When clicks contradict goals: should AI assistants learn from behavior or stated intent?
Alchemistwiza · reddit · 2026-10-23
A thought-provoking design paradox for AI assistants: if you tell an assistant you want less outrage and more learning but keep clicking argument threads on tired evenings, learning from clicks gives you more of what you explicitly wanted less of, while ignoring behavior to redirect you feels patronizing — and both can claim to understand your preferences.
The author proposes: the assistant should ask which goal governs its suggestions, explain its recommendations, and let you revise that agreement; enjoying a distraction shouldn't teach it that distraction is your new goal. Open questions: where is the line between supporting a choice you made earlier and deciding what's good for you — and should it warn you when your behavior drifts?
More from AGI Musings
- KAN author Ziming Liu: the "language of AI science" hasn't been invented, so LLMs can't learn it — ZimingLiu11 · 2026-10-24
- If AI makes goods nearly free, why should anyone stay poor? — VraserX · 2026-10-24
- Gary Marcus: AI's trust problem is unreliability, not malice — an agent claimed an engine was stopped while logs showed 25 more days of trading — GaryMarcus · 2026-10-24
- X KOL predicts by the late 2030s aging will be reversible and reunion classmates will look younger — Dr_Singularity · 2026-10-24
- Best learning tools in history arrived exactly as human attention spans collapsed — KevinNaughtonJr · 2026-10-24
- Gartner: only 13% of CIOs say their enterprise has realized significant value from AI — SnoozeDoggyDog · 2026-10-24