Two teachers grading the same student can reach very different conclusions? yet little is known about which teacher characteristics drive such disagreement. We investigate this question through a large-scale online experiment in which over 1,600 French secondary-school teachers independently evaluated the same set of fictitious student transcripts, generating hundreds of ratings per transcript. This design allows us to decompose grade variation into a component attributable to the student (transcript fixed effects) and one attributable to the teacher (rater fixed effects). After controlling for transcript fixed-effects, we find that teacher leniency accounts for about 40% of the remaining grade variance. These differences are largely unrelated to teachers' gender, subject, teaching style, or job satisfaction, but are positively associated with altruism (measured via an incentivized dictator game) and exhibit a U-shaped relationship with age, with middle-aged teachers being the most strict. An incentivized task reveals that most teachers perceive themselves as more lenient than their colleagues - in particular women - but middle-aged teachers less so. Finally, teachers' stated tolerance for inter-rater disagreement comfortably exceeds the disagreement we actually observe - in contrast with findings from other professional domains. These results speak to the design of fair assessment panels and suggest that differential leniency, while substantial, may be partly predictable and therefore correctable.

---

12 August 2026     Paper Number CEPDP2204

Download PDF - Teachers' leniency: experimental evidence

This CEP discussion paper is published under the centre's Education programme.