An AI-made chart caught its author’s date mistake
Ethan Mollick describes AI tools making charts and catching his own wrong month count. It is a useful demonstration, not a controlled test of accuracy or time saved.
AI AgencyRead story
Lead story
AI math feedback looked good. Did students learn more?
In two university math groups, reviewers liked the AI-made practice more than the feedback itself. Later exam comparisons found no clear difference. This is not evidence of a K-12 learning gain.
Source: arXiv · Oct 21, 2026Read original source
Covered by both lenses
Read story
More top stories
A critic presses Anthropic on its draft AI bill
Zvi Mowshowitz disputes the proposed limits on state rules and the process for blocking a risky AI model. This is his critique of a draft, not an independent legal finding.
AI AgencyRead story
Featured
Feeling sure is not the same as catching the AI error
In programming tasks, reviewers sometimes felt confident but missed quiet mistakes. The paper tested 12 people in an early study and 36 in a controlled study. A review needs a check that could expose a wrong result.
Beyond PromptingRead story
Models & Capabilities
Better AI feedback starts with an instructor’s correction
In a comparison with 16 instructors, ValiAssist turned edits to AI-written programming comments into rules people could inspect. The next check is where each rule fails, not how much feedback a tool can produce.
Beyond PromptingRead story
Culture & Human Agency
When AI holds back advice, users react differently
In a repeated financial-choice study, selective AI advice raised acceptance among people who saw themselves as expert, but tended to lower it among those who did not. More acceptance is not proof of better decisions.
AI AgencyRead story
