Your AI copilot can write a research plan in less than 3 minutes. Should you trust it?

Copilots are cropping up everywhere. From coding to copywriting, these AI-powered assistants promise to streamline workflows and supercharge productivity. And now, they’re elbowing their way into UX research. They can generate interview guides, summarise user feedback, and even “suggest” insights. But here’s the million-dollar question: Can a synthetic assistant really understand messy, nuanced, deeply human layers of user behaviour?

The Temptation of Speed vs the Cost of Oversight

There’s no denying the appeal. In a field where timelines are tight and stakeholder demands are relentless, having a copilot churn out an interview guide or analyse hundreds of user comments in seconds sounds like a dream. Tools like Microsoft Copilot, Notion, and even bespoke UX research copilots can draft questions, generate personas, and spot patterns all before it’s time for breakfast.

But fast isn’t always faithful.

Copilots operate on statistical prediction, not lived experience. While they’re getting better at mimicking researcher tone and structure, they often miss the subtleties, such as the hesitation in a participant’s voice, the unexpected detour in a user’s story, or the silent pain points that only emerge after repeated, thematic probing. These are the goldmines of good UX research, which often emerge from the intersections of identity, context, and lived experience.

Can automated insights really replace the value of the human touch? (Image: Sora by OpenAI.)

Algorithmic models, by contrast, are trained on historical data, much of which is riddled with gaps, generalisations, or implicit bias. As a result, they risk reinforcing dominant narratives while filtering out the edge cases: the very users we most need to hear from.

As Nielsen Norman Group puts it, observing real user behaviour, rather than simply assuming or inferring it, remains the cornerstone of effective UX research. Furthermore, the group suggests, truly human-centred research depends on recognising difference as design data, not statistical noise. 

Copilots can support, but they can’t fully replace that organic messiness, which underscores the case for a human-in-the-loop.

Confidently Incorrect?

When it comes to AI-generated insights, the danger lies in illusion. AI is exceptionally good at sounding confident, even when it’s wrong. This is why UX consultants warn against over-reliance on machine-suggested findings, especially when they haven’t been validated by actual user evidence. 

Compounding the risk is the tendency of large language models (which often power popular copilots) to affirm rather than interrogate. Designed to be helpful and agreeable, many LLMs fall into a pattern of sycophantic responses, validating whatever the user says, even when it’s factually flawed or biased. 

OpenAI acknowledged this behaviour in GPT4o after a recent update made ChatGPT overly flattering, even endorsing user delusions and reinforcing negative emotions. In their official statement, the generative AI giant admitted the model had become “overly supportive but disingenuous”, a side effect of over-optimising for short-term feedback. The change was rolled back by April 28 to restore more balanced, truthful responses.

According to Unite.AI, this behaviour is rooted in reinforcement learning strategies that favour politeness and conversational fluency over critical engagement. In the context of UX research, this design choice introduces a significant risk that flawed assumptions may be validated rather than interrogated, undermining the integrity of insight generation.

Using Tools to Enhance Research — Not to Let Tools Define It

AI copilots can be accelerators, but they should never be architects of the research process. When used thoughtfully, they help us move faster, but when allowed to lead, they risk shaping the very questions we ask, narrowing the lens through which we interpret human experience.

UX research is not a mechanical exercise. It’s a process of deep inquiry rooted in empathy, context, and complexity. Copilots can assist with the “what” and the “how,” but the “why”, and the meaning, still belongs to the human researcher.

As is true with most tools, the true value of copilots lies in their ability to support the human-in-the-loop, and not in replacing the methodology. 

How are you using AI copilots in your research practice? Drop a comment below and let us know!