The verdict on synthetic users is in, and it is not kind.
The largest reviews of the field conclude that AI-generated participants are too thin to rely on, that they flatter almost every idea put to them, and that they cannot tell you what a real person will actually do.
We agree. We also use them most weeks. Both of those things are true, because the jobs we give them are not the one the critics are arguing about.
The case against is strong
When Nielsen Norman Group tested synthetic users against three real studies last year, they found the responses too shallow to be useful. The synthetic participants cared about everything, reported the behaviour they thought they ought to have rather than the behaviour they had, and praised concepts a real user would have quietly ignored. Practitioners like Erika Hall and Jared Spool have been blunter still, arguing that designing services for people without talking to those people is indefensible.
For anyone working in the public sector, this lands on familiar ground. ISO 9241-210 builds human-centred design around an explicit understanding of real users and their involvement throughout. The GOV.UK design principles open with "Do research, analyse data, talk to users.". The service manual insists on real people with real needs.
There is a sharper edge for our world too: synthetic users are least reliable for under-represented and non-Western groups, the very people inclusive public services most need to hear from.
(Two caveats are worth keeping in view. Almost all of this evidence tests off-the-shelf models and naive prompting, and it tests them as stand-ins for participants, which is the use we have already ruled out. The field also moves quickly, so the verdict belongs to today's tools used carelessly, not to a permanent ceiling.)
The persona that nobody opened
Think about the last beautifully produced persona deck you saw. Researched, designed, signed off, and then filed somewhere nobody returned to. The static persona is the most reliably ignored artefact in design. It is correct, it is evidence-based, and it is dead on the page.
Now give the same proto-persona a name and a voice that a workshop can talk to. Something shifts. People who would never open a research report will happily argue with "Priya" about whether she would trust a Combined Authority's latest scheme, or what she would do when she's asked for a document she doesn't have.
The name travels. The context travels. The user-centred framing percolates out across a wider team, including the people who hold the budget and sign off the decisions. In a word, it is exciting, and excitement is the thing static personas never generate.
For evangelising user-centred design inside an organisation that has never really done it, a persona you can interrogate is worth more than a shelf of immaculate ones nobody reads.
What we actually use them for
So we use synthetic personas for three things, and none of them is producing findings.
The first is buy-in. In a workshop, a living persona gives a mixed room a shared language and a shared focus, fast. The conversation stops being about features and starts being about people. It's incredibly rewarding to see a room full of activity and buzz, laughter and discussion. Investment and momentum that can be harnessed.
The second is rehearsal. Before a research script ever meets a real participant, we run it past a synthetic persona to find the dead questions, the leading ones, and the gaps. It is a cheap way to arrive at the real interview better prepared. A prototype, if you will.
The third is part of wider suite of tools within our Platform to surface assumptions and hypotheses, feed them into the Assumptions Register, and generate experiment ideas worth testing with real users. The persona helps us shape what to ask or test. The full cohort of personas provides a common framing for the multiple lenses we can use. (This is invaluable in Portfolio reviews but more on that in future posts)
A persona is only as good as what sits behind it
This is where craft earns its keep, because a synthetic persona is only ever as good as the design and logic underneath it. Typing "pretend to be a 34-year-old carer" into a chatbot gets you a caricature. That is the thing the critics rightly dismiss, and it is not what we build.
The personas we use are assembled from real interview scripts drawn from real-life qualitative work. If that's not available, they are labelled as proto-personas, and treated as such until real research arrives to challenge or firm things up.
We map common and recognised biases and heuristics, so the behaviour they model has some grounding and we avoid building a legion of sycophants. We use regional dialects, English as a second language, swearing. None of that turns a persona into a participant.
A well-built persona is a thinking and framing tool. It is still not evidence.
The risk hides in the same place as the value
It would be disingenuous to leave it there, because the thing that makes these tools useful is also what makes them dangerous.
A dynamic persona is compelling for the same reason a chatbot feels like a friend: fluent, warm, confident language triggers us to project a real mind behind it. That is the ELIZA effect, named sixty years ago and stronger than ever.
Flimsy AI design can easily deliver the very believability that wins a room but masks how little is underneath. We've all had chats with off the shelf LLMs that answer confidently and wrongly. They can praise every idea, which makes it actively misleading for prioritisation. They also inherently look to smooth over conflict or contradiction, though this tends to differ by the underlying model (another variable to consider!).
The discipline, then, is seeing the line. A persona can help steer us towards what to go and find out. It can never tell us what the answer is. Treat its output as a hypothesis and it earns its keep. Treat it as data and you have built a strategy on potentially flattering hallucinations.
Don't lose the use because of the misuse
How to use AI ethically and responsibly is one of the biggest questions of the current age. We are working with central and local government teams to define where those boundaries are. And it's easy to conclude the risks outweigh the benefits. Change is hard. New is scary.
But there is a real tool here. Used to make a team care, to rehearse the questions, and to sharpen what you take to real people, a 'living' persona does work a static one cannot.
At Lighthouse, real users sit at the centre of every Lab. The synthetic persona is simply a tool that helps think and frame, and an artefact that gets more of the people that matter to pay attention.