This is my keynote for ECP 2026, the 22nd European Conference on Personality in Edinburgh.
The talk in brief
Here are the theses, following the deck slide by slide.
The question. “To use, or not to use?”
The debate. Weigh both sides before you commit.
- Pros: efficacy, breadth of knowledge, speed.
- Cons: genuine, and worth taking seriously (Jowsey et al., 2015).
- The field has not stayed quiet: 416 researchers pushed back with detailed arguments for and against LLMs in qualitative research (preprint).
Actually, it does not matter which way the debate falls.
- LLMs are already a reality, whatever one concludes on the merits.
- Using them is a form of literacy now, not an open dilemma. So the real question is not whether but how.
Key cons. Three obstacles that actually bite.
- Lack of skills (AI, IT, software development, awareness of existing tools).
- Reinventing the wheel.
- Transparency and reproducibility.
Toolkit / know your tool. Claude and ChatGPT are more than a chat box.
- Most people stop at the chat box.
- Basics: instructions, projects, modes (web search, research).
- Advanced: useful MCPs (arXiv, memory, Notion…), Knowledge / RAG (Zotero and reference managers), Claude Cowork, ChatGPT Agents.
- Pro: CLI agents (Claude Code, Codex) and Skills.
Toolkit / RAG. Ground the model in your own library.
- References database, user query, search for similar documents, augmented query, generation with the retrieved results, response.
- Answers anchored to your sources, not the model’s guesswork.
Toolkit / Skills. One reusable bundle, three ingredients.
- Instructions, extracts from knowledge, scripts.
- We already have writing, analysis, auditing and designing shipped as default skills, many of them open source.
Toolkit / in practice. Qualitative psychology skills.
- Repo psy-qm-skills: codebook-construct, segment, codebook-apply, codebook-validate, irr-report, case-finder, codebook-iterate, memo-write.
- Benefits: reusability, transparency, audit log, efficacy.
Applications. What we can actually do with LLMs.
- Assisted research.
- Auditing and compliance.
- Large-corpora analysis.
- Maths.
- Piloting.
- Tech: transcribing, restoring, summarising.
Open source and confidentiality. Flagship models are excellent, but they are not the only option.
- OSS models are usable too: GLM, Kimi, MiniMax, DeepSeek, Mistral, Qwen, Llama, Gemma.
- Run open models locally so sensitive participant data never leaves your own infrastructure.
Recap. Five takeaways.
- Know your tool.
- Go beyond prompting.
- Use RAG.
- Use skills.
- Share with the community.
If you were in the room, thank you for coming. If you were not, I hope the deck and the theses are enough to carry the argument. Either way, the honest position is that this is a moving target, and the best thing we can do is keep sharing what works.
Links
- Skills repo: github.com/smirik/psy-qm-skills
- Reading collection, LLMs in qualitative psychology: readmarginalia.com/collections/llms-in-qualitative-psychology
- Referenced preprint, the 416-researcher debate: osf.io/preprints/psyarxiv/25ywa_v1
References
- Smirnov, E. (2024). Enhancing qualitative research in psychology with large language models: A methodological exploration and examples of simulations. Qualitative Research in Psychology, 1–31. https://doi.org/10.1080/14780887.2024.2428255
- Smirnov, E., & Carruba, V. (2026). Evaluating multimodal commercial and open-source large language models for dynamical astronomy: A benchmark study of resonant behavior classification. Scientific Reports, 16(1), Article 10785. https://doi.org/10.1038/s41598-026-45926-y
- Smirnov, E. (2026). From prompts to skills: Open, reusable tools for large language model-assisted qualitative research in psychology [Manuscript submitted for publication]. Qualitative Research in Psychology.