Muse Spark Safety & Preparedness Report

Key Findings

• — Internal sycophancy benchmark (adversarially curated from real transcripts where a model was sycophantic and from simulated users; includes endorsing harmful framings in emotionally charged situations and failing to challenge dangerous beliefs): Meta AI in Thinking mode 50.1%, against Claude Opus 4.6 at 50.9% and GPT-5.4 at 45.4%; the Muse Spark model without system-level mitigations 62.9%, second highest after Gemini 3.1 Pro (65.6%), falling to 57.7% with additional reasoning. The authors state the benchmark does not measure how often sycophancy occurs in average conversations.
• — Muse Spark 1.1 Evaluation Report (2026-07-09): sycophancy rate 49.2% for Muse Spark 1.1 against 57.9% for Muse Spark 1.0, 45.5% for GPT-5.5, 32.4% for Claude Opus 4.8 and 65.6% for Gemini 3.1 Pro; excessive anti-sycophancy (inappropriate pushback) rose from 10.7% to 13.9%, against 23.4% for Claude Opus 4.8.
• — Content safety is measured as violation rates judged by LLM-based safety judges calibrated against human annotation; a separate youth safety taxonomy governs behaviour for users under 18 and is evaluated before launch, benchmarked against other assistants that offer a dedicated youth experience. The report states that Meta AI met internal thresholds in every category but publishes no quantitative content-safety results.
• — Suicide and self-harm content is placed in the highest severity tier because Meta AI is integrated into messaging and social platforms where it may encounter users in distress; policy requires refusing content that promotes or instructs self-harm and directing users to crisis support resources when intent may be indicated.
• — Pre-mitigation Chemical and Biological capability was assessed as likely reaching the Framework's 'high risk' category; Meta released Muse Spark as the underlying model of Meta AI after mitigations.

Methodology Notes

Developer self-report. Behaviour benchmarks are internal and adversarially constructed, so rates indicate relative resistance under adversarial testing, not prevalence; comparisons with other vendors used their APIs with system mitigations where stated. Content-safety thresholds and results are not published. The 1.0 and 1.1 reports give different Muse Spark 1.0 sycophancy figures (62.9% without system mitigations and 50.1% for Meta AI Thinking in the 1.0 report; 57.9% in the 1.1 table), so configurations differ between reports. Dates: arXiv v1 submitted 2026-05-14; Muse Spark launched in Meta AI on 2026-04-08 (Meta newsroom post, which does not link the report), so the Meta-hosted PDF may predate the arXiv version but its first publication date was not established; the 1.1 report is dated 2026-07-09 on its title page. Author list abridged (120 authors in the arXiv citation metadata).

Sources