The Scale of AI Governance
ConductAtlas tracks 872 provisions across 8 major AI platforms. 286, roughly one-third, are classified as high severity based on their operational impact on users and developers.
OpenAI alone maintains 360 provisions across its service terms, privacy policy, usage policy, data processing addendum, and API-specific agreements. Anthropic maintains 201. Perplexity has 139. Google Gemini has 79. Together AI, Groq, and Fireworks AI add another 93 combined.
These are not static documents. OpenAI has generated 36 governance events since ConductAtlas began monitoring. Google Gemini has generated 19. Perplexity has generated 9. The terms governing AI infrastructure are actively evolving.
What Just Changed: OpenAI, June 6-9
In a four-day span from June 6 to June 9, 2026, OpenAI modified its privacy policy and service terms multiple times. The most operationally significant change came on June 9: OpenAI removed explicit descriptions of ad personalization controls that were previously available to Free and Go users.
The privacy policy continues to authorize ad personalization for these tiers. What changed is the disclosure of user controls. Previously, the policy described specific mechanisms users could access to manage ad-targeting data through account settings. That language was removed without explanation or alternative disclosure.
This is a textbook example of governance shifting from explicit to implicit. The authorization remains. The documented control disappears. Users who relied on the prior disclosure to understand their options now face ambiguity about what controls exist and how to exercise them.
For developers building applications on OpenAI APIs, this matters because user data flowing through their applications is governed by the privacy terms users accepted with OpenAI, not the terms the developer sets. A change to how OpenAI handles ad personalization data affects every app that processes user interactions through the API.
How AI Terms Govern Downstream Apps
When a developer builds a product on an AI API, the relationship is not simply technical. It is a governance dependency. The AI platform's terms of service become operational constraints on the downstream application:
Usage restrictions flow downstream. If OpenAI's acceptable use policy prohibits certain content categories, every application built on OpenAI must comply. A startup that builds a product later prohibited by a usage policy update faces a choice between compliance and rebuilding on a different provider. Neither is costless.
Data handling terms are inherited. How the AI platform collects, stores, processes, and shares data is governed by its own privacy policy and data processing agreements, not the downstream developer's. When OpenAI changes its data handling terms, that change affects the data practices of every application built on its API.
Rate limits and pricing are governance decisions. When an AI platform changes rate limits, introduces usage caps, or restructures pricing tiers, every dependent application is affected. These are not just business decisions. They are governance decisions that determine which applications can operate and at what scale.
Enforcement is unilateral. AI platforms can suspend API access, restrict features, or terminate accounts under terms that grant broad enforcement discretion. ConductAtlas tracks the specific provisions that authorize these actions across every monitored AI platform.
The Arbitration Layer
Every major AI platform has adopted mandatory arbitration with class action waivers. ConductAtlas tracks these provisions in detail: OpenAI maintains 15 arbitration-related provisions, Anthropic has 8, Perplexity has 7, and Together AI has 5.
For developers, this means that disputes over API access, enforcement actions, pricing changes, or data handling are resolved through individual arbitration, not in court. When an AI provider suspends API access that a production application depends on, the developer's recourse is individual arbitration under terms the platform wrote.
This creates the same asymmetry ConductAtlas documented in the mandatory arbitration analysis: developers waive collective legal action while the platform retains broad enforcement discretion. The class action waiver means that even if a pricing change or enforcement action affects thousands of developers simultaneously, each must pursue their claim individually.
Comparing AI Platform Governance
Not all AI platforms govern identically. ConductAtlas data reveals structural differences:
OpenAI has the largest governance footprint: 360 provisions, 132 high severity, 36 events. Its terms are the most actively changing of any AI platform. The June 9 privacy policy revision demonstrates a pattern of frequent, incremental modifications.
Anthropic maintains 201 provisions with 61 high severity. It has generated fewer governance events (3), suggesting more stable terms. Its usage policy and acceptable use policy are more detailed than most competitors on specific prohibited applications.
Google Gemini has 79 provisions but has generated 19 events, the second-highest event count. This reflects the broader Google governance pattern identified in prior ConductAtlas analysis, where consumer-facing services can be modified without notice under terms that authorize broad service changes.
Perplexity AI has 139 provisions with 39 high severity, a dense governance footprint relative to its product scope. It has generated 9 events.
Smaller providers (Together AI, Groq, Fireworks AI) have 30-32 provisions each, reflecting simpler terms structures. For developers seeking governance simplicity, smaller providers currently present less complex terms, though less complex does not mean less restrictive.
What Developers Should Do
Map your AI governance dependencies. Every AI API your product uses is a governance dependency. Document which terms govern each integration: service terms, usage policies, data processing agreements, and privacy policies.
Monitor for changes. OpenAI changed its privacy policy 5 times in 4 days. These changes may be minor or material, but you cannot assess impact without knowing they occurred. ConductAtlas monitors all major AI platforms with same-day change detection.
Plan for provider portability. If your AI provider changes terms in ways that affect your product, how quickly can you migrate? API dependency governance risk is reduced by maintaining the technical ability to switch providers.
Review arbitration terms. Know the dispute resolution mechanism for every AI platform you depend on. When enforcement actions or pricing changes affect your business, your legal options are defined by terms you accepted when you created your API key.
Active Monitoring
ConductAtlas tracks governance changes across AI platform infrastructure:
- OpenAI: 360 provisions, 36 events (5 in the last 4 days)
- Anthropic: 201 provisions, 3 events
- Google Gemini: 79 provisions, 19 events
- Perplexity AI: 139 provisions, 9 events
- Together AI: 31 provisions, 2 events
- Groq: 32 provisions, Fireworks AI: 30 provisions
Primary Sources
OpenAI governance documents (360 provisions archived by ConductAtlas)
Anthropic governance documents (201 provisions archived by ConductAtlas)
Google Gemini governance documents (79 provisions archived by ConductAtlas)
Perplexity AI governance documents (139 provisions archived by ConductAtlas)
Every application built on AI infrastructure operates under governance terms set by the AI provider. Those terms change frequently, often without individual notice to downstream developers. The June 9 OpenAI privacy policy revision, removing documented user controls while maintaining the underlying authorization, illustrates how governance changes at the infrastructure layer flow downstream to every dependent application. Tracking these changes is not optional for anyone building on AI.