-hour odyssey through hyperlinks.

Curious people wandered and now curious people can interrogate more, in more depth, across more topics, in less time and ask more questions than the old web ever let them do with any hint of efficiency. So I think AI might have actually fed curiosity rather than killing it in this case.

But I get not everyone is particularly curious about things, they just want a decent answer. Before an AI overview in the search window, they went to Google, fired off a search, clicked the top link, read a bit and probably gave up after reading one blog/post/entry once their question was answered, waylaid or the only opinion was sufficiently validated. Or I guess they didn’t bother at all and stayed in the dark. But now they get an AI overview, which, admittedly, flattens the nuance (even if the sources are right next to the main result… just saying), and you could argue that for anyone who stops there, that’s a real loss. But it’s at least drawing on a spread of sources rather than one anonymous bloke’s slant. I reckon that’s a step up from clicking the first result and treating it as conclusive.

Byline’s are often fiction

There’s a worry that AI scootches itself between the reader and the truth, to which I say, maaaaaaybe. I’ve ghostwritten hundreds of technical ‘expert’ articles, so forgive me if I don’t clutch any pearls about authorship. The byline was always somewhat fictitious, shall we say, and really only useful if you’re reading an op-ed or you’re tracing how views on a topic shifted over time. Authorship in most industries tells you very little beyond how closely the named expert (or their marketing team) looked at the thing before it went out, because the framing and the voice were usually mine.

So I wouldn’t necessarily say AI created much of gap between the name on the page and the hand that wrote it because ghostwriting, with its spooky-but-paid-for ethereal hands, was there first. What does still matter, and what the AI overview genuinely lacks, is accountability. When (or if, in some cases) a human expert signs work off, someone real has put their name and their company’s credibility on the line and if it’s wrong, there’s a person at the other end with shit to lose. An AI overview has no one answerable for a specific claim, and if it turns out to be crap, it just is, and nobody’s job/rep is on the line. I do think that’s a loss worth worrying about, but that’s an accountability thing and we’re here to talk about curiosity.

Oh, and you also shouldn’t trust AI to sort out bias, obviously, as it pull from the same messy web as everyone else (LLMs use a search engine to find their raw stuff and so are susceptible to the same SEO rules as searches, it’s then AI’d into something more coherent, hopefully). But for someone who was only ever going to check a single top link or a Facebook post, it’s probably doing a better job than they’d have managed alone. On a subject you know nothing about, an average person is surely more likely to get an accurate run of answers from an amalgamation of knowledge than from picking their way through a dozen pages built with promotion in mind, with a bunch of them presenting pseudoscience, authoritatively, as fact. This is of course just speculation but you get the point.

Humans for humans, bots for facts

Subjective stuff gets a little more interesting or messy, depending on how to look at it. Someone said to be, on the matter of reviews, for example,

“The underlying content is wildly variable and often ill-considered, or may even be actively deceptive. The AI overview does a great job of smoothing out this chaos and providing what is ostensibly an aggregate view. Yet, I would rather go through the reviews until I find someone I trust, which I do based on signals like shared experience, tone of voice and other details that are nothing to do with the product. If I vibe with someone based on non-product factors, I then trust their views on the product itself. A context-free, author-less recommendation means nothing. I don’t just want advice – I need an advisor. “

And here I’ll agree that humanity is very lovely to have when it comes to subjective views on things like product or restaurant reviews, for example, but there’s a problem. If I’m reading a restaurant review that opens with

“once I’d parked my brand new white BMW in their crummy spaces, I finally sat down with the menu, and wow, let me tell you–”

I feel entirely comfortable stopping right there, because I don’t ‘vibe’ with that bloke, who I immediately assume has a gold chain hanging around his bare chest, a backwards cap and drives with seat back at at least 45 degrees and is therefore a certified bellend. Pure bias/prejudice on my part, and it cuts both ways because reviews are how I find places I love and how I build myself a nice little echo chamber of pleasant things. An AI overview on the other hand, telling me “most people praised the food (especially the mussels) and the waiting staff, though a few mentioned the small car park”, is arguably more useful for a potential customer, even if it is wholly dead inside.

(Better yet, let’s introduce standard deviations on reviews so we can see the spread!)

But research is much closer to fact-finding, and whether I happen to like someone’s tone has absolutely sod all to do with whether their account of a wave function is correct. Humans for humans, then, bots for facts. Though I’ll admit I still love it when humans do the facts with a bit of humanity in them; there’s just more to sift through when they do.

Scicomms and nobody measures the thing we’re arguing about

What I’d actually love, before anyone writes Curiosity’s obituary, is the data. AI companies are sitting on mountains of potentially anonymised query logs that could show whether one question leads to another, and another after that, which is essentially a measure of curiosity itself. In science communication (a field we all love, but, let’s be honest, it’s plagued with crappily defined objectives and outcomes nobody can measure) that would be gold. We spend our whole careers churning out words or making videos in the hope we’ve nudged public curiosity upward but almost never know whether we have. With such a humungous repository of what could be quantified as curiosity, we could start to assess trends over time and see if our work had any effect. Yes yes, I know ethics yada yada yada, and all that, but how cool would that be?!

Okay, I’m going to wrap this up before I ramble anymore. Is AI killing curiosity? I don’t think so. If anything, we gave the already-curious a sharper tool and gave everyone else a half-decent foundation to stand on instead of just diving into a top link. But I can’t prove that any more than Le Cunff can prove the reverse at a population level. It’s all speculation, based on real concerns and real issues. You can measure science curiosity in a sample well enough, (there are proper and validated scales for it) but watching curiosity move across a whole population, and pinning the shift on any one cause, least of all a chatbot, is a whole different challenge we’ve yet to resolve.