A room that was already on fire

I've mused on intellectual dilution for at least a year that I can remember. It always shocks me that so many people blindly believe things they read on social media or that someone they know told them, even when corrected by someone with more knowledge on the subject.

Is this just laziness — they don't want to research the topic themselves — or are we socially guided towards this level of acceptance?

The reason I'm finally writing it down is that today at college I sat in on a lecturer telling a room full of students that AI will stop people using their brains. I've heard the line a hundred times, but something about hearing it delivered from the front of a classroom, to students, as if it were established fact, tipped me over. Because here's the thing. Maybe for some people it will. But those people were already the ones cutting and pasting from the internet. They were already the ones who took the first Google result as gospel, who shared the screenshot without checking, who repeated what they heard down the pub and got irritated when you pushed back. The brain the lecturer was trying to protect from AI is not the brain most of his students are walking around with. And if we're going to have this argument seriously, we should probably start by being honest about that.

This panic is old

The first recorded version of "this new technology will rot our minds" is about 2,400 years old. In Plato's Phaedrus, Socrates argues that writing — writing, the thing you are doing right now — will "implant forgetfulness" in the souls of those who learn it. They'll stop using their memory. They'll have the appearance of wisdom without the reality of it. The irony that we only know this argument because Plato wrote it down has been pointed out so often it barely needs saying.

Fast forward to 1477. A Venetian humanist called Hieronimus Squarciafico is grumbling that the printing press is producing too many books, and that this "abundance of books makes men less studious; it destroys memory and enfeebles the mind." The 1970s panicked about pocket calculators. Parents and teachers were certain children would lose the ability to do mental arithmetic. Two large meta-analyses later — Hembree and Dessart in 1986, Ellington in 2003 — and the verdict was the opposite of what everyone feared. Calculator use, properly integrated into teaching, improved paper-and-pencil arithmetic. In 2008 it was Google. Nicholas Carr's essay "Is Google Making Us Stupid?" became the defining piece of the era. The argument, when you strip it down, is the same one Socrates made about writing. It always is.

Every dominant information technology arrives, an authority figure laments that it will atrophy the mental faculty it replaces, and that prediction turns out to be partly true — some skills do erode — but mostly wrong, because the technology enables new capacities that more than compensate. AI is the latest entry in a very old genre. That doesn't mean the worry is groundless. It does mean we should be suspicious of the certainty with which it's delivered.

So why do people blindly believe things?

Let me come back to the question I opened with. Is it laziness, or are we socially guided toward acceptance? The honest answer, after a year of turning it over and reading more cognitive psychology than I expected to, is: neither. Or rather, both, but the framing is wrong. People aren't lazy. They aren't dupes. They are running default human cognitive software in an environment that has been carefully engineered to exploit it. Three findings, in particular, do most of the work of explaining what we see.

The first is the illusory truth effect. Lisa Fazio and colleagues, in a 2015 paper in the Journal of Experimental Psychology, demonstrated that repetition increases the perceived truth of a statement even when the reader already knows the correct answer. Their conclusion was blunt: "knowledge does not protect against illusory truth." If you see a claim three times on Facebook, it starts to feel true. Not because you've reasoned your way to it. Because your brain uses fluency — how easily a thought comes — as a shortcut for truth. This is the engine of social media. The platforms didn't invent it. They just industrialised it.

The second is the continued influence effect, documented in a 2012 review by Stephan Lewandowsky and colleagues. Once misinformation is in your head, corrections largely fail to dislodge it. The original belief keeps influencing your reasoning even after you've explicitly accepted that it was wrong. This is why correcting someone on Facebook doesn't work, no matter how patient or well-sourced you are. The correction sits next to the misinformation in their head, and the misinformation does most of the steering.

The third, and to me the most uncomfortable, is identity-protective cognition. Dan Kahan at Yale has shown across a decade of work that on identity-laden issues — climate, guns, vaccines — more numerate and more analytical people are more polarised than less analytical ones, not less. Reasoning ability doesn't carry people toward truth on these questions. It carries them toward more sophisticated defences of their tribe. The careful, fact-checking, source-evaluating citizen the AI critics are trying to protect is, on the issues that actually matter to them, the one most likely to be lawyering for their team.

So the answer to my opening question is that intellectual dilution isn't a moral failing. It's how human cognition runs by default. Repetition feels like evidence. Tribe feels like truth. Corrections don't stick. None of this is new. None of this was caused by AI. All of it was being exploited at industrial scale for at least a decade before ChatGPT existed.

The damage was done before ChatGPT existed

If you want the single statistic that ought to settle the argument about whether AI broke our epistemic culture, here it is. In 2018, Soroush Vosoughi, Deb Roy and Sinan Aral published a paper in Science analysing 126,000 cascades of news stories on Twitter between 2006 and 2017. Their finding: falsehood "diffused significantly farther, faster, deeper, and more broadly than the truth in all categories of information." False news was roughly 70 percent more likely to be retweeted than true news. And the kicker — it wasn't bots. It was humans. People shared false news more than true news because false news was more novel, and humans are wired to share novelty.

That paper came out four years before ChatGPT. The epistemic environment was already broken.

Sam Wineburg's team at Stanford has spent a decade quantifying exactly how broken. In a 2016 study of 7,804 students across twelve US states, they found that 82 percent of middle-schoolers could not tell the difference between a piece of sponsored content and a real news article. In a follow-up, they put Stanford undergraduates, history PhDs and professional fact-checkers in front of unfamiliar websites and asked them to evaluate credibility. The undergraduates and historians were routinely fooled by professional-looking design. Only the fact-checkers got it right consistently, because only the fact-checkers used what Wineburg calls lateral reading — opening other tabs and checking the source from outside, instead of staring at the site itself looking for clues. "Very intelligent people," Wineburg wrote, "were bamboozled by the ruses that are part of the toolkit of digital deception today." That's elite university students. In 2017. Five years before AI.

Meanwhile, the place people were getting their information had already moved. Pew Research found that by 2025, more than half of US adults were getting news at least sometimes from social media; 43 percent of under-30s regularly got their news from TikTok, up from 9 percent in 2020. Fewer than 1 percent preferred AI chatbots for news. The algorithmic, fragmented, source-shorn information environment everyone is now blaming on AI was already in place. AI didn't build it. AI walked into a room that had been on fire for a decade and is now being blamed for the smoke.

And there's the cognitive side too, which I'll be briefer on. Betsy Sparrow's 2011 paper in Science — the so-called Google effect — found that when people expect to have future access to information, they remember the information itself less and remember where to find it more. Cognitive offloading to digital tools was empirically demonstrated eleven years before ChatGPT. Gloria Mark's longitudinal work at UC Irvine measured average focused attention on a single screen falling from about 150 seconds in 2004 to 47 seconds by 2016. Maryanne Wolf, a reading-brain neuroscientist, wrote a whole book — Reader, Come Home — about deep reading collapsing under screen use, including her own. None of this was caused by AI. All of it was the substrate AI arrived into.

Taking the critics seriously

I want to be fair to the lecturer at the front of the room. The "AI will stop people using their brains" claim is not stupid, and there is real research worth taking seriously. The most cited piece right now is a 2025 MIT Media Lab study by Nataliya Kosmyna and colleagues called "Your Brain on ChatGPT," which put EEG caps on people writing essays — some with ChatGPT, some with a search engine, some with nothing — and found that the ChatGPT group had the weakest neural connectivity, and that the effect persisted when the tool was taken away. The authors call this "cognitive debt." There's a parallel study from Microsoft Research and Carnegie Mellon at the CHI conference in 2025, surveying 319 knowledge workers across 936 uses of generative AI, which found that the more people trusted the AI, the less critical thinking they did.

These are real findings, and I'm not going to wave them away. But notice what they actually show. The Kosmyna result — offload a cognitive task to a tool, and the brain engages less — is structurally identical to what Sparrow found in 2011 with Google. The Microsoft result — trust your tool, check less — is structurally identical to what happens with every trusted tool humans have ever used. Your GPS. Your spell-checker. Your map. Your calculator. The novel claim would be that AI is qualitatively different from prior offloading tools, that it represents a discontinuity rather than a continuation. That claim has not been demonstrated. What's been demonstrated is that AI sits on the same continuum as every cognitive tool that came before it.

Which brings me to the bit the lecturer got most wrong. He told the students AI would stop them using their brains. But the students who will let AI think for them are the same students who were cutting and pasting from Wikipedia ten years ago, paraphrasing the first Google result five years ago, and copying off each other before any of that. The behaviour predates the tool. The tool just makes it faster. Someone who is genuinely using their brain — who understands the question they are asking the AI, who reads the answer critically, who knows enough about the topic to spot when the AI is bluffing — is doing more cognitive work with AI, not less. They are doing the kind of work the lecturer claims to be defending.

The bit nobody wants to talk about

Here's the move I think actually matters, and the reason I keep coming back to this in my own teaching. There is a body of educational research called Cognitive Load Theory, developed by John Sweller and refined by Fred Paas and Jeroen van Merriënboer over the last thirty-odd years. Its core claim is that working memory is a brutal bottleneck — roughly four chunks at a time, by current estimates — and that learning fails when that bottleneck is overloaded by stuff that isn't the actual learning. CLT distinguishes intrinsic load (the inherent difficulty of the material), extraneous load (load imposed by how the material is presented), and germane load (the load that actually builds understanding). Good teaching minimises extraneous load to protect germane load. That's not controversial. It's the boring consensus of half a century of educational psychology.

Now apply that to AI. Cognitive offloading is not the enemy. Writing was offloading. Notebooks were offloading. Calculators were offloading. Each was attacked when it arrived, and each turned out to enable new cognitive capacities when used well. The right question is not whether AI offloads cognition. It is what we do with the freed capacity.

If the freed capacity is absorbed by the attention economy — by infinite scroll, by autoplay, by the same engagement-optimised feeds that did the epistemic damage in the first place — then yes, AI will accelerate intellectual dilution. But that's a property of the surrounding environment, not of AI itself. That's a design problem and an education problem.

And here is where it lands hardest for me, because I see it every week in the Evolve sessions at Linkage. In SEN teaching in particular, the load of effortful reading and writing is often the binding constraint on whether higher-order work happens at all. A student who spends their entire cognitive budget decoding a paragraph has nothing left for the analysis, the creativity, or the actual production we are trying to elicit from them. If AI can reduce that extraneous load — reading a passage aloud, scaffolding a sentence, restructuring a paragraph — and if we then redirect the freed capacity toward the work we actually want them to do, that is not intellectual dilution. It is the opposite. It is finally letting them think.

I want to be careful here, because I know what the objection sounds like. "You're lowering expectations." No. I am refusing to let a bottleneck that has nothing to do with the actual learning outcome continue to be the thing that determines who gets to do creative and analytical work. The point of the reading isn't the reading. The point of the reading is what the reading enables. If AI can carry some of that load, and we use the saved capacity for harder cognitive work rather than for scrolling, we have not made the student less intellectually capable. We have made it possible for them to be more intellectually capable than the previous bottleneck allowed.

So what should we actually do?

If the problem isn't AI, and the problem is what people do with AI, then the answer isn't to keep AI out of classrooms. The answer is to teach people how to use it. We don't ban calculators from maths lessons. We teach kids when to reach for one and when to do it in their head. We don't ban Wikipedia from research. We teach lateral reading. We don't ban Google from essays. We teach citation and source evaluation. The fact that we have not, as a culture, done a particularly good job of teaching those things is a separate problem — and arguably it's the root problem behind everything I've described in this essay. But the solution to bad source-checking has never been to remove the source. It's to teach the checking.

What would AI literacy actually look like? Off the top of my head, and from my own teaching: understanding that an LLM produces plausible-sounding text by predicting the next token, not by knowing things; learning to recognise the smell of confident bluffing; reading the answer with the same scepticism you'd bring to a Wikipedia article or a Reddit thread; cross-checking specific claims against named sources; using AI to generate options to choose between, rather than answers to accept; treating it as a collaborator that needs supervision, not an oracle. None of this is hard. None of it requires a maths degree. It requires that we take AI seriously enough to teach it properly, instead of standing at the front of a classroom telling students it will stop them using their brains and then sending them out into a world that will reward them for using it well.

Closing

I'll end where I started. People blindly believe things they read on social media, or that someone they know told them, because repetition feels like truth, because tribe feels like evidence, because corrections don't stick, and because source-checking is a skill almost no one was ever taught. AI didn't cause any of that. AI walked into a room where all of it was already happening at scale.

The lecturer at the front of the room is worried about the wrong thing. The brain he's defending was already eroded by fifteen years of engagement-optimised feeds before any of his students opened ChatGPT. If we want to take intellectual dilution seriously — and I think we should, because it's real, and it matters — then the conversation we ought to be having is not whether AI will stop people thinking. It's whether AI, taught well, can be the first general-purpose cognitive tool that asks more of its users than the last one did. That's a much harder argument to have than the one the lecturer was making today. But it's the only one worth having.