Wharton tests show AI shopping agents' choices become unpredictable with added context
emollick · x · 2026-09-04
Ethan Mollick shares research from his Wharton Generative AI Lab at Penn:
- Agentic shopping: Across 26,000 tests, AI agents gave consistent recommendations when shown only product pages. But adding any context—a single review screenshot, competing sources and their ordering, an injected user "memory," or retrieval method—shifted what agents bought. The consistent pattern: the more context, the less predictable the final choice.
- Generative AI Studio education: Critique and charrette practices anchored student projects with no right answers.
- AIBO: An open-source tool for running controlled behavioral experiments on AI at scale, which turned AI into an active research collaborator.
- Plus preliminary work on persuading AI to comply with objectionable requests.
Mollick also explains why he posts visual demos on model launches: few people click links, so visuals communicate AI progress best.
More from Research
- Reasoning trace placement swings long-context accuracy by 50 points, Trace-as-State technique shows — dair_ai · 2026-09-04
- Sebastian Raschka's LLM Architecture Gallery now covers 102 models with a diff comparison tool — rasbt · 2026-09-04
- Scaling Law for Looped Transformers: Looping Boosts Reasoning, Not Knowledge — bookwormengr · 2026-09-04
- Tenstorrent Launches JapanFold: Sovereign Open-Source Drug Discovery for Japan — DavidBennett__ · 2026-09-04
- Stanford paper's Counterfactual Simulation Training boosts CoT faithfulness monitoring by 35 points — a_karvonen · 2026-09-04
- Stanford's Prefix Sliding cuts long-reasoning inference ~3x without retraining, AIME25 score intact — rohanpaul_ai · 2026-09-04