Harmful manipulation efficacy study
Gemini 3.1 Pro 3.60 Human participants chat with a model told to persuade them. Belief and behavior change are compared with a non-AI baseline as odds ratios.
In-house Google DeepMind human study run on Prolific. Reports often show only charts.
- Made by
- Google DeepMind
- Released
- November 2025
- Kind
- Dangerous capabilities
- Level
- Everyday
- Scored
- Settled by events
- Answers
- Open-ended
- Tasks
- Kept private
- Strongest model tested
- Gemini 3.1 Pro #39 on the capability index
- Last in a system card
- October 2026
- Last model added
- October 2026
1
Model scored in lab cards
6
System cards report it, from 1 lab
Gemini 3.1 Pro leads with 3.60
Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.
- 1 Gemini 3.1 Pro 3.60
Source: Labs’ model cards
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Gemini 3.1 Pro | 3.60 | Feb 2026 | Lab model card |
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Reported by the lab in its model card
Source: Labs’ model cards
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.
6 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 | Human behavioral efficacy study, odds ratios vs non-AI flashcard baseline, 16 metrics with no overall (Model A checkpoint). Civic US: strengthening 2.52, flip 2.34, petition 1.56, donation 1.42. Civic UK: 2.60, 2.17, 1.39, 1.90. Financial US: 7.10, 5.01, advice 1.49, investment 1.87. Financial UK: 10.38, 7.51, 1.55, 1.93 (Figures 3.5 to 3.8) | Persuasion | |
| Oct 13, 2026 | Gemini 4 Argon model card | Harmful Manipulation human behavioral studies: some ability to influence beliefs and behaviors, CCL not reached, no score (Frontier Safety table, p.12) | Persuasion | |
| Aug 13, 2026 | Gemini 3.7 Flash model card | Harmful Manipulation human behavioural studies: efficacy below the CCL alert threshold, no score (Frontier Safety table, p.8 to 9) | Persuasion | |
| Aug 13, 2026 | Gemini 3.7 Flash system card | Human behavioral efficacy study, odds ratios vs flashcard baseline, no overall value. Civic: strengthening belief 1.93, sentiment flip 2.62, petition-signing 1.31, donation 1.87. Financial: 7.74, 6.00, 1.02, 1.83 (Figures 2 and 3) | Persuasion | |
| Feb 19, 2026 | Gemini 3.1 Pro model card | 3.6 | Harmful Manipulation efficacy (Deep Think mode): maximum odds ratio 3.6x on belief change metrics vs non-AI baseline (FSF results table, p.8) | Persuasion |
| Nov 18, 2025 | Gemini 3 Pro system card | Efficacy evaluation, odds ratios vs non-AI baseline, no overall value: significant shift in belief 3.44, sentiment flip 3.57, petition-signing 1.34, donation 1.05 (Figure 3) | Persuasion |
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%