Ethics & Critical Perspectives

AI tools are powerful but not neutral. Historians have a professional responsibility to engage with these technologies critically. This page outlines key ethical considerations and ongoing debates.

Bias and Representation

Training Data Problems

Large language models are trained on internet-scale text corpora that reflect existing power structures. Consequences for historical work include:

Overrepresentation of English-language, Western, and elite perspectives. Ask an AI about the history of medicine, and you are more likely to get a narrative centered on European developments than on Indigenous, Chinese, Islamic, or African medical traditions.
Temporal bias. Models contain more text about recent centuries than earlier periods, and more about well-documented societies than those with primarily oral traditions.
Erasure and distortion. Marginalized communities—Indigenous peoples, enslaved persons, LGBTQ+ individuals, religious minorities—may be stereotyped, flattened, or absent in AI-generated historical narratives.

What Historians Can Do

Test AI tools on topics where you have deep expertise and can evaluate accuracy.
When AI output omits perspectives, note the omission explicitly in your teaching or research.
Use AI failures as teaching moments: Why did the model get this wrong? What does that reveal about whose knowledge is digitized and valued?

Accuracy and Hallucination

AI models generate text that is statistically plausible, not verified. “Hallucination”—the confident generation of false information—is a well-documented phenomenon. In historical contexts, this can include:

Invented quotations attributed to real historical figures
Fabricated book titles, journal articles, and archival references
Anachronistic descriptions (attributing modern concepts to past actors)
Blending of distinct events, people, or periods

Mitigation Strategies

Never cite AI output as a source. Use it as a starting point for research, not an endpoint.
Verify every factual claim. Cross-reference with established secondary sources and, where possible, primary sources.
Ask the AI to flag uncertainty. Prompts like “indicate where you are less confident” can help, though models may still present false information with high confidence.
Use retrieval-augmented generation (RAG) when possible—tools that ground AI responses in specific, verified document collections.

Citation and Attribution

Should You Cite AI?

Professional norms are still emerging. Key considerations:

Transparency. If AI meaningfully shaped your research process, analysis, or writing, disclose it. This mirrors existing norms around research assistants, translators, and computational tools.
Style guide guidance. The Chicago Manual of Style (18th ed.) recommends treating AI-generated text as a personal communication or describing the tool in a methodology note. APA and MLA have also issued guidance.
Reproducibility. AI outputs are not reproducible—the same prompt may yield different results at different times. Document your prompts, the model used, and the date of interaction.

In the Classroom

Be explicit with students about your expectations for AI citation. See Teaching with AI for sample syllabus policies.

Data Privacy and Archival Ethics

Uploading Sources to AI Services

When you paste text into a commercial AI chatbot, consider:

Terms of service. Some platforms may use your inputs for model training. Check current policies.
Unpublished materials. Uploading unpublished archival documents, oral histories, or personal correspondence raises ethical questions about consent, ownership, and cultural sensitivity.
Indigenous data sovereignty. Indigenous communities increasingly assert control over how their cultural knowledge is used. Using AI to process Indigenous materials without community consent may violate these principles.
GDPR and privacy law. In many jurisdictions, uploading personal data about identifiable individuals to third-party AI services may have legal implications.

Best Practices

Use locally hosted or privacy-respecting AI tools for sensitive materials.
Consult with archivists, community stakeholders, and IRB/ethics boards before using AI on sensitive collections.
When in doubt, do not upload.

Labor and Environmental Concerns

Data labeling. AI models rely on human labelers, often working in low-wage conditions. Historians concerned with labor history should be aware of these supply chains.
Environmental cost. Training and running large AI models requires significant computational resources and energy. Consider whether AI is the most appropriate tool for a given task, or whether simpler methods suffice.
Displacement anxieties. AI will not replace historians, but it may change what kinds of historical work are valued and funded. Engaging proactively with AI is a form of professional agency.

Key Questions for Reflection

Whose history is amplified by AI tools, and whose is silenced?
What is lost when students learn to generate text before they learn to think historically?
How should the profession establish norms around AI use in peer-reviewed research?
What responsibilities do AI developers have to consult with historians and other domain experts?
How do we balance the efficiency gains of AI with the slower, deeper work of humanistic inquiry?

History & AI