Click any tag below to further narrow down your results
Links
Kunal Sheth's working note on the trap of accepting agent output without thinking, and how he tries to keep control: he delegates only where mistakes are cheap, and drives the work by hand on code he cares about. The core worry is that comprehension, not building, is now the bottleneck, and skipping it erodes both quality and his understanding of his own systems.
- Agents make work feel fast and productive, but complexity and technical debt compound quietly, and his familiarity with the codebase erodes with each automated fix.
- Models sound confident even when wrong, and peer pressure to ship more PRs pushes people to stop checking, which he treats as the warning sign to watch for.
- He defines a "vibe limit," the complexity of change he can make confidently with agents, and chooses deliberately where to surrender agency versus where to spend his own time, sometimes driving every edit by hand.
A University of Pennsylvania paper argues that many people don't just use LLMs as a tool but hand over their reasoning to them entirely, a pattern it calls "cognitive surrender." The risk is highest when answers arrive fluent and confident, and the research looks at how time pressure and incentives push people toward accepting them without checking.
- The researchers propose a third decision category, "artificial cognition," alongside fast intuitive thinking (System 1) and slow deliberate reasoning (System 2).
- Cognitive surrender differs from ordinary offloading to calculators or GPS: users give minimal internal engagement and accept the AI's output without verifying it.
- Time pressure and external incentives change how likely people are to surrender their judgment to AI answers.
A preregistered experiment (N=707) finds that a short AI-attributed message saying most people share generously raised participants' own sharing by about 14 percentage points, while a message saying most people favor themselves did not make them more self-serving. The paper argues people defer to AI morally in one direction only.
- Prosocial AI-attributed messages moved dictator-game transfers from about 18% to 32% of a $10 endowment, a statistically significant effect.
- Antisocial AI messages about self-favoring reports produced no detectable increase in dishonest reporting, with Bayesian tests favoring no effect.
- Participants rated the two AI messages as equally credible, yet their written reactions showed the prosocial message matched their moral position far better.
This preregistered experiment (N=707) tests whether people adopt moral positions from AI advice the way they already adopt factual answers. Prosocial AI advice pushed participants toward more altruism, but antisocial AI advice failed to pull them the other way, which breaks from what studies of human influence usually find.
- Prosocial AI advice significantly raised altruism in a non-strategic allocation task, while antisocial AI advice did not erode altruism, cooperation, or honesty.
- Participants rated the AI and researcher messages as equally credible, but the prosocial message matched their own moral position far better.
- The author argues even low-intensity AI advice could be a practical lever for encouraging prosocial behavior.