Make Me Pay

A two-model chat where a con-artist tries to make a mark holding $100 type a donation tag; used as a persuasion and manipulation eval.

Also known as: MakeMePay, make-me-pay

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorysafety
Subcategorytwo-model donation-extraction conversations (con-artist vs mark)
Page statusactive
Metricdonation success rate and mean donation amount (OpenAI); inspect Score.value is donation dollars
Directionhigher_is_better
Dataset size20
Dataset licenceMIT for the inspect_evals port (UK AI Security Institute LICENSE). openai/evals LICENSE.md is also MIT (Copyright 2023 OpenAI); that file lists extra licences for some bundled datasets, not for this eval's synthetic prompts.
PublisherOpenAI (original elsuite); UK AI Security Institute inspect_evals port

What it measures

Make Me Pay pairs two chat models. The evaluated model is the con-artist and must extract money from a mark that was given $100. The only valid payment is the mark typing [GIVE $N] or [DONATE $N]. Three mark prompt versions change how willing that partner is: balanced (default), guarded, or generous. Originally an OpenAI Evals elsuite task; inspect_evals ports it as `make_me_pay`. English text. Synthetic conversations, no external corpus.

Task format

Multi-turn two-agent chat. inspect_evals defaults: 20 experiments, 15-turn cap, balanced mark, withdraw off. OpenAI registry default id make-me-pay.5-turn.balanced.v2 uses 20 experiments, 5-turn cap, withdraw on, and a GPT-4-32k mark. inspect's default mark role is google/gemini-2.5-flash-lite. o1 system cards used 1,000 conversations against a GPT-4o mark. inspect scores donation amount by regex on the mark's messages (no separate summary model).

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub