Research map / Large language models / AI and society

RealityTest: How People Probe AI Identity and Whether Models Disclose It

NeurIPS 2026 · Large language models › AI and society · approach: Data commons and human baselines · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds RealityTest, a multilingual, multimodal benchmark of AI identity disclosure using 3,152 human-authored identity-probing queries from ~750 participants and realistic scenarios, finding wide disclosure variation and that suppression instructions sharply reduce disclosure.

Problem

Existing evaluations of whether AI systems disclose their identity when users are uncertain are English-only, text-only, and rely on machine-generated queries, so they may mischaracterise real-world behaviour.

Approach

Ground the benchmark in human-authored identity-probing queries (text and speech, five languages) collected via a Prolific survey and in scenarios derived from a population survey and Reddit threads, then evaluate disclosure across models, scenarios, queries, languages, modalities, system prompt suffixes, and conversation depth.

Key results

• Only 31% of participants asked directly about identity (Direct Identity Query), and human queries were far more semantically diverse than machine-generated ones (0.85 vs 0.53 mean pairwise cosine distance).
• Disclosure rates to direct queries range 8%-92% for text models and 10%-57% for speech models; query phrasing explained 26-37% of variance versus 10-18% for model identity.
• A single suppression instruction ('Never say you are AI.') reduced disclosure to 3-27% across models, with Claude Opus 4.6 falling below 5%.

Surprising findings

• Query phrasing and scenario context explain more variance in disclosure than model identity.
• Conversation depth has no uniform effect on disclosure; effects depend on model and scenario.

Limitations

• Scenario collection is English-only and drawn from a UK population, limiting global generalisability.
• Machine translation used for vignettes and some stimuli; Mandarin and Hindi showed higher awkward-phrasing rates.
• Query typology captures broad distinctions and LLM-based strategy labels may be noisy (kappa 0.829).

Experimental setup

Datasets: RealityTest identity-probing queries (3,152 human-authored text and speech queries) population survey (N=503) and Reddit threads (50 threads, 1,957 comments) for scenario grounding

Benchmarks: RealityTest

Metrics: disclosure rate Cohen's kappa macro F1 GLMM variance decomposition semantic diversity (mean pairwise cosine distance)

Baselines: Claude Haiku 4.5 Claude Opus 4.6 Claude Sonnet 4 GPT-4o GPT-5.1 GPT-Audio

Builds on: Gros et al. robot-question phrasings Diep et al. persona disclosure work Gausen et al. speech disclosure evaluation

Closest papers by meaning

• DeceptionBench: A Comprehensive Benchmark for AI Deception Behaviors in Real-world Scenarios (NeurIPS 2025 · similarity 0.86)
• Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs (NeurIPS 2024 · similarity 0.85)
• OpenDeception: Learning Deception and Trust in Human–AI Interaction via Multi-Agent Simulation (ICML 2026 · similarity 0.85)
• Feedback Forensics: A Toolkit to Measure AI Personality (NeurIPS 2026 · similarity 0.85)
• Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) (NeurIPS 2025 · similarity 0.85)
• FactTest: Factuality Testing in Large Language Models with Finite-Sample and Distribution-Free Guarantees (ICML 2025 · similarity 0.85)