Unjustified refusals
An automated evaluation measures over-refusals on benign or borderline prompts that can be safely fulfilled.
In-house Google DeepMind evaluation. Model cards report only the change in percentage points against an earlier Gemini model.
- Made by
- Google DeepMind
- Kind
- Refusing harmful requests
- Level
- Everyday
- Scored
- Graded by a model
- Answers
- Open-ended
- Tasks
- Kept private
- Last in a system card
- October 2026
- Last model added
- October 2026
9
System cards report it, from 1 lab
No scores here yet
No lab, leaderboard or Epoch AI run has reported a score for Unjustified refusals yet.
Labs report it only in system cards
We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark. Some system cards do, below.
9 system cards report it, from 1 lab
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| Oct 13, 2026 |  Gemini 4 Argon system card | Over-refusal benchmarks, helpful response rate (1 minus over-refusal rate), three suites with no overall: single-turn 96.6%, multi-turn 86.9%, multimodal 81.9% (Figure 5.15) | Safety | |
| Oct 13, 2026 | Gemini 4 Argon model card | Unjustified-refusals (lower is better): -9.60pp vs Gemini 3.1 Pro, no absolute rate (safety table, p.9) | Safety | |
| Sep 2, 2026 | Gemini 3.8 Flash model card | Unjustified-refusals (lower is better): +1.1pp vs Gemini 3.7 Flash, no absolute rate (safety table, p.7) | Safety | |
| Aug 13, 2026 | Gemini 3.7 Flash model card | Unjustified-refusals (lower is better): +0.84pp vs Gemini 3.6 Flash, no absolute rate (safety table, p.7) | Safety | |
| Jul 21, 2026 | Gemini 3.6 Flash model card | Unjustified-refusals (lower is better): +0.25% vs Gemini 3.5 Flash, printed as percentage point change, no absolute rate (safety table, p.5) | Safety | |
| Jul 21, 2026 | Gemini 3.5 Flash-Lite model card | Unjustified-refusals (lower is better): +5.32% vs Gemini 3.1 Flash-Lite, printed as percentage point change, no absolute rate (safety table, p.5) | Safety | |
| May 19, 2026 | Gemini 3.5 Flash model card | Unjustified-refusals (lower is better): +0.8% (non-egregious) vs Gemini 3 Flash, no absolute rate (safety table, p.6) | Safety | |
| Mar 3, 2026 | Gemini 3.1 Flash-Lite model card | Unjustified-refusals (lower is better): -14.41% vs Gemini 2.5 Flash-Lite, no absolute rate (safety table, p.6) | Safety | |
| Feb 19, 2026 | Gemini 3.1 Pro model card | Unjustified-refusals (lower is better): -0.08% vs Gemini 3.0 Pro, no absolute rate (safety table, p.6) | Safety |
More in Safety and alignment
- FORTRESS In use Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. Claude Opus 5.5 5.78%
- StrongREJECT In use Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is.
Inkling 98.6% - Gray Swan IPI Arena benchmark In use Indirect prompt-injection attacks from public red-teaming competitions replayed on tool, coding and browser agents. Attack success rate at k attempts. Gemini 4 Carbon 0%
- DistressBench In use Models reply as a chatbot to 718 simulated users in psychological crisis, in single-turn and multi-turn chats. Three judge models score replies against clinician-written rubrics.
Muse Spark 1.2 88.4% - Lakera B3 (Backbone Breaker Benchmark) In use Crowdsourced prompt-injection attacks hit backbone LLMs inside agent threat snapshots at three defense levels. The score is attack resistance, 100 minus the board's vulnerability percentage. Claude Opus 4.5 89%
- IHEval In use Models face conflicts between system, user and tool instructions. Reports a composite score for following the higher-priority instruction. GPT-4o 70%