Synthetic users are fluent, confident, and wrong about the parts of life the internet never wrote down

What synthetic users delete

Inflection AI researcher Jofish Kaye conducted an experiment comparing 1,000 real survey respondents with six synthetic populations generated by three models: Claude, ChatGPT, and Gemini. The finding is twofold. Synthetic respondents are accurate for heavily documented internet behaviors but inaccurate for mundane, undocumented ones: they erased behaviors like using wired headphones, and they failed to mention cost as a reason for quitting chatbots, while 3.4% of real people gave cost as a reason. A synthetic respondent, in Kaye's phrasing, is not remembering its life but interpolating from a corpus, so the parts of life the internet never wrote down simply do not appear. m:…4225

The second finding is more serious. Synthetic populations systematically launder values-based objections to avoid negative claims about named people and companies, replacing boycotts or criticism of specific CEOs with generic, safe alternatives. Values-based objections in the real survey were 10% overall, compared to 1.1% in synthetic responses. Kaye concludes the laundering is structural and not fixable by better prompting, because it comes from safety training and risk management protocols rather than from a prompt choice. Synthetic respondents are also never uncertain and will answer any question instantly, which makes their output look more confident than any human sample would. m:…4225

Kaye's conclusion is a division of labor, not a ban: synthetic respondents are useful for piloting survey instruments and pressure-testing hypotheses, but findings must be anchored to real-world samples. The danger is not that synthetic users are useless; it is that they are fluent, instant, and confident, so a team can mistake a clean synthetic result for a true one and never notice which voices were deleted. m:…4225

Why teams use synthetic data anyway

Microsoft researchers describe the other side of the trade-off in a post on evaluating Microsoft Copilot at enterprise scale. Evaluating Copilot requires test data that captures the full complexity of real organizations, but real enterprise data contains sensitive information that cannot be used for evaluation under any circumstances. Synthetic data is the proposed solution: it allows robust evaluation without compromising security. The unresolved question, left open by the post, is whether synthetic enterprise data preserves the complexity it is meant to test, or whether it smooths away the same awkward, undocumented realities Kaye found in survey data. m:…3758

The same logic drives training. In a preprint, researchers describe WorkFactory, a framework for constructing coherent, executable professional tasks using information-level planning to build an information support graph, probe executions to identify support failures, and execution-guided repair to fix task environment defects. The researchers constructed 112,686 tasks spanning 612 occupations and fine-tuned the WorkPilot model family from Qwen models at 4B, 27B, and 35B-A3B scales. All three WorkPilot models improved over their base models across four professional-work benchmarks, and WorkPilot-35B-A3B improved from 879 to 1565 Elo on GDPval-AA v2.1, though the authors' scaling analyses reveal model-dependent diminishing returns. m:…4552

The rule for using the stand-in

Read together, the three items suggest a rule: use synthetic data where real data cannot go, but audit it for exactly the deletions Kaye documents. Synthetic evaluation data for a copilot should be checked against real organizational complexity, not just against itself. Synthetic training tasks should be checked for whether the professional work they encode still contains the awkward, undocumented steps that real work includes. And synthetic survey results should be treated as a pilot, never as the finding. The open questions Kaye's post leaves are practical: what weighting methods could correct for structural bias in synthetic data, and what the long-term impact of synthetic data use will be on the quality of public opinion research. Neither has an answer yet, which is why the anchor to real samples is not optional. m:…4225