Study finds AI assistance reduced persistence after tool removal

Researchers recruited 1,222 people online for randomized controlled trials.[1] The study included three experiments involving mathematical reasoning and reading comprehension.[2] Berkeley News reports that the team presented an updated, peer-reviewed version of the paper at the Conference on Language Modeling.[1] The paper, titled “AI Assistance Reduces Persistence and Hurts Independent Performance,” reports causal evidence that AI assistance can reduce persistence and impair performance when people later work without it.[2]

The first experiment involved 354 participants solving fraction problems.[1] One group worked without help; another could use ChatGPT and ask it for answers.[1] Participants with AI access initially answered more accurately than those without it.[1] After participants completed 12 problems, researchers removed access to AI.[1] The AI-assisted group’s accuracy fell almost immediately, and some participants got answers wrong or gave up.[1] In the second experiment, participants who had not used AI persisted and were more successful by the end.[1]

The third experiment involved 201 participants completing an SAT reading-comprehension prompt.[1] When researchers removed the AI tool, both persistence and accuracy fell.[1] Across the findings, the paper says AI improved short-term performance but left participants performing worse and giving up more often when they had to continue without the tool.[2] That distinction matters for interpreting assistance: a stronger result while AI is available did not mean participants were better prepared to work independently afterward.[1][2]

The researchers suggest that AI may reduce persistence by encouraging expectations of immediate answers and leaving users with fewer chances to work through challenges themselves.[2] The paper cautions that, if such effects accumulate with sustained use, current AI systems could erode capabilities they are meant to support.[2] That is a possible longer-term consequence, not a finding that this study established over sustained use.[2] Brian Christian, a UC Berkeley research fellow, said he experiences the finding too and argued that AI should augment human abilities rather than replace them.[1] He suggested more instructive, tutor-like responses instead of defaulting to quick answers.[1]