Is personalized AI more persuasive than humans?

01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?

Is GPT-4 given personal information more persuasive than humans in short discussions?

Salvi and his colleagues investigated whether large language models not only produced persuasive text but also adapted their arguments more effectively when they received basic sociodemographic information about the other person.

02 / HOW WAS THE RESEARCH CONDUCTED?

How did they test the question?

In a preregistered experiment, 900 participants had a brief, multi-round online discussion with a human or GPT-4. In some conditions, the opponent received basic information such as age, gender, education, employment status, ethnicity and political affiliation. The design created 12 conditions based on whether the opponent was human or AI, had access to personal information and how strongly the participant initially held the view.

03 / RESEARCH FINDINGS

Accessing personal information, GPT-4 showed a distinct advantage in the comparison.

In comparisons where humans and AI were not equally persuasive, personalized GPT-4 had a 64.4% probability of being the more persuasive side. This corresponded to an 81.2% relative increase in the likelihood of higher agreement with the advocated view after discussion. The 95% confidence interval was +26.0% to +160.7%, and the sample included 900 people. GPT-4 without personal information did not show the same clear advantage over humans.

GPT-4 in unequal matches

64.4%Probability that personalized AI was more persuasive in unequal matches; not a success rate for all conversations.

Relative probability increase

81.2%Relative increase in the likelihood of higher post-discussion agreement with the advocated view; not an increase in sales.

Source: Nature Human Behavior: summary, method, findings and limitations ↗

04 / DIJITALPI'S EXPLANATION

How should we interpret these findings?

This study suggests that personalization may be more powerful than just putting the person's name in the message. The model is able to reflect basic information in argument selection and expression. For business, the question is not to collect more data but to test what information actually makes the message more useful.

A wide confidence interval indicates that there is uncertainty about the size of the effect. We cannot use the central figure of 81.2% as a definitive rate that can be repeated in every environment. Different models, languages, products, message lengths and target audiences require a new experiment.

If personal data increases persuasive power, it also increases ethical and legal responsibility. Explicit consent, data minimization and avoiding unexpected inferences must be considered alongside performance. This section is DijitalPi's practical interpretation; the article did not test KVKK-compliant campaign design in Turkey.

05 / CONCLUSION AND OPEN QUESTIONS

What did we learn, and what do we still not know?

The study shows that GPT-4 can gain a persuasive advantage over human opponents when it receives basic personal information in short online discussions. The result is limited to political debates, specific model versions and short-term attitude change; it does not establish effects on ad conversion, purchases or long-term trust.

Would the same effect be seen in Turkish messages, different product categories and real customer behavior? To know this would require a separate field experiment with a control group without personal data, predetermined behavioral measures, and data use limits.