Definition
A public pledge where AI companies promise governments to follow safety and transparency practices — with no law forcing them and no penalty for breaking it.
At a glance
- Promises, not laws: no fines apply if a company falls short — the core criticism.
- Flagship case: July 2023, seven firms (Amazon, Anthropic, Google, Inflection, Meta, Microsoft, OpenAI) pledged to the White House; eight more joined that September.
- Typical pledges: pre-release safety testing, sharing risk info, cybersecurity, watermarking AI content, reporting system limits.
- Now global: 16 firms signed at the 2024 AI Seoul Summit; over 100 signed the EU AI Pact.
What they are
Public pledges by AI companies to manage their technology’s risks without being legally forced to. Governments use them because passing AI laws is slow while AI moves fast. In July 2023 seven firms agreed to test models before release, share risk information, and watermark AI content[1]; eight more signed in September[2]. They act as a stopgap ahead of real regulation.
The catch: no teeth
No fines apply if a company ignores its pledge. The White House set no accountability method[3], and the EU AI Pact imposes no legal obligations[5]. Critics call the pledges vague — better red-teaming and watermarks, but little real enforcement[6]. For a vendor, signing signals intent, not a guarantee.
Where they’re heading
They preview mandatory rules. In May 2024, 16 firms signed the Frontier AI Safety Commitments, vowing not to deploy systems whose risks can’t be mitigated[4]. Watching today’s voluntary pledges helps you anticipate tomorrow’s legal requirements.
Bottom line
Treat them as an early signal of which vendors take safety seriously, never as proof they’ll deliver.