AI Disclaimer

Last updated: 1 January 2025 · Unpinned B.V. · Utrecht, Netherlands

1. What the AI does

When you run the AI proposer, FieldNotes sends each unprocessed transcript excerpt to a large language model (LLM). The model reads the excerpt, considers your existing codebook and project description, and suggests a code label with a brief rationale.

This proposal is then placed in your review queue. Nothing happens to your codebook or your analysis until you act on it.

2. What the AI does not do

3. Known limitations of AI-generated code proposals

You should be aware of the following limitations when reviewing AI proposals:

• Hallucination: LLMs can generate plausible-sounding but incorrect or contextually inappropriate code suggestions. Always evaluate proposals against your data and theoretical framework.
• Context insensitivity: The AI processes excerpts sequentially and may not fully account for the broader context of an interview or your study's epistemological stance.
• Bias: LLMs are trained on large corpora of text that may embed cultural, linguistic, or demographic biases. These can surface in code proposals, particularly when analysing data from marginalised communities or non-Western contexts. Published evaluations report substantially lower human–AI agreement on culturally or emotionally nuanced themes than on descriptive ones — review proposals on such material with particular care.
• Consistency: The same excerpt may receive different proposals on different runs due to the probabilistic nature of LLM outputs.
• Language: Performance may vary for transcripts in languages other than English. Non-English transcripts should be reviewed with additional care.

4. Your responsibilities as a researcher

Using AI assistance in qualitative research does not reduce your methodological responsibilities — it changes some of them. You remain responsible for:

• Reviewing every AI proposal independently and applying your own analytical judgement.
• Disclosing the use of AI assistance in your methods section, including the tool name, provider, and the human-in-the-loop review process.
• Ensuring your use of AI-assisted coding is consistent with your institution's ethics approval and any relevant funding body guidelines.
• Verifying that participant data is appropriately anonymised or pseudonymised before upload, where required.
• The scientific validity, rigour, and integrity of your final analysis and any publications derived from it.

5. Transparency and reproducibility

FieldNotes is designed to support transparent and reproducible qualitative research:

• Every AI run logs the provider and model name used, included in your XLSX export.
• Every AI run archives the full, exact prompt text sent to the model — viewable on the project's AI audit page — so you can disclose precisely what the model was asked.
• Every proposal — approved, merged, or rejected — is recorded in the decision audit trail, append-only and unedited.
• The AI audit page reports acceptance rates per run and per model, computed from your own review decisions, as an ongoing verification record of AI output quality.
• The full decision log is exportable and can be provided as a supplementary file in your publication.
• A ready-to-use methods section paragraph is available on the homepage.

These records map to the disclosure items proposed in recent peer-reviewed guidance on AI-assisted qualitative analysis, including Gupta et al. (2026) in Nursing Inquiry and the trustworthiness criteria of Lazarus et al. (2026) in Anatomical Sciences Education.

6. Data use by AI providers

Excerpt text is sent to an LLM API provider to generate proposals. We use providers whose API terms explicitly prohibit using API input data for model training. Your research data is not used to train AI models. For the specific provider in use, check the model name in your application settings or export file. Enterprise users may configure their own API key (BYOK) to use a provider of their choosing.

7. EU AI Act compliance

FieldNotes uses AI as a decision-support tool in a context where all outputs are reviewed by a human before any effect occurs. Under the EU AI Act, this human-in-the-loop design is a mitigating factor for risk classification. We will update our compliance documentation as the EU AI Act's implementing regulations take effect.

8. Contact

Questions about our AI use or this disclaimer:
azadali@unpinned.nl
Unpinned B.V. · Utrecht, Netherlands