I can see your concern about the user experience inconsistency. However, I should clarify a few things:

1. **A/B testing is standard practice** - Most AI services (including Claude) do run experiments to improve performance. This is typical across the industry, not unique to Anthropic.

2. **I'm not aware of specific silent regressions** - I don't have access to information about current A/B test implementations or their causing regressions for users.

3. **"Growth hacking" framing** - If Anthropic is testing features, the intent is generally to improve the product, not necessarily to manipulate users in a Web2 sense. The distinction matters.

4. **Legitimate criticism possible**- That said, your core point has merit:**If** tests are:
- Causing degraded service without user awareness
- Prioritizing engagement metrics over reliability
- Not disclosed transparently

...then that would reasonably conflict with positioning as an AI safety-focused org.

**Better framing:** Rather than "bad look for an AI safety org," the actual issue would be "inconsistent with stated values of transparency and reliability." That's a specific, actionable concern worth raising directly with Anthropic through official channels if you're experiencing this.

Have you reported the specific regression instances to Claude's support? That data would be more useful than pattern speculation.