MIT tested 41 AI models on 11,000 real tasks. The "good enough" problem is worse than you think.

A new MIT study has quantified a growing concern in professional AI adoption: the 'good enough' problem. Researchers put 41 different AI models through 11,000 real-world tasks, simulating work in fields like writing, analysis, and coordination. The results were stark. While 65% of basic text generation tasks met a minimal 'acceptable' quality bar, a sobering 0% of models could reliably deliver 'superior' work on complex tasks. Performance on work requiring management, judgment, and coordination was particularly weak, with only a 53% success rate.

The study argues the core issue isn't just model capability, but flawed human workflows. Tools like ChatGPT and Claude deliver outputs with uniform confidence, whether correct or hallucinated. The research documents real-world consequences, including a consulting firm submitting hallucinated reports to government clients, law firms filing briefs with fake citations, and media outlets publishing under fake bylines—all cases where a human had ostensibly 'reviewed' the AI's work. The findings expose a critical gap: most organizations lack a structured, systematic validation process for AI-generated content, relying instead on passive review that often misses subtle errors.