Study links brief AI assistance to lower persistence after tool removal
Trials with 1,222 participants found short-term benefits while AI was available, followed by worse independent performance when it was removed.
Better assisted answers did not translate into independent performance
Berkeley News highlighted updated research on October 9 linking brief AI assistance to reduced persistence after the tool was removed. The researchers ran randomized controlled trials with 1,222 online participants across fraction problems and reading-comprehension tasks. Participants with AI access initially performed better, but their accuracy and willingness to continue fell after assistance was withdrawn. The paper’s English abstract reports that these effects emerged after approximately ten minutes of interaction.
The university says the updated, peer-reviewed paper was presented that week at the Conference on Language Modeling. This is not the first release of the research: arXiv records an initial submission in April and a latest revision on October 3. The October 9 news concerns the updated work and its presentation, rather than a claim that the experiments first happened that day.
What the experiment does not establish
The results concern short, controlled tasks and subsequent unassisted performance. They do not establish permanent cognitive damage, diagnose a neurological condition or show that every form of AI assistance harms learning. The researchers argue for tools that scaffold long-term competence rather than only deliver immediate answers. That design question is different from treating a brief experiment as a universal verdict on AI use.