What happens to knowledge when research runs on copies?

Synthetic users promise research without the inconvenience of people. But when data becomes a copy of a copy, something happens to knowledge itself, and to the possibility of making anything new.

Abstract

Synthetic data is arriving in user research the way it arrived everywhere else, as a cost argument. Why recruit twelve participants when a model can simulate a thousand? The question this essay asks is older than the technology. What is knowledge worth when its source is a copy?¹1I build with synthetic data myself; my Voice Of tool simulates stakeholder voices, grounded in real interviews. The grounding is the point. For user-centered practitioners, whose entire discipline rests on contact with actual people, the answer decides whether the discipline survives its own tooling.

I. Research without people

The pitch is seductive because the pain is real. Recruiting is slow, incentives are expensive, access to specialists is brutal. I've lived this. Getting time with twelve pathologists across three hospitals was the hardest constraint of the engagement. A synthetic pathologist never cancels.

And synthetic users are not useless. They're a reasonable way to pressure-test an interview guide, to explore the space of plausible objections, to rehearse. The trouble starts when simulation stops being rehearsal and starts being cited as evidence. A model's rendering of "a nurse" is a statistical composite of everything written about nurses, weighted toward whatever got written down most. It's an average with good manners. The participants who change a product's direction are almost never the average; they're the outlier whose workaround reveals the real workflow, the expert whose irritation reveals the real stakes.

II. Copies of copies

Baudrillard gave this a name forty years ago, the simulacrum, a copy that no longer refers to any original.²2Baudrillard (1981): the successive phases of the image, ending in signs that refer only to other signs. A synthetic user is textbook. It precedes and replaces the person it claims to represent. Borges got there earlier with the empire whose map grew until it covered the territory exactly.³3Borges, “On Exactitude in Science” (1946), a one-paragraph story doing the work of a shelf of media theory. Synthetic data is the map declaring the territory redundant.

This isn't only philosophy anymore; it has an empirical result. Models trained recursively on their own outputs degrade, forgetting the tails of the original distribution first. The phenomenon has a technical name, model collapse,⁴4Shumailov et al., “AI Models Collapse When Trained on Recursively Generated Data,” Nature (2024). and a familiar shape. Each generation of copies keeps the center and loses the edges. Research built on synthetic users inherits exactly this failure mode, because the edges are what research exists to find.

III. What is knowing worth?

Epistemology has always asked what separates knowledge from confident belief, and one classic answer is the causal one. To know something, your belief has to be appropriately connected to the fact itself, and coherence with other beliefs is not enough.⁵5Goldman, “A Causal Theory of Knowing,” The Journal of Philosophy 64, no. 12 (1967): 357–372. Synthetic research swaps this out. Findings cohere beautifully with everything the model has read, and connect to no one.

Floridi's philosophy of information offers a useful line here. Data only becomes information when it's meaningful and truthful about something.⁶6Floridi, The Philosophy of Information (2011), on the semantic conditions data must meet. A synthetic quote can be meaningful and still be about nothing. For a user-centered practitioner this is not an abstract worry. Our authority in a product organization rests on one claim. We talked to the people, and here's what they said. Replace the people and the claim doesn't weaken. It becomes false.

None of this means the practice is doomed. It means provenance becomes the discipline. Knowing, in a synthetic-data world, is worth exactly as much as your ability to trace a finding back to a person who exists.

IV. Innovation needs an outside

Innovation theory has been consistent about one thing for a century. The new is made from recombination, but recombination needs fresh material. Schumpeter defined innovation as new combinations of existing elements,⁷7Schumpeter, The Theory of Economic Development (1934): development as the carrying out of new combinations. and Brian Arthur described technology as combinatorial evolution that periodically has to reach outside itself, into newly captured natural phenomena, to keep producing anything new.⁸8Arthur, The Nature of Technology (2009).

A research practice that only recycles what's already been written has cut off its outside. The surprising observation, the thing no one thought to write down, is the raw material of every new combination, and it lives in the territory rather than the map. If we keep recycling, we don't get less innovation at the margin. We get recombinations of recombinations, novel in arrangement and empty of news.

V. The steelman

The oldest argument in the Western canon is on the other side of this essay, and it deserves its full weight. Plato banished imitation from his republic because a copy stands, as he put it, at three removes from the truth.⁹9Plato, Republic, Book X, on mimesis as thrice removed from the real; Aristotle, Poetics, on imitation as the medium of learning. Aristotle answered that imitation is how human beings learn at all. We rehearse the world in copies before we act in it. If Aristotle is right, the synthetic user is not an epistemic scandal. It is a rehearsal space, the newest member of a family that includes the anatomical model, the flight simulator, and the role-played interview, and refusing it outright would be a kind of superstition.

I side with Aristotle further than the earlier sections suggest. Simulation legitimately extends research wherever rehearsal is the honest description of the work, whether that means exploring a hypothesis space before spending scarce participant hours, stress-testing an instrument, or working where real data cannot leave its enclosure, which is exactly why my own Voice Of tool exists. But rehearsal has a boundary, and Benjamin marked it a century ago. What the copy cannot carry is the original's presence, its embeddedness in a particular time and place, what he called its aura.¹⁰10Benjamin, “The Work of Art in the Age of Mechanical Reproduction” (1935). In research, the aura is not mystique. It is the situated, resistant particularity of a person who can contradict you.

And that is where the argument becomes ethical rather than methodological. Levinas placed the ground of ethics in the encounter with the face of the Other, the one thing that exceeds every representation I can form of it.¹¹11Levinas, Totalité et Infini (1961): the face of the Other as what resists totalization into my categories. A synthetic user is a representation with no remainder; it can never exceed what the model already contains, which means it can never make a claim on the researcher. User research is more than a measurement practice. It is a relationship of answerability to people who can be wronged, and no simulation can stand in that relation. The steelman, followed to its end, concedes the method and returns the obligation intact.

VI. What practitioners can do

Three working rules, none of them anti-synthetic. Use synthetic users for rehearsal and never for evidence; the line is whether the output shows up in a findings deck. Treat provenance as a first-class property of every insight, the way a journalist treats sourcing. And protect a fixed budget of real human contact per project, because that contact is not a cost to optimize away. It's the input the whole system, synthetic layers included, ultimately runs on.

Notes

1. See “Voice Of” in the Lab section: a local RAG pipeline that simulates stakeholder voices, deliberately grounded in real interview transcripts.
2. Jean Baudrillard, Simulacres et Simulation (1981); English trans. Simulacra and Simulation, University of Michigan Press, 1994.
3. Jorge Luis Borges, “Del rigor en la ciencia” (“On Exactitude in Science,” 1946).
4. Ilia Shumailov et al., “AI Models Collapse When Trained on Recursively Generated Data,” Nature 631 (2024): 755–759.
5. Alvin I. Goldman, “A Causal Theory of Knowing,” The Journal of Philosophy 64, no. 12 (1967): 357–372.
6. Luciano Floridi, The Philosophy of Information (Oxford University Press, 2011).
7. Joseph A. Schumpeter, The Theory of Economic Development (Harvard University Press, 1934).
8. W. Brian Arthur, The Nature of Technology: What It Is and How It Evolves (Free Press, 2009).
9. Plato, Republic, Book X; Aristotle, Poetics, chs. 1–4.
10. Walter Benjamin, “The Work of Art in the Age of Mechanical Reproduction” (1935), in Illuminations, trans. Harry Zohn (Schocken, 1968).
11. Emmanuel Levinas, Totality and Infinity, trans. Alphonso Lingis (Duquesne University Press, 1969; orig. 1961).

Works Cited