<p>Demands for algorithmic accuracy and evaluations of racial and gender bias have emerged in the wake of extensive public controversies and civil society campaigns. Public attention to AI errors has also gained momentum with the rollout of facial recognition from borders to supermarkets and from airports to public squares. Yet, despite some early wins and even limited bans, public and private actors subsequently resumed or even intensified their use of facial recognition technologies. This paper argues that critiques of AI errors have shifted from ‘matters of concern’ to ‘matters of value’. Drawing on Luc Boltanski and Arnaud Esquerre’s analysis of the enrichment economy, I show that facial recognition bias has been articulated as algorithmic enrichment, whereby calculations of error increase the value of algorithms that perform better on evaluations. Through distinctions between exceptional and standard algorithms, evaluations of accuracy and error turn some facial recognition technologies into ‘enriched things’. In response to civil society critiques, private companies and police forces in the UK and the USA justify their deployment of facial recognition technologies by citing highly ranked algorithms. Critical interventions need to address how enriched AI is made more valuable through, rather than despite, its errors.</p>