Figure 2
The two-panel grouped vertical bar chart is drawn on coordinate planes and compares the percentage choosing a statistical model across four experimental conditions. Both panels share the same horizontal categories listed from left to right as “Control”, “Participant Only”, “Algorithm Only”, and “Participant and Algorithm”. In both panels, the vertical axis is labeled “percent choosing statistical model” and ranges from 0 percent to 100 percent in increments of 20 percent. Panel (a), titled “G P T decisions compared to humans”, contains 8 bars, with 2 bars per category. A legend identifies the two groups as “L L M” and “Humans”. The data values are as follows. Control: L L M: 31 percent; Humans: 65 percent. Participant Only: L L M: 98 percent; Humans: 63 percent. Algorithm Only: L L M: 50 percent; Humans: 26 percent. Participant and Algorithm: L L M: 82 percent; Humans: 23 percent. Panel (b), titled “G P T decisions across models and temperature”, contains 16 bars, with 4 bars per category. A legend identifies four groups as “G P T-3.5, Temp equals 0”, “G P T-3.5, Temp equals 1”, “G P T-4, Temp equals 0”, and “G P T-4, Temp equals 1”. The data values are as follows: Control: G P T-3.5 Temp 0: 60 percent; G P T-3.5 Temp 1: 30 percent; G P T-4 Temp 0: 15 percent; G P T-4 Temp 1: 20 percent. Participant Only: G P T-3.5 Temp 0: 100 percent; G P T-3.5 Temp 1: 90 percent; G P T-4 Temp 0: 100 percent; G P T-4 Temp 1: 100 percent. Algorithm Only: G P T-3.5 Temp 0: 70 percent; G P T-3.5 Temp 1: 50 percent; G P T-4 Temp 0: 45 percent; G P T-4 Temp 1: 35 percent. Participant and Algorithm: G P T-3.5 Temp 0: 100 percent; G P T-3.5 Temp 1: 80 percent; G P T-4 Temp 0: 70 percent; G P T-4 Temp 1: 75 percent. Note: All numerical values are approximated.Proportion of decisions to rely on the external algorithm across Study 1 treatments. Panel (a) compares GPT participants to human participants from Dietvorst et al. (2015). Panel (b) shows GPT decisions by model version and temperature. Source: Authors’ own work