How to ethically clean up user interview transcripts with LLMs (without fabricating insights)

The temptation of the "Summarise" button

We have all been there. You have just wrapped up six hours of back-to-back user interviews. Your brain is mush, your notebook is a chaotic web of arrows and exclamation marks, and you have a transcript that looks like a tragic modern play. The temptation to dump the whole lot into Claude and type "Summarise the key pain points" is almost overwhelming.

And why shouldn't you? It takes seconds. But here is the problem: LLMs are people-pleasers by design. They like structure, symmetry, and clean narrative arcs. Human beings, especially when complaining about software, are messy, contradictory, and delightfully illogical.

When you ask an AI to clean up and synthesise your user transcripts, you risk stripping away the exact friction points that make user research valuable. If you are not careful, you end up with a sanitised, hallucinated version of user feedback that tells you exactly what you want to hear. Here is how to use LLMs to speed up your qualitative analysis without losing the raw, unfiltered truth of what makes your users tick.

The hidden danger: How LLMs "optimise" human frustration

When a user struggles with a UI, they do not speak in neat bullet points. They mumble, they sigh, they use colourful language, and they blame themselves for things that are clearly your design team's fault. This messiness is where the gold is hidden.

If you feed a raw, rambling transcript into a standard LLM without strict boundaries, the model will translate that raw emotion into sterile, corporate prose. A user saying, "God, this button is tiny, I keep clicking the wrong thing, whatever, I guess I'm just stupid," gets summarised as: "The user experienced minor accessibility issues with target size."

That translation is a lie. It sanitises the emotional gravity of the issue. It turns a screaming red flag into a polite suggestion. To avoid this, you need to treat the AI as a highly literal sorting clerk, not an editorial director.

Step 1: Privacy first (Stop feeding customer data to public models)

Before you upload a single line of a transcript, we need to talk about data ethics. Your users did not consent to have their frustrations, names, company details, or proprietary workflows trained on by commercial models.

Before processing any transcript, you must anonymise it. Strip out names, company names, and hyper-specific identifying details. Furthermore, make sure you are using a workspace environment where data training is turned off. If you are using Claude, check their data privacy settings on Claude Support to ensure your inputs are not being used to train future models. The same goes for enterprise accounts on OpenAI. If you cannot guarantee data privacy, you should not be using the cloud API at all. Period.

Step 2: Write constraints that forbid "smoothing"

If you want raw truth, you have to build a prompt that actively discourages the LLM from being helpful or creative. You want the model to be cold, clinical, and literal.

Instead of asking for a summary, ask the model to extract and categorise verbatim quotes. Your prompt should instruct the LLM to preserve the exact phrasing of the user, including sighs, pauses, and expressions of frustration.

If you want to save yourself some setup time, you can use our prompt generator to draft custom extraction prompts that enforce strict verbatim compliance.

Here is an example of an ethical extraction constraint:

> "Your task is to extract direct quotes related to user frustration. You must not rephrase, summarise, or correct the grammar of the user. If the user says 'this is rubbish,' you must output 'this is rubbish.' Do not translate this to 'the user expressed dissatisfaction with the interface.' If no direct quotes match, output 'None'."

Step 3: Use LLMs to find patterns, not to define them

One of the best ways to ethically use AI in user research is to use it for first-pass categorisation, not final synthesis.

Let the LLM tag your transcripts based on a taxonomy you have already defined. For example, ask it to label paragraphs as [Navigation Issue], [Pricing Complaint], or [Feature Request].

Once the AI has tagged the transcript, you—the human researcher—must do the actual reading and analysing. The AI has done the grunt work of organising the filing cabinet, but you are the one who needs to read the files. This keeps your qualitative research grounded in reality, preventing the AI from hallucinating a collective user sentiment that does not actually exist.

The litmus test for ethical qualitative AI

Whenever you use an AI to process user research, ask yourself this single question: Could I show this summary to the user who gave the interview, and would they agree that it accurately represents their feelings?

If your AI-generated summary has turned a chaotic, frustrating thirty minutes of testing into a neat, three-point list of minor UI optimisations, you have failed the test. You have let the AI lie to you to make your job easier. Keep the raw edges. Keep the frustration. That is where your next great feature is hiding.