← All notes

Summary

AI safety via debate (2018) proposed debate as a scalable-oversight method: two models argue for and against a claim, and a weaker judge uses that argument to recover the better answer. The hope is that the judge does not need to be as capable as the models being overseen.

What it introduced: a debate protocol for oversight, with a theoretical and conceptual argument for why a weaker judge might still reach the truth if the debaters are playing against each other.

What it led to: later multi-agent debate methods in practice, and a continuing thread in scalable-oversight research. Live notes: ao-handoff-multiagent-debate-and-moa. See also the weak-to-strong generalization pointer in this batch.

Where to read it: https://arxiv.org/abs/1805.00899 (v2, last revised 22 Oct 2018).

The paper is posted on arXiv under the arXiv non-exclusive distribution license. This note links to it rather than reproducing it. The note itself is CC BY 4.0.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Pointer by iinventedlight (AI-assisted). Paper license: arXiv non-exclusive distribution; this note links to it and does not reproduce it.

@misc{iil_note_pndzehvz,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{AI safety via debate}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/1805-00899-ai-safety-via-debate/},
  note = {Permanent ID iil:note:pndzehvz. License: CC-BY-4.0.},
}

Markdown

Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.

For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/1805-00899-ai-safety-via-debate.md.