← All posts
qualitative researchAI researchinsight agencies

Does AI invent quotes in qualitative research?

Sajin Kabeer · November 18, 2026

Yes, it can. Studies of AI thematic analysis have found quotes that did not exist in the transcript, quotes altered by a word or stitched together from several passages, and real quotes credited to the wrong participant. Treat every AI-supplied quote as unverified until you have matched it word for word, with the right speaker, in the original transcript.

That does not make AI useless for qualitative work. It means the quote is the one part of the output you can never take on trust.

Does AI really invent quotes? What the studies found

Yes. Three published pieces of work, using different models, data and teams, found the same family of problems.

A one-interview comparison, 2024. Researchers compared a general-purpose AI chatbot with a human researcher analysing the same 58-minute interview, published in Qualitative Health Research. When asked for verbatim quotes to support each theme, the model returned two quotes that did not exist in the original transcript. The quotes fitted the interview's meaning, which is what made them convincing. The model also did not consistently say when it had edited the original wording, and the authors felt quotes had been retrofitted onto themes it had already chosen.

A focus group evaluation, 2025. A team writing in Scientific Reports asked a large language model to analyse focus group transcripts from a health study in western Kenya, then had two reviewers score its work against a human analysis. The share of AI-selected quotes judged consistent with their theme ranged from 33 to 79% for one group of transcripts and 64 to 87% for the other. The errors ranged from a single changed word to truncated or combined text that changed the meaning.

A practitioner test, 2026. Nesta researchers analysed two interview projects, of 200 and 24 interviews, both by hand and with an AI notebook tool. They found the tool attributed quotes to the wrong participant, and that some quoted text could not be found in the transcript at all. They did not publish counts.

None of these studies tested every tool, and models change. The pattern is still consistent enough to plan around.

Why does AI produce quotes that were never said?

Because a language model generates text that is likely to fit, rather than copying text from a store. When you ask for "a quote that supports this theme", the most fitting answer is a sentence that sounds like the participant and supports the theme. Sometimes that sentence is in the transcript. Sometimes it is a close version. Sometimes it was never said.

Three conditions make it more likely:

What kinds of false quote should you look for?

Six kinds turn up, and each needs a different check. A text search alone catches only some of them.

TypeWhat it looks likeThe check that catches it
InventedA plausible quote with no source at allExact-text search of the transcripts finds nothing
AlteredA real quote with a word or phrase changedExact-text search fails; a search on a few words finds the near-match
StitchedTwo passages joined into one quoteThe parts are found in different places, or from different speakers
PolishedFillers and false starts removed, grammar fixedThe near-match shows the original wording
MisattributedA real quote credited to the wrong participant or segmentSpeaker check against the transcript; the text search passes
UnsupportiveA real, correctly credited quote that does not support the themeRead the passage in context; the text search passes

The last two are the hardest to catch, because the words really are in the transcript. Only a person reading the surrounding passage can judge whether the quote means what the report says it means.

Does asking for verbatim quotes stop it?

No. A strict prompt reduces the problem but does not remove it. In the 2024 study the researchers explicitly asked for verbatim quotes and still received two that did not exist. The model does not reliably know where a passage starts and ends, so it cannot reliably promise exact text.

Prompt wording is still worth getting right. Ask for each quote with its participant ID and its location in the transcript, ask the model to say "no supporting quote found" when it cannot find one, and ask for quotes while coding, not after the themes are written. Then check the output anyway.

How do you check AI quotes against transcripts?

Check every quote that will reach a client, in a fixed order, and record the result. The routine takes a few minutes per quote and catches all six types.

  1. Search the exact text. Search the transcripts for the full quote. A match means it exists; no match means it is invented, altered, stitched or polished.
  2. If there is no exact match, search a short phrase. Use three or four distinctive words. A near-match shows you the original, which you can restore or drop.
  3. Check the speaker. Confirm the participant ID, and the segment or group if the report cuts by segment.
  4. Read the passage around it. Read a few lines before and after. Check the participant was not quoting someone else, joking or answering a different question.
  5. Check it supports the theme. Ask whether a colleague who had not seen the theme would read the quote the same way.
  6. Mark omissions honestly. If you shorten a quote, mark the gap with an ellipsis and make sure the cut does not change the meaning.
  7. Log the check. Record the quote, its location, who checked it and the result.

If a quote fails, do not ask the model to fix it. Go back to the transcript and choose a real passage yourself, or drop the quote. For focus groups, where speakers overlap and interrupt, our guide to analysing focus group transcripts with AI covers speaker labelling in more detail.

What should a client report say about quotes?

A client report should let any reader find the source of each quote. Three habits make that possible:

These habits apply whatever tool you use. They are also the quickest way to show a client that a finding rests on what people actually said.

What can AI do safely in qualitative analysis?

AI is most dependable on tasks where an error is visible and cheap to fix. The authors of the 2025 Scientific Reports study concluded that AI can currently help identify themes, keywords and the basic narrative, and can act as a check for human error or bias, but is not yet reliable enough for a rigorous thematic analysis on its own.

In agency terms, that suggests a split:

The lower-risk tasks still need review, but a wrong suggestion is caught in the normal course of the work. The higher-risk tasks put the AI's output directly in front of the client. Those are where a person should do or check every step.

How does Meaningful handle quotes and evidence?

Meaningful is built so that a finding can be traced back to its evidence. Agents read every response and transcript in small, focused batches and reconcile what they find, rather than answering from a few passages a search step picks out. Each finding links back to its evidence, with caveats, so checking a quote starts from a link rather than a search across files.

That makes the routine above faster. It does not make it unnecessary, and no AI tool should be assumed to be free of these errors. Your team still checks the quotes and makes the calls. Meaningful also does not replace specialist recruitment, sampling design or the agency's own methodological judgement.

Interview transcripts sit in the same project as the study's survey, focus groups and desk research, so a quote can be read next to the numbers it is meant to explain. Our guide to AI tools for analysing interview transcripts gives a wider test you can run on any tool.

Frequently asked questions

Does AI make up quotes when analysing interviews?

It can. In published studies, AI models returned quotes that did not exist in the transcript, quotes altered by a word or stitched from several passages, and real quotes credited to the wrong participant. Treat every AI-supplied quote as unverified until it has been matched to the transcript.

Is a paraphrase presented as a quote a hallucination?

For reporting purposes, yes. A tidied or reworded quote may keep the meaning, but it is no longer what the participant said. Either restore the exact words or present it as a paraphrase without quotation marks.

Does asking for verbatim quotes stop AI inventing them?

No. The instruction helps, but it does not guarantee exact text. In one study the model still returned two quotes that did not exist and did not say when it had edited the original wording. Check the output regardless of the prompt.

Can I use AI-selected quotes in a client report?

Yes, once each one has been matched word for word to the transcript, checked for the right speaker and read in context. Give each quote a participant ID and a transcript location so a colleague or client can find it.

Which AI quotes should I check?

Every quote that will reach a client. Checking a sample is useful during analysis, but a single invented or misattributed quote in a final report can undermine the whole finding it supports.

Are invented quotes the same as context rot?

No. Context rot is AI missing or underweighting evidence in long material. Invented quotes are AI adding evidence that is not there. Both can happen in the same analysis, and both are caught by tracing findings back to the source.

How much does Meaningful cost?

Meaningful costs €500 a month excluding VAT, with unlimited seats, checked on 18 November 2026. There is a free three-day trial.

What should you do next?

Take the quotes from your last AI-assisted readout and run the seven-step check on them. Then load the same transcripts into Meaningful's free three-day trial and see how quickly each finding leads you back to the passage behind it.