They also drift on very large tasks, losing track of earlier decisions, and they can quietly introduce security or performance issues that look fine in isolation. They're pattern-matchers, so genuinely new problems, where there's nothing similar in their training, are where they're weakest.
The practical takeaway is to use them for what they're good at (well-defined, self-contained work with clear success criteria) and to keep human oversight on design, security, and the decisions that carry real consequences.
Updated July 2026