You even use the term "measurable erosion" to describe the impacts of AI. If that's not saying that "AI use will erode thinking" then I'm not sure what is.

I don't understand. This is something that your paper explicitly states - that 10 minutes of AI use results in significant impairments in "independent performance and persistence" and then speculation that the cumulative effects may be "profound and difficult to reverse".

More specifically, they say this (and similar) which is inaccurate based on their data. As in, they've not shown this at all, but it's a central claim of their paper and public statments.

Ok, I've looked at their data file. A "solve" is getting it right, and a "skip" is whether the person clicked the skip button. So a "skip" is also a question that the person fails to solve. If the effect is due to skips, then the statements about performance are untrue!

Sorry, no *MORE wrong answers

Sure, but their entire public claim is that using AI impairs performance. It's also the main claim of the paper itself. And as far as I can tell, that claim is not supported by their own data, which seems to show only that AI use causes slightly more skips (but no fewer wrong answers)

I don't think it's small and noisy though. If you removed skips from the solve rate analysis - assuming they're there in the first place - I suspect you'd find no difference in solves. Which calls into question the idea that this has anything to do with performance.

Sure, but also their claim about AI reducing unassisted performance is predicated on the lower solve rate. But if the lower solve rate is mostly skips, then there is no causal impact on unassisted performance, it's just that people who had AI and lost it were more likely to skip.

Oh wow. Not really a fair experiment at all! I also don't understand the bit about unassisted performance. If I'm reading the graphs right, the lower solve rate from exp2/3 is just due to skips, no?

That is, there is a 10% increase in skip rates for those questions and a 10% decrease in solve rates. If it's just skipping questions then the statement "The lower solve rate of participants in the AI condition demonstrates that AI assistance impairs unassisted performance" is not true.

It's also unclear to me how the authors dealt with skipping. They present "solve" rates and "skip" rates. Are the skips excluded entirely from the solved graphs? If not, the effect from experiment 3 seems to be entirely due to skipped questions

Hahaha yes that's pretty wild. You can see the obvious effect from the exclusions - 6/191 vs 41/163, p=0.000000001 on chi-square. There is a 22% absolute increase in failing the initial questions with no AI, from 3% up to 25%

I would be interested in a replication study where two groups were asked to solve trig problems, one with calculators the other with old-style tables. If the people using calculators have to switch to tables at question 12, and do worse, would that imply cognitive damage caused by calculators?

I’ve seen a few folks comment that “Guinea-Bissau is a low income country and cannot afford the hep B vaccine birth dose, so why is this study unethical?” It’s a good question. Let me have a go at answering it. In a nutshell, researchers cannot ethically exploit poverty & scarcity to do a trial

I do try my best not to hate people (my parents taught me that "hate eats you up"), but, well, when it comes to RFK Jr, who is motivated both by his eugenicist beliefs & his truly horrific financial conflicts of interest to cause as much pain, suffering, & death as he can, I'm afraid I do feel hate

Thinking about how Russia spent much of the pandemic lying about their COVID-19 numbers and what that says about their public health responses. Graph from Dmitry Kobak

I very much understand coming from a place where a terrible thing has happened to you and feeling more cautious because of it - we all have trauma - but I don't think that the data supports people being concerned of LC in their daily lives in 2026.