What's this about?
OpenAI published research on how organizations actually use ChatGPT, reporting that the top 10% of enterprises use plugins twice as often and skills six times as often as typical firms. OpenAI's framing: "these frontier firms are not ahead by accident." Ethan Mollick added the uncomfortable implication — the firms using AI most heavily also tend to have the most productive employees, so early adopters who were already doing well may now start pulling further ahead.
What exactly is the selection-effect problem here?
The data shows heavy AI use and high performance occurring together. It cannot show which causes which. Productive, well-run firms may adopt new tools early because they're well-run — in which case AI is a symptom of competence, not a source of it. Distinguishing the two requires comparing similar firms where adoption varied for reasons unrelated to quality, which this data doesn't do. Mollick is careful to say "to the extent that AI use boosts firm performance," which is exactly the right hedge.
Why is this worth noticing?
- The specific gap — 6x on skills, 2x on plugins — suggests the differentiator isn't access to models, which everyone has, but the organizational work of encoding your own processes into reusable form.
- It's vendor-published research about the vendor's own product, which is a real reason for caution, but the underlying usage data is the sort only the provider can see.
- If the divergence is real, it argues AI widens gaps between firms rather than levelling them — the organizational mirror of the "AI is an amplifier, not an extender" argument that dominated yesterday's engineering-jobs debate.