1: \begin{abstract}

2:   As algorithms are increasingly applied to screen applicants for high-stakes

3:   decisions in employment, lending, and other domains, concerns have been raised

4:   about the effects of {\em algorithmic monoculture}, in which many

5:   decision-makers all rely on the same algorithm. This concern invokes analogies

6:   to agriculture, where a monocultural system runs the risk of severe harm from

7:   unexpected shocks. Here we show that the dangers of algorithmic monoculture

8:   run much deeper, in that monocultural convergence on a single algorithm by a

9:   group of decision-making agents, even when the algorithm is more accurate for

10:   any one agent in isolation, can reduce the overall quality of the decisions

11:   being made by the full collection of agents. Unexpected shocks are therefore

12:   not needed to expose the risks of monoculture; it can hurt accuracy even under

13:   ``normal'' operations, and even for algorithms that are more accurate when

14:   used by only a single decision-maker. Our results rely on minimal assumptions,

15:   and involve the development of a probabilistic framework for analyzing systems

16:   that use multiple noisy estimates of a set of alternatives.

17: \end{abstract}

18: