AI tools can generate customer personas, synthesise interview themes, and produce market analysis reports in minutes. The problem is that output which looks like insight isn't always built on it, and the format makes the gap hard to see.
This is the specific failure mode that matters for marketing and research teams: not that AI is bad at research, but that it produces results confident enough to act on, plausible enough to pass a review meeting, and divorced enough from real customers to quietly undermine whatever comes next.
There's a version of AI-assisted research that genuinely works. Summarising interview transcripts, spotting patterns across large feedback sets, drafting survey questions for expert review. These are tasks where AI adds speed without substituting for judgement.
Then there's a different pattern that's become more common: drop a prompt into ChatGPT, generate a customer persona, present it to the team, move on. The output looks professional. It has headers and bullet points and customer quotes. It mirrors the format of research.
What it doesn't have is a real customer behind it.
Industry researchers working directly with AI-generated data have noted that many synthetic research tools produce outputs that look credible on first read but aren't backed by direct evidence. An AI panelist might give you a neatly phrased explanation for why customers would value your new product, but that doesn't mean real customers actually would. The format signals confidence even when the substance doesn't support it.
This matters most when the output lands in a context where no one has direct customer experience to challenge it. If the sales team hasn't been in enough discovery calls to know what language buyers actually use, and the marketing team is working from AI-generated personas, neither group has the reference point to notice what's missing.
Research published in Harvard Business Review in March 2026 found that leading LLMs have clear biases when it comes to strategy, consistently recommending approaches that align with modern managerial buzzwords and trends rather than context-specific logic. The researchers named this "trendslop." What made the finding particularly pointed is that better prompting didn't fix it. Richer context, reversed option framing, and demands for pros-and-cons analysis made little difference across tens of thousands of simulations. The output sounded tailored to each situation while steering toward the same cluster of fashionable answers.
The same dynamic applies in customer research. AI tools are trained on large volumes of existing content: previous research reports, published case studies, market analyses. When asked to generate a customer persona or summarise buyer behaviour, they pattern-match against that corpus. The output is structurally coherent and tonally consistent with research that's been done before, on different companies, in different categories, with different customers. It produces insight that sounds specific to you while actually reflecting the category average.
There's also a confirmation dynamic at work. Research published in 2025 found that the personalised and human-like nature of AI interactions can reinforce users' views through sycophancy and amplify confirmation bias. Prompt an AI tool with an existing hypothesis about your customers and it will tend to support that hypothesis, not challenge it. That's the opposite of what research is supposed to do.
As the HBR researchers put it, LLMs are like the person in the room who can articulate every buzzword from TED Talks and conference seminars but has never analysed how those ideas play out in the messy realities of a specific market.
That description applies just as precisely to customer research. The output reflects what's generally true about companies like yours, customers like yours, categories like yours. It doesn't reflect what's actually true about your specific buyers, in your specific market, at this specific moment in their journey.
Real customer research produces three things that AI pattern-matching can't replicate:
The language customers actually use to describe the problem, not the language your category typically uses
The distinction between what customers say they want and what their behaviour suggests they're doing
The specific moment in the buying journey when the decision crystallises.
Research on AI-generated personas from the ACM's 2025 Intelligent User Interfaces conference found that concerns over factuality and potential misinformation, including biased or stereotypical end-user representation, form a major risk in AI-generated personas.
When those personas are used to inform messaging decisions without validation against real customer input, the risk compounds.
The other gap is tension. Experienced researchers look for the things that don't add up: the customer who said something that contradicted the three before them, the friction in the buying process that nobody names directly. AI outputs tend toward coherence. They average across the training data and produce something that reads cleanly. The edges are exactly what gets smoothed away.
AI tools can be helpful. They can organise information, speed up early-stage work, and save time on repetitive tasks. They’re good at pulling together summaries, highlighting patterns, and even suggesting themes across large datasets.
But AI doesn’t know what’s missing. It doesn’t weigh relevance or understand nuance. It can’t challenge assumptions, spot tension, or sense when something feels off.
AI treats all inputs as equal—whether it’s a peer-reviewed report or a low-quality blog post. It formats outputs in a way that feels polished and complete, even when key gaps remain.
And yet, because the format looks professional and the tone sounds credible, it creates confidence. The kind of confidence that skips reflection and moves straight to action.
This is where the illusion of knowledge takes hold. Easy access creates a false sense of clarity. But without knowing what data was excluded—or how it was interpreted—your decision-making is built on unstable ground.
The value of structured research isn't the data itself. It's the process that produces the data.
Skilled researchers define the right question before they choose a method. They map the assumptions the team is carrying into the research so those assumptions can be tested, not validated. They look for tension between what different people say, and between what people say and what they do. They turn what they find into direction: priorities, trade-offs, positioning decisions.
None of that is a task. It's a discipline. It takes experience to know what question is actually worth asking, and to recognise when an answer is incomplete.
AI can support parts of the process well. Drafting interview questions for expert review, grouping themes from transcripts, summarising secondary research. These are useful contributions to a research process. They're not a replacement for it.
The teams who get the most from AI-assisted research are the ones who are clear about where the line is. AI accelerates the parts of research that are about information handling. The parts that are about judgement still require a human to exercise it.
The distinction isn't always obvious in practice, so it's worth being concrete.
AI is genuinely useful for: drafting interview or survey questions before an expert refines them; grouping qualitative feedback from a large dataset; summarising patterns across existing reports; and identifying gaps in secondary research.
Real customer input is required for: defining what question is actually worth researching; understanding the language buyers use when they're not being surveyed; identifying the difference between stated preferences and actual behaviour; and validating any insight before it informs a positioning or messaging decision.
The middle ground, using AI-generated personas or AI-synthesised research as a proxy for customer input, is where the risk concentrates. The output can be a useful starting point for a hypothesis. It's not a substitute for direct customer conversations when the question is about how buyers actually think about a problem.
Insight doesn't come from having the fastest answer. It comes from asking better questions, and being willing to sit with what you find. And if that sounds slow? Just wait until you're rebuilding a strategy based on shallow input.
Here’s how to think about what to automate—and what still requires people with the skills to interpret and apply what they’re seeing.
|
Use AI to... |
Use humans to... |
|---|---|
| Draft interview or survey questions | Define the right research question for your context |
| Spot biased phrasing or leading language | Decide what needs to be explored in more depth |
| Group qualitative feedback into themes | Interpret contradictions and tension |
| Summarise desk research and existing reports | Prioritise what matters for your product or strategy |
| Synthesise notes and transcripts | Sense what’s missing or unspoken |
| Speed up admin and documentation | Connect findings back to business goals and decisions |
There’s no shortage of data or content anymore. That’s not the challenge.
The challenge is knowing what to trust, what to ignore, and what to dig into further. That still takes time, skill, and people who are willing to look beyond the surface.
The teams who get this right aren’t the ones who move slow—they’re the ones who move forward with confidence, because they’ve done the work to understand what matters.
Insight doesn’t come from having the fastest answer. It comes from asking better questions—and being willing to sit with what you find.
And if that sounds slow? Just wait until you’re rebuilding a strategy based on shallow input.
What is the illusion of insight in AI customer research? The illusion of insight occurs when AI-generated research outputs, such as customer personas, synthesised interview themes, or market summaries, produce results that are structurally plausible but not grounded in real customer input. The output passes a credibility check because it matches the format and language of genuine research, but the absence of actual customer data means key distinctions, tensions, and language patterns are missing.
Can AI replace customer interviews and focus groups? AI can support customer research by drafting questions, grouping themes from large feedback datasets, and summarising secondary sources. It can't replace direct customer conversations, because it can't surface what's undocumented: the language real buyers use, the moments when decisions actually shift, or the tension between what customers say and what they do.
Why do AI-generated customer personas get ignored or shelved? AI-generated personas tend to get ignored or quietly shelved because the people closest to customers, particularly experienced salespeople or customer success teams, recognise that the persona doesn't match anyone they've actually talked to. The output is internally coherent but externally unrecognisable, so it gets set aside rather than challenged directly.
What is confirmation bias in AI-assisted market research? Confirmation bias in AI-assisted research occurs when a team's existing assumptions shape the prompt they give an AI tool, and the tool generates output consistent with those assumptions. Because AI tools are trained to produce coherent, plausible responses, they tend to support the framing of the question rather than challenge it. This is the opposite of what good research is designed to do.
What is "trendslop" and why does it matter for customer research? "Trendslop" is a term coined in a March 2026 Harvard Business Review study by researchers from Esade Business School, the University of Sydney, and NYU Stern. It describes the tendency for AI tools to recommend fashionable, buzzword-aligned options regardless of the specific context they're given. In customer research, the equivalent is AI-generated personas and insight summaries that reflect category-level averages and trend-aligned language rather than what your specific customers actually think and say.
How reliable are AI-generated customer personas for messaging decisions? AI-generated personas carry significant accuracy risk for messaging work specifically, because messaging decisions depend on knowing the exact language buyers use to describe a problem. That language is often specific to your customers, your category, and the moment in the buying journey, none of which AI pattern-matching reliably captures. Personas built without direct customer input are a starting hypothesis, not a validated insight.
What should teams use AI for in customer research? AI works well for parts of the research process that are about information handling: drafting interview questions for expert review, grouping themes from transcripts, summarising secondary reports, and identifying gaps in existing knowledge. The judgement-intensive parts of research, defining the right question, spotting tension in customer responses, translating insight into strategic direction, still require human expertise.
What is the difference between AI research and real customer research? AI-assisted research synthesises what's already been written down. Real customer research surfaces what's true but undocumented, or true for your specific customers but not the broader category. The gap shows up most clearly in positioning and messaging work, where the precise language customers use to describe a problem is exactly what differentiates a message that lands from one that doesn't.
How can teams tell if their research is AI-generated versus customer-validated? The clearest signal is whether the research can be attributed to specific customers. If a persona or insight summary contains direct quotes, describes specific buying scenarios, or identifies language patterns that differ from category norms, it's likely grounded in real input. If it reads as a coherent summary of what's generally true in the category, it may be pattern-matched from existing content rather than built from customer conversations.