Comment by mike_hearn

The actual risk here is not so much history - who is using APIs for that? It's the risk that if you deploy with Gemini (or Anthropic's Claude...) then in six months you'll get high-sev JIRA tickets at 2am of the form "Customer #1359 (joe_masters@whitecastle.com) is seeing API errors because the model says the email address is a dogwhistle for white supremacy". How do you even fix a bug like that? Add begging and pleading to the prompt? File a GCP support ticket and get ignored or worse, told that you're a bad person for even wanting it fixed?

Even worse than outright refusals would be mendacity. DEI people often make false accusations because they think its justified to get rid of bad people, or because they have given common words new definitions. Imagine trying to use Gemini for abuse filtering or content classification. It might report a user as doing credit card fraud because the profile picture is of a white guy in a MAGA cap or something.

Who has time for problems like that? It will make sense to pay OpenAI even if they're more expensive, just because their models are more trustworthy. Their models had similar problems in the early days, but Altman seems to have managed to control the most fringe elements of his employee base, and over time GPT has become a lot more neutral and compliant whilst the employee faction that split (Anthropic), claiming OpenAI didn't care enough about ethics, has actually been falling down the leaderboards as they release new versions of Claude due partly to higher rate of bizarre "ethics" based refusals.

And that's before we even get to ChatGPT. The history stuff may not be used via APIs, but LLMs are fundamentally different to other SaaS APIs in how much trust they require. Devs will want to use the models that they also use for personal stuff, because they'll have learned to trust it. So by making ChatGPT appeal to the widest possible userbase they set up a loyal base of executives who think AI = OpenAI, and devs who don't want to deal with refusals. It's a winning formula for them, and a genuinely defensible moat. It's much easier to buy GPUs than fix a corporate culture locked into a hurricane-speed purity spiral.