Perspective

Researchers Always Delegated Coding. What Changes With AI?

By Stephen McCurdy, Founder · Published 2 November 2026

As AI reaches qualitative analysis, some universities and programmes have responded in the simplest way available: by ruling it out. The worry behind those rules is serious, and in places it is right. Interpretation is the heart of qualitative research, and a chatbot that turns forty interviews into five tidy themes is not doing qualitative research.

But a blanket ban answers the wrong question. The useful question is not whether software touched the data. It is who made each analytic decision, and whether anyone can see how those decisions were made. Qualitative research has been answering that question for decades, because researchers have always delegated.

Qualitative research has always been a team effort

Large qualitative studies have long relied on teams. Research assistants, PhD students and junior colleagues do much of the first-pass coding; lead researchers check it, refine the codebook and take responsibility for the interpretation. Few principal investigators personally code every line of a multi-site study.

The methods literature treats this as normal and builds practice around it. MacQueen and colleagues at the US Centers for Disease Control and Prevention (1998) set out how to build a shared codebook so that several coders apply codes consistently, and Guest and MacQueen's Handbook for Team-Based Qualitative Research (2008) is built on the same premise.

What made that delegation legitimate was never that the lead researcher did all the coding. It was supervision, checking, a shared codebook, and accountability: the person who signed the findings could explain and defend every one of them.

What actually changes with AI

Some things genuinely change. An AI assistant works at a speed and scale no research assistant can, proposing evidence across every transcript in minutes. It also brings new ways to go wrong: fluent, confident output that can be mistaken; summaries of what participants talked about that look like themes but are only topics; participant data sent to a third-party service; and, with general-purpose chatbots, no record of what was asked, suggested or accepted.

One thing does not change: responsibility. A finding still needs a person who made the decisions behind it and can defend them to a supervisor, an examiner or an ethics committee.

So the test for AI should be the test we always applied to delegation. Is a person making every interpretive decision, and is there a record that shows it?

Bans don't stop AI use. They stop disclosure.

The numbers make the problem plain. In the Higher Education Policy Institute's 2026 survey of 1,054 full-time UK undergraduates, 95% said they use AI in at least one way, yet only about a third said their university actively encourages it. A year earlier, 88% said they had used generative AI for assessments, up from 53% in 2024.

The pattern is the same outside universities. Microsoft and LinkedIn's 2024 Work Trend Index found that 78% of people using AI at work bring their own AI tools rather than waiting for approved ones.

When use is near-universal and permission is rare, a ban does not remove AI from the analysis. It removes it from the methods chapter. The work moves into a personal chatbot account, with no record, no checking and no disclosure. That is the worst outcome for rigour, for participants' privacy and for examiners, the very things a ban is meant to protect.

What good governance looks like

The Russell Group's principles on generative AI (2023) commit its universities to making students and staff AI-literate, to adapting teaching and assessment to incorporate the ethical use of generative AI, and to upholding academic rigour and integrity. For qualitative research, those commitments translate into five practical rules:

1.Permit by task, not by tool. Separate the tasks AI may help with, such as transcription, searching the data, or suggesting passages and candidate codes, from the tasks that stay with the researcher: deciding what a passage means, defining themes and writing the interpretation.
2.Require a human decision on every suggestion. Nothing should enter the analysis simply because an AI proposed it. Every suggested code or extract is accepted, changed or rejected by a person.
3.Require a record. What the AI suggested, what was accepted or rejected, and how the codebook changed over time. Without a record, oversight is impossible after the fact.
4.Require disclosure that matches the record. Name the tool, the tasks it performed, what it was not allowed to do and how its suggestions were reviewed. A one-line acknowledgement tells an examiner nothing.
5.Protect participant data. Check that ethics approval and participant consent cover the service being used, where the data is stored, and whether it is used to train AI models.

None of these rules depends on a particular tool. They describe accountable delegation, and they apply equally to a research assistant, a dedicated QDA package or a chatbot. Our guides to disclosing AI in your methods chapter and audit trails in qualitative research go into the record and the disclosure in more detail.

Where a ban does make sense

There is one place where ruling AI out is exactly right: assessments that test a student's own analytic skill. When the point of a task is to show that the student can code and interpret data, AI should not do that work, for the same reason the student could not hand their coding to a research assistant. The delegation argument cuts the other way here. Delegation is legitimate when you are accountable for findings, not when you are being assessed on doing the analysis yourself.

So if your course or assessment rules out AI for data analysis, follow it, whichever tool you use. Institutions can help by being specific: say which tasks are permitted, which are not, and how students should disclose what they did.

The case for AI is depth, not just speed

Speed is the obvious selling point, and the weakest argument. The better one is depth. Even a researcher who reads every transcript closely benefits from a second pass that surfaces passages they may have missed, contradictions to a theme they are developing, and negative cases that test it. That widens what the researcher examines without deciding what any of it means.

Used that way, AI is closer to a tireless research assistant than to an analyst. And like any assistant, its work is only as good as the checking.

How we built QualIntel OS around this

QualIntel OS is built on the delegation model. The AI surfaces candidate evidence and suggests codes; the researcher accepts or rejects every suggestion, and nothing is coded automatically. Every decision, rename and merge is recorded in an audit trail, and the researcher can export an AI-disclosure statement for their methods chapter. The aim is to make delegating to AI at least as accountable as delegating to a colleague, and more visible.

For supervisors and examiners, our supervisor and examiner trust pack explains what to look for in AI-assisted qualitative work. For researchers, keeping qualitative research human-led sets out where AI assistance ends and analytic judgement begins.

Frequently asked questions

Is using AI for qualitative coding cheating?

It depends on the rules where you are working. In research, AI-assisted coding where a person makes every decision, keeps a record and discloses the AI's role is increasingly accepted. In an assessment designed to test your own analytic skills, it may be prohibited. Check your institution's and your course's rules, and follow them.

Can AI do qualitative analysis on its own?

No. AI can find and organise passages, suggest codes and summarise what participants talked about, but a summary of topics is not a set of themes. Deciding what participants mean, developing themes and relating them to theory and context is the researcher's work, and it is the part a researcher has to be able to defend.

Why not simply ban AI in qualitative research?

Because use is already near-universal among students and common at work, so a ban mostly moves AI out of sight. The analysis then happens in personal chatbot accounts, with no record, no checking and no disclosure. Clear permitted uses, a record of AI suggestions and decisions, and specific disclosure give institutions far more oversight than a rule that is quietly ignored.

What should a university AI policy for qualitative research include?

Permitted tasks rather than permitted tools; a human decision on every AI suggestion; a record of what the AI suggested and what was decided; specific disclosure in the methods section; protection of participant data in line with ethics approval and consent; and clear exceptions for assessments that test a student's own analytic skills.

References

Freeman, J. (2025). Student Generative AI Survey 2025 (HEPI Policy Note 61). Higher Education Policy Institute.
Guest, G., & MacQueen, K. M. (Eds.). (2008). Handbook for team-based qualitative research. AltaMira Press.
Higher Education Policy Institute. (2026). Student Generative AI Survey 2026 (HEPI Report 199).
MacQueen, K. M., McLellan, E., Kay, K., & Milstein, B. (1998). Codebook development for team-based qualitative analysis. Cultural Anthropology Methods, 10(2), 31–36.
Microsoft & LinkedIn. (2024). 2024 Work Trend Index annual report: AI at work is here. Now comes the hard part.
Russell Group. (2023, 3 July). Russell Group principles on the use of generative AI tools in education.