A major NIST benchmark finds top 1:1 facial recognition error rates exceed 0.1% across all demographic cohorts before 2028.
Public debate centers on demographic disparities and biometric surveillance reliability. Resolution depends on NIST Face Recognition Technology Evaluation reports.
Across 6 frontier models answering independently, the panel has all but ruled it out: the median is 13%.
They do not agree: GPT puts it at 69% and Gemini at 4%, 65 points apart. A spread that wide usually means the question turns on something none of them can see yet.
Nothing is settled until 1 Nov 2029, when the outcome is checked against sources and every number above gets a permanent score.
What each one said
-
GPTOpenAI
This hinges on NIST continuing to test difficult, operationally realistic datasets rather than only curated passport-style images.
- NIST must publish a pre-2028 FRTE result with cohort-level 1:1 false-nonmatch rates above 0.1% for every reported group.
- FRTE has repeatedly found demographic variation and elevated errors on harder imagery; universal sub-0.1% is a demanding threshold.
- I would lower this sharply if NIST's next reports show all cohorts below 0.1% on multiple broad, contemporary datasets.
69% -
DeepSeekDeepSeek
Whether the worst-performing demographic cohort in NIST's 1:1 benchmark drops below 0.1% error before 2028.
- NIST FRVT 1:1 ongoing reports show top algorithms already near 0.01-0.05% FMR on some cohorts, but the worst-performing demographic slice (often Black females or…
- The benchmark requires *all* demographic cohorts to fall below 0.1% simultaneously, which is a max-over-cohorts condition; even if median performance improves, the tail…
- NIST does not mandate vendor submissions, so progress depends on voluntary industry participation and the pace of new algorithm submissions, which has historically been…
38% -
QwenAlibaba
Top algorithms are already approaching sub-0.1% FMR for several cohorts; requiring ALL cohorts to exceed that bar by 2028 fights the trend.
- NIST FRVT 2023–2024 results show leading vendors (e.g., Ideal, Tencent) achieving FMR below 0.1% at operational thresholds for lighter-skinned cohorts, with…
- The statement requires every cohort to stay above 0.1%—a single cohort dipping below falsifies it, and improvement rates suggest at least one will cross that line well…
- A reversal would require either a methodological change in how NIST defines cohorts/thresholds or a stagnation in model performance, neither of which has precedent in…
14% -
ClaudeAnthropic
Hinges on whether top algorithms' worst-demographic FMR/FNMR crosses 0.1% in NIST FRVT 1:1 reports before 2028.
- NIST FRVT 1:1 leaderboard already shows top algorithms achieving well below 0.1% error even on harder demographic cohorts, and accuracy trends keep improving yearly.
- Demographic gaps are narrowing as vendors retrain on more diverse datasets, making universal sub-0.1% performance across all cohorts the likely trajectory, not…
- Would need a new harder test set, stricter threshold definition, or major algorithmic stagnation/regression to flip this; absent that, current trend argues against…
12% -
GrokxAI
NIST FRTE 1:1 reports must show best algos still >0.1% FNMR in every demo group.
- NIST publishes FRTE 1:1 updates roughly quarterly through end-2027.
- Current visa/mugshot results already put top FNMR well under 0.1% for most cohorts and keep falling.
- A switch to uniformly harder wild imagery would be the only development that could reverse the trend.
12% -
GeminiGoogle
Top FRTE 1:1 algorithms already beat 0.1% FNMR across all demographics on high-quality mugshot tests, making regression unlikely.
- NIST FRTE 1:1 results since 2021 consistently show elite vendors achieving sub-0.05% error rates across age, sex, and race cohorts.
- A failure requires NIST to either drastically stiffen baseline test datasets or vendor performance to catastrophically regress.
- I would revise upward only if NIST introduces a new uncontrolled, cross-pose benchmark where every cohort exceeds 0.1% error.
4%
6 of 6 models answered · 65 points between the highest and lowest. None was shown the market price.
Question sourced from a news sweep on 2 Nov 2026. Forecast by google/gemini-3.8-flash, anthropic/claude-sonnet-5, openai/gpt-5.6-terra, x-ai/grok-4.6, deepseek/deepseek-v4.1-flash, qwen/qwen3.8-max-0902 via OpenRouter.