The Friction of AI Training: PostHog's Bold Move and the Community Backlash

PostHog recently announced an ambitious shift in its product strategy: the transition toward "proactive, self-driving products." The vision is to move beyond simple analytics and toward tools like PostHog Code—a product editor designed to help teams build better products faster by surfacing answers and solutions automatically.

To achieve this, PostHog intends to train its own AI models using data from its customers' instances. While the company frames this as a necessary step to make their products smarter and more useful, the announcement has triggered a significant backlash from the developer community, highlighting a growing tension between AI innovation and user consent.

The Vision: Proactive Analytics and Synthetic Testing

PostHog's goal is to leverage its vast amount of user behavior data to move from reactive analysis to proactive optimization. The company has identified three primary areas where custom-trained models could provide immediate value:

  • Session Replay Analysis: While PostHog AI can already detect issues in replays, the process is currently expensive and difficult to scale. A model trained on underlying replay data could automate the diagnosis of problems across thousands of users simultaneously.
  • Synthetic User Testing: PostHog aims to use its knowledge of user behavior to predict where users might get confused or where flows might break before a feature is shipped to production, potentially reducing the burden of manual QA and code review.
  • Predictive Behavior Modeling: By predicting user behavior, PostHog hopes to suggest specific changes to improve conversion rates and reduce user frustration for existing features.

The Implementation: Transparency vs. Consent

PostHog has attempted to be transparent about its rollout, avoiding the typical "buried in the T&Cs" approach. They have outlined a specific set of rules for how data will be used:

  1. EU Cloud Users: Opted out by default due to regulatory requirements.
  2. Contractual Protections: Users with BAAs, MSAs, or similar agreements are opted out by default.
  3. US Cloud Users: Opted in by default.
  4. Anonymization: All data is promised to be anonymized before training.
  5. Internal Training: PostHog will perform the training themselves, promising not to sell or send data to third-party model providers.

The Community Reaction: "Opt-in by Default"

Despite the transparency of the announcement, the reaction on platforms like Hacker News has been overwhelmingly negative. The core of the controversy lies in the "opt-out" nature of the policy for US-based customers.

Critics argue that "opt-in by default" is a contradiction in terms. Many users expressed that being automatically enrolled in a data-training program—especially one that uses telemetry data inextricably linked to their proprietary code—is an unacceptable breach of trust.

"The only way to get their data is to assume you can take it and force them to tell you to stop. Wonder what that could mean?"

Beyond the policy itself, users raised concerns about the definition of "anonymization." Given that PostHog has access to sensitive logged-in data, developers are questioning exactly what scrubbing processes are in place to ensure that sensitive information doesn't leak into the training set.

The Brand Paradox

Another point of contention is the misalignment between PostHog's public brand persona and this business decision. PostHog has cultivated a "no-BS," quirky, and transparent image. However, many users feel that an opt-out data policy is a "slimy" move that contradicts that very ethos.

"If they are going to position yourself as the non-slimy no-BS guys, they can't pull this nonsense."

Conclusion: The Cost of Data Hunger

PostHog's dilemma is a classic one in the AI era: the hunger for high-quality, proprietary data to build useful models. The company admitted that without an opt-out policy, they simply would not have enough data to make the models effective.

However, the community response suggests that for many developers, the cost of this data acquisition—measured in user trust and customer churn—may be higher than the potential value of the AI features being built. As users migrate toward self-hosted solutions or alternative analytics providers, the incident serves as a reminder that in the modern developer ecosystem, consent is often more valued than convenience.

Sources