RoundCraft Memory Gym

Carefully crafted quiz rounds, played the pub-quiz way

Is AI making us dumber?

What the research really says about memory

RoundCraft · 8 October 2026

A phone beside a stack of handwritten flashcards and a pencil on a wooden desk

Every few weeks a new headline asks whether ChatGPT and its rivals are rotting our brains. The honest answer is more interesting than yes or no: there's no good evidence that using AI weakens your ability to remember, but there's good evidence that you only keep what you practise. Here's what the research actually shows, where it's weaker than the headlines suggest, and what you can do about it in five minutes a day.

Handing thinking to tools is nothing new

Psychologists call it cognitive offloading: using something outside your head to reduce the mental effort a task needs, like writing a shopping list, setting a phone reminder or doing a sum on a calculator (Risko & Gilbert, 2016). We have always done it, and it's often smart. In one set of experiments, people who saved a file before studying a second one remembered more of the second, as if saving freed up room for what came next (Storm & Stone, 2015).

The catch is that what you offload, you don't practise. When people photographed objects on a museum tour instead of just looking at them, they later remembered fewer of those objects and fewer of their details. Interestingly, when people zoomed in to photograph one detail, the effect disappeared: paying attention is what counts (Henkel, 2014).

The "Google effect", and a word of caution

A much-cited 2011 study reported that people were less likely to remember information they expected to be able to look up later, and better at remembering where to find it (Sparrow, Liu & Wegner, 2011). It became known as the "Google effect".

It deserves some caution, though. When a large project re-ran 21 experiments from Science and Nature, the Sparrow study was among those that didn't replicate, although the experiment chosen was a related priming task, not the memory test itself, and the original authors later pointed to differences between the replication's design and their own (Camerer et al., 2018). The broad idea, that we lean less on memory when we trust an outside source, still fits the offloading research above. But it's a reminder to treat single, striking studies with care, including the newer ones about AI.

What the AI studies actually found

Writing essays with ChatGPT. In a study from MIT Media Lab, 54 people wrote essays over several sessions using ChatGPT, a search engine, or nothing but their own heads, while their brain activity was recorded with EEG. Brain connectivity was strongest in the unaided group and weakest in the ChatGPT group, and the ChatGPT users struggled to quote from essays they had written minutes earlier. The authors call the long-term cost of leaning on AI "cognitive debt" (Kosmyna et al., 2025). It's a small, early study (only 18 people took part in the final session) and it was released as a preprint (updated December 2025); as of October 2026 it still hasn't been published after peer review, and other researchers have argued that some of its results could be interpreted more conservatively (Stankovic et al., 2026). Its authors are clear that it does not show AI is making people "dumber".

Learning from AI summaries. Across seven experiments with more than 10,000 people, those who learned about a topic from an AI summary (ChatGPT or Google's AI Overview) reported learning less deeply than those who used an ordinary web search, even when the facts shown were identical, and the advice they then wrote was shorter, less factual and less original (Melumad & Yun, 2025). Doing the work of finding and piecing things together yourself is part of how you learn them.

Learning maths with and without guardrails. In a field experiment with nearly a thousand high-school students in Turkey, those given a plain ChatGPT-style assistant did much better on practice problems (48% better), but when the assistant was taken away they did 17% worse on the exam than students who never had it. A version designed as a tutor, which gave hints instead of answers, largely avoided that drop (Bastani et al., 2025).

Thinking at work. In a survey of 319 knowledge workers, people who trusted AI more reported applying less critical thinking to its output, while people more confident in their own skills reported applying more (Lee et al., 2025). Another survey, of 666 people, found that heavier AI use went with lower critical-thinking scores, and that younger people both used AI more and scored lower (Gerlich, 2025). Both are surveys, so they show associations, not cause and effect.

Put together, the picture is consistent and fairly reassuring. A short 2026 commentary by cognitive psychologists, titled "Is AI making us stupid?", argued that skills and knowledge you hand over to AI entirely can fail to develop or fade with disuse, while suggesting, more tentatively, that the basic mental abilities underneath them are probably more resilient (Cash et al., 2026). In other words, AI isn't shown to weaken your memory as such. But if a tool does the remembering or the thinking for you, you don't learn that thing, just as with the museum photos. How you use it matters: the maths students who were made to think for themselves kept their gains.

What actually makes things stick

Memory research has two findings that are about as solid as psychology gets, and both are free. A major review of ten popular study techniques rated just two of them as highly useful: practice testing and spreading practice out over time (Dunlosky et al., 2013).

Recall beats rereading (active recall). Students who studied a passage once and then practised recalling it three times remembered 61% of it a week later; students who studied it in four sessions remembered 40% (Roediger & Karpicke, 2006). In another experiment, students learning Swahili vocabulary remembered about 80% a week later if they kept testing themselves, against about 35% if they stopped testing once they had got a word right, even though they kept studying it (Karpicke & Roediger, 2008). The effort of pulling an answer out of your own head is what strengthens it.

Does the question format matter? Less than you might think. In experiments with educational material, including a real college course, quizzes where students wrote the answer themselves led to better memory later than multiple-choice quizzes (McDaniel, Roediger & McDermott, 2007). Another study found the same, but only when people were shown the correct answer afterwards; without that feedback, multiple choice did better, possibly because people got far more of the typed answers wrong (54% right, against 86% for multiple choice) and so successfully recalled less (Kang, McDermott & Roediger, 2007). Feedback matters for multiple choice too: without it, people can come away remembering a wrong option as the answer (Butler & Roediger, 2008). Large meta-analyses disagree on format: one found that recall-style tests give a bigger boost (Rowland, 2014), another that multiple-choice practice helps at least as much (Adesope, Trevisan & Sundararajan, 2017). That's why we don't claim our typed answers are better for your memory; we use them because that's how a pub quiz works. The big win is quizzing yourself at all, rather than rereading.

Curiosity helps. When people were more curious about the answer to a trivia question, they were more likely to remember it later, especially when the answer surprised them, and brain scans showed reward-related areas lighting up while they waited for the answer (Kang et al., 2009; Gruber, Gelman & Ranganath, 2014). In one study, people even remembered unrelated things they saw while curious better (Gruber, Gelman & Ranganath, 2014). A question you actually want to know the answer to is easier to learn from: one more reason quizzes should be fun.

Spacing beats cramming (spaced repetition). We forget quickly at first and more slowly later, a pattern first measured by Hermann Ebbinghaus in the 1880s and confirmed by a careful modern replication (Murre & Dros, 2015). Spreading reviews out over days leads to memories that last far longer than the same time spent in one go, and the longer you want to remember something, the longer the best gap between reviews: findings repeated across hundreds of experiments (Cepeda et al., 2006).

What doesn't: "brain training"

It's tempting to fix this with brain-training games. A major review of the evidence found that they mostly make you better at the games themselves, with little sign that the benefit spreads to everyday memory or thinking (Simons et al., 2016). Practice works on what you practise. So if you want to remember real things, practise remembering real things.

What about later life?

Researchers use the term cognitive reserve for something a bit like fitness for the mind: differences in how well people's brains cope with ageing or disease, built up over a lifetime by things like education, demanding work and mentally active leisure (Stern et al., 2020). People who stayed mentally active across their lives tend to show slower decline in memory and thinking in old age, even after accounting for the changes in their brains (Wilson et al., 2013), and the Lancet Commission on dementia counts less education in early life among 14 risk factors that could be changed (Livingston et al., 2024).

There's good news in the data too: while some abilities, like speed, peak young, vocabulary and general knowledge keep growing well into middle age and beyond, according to a study drawing on nearly 50,000 people (Hartshorne & Germine, 2015). Quizzes play to a strength that grows with age.

The muscle comparison only goes so far, though. Most of this evidence comes from observing people, not from experiments, and it can run the other way: people whose thinking is starting to slip may simply do fewer mentally demanding things. The longest experiment so far enrolled about 2,800 older adults and tracked some 2,000 of them through 20 years of health records. Neither a short memory-training course nor, overall, a computerised speed-of-processing course was shown to lower dementia diagnoses, although people who also got later "booster" sessions of the speed training were diagnosed less often: a promising hint, not proof (Coe et al., 2026). So keeping your mind busy throughout life is a sensible habit, but no quiz, ours included, has been shown to protect against dementia.

A five-minute habit for the AI age

  • Try first, then look it up. Before you search or ask an AI, spend ten seconds trying to recall the answer. Even a failed attempt helps the answer stick when you then see it (Kornell, Hays & Bjork, 2009).
  • Make AI quiz you, not tell you. Ask it for questions on what you're learning, or for hints rather than answers: the approach that protected the maths students.
  • Come back later. Revisit what you want to keep after a day, then after a few days, then after a few weeks.
  • Keep it small and daily. A few minutes every day beats an hour once a month.

That's exactly what Memory Gym does: one quiz a day on a topic you'll enjoy, ten answers to learn and recall without help, and earlier answers brought back after a day, then five days, then twenty, so they should last much longer. It's free and needs no sign-up. It won't make you a genius, but it should make you better at remembering things you choose to remember. And whatever the research says, above all we want it to be fun.

Sources

Try Memory Gym: one quiz a day, answers that stick