AI-Assisted Recruitment vs. AI-Generated Participants

Finding someone to answer a question and generating their answer can both sit under the heading of AI in research. Lyssna’s September 2026 recruitment guide separates those uses. It covers refining screeners, sourcing people from customer data, automated vetting conversations, synthetic participants, fraud detection and outreach administration. It explicitly distinguishes synthetic participation from recruiting real people. I give it credit for making that separation rather than treating everything with an AI label as one kind of recruitment.

My view is that you should judge AI by where it enters the research process, and require a clear account of who supplied the answers before treating a report as customer evidence.

I’m happy for software to help arrange a conversation. I want a different level of scrutiny when it changes what someone said. And if a system generates the response itself, I won’t count that response as evidence that a customer thinks or does something.

Where the answer came from

Suppose you sell refillable cleaning products and want to understand why customers haven’t placed a second order. You could use software to find eligible customers in your records, send invitations and handle scheduling. The people who reply still supply the evidence about their experience.

You could also ask a system to generate responses from synthetic versions of those customers. Maybe the answers mention storage space, fragrance or the trouble of remembering to reorder. Those could be useful possibilities to investigate. But nobody in that exercise has reported what happened in their home.

I’d be tempted by a neat answer here too. If I already suspected that reordering was awkward, a convincing synthetic account could feel like confirmation. It would let me move straight to fixing the subscription flow without doing the uncomfortable part: asking customers and finding out my explanation was wrong.

That’s where I draw a harder line than a broad catalogue of AI uses might suggest. Synthetic participation belongs in a separate evidence category, even when it appears alongside recruitment methods in the same guide. I can use it to think of possibilities. I wouldn’t use it to claim that buyers have validated a decision.

Real people still need scrutiny

Knowing that a person answered doesn’t settle every question about a study. They might fall outside your target audience. They might misunderstand a question or rush through it. A genuine customer might give a polite explanation that leaves out the reason they actually stopped buying.

AI-assisted vetting doesn’t remove those possibilities by definition. Neither does the word “vetted.” You need to know what was checked and what remains uncertain.

There are also several steps between a person speaking and a finding appearing in a report. An automated moderator might choose the follow-up questions. Translation might change the strength of a complaint. A summary might turn a hesitant comment into a firm preference.

I’m open to all of those uses when they help a study run. I still want to know what happened. A human-origin response can be distorted in processing, just as a human researcher can misread it. Being able to return to the original gives you a way to challenge the interpretation.

For a founder, that comes down to a practical question: can you trace the claim you’re about to act on back to the evidence behind it?

Questions to ask a research vendor

These are due-diligence questions. They aren’t features I assume every vendor offers, and a sales page that says “AI-powered research” doesn’t answer them.

1. Who supplied the answers? Ask whether responses came from your customers, recruited people, synthetic participants or a mixture. If people were recruited, ask where they came from. People who buy your category can help with some decisions, but they can’t explain their experience of your checkout unless they’ve used it.

2. How was eligibility assessed? Ask which requirements were self-reported and which, if any, were checked against other evidence. If automated vetting was involved, ask what it assessed. A claim that someone fits your audience needs more explanation than a label beside their response.

3. What did AI do to the responses? Did it moderate, translate or summarise human answers? Did it generate any answers itself? Ask whether gaps or incomplete responses were filled in. You want the actual sequence of steps, rather than a general assurance that humans were involved.

4. Are original responses retained for inspection? Ask what you can review, in what format and for how long, subject to participant consent and privacy restrictions. Original survey text, recordings or transcripts offer different ways to inspect a conclusion. If you only receive a summary, know that limit before buying.

5. Are human and synthetic materials labelled separately? Ask whether that separation survives exports and presentations. A label inside a platform won’t help the colleague reading a downloaded report if it disappears there. Generated material shouldn’t quietly enter response counts or appear as a participant quote.

6. What quality checks and replacement rules apply? Ask what can cause a response to be excluded, how suspected fraud is handled and what happens if a recruited participant fails the agreed requirements. Get the replacement or refund terms before launch. Don’t assume “quality checked” means every questionable response will be replaced.

I’d rather get a narrow, specific answer to these questions than a confident promise of trustworthy insights. An honest limitation lets you decide whether the study is suitable for the decision you have to make.

A small study with a visible evidence trail

You don’t need an elaborate procurement exercise for every customer survey. You do need enough clarity to describe the study honestly to someone who wasn’t there.

For the repeat-order question, I’d start with customers who bought once and haven’t bought again within a relevant period for the product. I’d ask about their experience and what they did afterward, leaving room for the possibility that they simply haven’t used it up yet. Then I’d inspect the answers behind any theme I planned to act on.

If recruitment software helped find those people, fine. If automated summaries helped me navigate their replies, I’d check the passages supporting my conclusion. If synthetic responses helped me plan the study, I’d keep them out of the customer findings.

At Probe, you can send a study to your own audience for free up to 10 questions and 25 responses. If you need recruitment, participants come from our recruitment partner’s vetted panel, with one all-in price per study shown before launch. Results come back as charts, themes and quotes. Those details describe the service; they don’t establish that every response will be useful or every interpretation correct.

Before your next study, ask the vendor to show you how one answer becomes a finding. Who supplied it? What happened to it? What can you inspect? If that path is unclear, resolve it before the finding becomes a reason to spend money.