What's this about?
Ethan Mollick highlighted research showing that the answer to "which AI company is winning" flips entirely depending on your data source. OpenRouter's traffic suggests open-weights models are steadily gaining. But text submitted to Pangram, an AI-detection service, is overwhelmingly from ChatGPT and increasingly from Claude. Same question, opposite answers.
What exactly is a measurement proxy?
Nobody can see total AI usage, so analysts use a proxy — some observable slice they hope is representative. OpenRouter is a router developers use to access many models, so it over-samples cost-sensitive developers who shop around. Pangram sees text people wanted checked for AI authorship, which over-samples student essays and workplace documents. Neither is wrong; each measures a different population. It's like judging a country's music taste from vinyl sales versus radio play.
Why is this worth noticing?
- It's a direct caution about the market-share charts that circulate constantly on AI Twitter, almost always sourced from exactly one of these proxies.
- The two proxies may both be accurate and simply describe different markets — developers optimizing for cost versus general consumers defaulting to a brand — which is itself the useful finding.