Research map / Large language models / AI and society
Confirmation Bias: A Challenge for Scalable Oversight
Gabriel Recchia, Chatrik Singh Mangat, Jinu Nyachhyon, Mridul Sharma, Callum Canavan, Dylan Epstein-Gross, Muhammed Abdulbari
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LLM-generated research guidelines to open consultancy plus online research (ablation) | open consultancy with gpt-4-0613 · Main study, both waves combined | Accuracy: 0.74 -> 0.73 (p = 0.19) | No clear effect |
| add LLM-generated research guidelines to open consultancy (ablation) | open consultancy with gpt-4-0613 · Main study, Wave 1 | Accuracy: 0.74 -> 0.72 (p = 0.027) | Hurt |
| add LLM-generated research guidelines to open consultancy (ablation) | open consultancy with gpt-4-0613 · Main study, Wave 2 | Accuracy: 0.74 -> 0.75 (p = 0.54) | No clear effect |
| add LLM-generated research guidelines (ablation) | open consultancy with gpt-4-0613 · Main study, Wave 1 | Logit of final probability for correct answer: 1.33 -> 1.05 | Hurt |
| add LLM-generated research guidelines (ablation) | open consultancy with gpt-4-0613 · Main study, mixed model | Group x timepoint interaction on logits: B = -0.16, OR = 0.85 | Hurt |
| add one-turn debate arguments to unstructured LLM interaction (ablation) | gpt-3.5-turbo-0301 assisted participants · Preliminary experiment | Accuracy: 69% -> 73% (p = 0.37 vs LLM alone at 75%) | No clear effect |
| human-LLM interaction instead of LLM alone | gpt-3.5-turbo-0301 · Preliminary experiment | Accuracy: 75% -> 69% (p = 0.003) | Hurt |
| conduct online research to verify the LLM consultant's answer | participants evaluating gpt-4-0613 answers · Main study, mixed model | Logit of probability on correct answer: +1.17 logits (OR = 3.21) | Improved |
| model answer is incorrect vs correct | participants' updating towards the consultant · Main study | Update toward consultant after research: B = 0.72, OR = 2.06 | Mixed |
| more conversation turns with the LLM in Bowman et al. data | Human + Model condition · MMLU reanalysis | Odds of answering correctly: OR = 0.93 per additional turn | Hurt |
| more conversation turns with the LLM in Bowman et al. data | Human + Model condition · QuALITY reanalysis | Accuracy | No clear effect |
| resetting the conversation history at least once | Human + Model condition · MMLU reanalysis | Probability of answering correctly: chi2 = 4.24, p = 0.039 | Hurt |
Closest papers by meaning
- Scaling Laws For Scalable Oversight (NeurIPS 2025 · similarity 0.88)
- Great Models Think Alike and this Undermines AI Oversight (ICML 2025 · similarity 0.88)
- LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical Study (AAAI 2025 · similarity 0.87)
- Calibrating Conservatism for Scalable Oversight (ICML 2026 · similarity 0.87)
- ACED-Bench: Evaluating How LLM Agents Acquire and Act on Causal Evidence (NeurIPS 2026 · similarity 0.86)
- On scalable oversight with weak LLMs judging strong LLMs (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
