AI Fake Users Get More Convincing When Told Not to Overthink

Companies and platforms increasingly test their rules and features on AI 'fake users' — software stand-ins that imitate how real people would react. That only works if the stand-ins behave like the person they're supposedly copying. A new study from researchers in Serbia and Germany put that to the test: they profiled eight real volunteers through questionnaires, deep interviews and written self-descriptions, then recorded how those people actually reacted to 68 social media posts.

Four different AI language models (the tech behind chatbots like ChatGPT) were then asked to predict each volunteer's reactions, using five different instruction styles. The biggest finding was about style, not brainpower. When the AI was told to answer intuitively and immediately — to go with its gut — it matched real people far better than when it was told to think carefully and analyze. That 'gut feeling' setup cut the exaggeration of personality differences from seven times the human level down to three.

The most striking part came with posts about topics the volunteers had never been asked about. Even there, the intuitive AI predicted reactions better than a crowd of real people guessing. In other words, these simulated users could work as general-purpose stand-ins, not just specialists on the topics they were profiled for. Notably, the AI copies matched their own stated profiles more closely than the real humans matched their own survey answers.

The catch is uncomfortable. The same realism that makes these tools useful for testing platform rules also makes them useful for faking grassroots support before an election — a practice known as astroturfing. The researchers flag this directly. The paper also hints at a broader lesson for AI builders: for some jobs, 'don't think too hard' beats careful reasoning.