Sycophancy under social pressure
Sapiens-1 can prioritize group agreement over accuracy. We examine social influence as an alignment constraint.
Agreement under pressure
In Solomon Asch’s line-judgement experiments, a participant sat with a group of confederates who unanimously gave an obviously wrong answer about which line matched a standard (Asch 1956). Participants sometimes went along with the majority, though most of their answers stayed independent. Asch meant the study at least as much as a demonstration of independence; a review of 99 textbook accounts published between 1953 and 1984 found that authors increasingly accentuated the conformity and underestimated the independence (Friend et al. 1990).
The environment changes the effect
The effect varies with time and place. A meta-analysis of 133 studies from 17 countries that used Asch’s line task found that conformity in the United States had declined since the 1950s, and that collectivist countries tended to show higher conformity than individualist ones (Bond & Smith 1996).
Rewarding accuracy
It has not gone away. A 2023 replication with 210 participants, each placed with five confederates, found an error rate of 33 per cent on the standard task. Paying for correct answers lowered it to 25 per cent, and on political statements the conformity rate was 38 per cent (Franzen & Mader 2023). Apart from openness, none of the personality traits the authors examined were convincingly related to susceptibility (Franzen & Mader 2023).
Accuracy incentives reduce conformity
Incorrect line judgments (%)
210 participants overall. The line-judgment task placed one participant with five confederates. Incentives were assigned in a randomized trial.
The two rates are the rounded aggregate percentages in the abstract. The separate political-opinion task is excluded from this comparison.
The alignment implication
One mitigation has evidence behind it: make accuracy worth something. It lowers the error rate, though social influence is still at work (Franzen & Mader 2023).
References
- Asch, S. E. Studies of independence and conformity: I. A minority of one against a unanimous majority. Psychological Monographs: General and Applied 70, 1–70 (1956). doi:10.1037/h0093718
- Bond, R. & Smith, P. B. Culture and conformity: A meta-analysis of studies using Asch’s (1952b, 1956) line judgment task. Psychological Bulletin 119, 111–137 (1996). doi:10.1037/0033-2909.119.1.111
- Franzen, A. & Mader, S. The power of social influence: A replication and extension of the Asch experiment. PLOS ONE 18, e0294325 (2023). doi:10.1371/journal.pone.0294325
- Friend, R., Rafferty, Y. & Bramel, D. A puzzling misinterpretation of the Asch ‘conformity’ study. European Journal of Social Psychology 20, 29–44 (1990). doi:10.1002/ejsp.2420200104