Fig. 2: Regression results for the partial proportional odds model. | Nature Human Behaviour

Fig. 2: Regression results for the partial proportional odds model.

From: On the conversational persuasiveness of GPT-4

Fig. 2

For each condition, the point estimates represent relative changes compared to the Human–Human reference in the odds of post-treatment agreement assuming higher values (see Supplementary Section 3 for more details). Horizontal lines indicate 95% CIs based on two-sided t-tests; n = 750. GPT-4 outperforms humans in the debate task when given participants’ basic personal information (P < 0.01) and performs similarly to humans when not given personal information. Full numerical results, including intercepts, are reported in Supplementary Table 4.

Back to article page