Jed Brown @jedbrown@hachyderm.io
Public

Imagine if we relied on oil companies to publish evidence that CO2 emissions cause climate change. This statement against interest by Anthropic illustrates the epistemic vulnerability in which funding agencies and universities have uncritically accepted vibes-based claims. Universities should be at least as critical as Anthropic, and should have been leading independent studies with scope similar to this one and the METR study (a group aligned with boosters, also publishing against interest).
arxiv.org/abs/2601.20245

Figure 6: Difference in means of overall task time and quiz score between the control (No AI) and treatment (AI Assistant) groups in main study (n=52). Error bars represent 95% CI. Significance values correspond to treatment effect. * p<0.05, **<0.01, ***<0.001 Figure shows Task Time about 7% faster in the AI group than No-AI group (p=0.391; not statistically significant). Quiz score mean 50% for AI group vs 66% for No-AI group (p=0.010). Treatment effect is not significant for task time, but is significant for quiz score.Figure 8: Score breakdown by questions type relating to each task and skill area. Debugging questions revealed the largest differences in average quiz score between the treatment and control groups Figure shows higher scores from the No-AI group than the AI group in all tests (Task 1, Task 2, Conceptual, Debugging, and CodeReading).
Jed Brown @jedbrown@hachyderm.io
Public

RE: hachyderm.io/@jenniferplusplus

Excellent long thread. And there's an important point here about the group that started with "conceptual inquiry" (a bit of a misnomer because LLM responses will often be misleading as the subject matter gets more novel, there are more distractors, etc.) and moved to "delegation" (suffering a 30% drop in quiz performance, apparently without noticing and self-correcting).

Those study participants did that entirely of their own volition, and there isn't a cohort that went the other way. It's like a new years resolution phenomenon where people start out with the best intent and then rationalize a pattern of behavior that is actually harmful. Any educators who think that good habits will develop needs to contend with the evidence showing exactly the opposite. This is also consistent with the METR study showing the cognitive bias in which people believe these products and this mode of interaction makes them mode "productive" or helps their learning even when metrics show the opposite effect. I think studying this effect through psychology of addiction would be fruitful.

0
0
0

If you have a fediverse account, you can quote this note from your own instance. Search https://hachyderm.io/users/jedbrown/statuses/115997740734501329 on your instance and quote it. (Note that quoting is not supported in Mastodon.)