Research should be conducted and communicated with scientific integrity.

Participants should:

• report methods, data, results, and limitations accurately;
• distinguish evidence from interpretation, speculation, or advocacy;
• avoid fabrication, falsification, plagiarism, and misleading presentation of results;
• acknowledge uncertainty where it materially affects a conclusion;
• credit collaborators and prior work appropriately; and
• make claims that are proportionate to the evidence presented.

FAIR values negative results, replication, and careful analysis. Researchers should not feel pressure to overstate novelty, capability gains, or the significance of a result.

Model training and evaluation often depend on choices that materially affect results.

Where relevant, researchers should explain important decisions involving:

• training data and data provenance;
• evaluation datasets and benchmarks;
• prompts, rubrics, graders, and judges;
• reward functions and optimization objectives;
• human evaluators and annotation procedures;
• model selection and checkpoint selection;
• filtering, exclusion, or sampling decisions;
• statistical methodology; and
• limitations affecting reproducibility or interpretation.

Research should make it possible for others to understand not only what happened, but how the result was produced.

People involved in AI research should not be treated as interchangeable components of a pipeline.

This includes annotators, evaluators, research participants, subject-matter experts, contractors, and individuals represented in datasets.

Researchers should consider:

• appropriate consent and oversight;
• fair and transparent compensation;
• working conditions;
• privacy and confidentiality;
• exposure to disturbing or sensitive material;
• cultural and linguistic context;
• evaluator training and calibration; and
• reasonable mechanisms for raising concerns.

Where human judgment materially influences a result, that contribution should be recognized as part of the research methodology.

Researchers should understand where important data and research materials came from and whether they have an appropriate basis for using them.

Particular care should be taken with:

• personal or sensitive information;
• proprietary or confidential data;
• data collected without clear consent;
• datasets involving vulnerable populations;
• scraped or aggregated data;
• copyrighted material; and
• datasets whose provenance is uncertain.

Technical availability should not automatically be treated as ethical permission to use data.

AI research can produce both beneficial capabilities and new risks.

Researchers should consider material and reasonably foreseeable harms associated with their work, including risks involving:

• safety;
• privacy;
• security;
• discrimination;
• deception;
• surveillance;
• manipulation;
• misuse;
• loss of human agency; and
• other significant downstream effects.

Research into these areas is not discouraged. Work that identifies dangerous behavior, reveals vulnerabilities, tests safeguards, or improves understanding of risk can be valuable.

Where research creates material risk, authors should explain how that risk was considered and, where appropriate, mitigated.

A system should not be characterized only by its strongest results.

Researchers should make reasonable efforts to understand where models fail, regress, behave inconsistently, or produce unintended behavior.

This is particularly important when:

• improvements in one capability may reduce performance elsewhere;
• benchmark scores mask meaningful failure modes;
• automated evaluators disagree with human judgment;
• reward signals can be gamed;
• results vary across languages, cultures, or populations; or
• deployment conditions differ materially from research conditions.

Evaluation should help reveal the boundaries of a system, not merely demonstrate its strengths.

Researchers should consider whether their methods, datasets, evaluations, or systems create or reinforce unfair disparities.

Where relevant, work should account for differences across:

• languages;
• cultures;
• regions;
• demographic groups;
• accessibility needs; and
• contexts of deployment.

Fairness does not require identical treatment in every setting. It requires careful attention to who benefits, who bears risk, and whose experiences may be missing from the research.

Researchers should credit the work required to produce ideas, datasets, benchmarks, code, models, annotations, evaluation frameworks, and other research artifacts.

Participants should respect:

• authorship and attribution;
• licenses;
• intellectual property;
• confidential information;
• unpublished research; and
• contributions that may not traditionally appear as academic authorship.

This includes meaningful contributions from engineers, annotators, evaluators, data creators, and other contributors to the research process.

Researchers, reviewers, and organizers may encounter confidential or sensitive information.

This may include:

• unpublished manuscripts;
• private datasets;
• research participant information;
• proprietary model details;
• confidential company information;
• reviewer discussions; and
• security vulnerabilities.

Such information should not be disclosed or used outside the purpose for which access was granted unless required by law or necessary to report a serious ethical or safety concern through an appropriate channel.

Participants should disclose material conflicts that could reasonably call their judgment into question.

These may include:

• financial relationships;
• employment relationships;
• investments;
• advisory roles;
• sponsorships;
• close collaborations;
• personal relationships; or
• direct competitive interests.

Disclosure does not automatically disqualify participation. Its purpose is to allow conflicts to be managed appropriately.

AI tools may be used in research and research communication, subject to FAIR’s AI Use Policy.

Authors remain responsible for all work submitted under their names.

Researchers should verify AI-assisted claims, citations, code, analysis, and other substantive outputs. Material use of AI in research methods, data generation, analysis, or interpretation should be disclosed.

AI systems should not be used to evade authorship responsibility, fabricate evidence, generate false citations, or obscure the provenance of research.

FAIR Summit brings together researchers, companies, sponsors, and infrastructure providers.

Commercial relationships should not determine research conclusions.

Authors, reviewers, and organizers should preserve their independence when evaluating evidence, selecting research, or communicating findings.

Sponsorship does not confer influence over:

• acceptance decisions;
• reviewer selection;
• research conclusions;
• technical findings; or
• the interpretation of scientific results.

Participants are encouraged to identify ethical concerns, limitations, and risks rather than conceal them.

The existence of a difficult ethical question is not itself evidence that research should not be conducted or presented.

Good research can clarify uncertainty, expose tradeoffs, and make risks easier to understand.

Where a concern is material, participants should raise it through the appropriate review, program, or organizational process.

This Code should be read alongside the FAIR Summit Code of Conduct and the Code of Conduct for Research Submissions.

FAIR Summit may investigate credible concerns involving research integrity, ethical violations, or misuse of the review or conference process.

Depending on the severity and context of the issue, FAIR Summit may:

• request clarification;
• require correction or disclosure;
• remove material from the program;
• reject or reconsider acceptance of a submission;
• restrict participation; or
• take other proportionate action.

Concerns will be considered in context.

Controversial findings, disagreement, or the presence of risk are not by themselves violations of this Code.