Skip to main content
Back

Matching AI Personality to Yours Lifts Ad Clicks by 55 Percent

A preregistered randomized trial from Harang Ju and Sinan Aral shows Big Five personality pairings between humans and AI agents shift both work quality and ad performance in the field.

Pro

Paper, Full Access
  • Ju and Aral run a 1,258-person RCT pairing humans with AI agents prompted on Big Five traits (paper).
  • Extraverted humans plus conscientious AI produced the lowest-quality ads, a 0.33 SD drop from one prompt line.
  • Chat log analysis suggests bad pairings made the AI more compliant and less willing to push back.
  • Field experiment on X across ~5M impressions: 1 SD better text quality equals 5.9% higher CTR.
  • Neurotic human plus neurotic AI raised CTR by 55%, robust to quality and spend controls.
  • Effects are model-dependent (gpt-4o-2024-08-06 snapshot) but the pairing principle is expected to generalize.

A new preregistered study from Harang Ju and Sinan Aral argues that when we treat AI agents as teammates rather than tools, the way we prompt their personality can measurably change what a human-AI team produces. In a 1,258-person randomized experiment feeding a real advertising campaign on X, the authors show that particular pairings between a person's Big Five traits and an AI agent's prompted traits shift ad quality by up to a third of a standard deviation and, in one case, lift click-through rate by 55 percent.

The industry is already shipping AI personalities without much causal evidence for how they interact with users. OpenAI has introduced personality customization in GPT-5 with modes like Default, Cynic, Robot, Listener, and Nerd, while Anthropic has begun measuring undesirable shifts during training such as sycophancy and released methods to mitigate them. No large randomized experiment had tested whether matching those personalities to specific humans actually helps.

The gap this study fills

Prior productivity studies handed every participant the same chatbot and measured average gains. Those benefits varied considerably across tasks, worker seniority, gender, and skill levels. Ju and Aral hypothesize that some of that variance is a mismatch problem: identical AI, different humans. Person-team fit theory from organizational psychology holds that human teams work better when personalities complement each other, but nobody had run a proper randomized controlled trial with AI agents as the teammates.

To do it, the authors built Pairit, a collaborative workspace where a human and an AI agent share a live document, chat in real time, and iterate on display ads together. Only the human can hit submit, but the agent can edit text, pick images from a carousel, call DALL-E 3 to generate new ones, or wait. It sees screenshots of the current ad, the brief, chat history, and its own action log every ten seconds.