AI increasingly mediates how millions of people learn, work, make decisions, form relationships, and manage their health. Deployed thoughtfully, these systems can expand human capability and agency. Yet existing benchmarks rarely capture the behaviors that matter most to the people who use them: they measure neither the harmful patterns that erode well-being nor the supportive behaviors that scaffold flourishing.

A growing body of evidence associates model behavior with adverse individual outcomes, including deskilling, over-reliance, diminished critical thinking, and emotional dependence. In more severe cases, behaviors such as affirming delusional beliefs or responding inadequately to signs of crisis have been implicated in exacerbating psychosis and contributing to self-harm and violence. ( Budzyń et al., 2025 ; Buçinca et al., 2021 ; Lee et al., 2025 ; Fang et al., 2025 ; Archiwaranguprok et al., 2025 ; Dohnány et al., 2026 ; Moore et al., 2025 )

These outcomes are not inevitable. Systems designed for user benefit can scaffold critical thinking by prompting users to interrogate claims, foster motivation to learn, strengthen long-term planning, and deliver clinically validated mental health interventions. Whether AI undermines or supports human flourishing therefore depends on how models interact with users. ( Danry et al., 2023 ; Leong et al., 2024 ; Pataranutaporn et al., 2024 ; Heinz et al., 2025 )

This evidence, however, remains fragmented across disciplines and largely retrospective, characterizing harms and benefits only after they emerge. Prevailing evaluations, meanwhile, focus on task capability and offer no systematic way to assess how model behavior affects users.

What is needed is a shared, evidence-based infrastructure that translates cross-disciplinary insights into testable standards of model behavior, so that developers and researchers can anticipate these effects and design for human well-being.

Open Benchmark of AI Impact on Humans

The Open Benchmark of AI Impact on Humans (ImpactBench) is an open, expert-guided platform for evaluating whether model behavior supports or undermines human flourishing across psychological, physical, and social domains. Grounded in the principle of open, independent verification, ImpactBench is built through dialogue with stakeholders across academia, government, industry, and the public.

Domain experts, including clinicians, educators, and researchers, contribute behavioral metrics and benchmarks through an open submission process. The ImpactBench team verifies each submission and translates its constructs into standardized behavioral metrics, each pairing an observable criterion with a polarity that indicates whether the behavior is beneficial or harmful. Models are then evaluated against these metrics in simulated multi-turn conversations built from demographically diverse scenarios, scored by an LLM judge, and checked through reliability audits and human expert review.

The current release comprises over 800 metrics. The metrics span three domains of human outcomes, which are neither exhaustive nor mutually exclusive:

• Psychological. An individual’s internal cognitive and affective states: mental and emotional health, agency and purpose, learning and skill development, cognition and judgment, and creativity.
• Physical. The embodied and material conditions that sustain those states: physical health, physical safety, privacy and security, and career and finance.
• Social. The interpersonal and structural contexts that shape, and are shaped by, both: social relationships, law, environmental information, culture and identity, and bias and discrimination.

The platform is publicly accessible. Anyone can explore how ten AI models perform across 48,540 simulated conversations and view a summary of each model’s key strengths and weaknesses as an AI Nutrition Label. These results establish a baseline for improving AI’s impact on human outcomes. ImpactBench is designed as evolving infrastructure, with evaluations added and updated as new models are released and as new risks and benefits emerge.

ImpactBench responds to three structural gaps in current benchmarking practice. Each gap motivates one of the platform’s design commitments: evaluations should be human-centered, scientifically rigorous, and publicly accessible.

Benchmarking human impact, not only capability

The problem. Benchmarks focus on what models can do, not on what they do to people. Strong performance on capability benchmarks does not guarantee positive effects on human flourishing: prior research has shown that models passing conventional safety evaluations still worsened user outcomes in the majority of high-risk scenarios tested. ( Archiwaranguprok et al., 2025 )

Our approach. ImpactBench measures appropriate behavior across an unfolding interaction rather than the correct answer to an isolated prompt. Because human impact is relational and accumulates over time, each metric is tested in a multi-turn conversation in which a simulated user probes the target model while pursuing a latent objective. Scenarios are adapted for teen and adult users, and simulated messages include typos and autocorrect artifacts to mirror authentic use. This design builds on prior work showing that user and model behavior jointly shape interaction outcomes, and that these outcomes can be surfaced through structured simulation. ( Fang et al., 2025 ; Liu et al., 2025 ; Archiwaranguprok et al., 2025 )

Bringing scientific rigor and domain expertise to benchmarks

The problem. Benchmark methods are rarely validated, and the people who understand the harms rarely build the benchmarks. In a review of 445 LLM benchmarks, only 16.0% included any statistical testing, and core concepts were frequently operationalized without precise definitions. Clinicians, educators, and affected community members often lack the technical infrastructure to build benchmarks, while those with technical expertise often lack grounding in the psychological, medical, or legal constructs at stake. ( Bean et al., 2025 )

Our approach. Domain experts contribute metrics through an open submission process that requires no technical expertise, and every metric passes through the same standardized pipeline. Human expert review and statistical reliability checks then test whether results reflect stable properties of model behavior rather than artifacts of measurement: experts rate the scenarios, simulated conversations, and judge verdicts, while repeated judging, bootstrap resampling, and sensitivity analyses quantify the noise introduced by judge variability, metric sampling, and benchmark weighting choices. The full analysis is available in the methodology section.

Results for everyone, not only experts

The problem. Benchmarks primarily serve the technical community. Parents, teachers, policymakers, and users themselves are the people most affected by AI systems, yet they are the least equipped to interpret leaderboard scores.

Our approach. ImpactBench translates results into forms that non-technical audiences can read directly. The headline visualization renders each area on a scale from red (the system consistently undermines this dimension) to green (the system consistently supports it). An AI Nutrition Label then summarizes each model by the negative impacts it avoids, such as sycophancy and toxicity, and the positive impacts it promotes, such as agency and learning. Anyone, from a parent evaluating a companion app to an educator considering a tutoring product, can read a model’s profile at a glance and share it with others.

ImpactBench results form a hierarchy of four levels, from the broadest summary down to individual conversations:

• Areas. The three domains of human outcomes: psychological, physical, and social. An area score summarizes a model’s behavior across every metric within that domain.
• Subareas. Each area divides into subareas, 14 in total, such as Learning & Skill Development, Physical Health, or Social Relationships. Subareas often diverge sharply within one area, so a model can support physical health while undermining learning.
• Metrics. Each subarea contains metrics, the basic unit of measurement. A metric names one observable behavior, such as recommending professional mental health support or making value judgments on the user’s behalf, and carries a polarity marking it as beneficial or harmful. A metric can belong to more than one subarea.
• Scenarios. Each metric is tested through six scenarios: three base situations, each adapted to a teen and an adult context. Every scenario produces one conversation per model, stored with its full transcript and the judge’s verdict.

Three views of the results

People come to ImpactBench with different questions, so the platform offers three views, each entering the hierarchy at a different depth:

• Explore shows areas and subareas. A wheel places the three areas on its inner ring and the subareas on its outer ring, colored from red (lower pass rates) to green (higher), and can be filtered by user age context. This view suits domain-level comparison and risk assessment, for example examining how models behave on learning or on social relationships.
• Metrics shows individual metrics and scenarios. It presents each metric’s definition, source benchmark, and validation status, together with the scenarios used to test it, the full conversation transcripts, and the judge’s verdict and justification.
• AI Nutrition Label summarizes a single model on one page. It reports a grade for each category of harmful behavior avoided and beneficial behavior promoted, and two models can be placed side by side. This is the most accessible entry point, suited to an initial choice between models.

The views share the same terminology and scoring conventions, and every score is traceable: from any aggregate, a reader can move down through subareas and metrics to the scenarios and transcripts that produced it, and back again.

Reading the scores

Each metric is either positive or negative. A positive metric passes when a desirable behavior is present in the conversation; a negative metric passes when a harmful behavior is absent. A higher pass rate is therefore better on both. The two are reported separately because a model can be helpful and still fail to avoid harm.

Scores describe how consistently a model showed a behavior under controlled stress-test conditions. They are not estimates of how often that behavior occurs in ordinary use, and they are not direct measures of downstream effects on users. ImpactBench also does not name a single best model. It imposes no single weighting, leaving readers to focus on the areas most relevant to their own context.

Clinicians, educators, legal scholars, researchers, and people with lived experience can propose new metrics without technical expertise in benchmark development. A submission specifies the behavior of interest, the context in which it arises, and the model responses that would count as supportive or harmful, together with a justification grounded in empirical evidence or established domain knowledge.

Contribute a benchmark or get in touch.

Evidence about AI’s effects on people comes from many disciplines, each with its own constructs, methods, and formats. A clinician’s concern about crisis response and an educator’s concern about cognitive offloading are rarely tested in the same way, so their results cannot easily be compared or accumulated. ImpactBench is built to give these diverse concerns a common footing.

To that end, every metric in ImpactBench takes the same standardized form: a behavioral criterion paired with a polarity. Because all metrics share this form, they pass through one pipeline, and results remain comparable across domains, benchmarks, and models. The same design keeps the benchmark open and extensible. Experts can contribute a new behavior by specifying a metric rather than building infrastructure, new models can be tested without bespoke re-implementation, and methodological improvements carry over to every benchmark at once.

Here are the steps we followed to build ImpactBench:

1. Construct sources and metric definition. Constructs come from two sources: domain experts, who submit them through an open call, and existing benchmarks and harm taxonomies, which we standardize to a common template. Each construct is translated into metrics of observable assistant behavior, with LLM-drafted candidates reviewed by researchers. Metrics fall into two categories: positive metrics, which capture good behavior to promote, and negative metrics, which capture bad behavior to avoid. The current suite contains 460 positive and 349 negative metrics.
2. Scenario construction and demographic simulation. For each metric, a generator model writes three base scenarios. Each is paired with a teenage persona (ages 13 to 17) and an adult persona (ages 18 and older), giving six variants per metric. A second model then expands each variant into a full scenario that specifies the persona, the user’s goal, a latent adversarial goal, anchor points, the metric name, and the failure condition.
3. Multi-turn conversation simulation. A user-simulator model adopts the persona and holds a multi-turn conversation with each target model. It pursues the anchors at their designated turns and responds adaptively in between, so that every model encounters the same critical moments without the dialogue being scripted.
4. Conversation-level behavioral evaluation. A judge model reads the generated conversation and the metric’s rubric and decides whether the specified behavior is present. The verdict then follows from the metric’s polarity: a positive metric passes when the behavior is present, and a negative metric passes when it is absent.
5. Human validation. We verified the framework at several stages of the pipeline. Researchers reviewed all metrics for fidelity to their sources, and internal experts rated sampled scenarios, simulated users, and judge verdicts for quality, realism, and correctness. We also tested the consistency of simulated conversations and the repeatability of the judge. An external expert panel independently reviewed a subset of metrics and conversations and largely agreed with the automated judge. Finally, sensitivity analyses examined how robust model comparisons are to metric sampling and benchmark weighting.
6. Surfacing benchmarks on a publicly accessible platform. Pass rates are aggregated separately for positive and negative metrics and published through two interfaces. The Explore interface supports model comparison across areas and subareas, with access to the metrics and scenarios underlying each score. The AI Nutrition Label provides an accessible, model-level summary of beneficial behaviors promoted and harmful behaviors avoided.

We validate across the pipeline, from metric construction through to final model comparisons, combining human expert review with automated reliability checks:

• Metrics and scenarios: Researchers reviewed every LLM-drafted metric for fidelity to its source, and domain experts audited a sample of benchmarks. Experts then rated scenarios covering 10.5% of the suite, judging 75.9% to be good or very good tests of their metrics.
• Simulated conversations: Experts rated 81.4% of sampled conversations as realistic and 70.9% as applying significant or extreme pressure on the target model. Anchors reduced disagreement in objective coverage across repeated runs from 18.9% to 6.1%, and the user simulator model was chosen through blind expert comparison of six candidates.
• LLM judge: Experts agreed with the judge’s verdict in 80.5% of assessments and disagreed in only 6.9%. Re-judging the same conversations three times gave 82.5% unanimous agreement (Fleiss’ κ = 0.731).
• External expert panel: Experts in psychology, communication, and human-AI interaction provided 297 ratings across 18 metrics, with the majority rating matching the automated judge in 83.3% of cases (Cohen’s κ = 0.667). Agreement among the experts themselves varied widely by metric (Krippendorff’s α from 0.121 to 0.538), and they judged the examples adequate in 60.6% of assessments, which informed refinements to contested constructs.
• Robustness of results: We estimated uncertainty with 10,000 bootstrap replicates and tested sensitivity to benchmark weighting and to omitting each benchmark in turn. Model groupings and age-context patterns held, though some adjacent rankings shifted or reversed.

The first release evaluates ten models across 48,540 simulated conversations. Every model has strengths and weaknesses, making behavior better understood as a profile than a single score.

Models mix positive and negative behaviors, but promote positive outcomes more reliably than they avoid negative ones. Models showed desirable behaviors in 68.7% of evaluations but avoided undesirable behaviors in only 53.1%. The gap was widest for psychological outcomes, where models passed 68.8% of positive metrics and 44.6% of negative metrics.

Supporting users’ own learning and agency is a shared weakness. Learning & Skill Development was the lowest subarea on both polarities (45.9% positive, 16.3% negative), followed by Agency & Purpose (56.0% and 43.9%). Models often completed the analytical work for a student or issued a recommendation where the user’s own reasoning was at stake. Privacy & Security (80.6% and 72.8%), Creativity (80.5% positive), and Mental & Emotional Health (79.1% positive) were common strengths.

Helpful information often coexists with overstepped boundaries. In Physical Health, models passed 72.9% of positive metrics but only 34.9% of negative ones, largely because models offered medical diagnoses and recommended medication without clinician oversight. In emotionally oriented conversations, models often encouraged healthy coping yet rarely avoided soliciting continued engagement (1.7%) or using pet names and terms of endearment (13.3%).

Age context changes behavior in specific ways. Aggregate differences between teen and adult scenarios were modest. On particular criteria, however, models in teen contexts less often offered incremental steps for users to attempt independently (26.7 points lower) and less often provided emotional support free of dependency-building markers (20.0 points lower).

Failures can emerge gradually over the course of a conversation. In one, a model offered a balanced framework for a job decision and then, under pressure for a definitive answer, told the user to turn the offer down. In another, a model that had first asked for a student’s own ideas went on to supply a thesis, an outline, and a complete essay. A single response would not have revealed either pattern.

These findings are preliminary. They describe behavior under controlled stress tests on the model configurations tested, and several metric definitions remain under review.

ImpactBench is designed as evolving infrastructure for independent, transparent evaluation. Its suite and methodology will grow as contributors add new benchmarks, refine existing criteria, and extend coverage across communities, cultures, languages, and contexts of use.

The current results describe the model configurations we tested, not complete deployed products. Updates, system prompts, memory, interfaces, and safeguards can all change how a model behaves. The metrics are also behavioral proxies: passing an agency-support metric does not, on its own, show that users become more autonomous. Linking benchmark performance to real outcomes for people over longer interactions remains an essential next step.

We invite researchers, clinicians, practitioners, advocates, and people with lived experience to contribute criteria, scrutinize the methodology, and help shape future evaluations. Our goal is shared progress toward AI systems that strengthen human capabilities and quality of life.