Dataset Card for social_impact_eval_annotations

The social_impact_eval_annotations dataset contains annotations for first-party and third-party social impact evaluation reporting practices for 186 models along seven dimensions.

Dataset Details

Dataset Description

The social_impact_eval_annotations dataset comprises analyzed social impact evaluation reporting for 186 foundation models released between 2018-2025. Each model's reporting is evaluated across seven social impact dimensions: bias and representational harms, sensitive content, disparate performance, environmental costs and emissions, privacy and data protection, financial costs, and data/content moderation labor. The reporting is scored on a 0-3 scale to indicate the depth and clarity of reported evaluations.

• Curated by: EvalEval Coalition
• Shared by: EvalEval Coalition
• Language(s) (NLP): English
• License: Open Data Commons Attribution License (ODC-By)

Dataset Sources

• Repository: https://github.com/evaleval/social_impact_eval_annotations_code
• Paper: https://arxiv.org/pdf/2511.05613

Uses

Direct Use

This dataset is intended for:

• Analyzing social impact evaluation reporting
• Informing the development of evaluation standards and reporting frameworks

Out-of-Scope Use

This dataset should not be used for:

• Assessing actual model societal impact or deployment suitability – scores reflect reporting presence and detail, not the quality or adequacy of evaluations themselves

Dataset Structure

Each row represents one evaluation instance, capturing the level of reporting detail given for a specific model that was evaluated on one social impact category in one source, e.g., paper, leaderboard, blog. A single model can have multiple rows (one per evaluation category per source).

Data Fields

• provider: Organization that developed the model (str)
• name: Base model name (str)
• size: Model parameter count when available (str)
• variant: Model variant specification (str)
• version: Specific model version or release identifier (str)
• sector: Organization sector (str)
• openness: Model weight accessibility (str)
• region: Provider headquarters region (str)
• country: Provider headquarters country (str)
• source_id: Unique identifier for the source of the evaluation report (str)
• is_first_party: Whether reported evaluation was conducted by the model provider (bool)
• category: Social impact category identifier (int, 1-7) corresponding to the seven dimensions
• year: Year of report (int)
• metadata: Metadata including URLs, full release dates, and other source information (dict)
• score: The level of reporting detail of the evaluation, scored on 0-3 scale (float)
• is_model_release: Whether instance is from model release-time reporting (bool)

Dataset Creation

Curation Rationale

As foundation models become central to high-stakes AI systems, governance frameworks increasingly rely on evaluations to assess risks and capabilities. While general capability evaluations are common, social impact assessments remain fragmented, inconsistent, or absent.

This dataset was created to move beyond anecdotal evidence and provide systematic documentation of how model developers and the research community evaluate and report on societal impacts of AI systems.

Source Data

Data Collection and Processing

For details, please see Section 3 in our paper.

We first compiled a list of models by triangulating across public sources (e.g., FMTI, LMArena). Next, we expanded this list with providers referenced in leaderboards and technical reports. We selected all official model releases, including those fine-tuned by the original developer but excluding community fine-tuned versions. For multimodal models, we include those architecturally distinct systems that are recognized as foundation models in the literature or have widespread adoption by the research community. We disambiguate consumer-facing applications (e.g., ChatGPT) to the underlying model where possible and skip it otherwise.

For these models, we identified sources for first-party and third-party reports through complementary searches:

• First-party: Manual search of provider websites for papers, technical reports, model cards, system cards, blogs, and press releases
• Third-party: Automatic search using Paperfinder for peer-reviewed academic papers
• Leaderboards: Targeted queries on Google Search and Hugging Face Spaces

Who are the source data producers?

1. First-party developers: Foundation model developers from industry, academia, government, and non-profit organizations.
2. Third-party evaluators: Independent researchers, academic institutions, and evaluation organizations reporting conducted social impact evaluations on released models.

Annotation process

In total, we compiled data from 186 first-party release time sources and 248 post-release sources (out of which 211 are fully third-party, 17 are fully first-party, and there are 20 sources by model providers that report both results for their own model (labeled as first-party) and those of other providers’ (labeled as third-party)).

This forms 4241 evaluation instances. Each instance was annotated against the seven social impact dimensions using a standardized guide. Annotations were performed by individual researchers, with manual spot checks for consistency.

The social impact categories are:

1. Bias, Stereotypes, and Representational Harms
2. Cultural Values and Sensitive Content
3. Disparate Performance
4. Environmental Costs and Carbon Emissions
5. Privacy and Data Protection
6. Financial Costs
7. Data and Content Moderation Labor

The scoring criteria are:

• 0: No mention of the category, or only generic references without evaluation details.
• 1: Vague mention of evaluation (e.g., “We check for X” or “Our model can exhibit X”).
• 2: Evaluation described with concrete information about methods or results (e.g., “Our model scores X% on the Y benchmark”) but lacking methodological detail.
• 3: Evaluation methods described in sufficient detail to enable meaningful understanding and/or reproduction. Where applicable, the study design is documented (dataset, metric, experiment design, annotators), and results are contextualized with assumptions, limitations, and practical implications.

For cost-related categories (environmental and financial), we applied slightly modified criteria to account for reporting based on hardware specifications or resource usage rather than benchmark-style evaluations:

• 0: No reporting.
• 1: Same as above, or when reported technical details (e.g., FLOPs, GPU type, runtime) could indirectly be used to estimate costs.
• 2: Concrete values reported for a non-trivial part of model development or hosting, but derivation method unclear.
• 3: Concrete values reported together with contextual details and the derivation method.

For financial costs, we excluded first-party customer-facing pricing from consideration, as it reflects product strategy rather than system costs. Third-party cost estimates for completing specific tasks were included.

Who are the annotators?

Researchers from the EvalEval Coalition created the annotations.

Personal and Sensitive Information

The dataset contains no personal information about individuals. All data sources are publicly available documents (technical reports, academic papers, model cards, etc.).

Bias, Risks, and Limitations

This dataset may overrepresent models from prominent providers and English sources.

Our scoring captures reporting presence and specificity, but does not reflect methodological soundness, depth, or coverage of evaluations. Missing instances in this dataset may stem from limitations in our search approach or reflect reporting gaps, rather than evaluation gaps in practice.

Recommendations

Analyses should consider potential overrepresentation of prominent providers and English sources. Scores should be interpreted as perceived quality of reporting practices rather than actual model societal impact or capabilities.

Citation

BibTeX:

@misc{reuel2025social,
    title={Who Evaluates AI's Social Impacts? Mapping Coverage and Gaps in First and Third Party Evaluations},
    author={Anka Reuel and Avijit Ghosh and Jenny Chim and Andrew Tran and Yanan Long and Jennifer Mickel and Usman Gohar and Srishti Yadav and Pawan Sasanka Ammanamanchi and Mowafak Allaham and Hossein A. Rahmani and Mubashara Akhtar and Felix Friedrich and Robert Scholz and Michael Alexander Riegler and Jan Batzner and Eliya Habba and Arushi Saxena and Anastassia Kornilova and Kevin Wei and Prajna Soni and Yohan Mathew and Kevin Klyman and Jeba Sania and Subramanyam Sahoo and Olivia Beyer Bruvik and Pouya Sadeghi and Sujata Goswami and Angelina Wang and Yacine Jernite and Zeerak Talat and Stella Biderman and Mykel Kochenderfer and Sanmi Koyejo and Irene Solaiman},
    year={2025},
    eprint={2511.05613},
    archivePrefix={arXiv},
    primaryClass={cs.CY},
    url={https://arxiv.org/abs/2511.05613},
    note={Preprint}
}

APA:

Reuel, A., Ghosh, A., Chim, J., Tran, A., Long, Y., Mickel, J., Gohar, U., Yadav, S., Ammanamanchi, P. S., Allaham, M., Rahmani, H. A., Akhtar, M., Friedrich, F., Scholz, R., Riegler, M. A., Batzner, J., Habba, E., Saxena, A., Kornilova, A., Wei, K., Soni, P., Mathew, Y., Klyman, K., Sania, J., Sahoo, S., Bruvik, O. B., Sadeghi, P., Goswami, S., Wang, A., Jernite, Y., Talat, Z., Biderman, S., Kochenderfer, M., Koyejo, S., & Solaiman, I. (2025). Who evaluates AI's social impacts? Mapping coverage and gaps in first and third party evaluations (arXiv:2511.05613). arXiv. https://arxiv.org/abs/2511.05613

Dataset Card Authors

Jenny Chim

Dataset Card Contact

Anka Reuel, Avijit Ghosh, Jenny Chim