MIT Researchers Back Ensemble Algorithms as Cure for AI Monoculture Risks

2 articles · Updated · MIT News · Sep 29

MIT researchers argue algorithmic monoculture is not inherently harmful, after reviewing common objections to one AI system dominating decisions in areas such as hiring.
Their analysis found fears of systematic exclusion are often overstated: if firms use the same screening tool, jobs still get filled, and competition for the same candidates can even lift wages.
The sharper risk, they say, is informational homogenization—one algorithm can create echo chambers that reduce exploration and make firms less likely to discover the best candidates.
Simulations showed an ensemble algorithm that averages multiple hiring models can sometimes match or beat a polyculture of different firm-specific systems, though the researchers say real-world feasibility remains untested.
The study points to lending and hiring as current monoculture examples, while warning that impacts will vary by domain and may be more problematic in areas like generative AI or AI-guided research.