PostHog trains models on customer data

Hot take: this is the kind of AI moat most dev tools want but few can justify, because first-party product telemetry is exactly the data generic model vendors do not have.

• –Training on replay and usage data could make PostHog materially better at surfacing product issues, predicting confusion, and suggesting fixes
• –PostHog Code in beta reads like a shift from “AI features” to an AI-native product editor, not just a thin copilot layer
• –The trust tradeoff is the whole story: opt-out-by-default for US cloud users is defensible on utility grounds, but it will draw scrutiny
• –Keeping training in-house and avoiding third-party model providers is a meaningful differentiator if PostHog can prove the anonymization story is solid
• –This is a strong fit for PostHog’s stack, because the company already sits on the behavioral data needed to make the models useful