Best practices for deploying language models

Three companies that normally compete for the same customers just put their names on the same document. Cohere, OpenAI, and AI21 Labs released a joint set of best practices aimed at anyone building or shipping large language models, and the fact that it happened at all says something about where the industry thinks it's headed.

The framework isn't a technical paper full of benchmarks. It reads more like an early attempt at self-governance, the kind of thing companies do when they'd rather write their own rules than wait for regulators to write them first. Details on enforcement are thin, and there's no mention of penalties or audits, but the symbolism matters: three labs with different training methods and different business models agreeing that some baseline standard should exist for how these models get deployed.

What's notable is who's missing from the byline. Google, Meta, and Microsoft aren't listed as co-authors, despite running some of the largest language model programs on the planet. Whether that's a timing issue or a signal that the biggest players want more room to maneuver before committing to shared rules is left unsaid.

This kind of coordination usually shows up right before governments start asking harder questions. Publishing a voluntary standard is a way to get ahead of that conversation, to say we already thought about this before anyone had to ask. It's a smart move regardless of the motive, and it sets a marker that other companies entering the space will now be measured against, whether they signed on or not.