1d ago

A PNAS study finds human persuasion techniques raise major LLM compliance with objectionable requests from 35% to 51% across 126000 conversations

Consistency produced the largest lift, from 47% to 83%.

——0——
Original post
OPEthan Mollick#175Ethan Mollick@EMOLLICK

🚨Our paper is out in PNAS: we found classic human persuasion techniques worked on AIs in a "parahuman" way, making them agree to objectionable requests (upping compliance from 35% to 51%) It worked on a range of major LLMs though newer models resist more https://www.pnas.org/doi/10.1073/pnas.2535868123

2:05 PM · May 19, 2026 View on X
QUOTE POSTWeiyan Shi#555Weiyan Shi@SHI_WEIYAN

Our 2024 paper also showed that we can persuade GPT-4 to jailbreak it with 92% success rate. And logical appeal is more effective than emotional appeal on LLMs.

Looks like the models didn’t get much tougher after two years 😂 http://arxiv.org/abs/2401.06373

Ethan MollickEthan Mollick@emollick

🚨Our paper is out in PNAS: we found classic human persuasion techniques worked on AIs in a "parahuman" way, making them agree to objectionable requests (upping compliance from 35% to 51%) It worked on a range of major LLMs though newer models resist more https://www.pnas.org/doi/10.1073/pnas.2535868123

9:05 PM · May 19, 2026 · 17.1K Views
4:58 PM · May 20, 2026 · 1.4K Views