Proposed AI Safety Protocols & Standards

Draft for public comment from the AI Guardrails Pledge. These are proposals, not law or legal advice.

Not every AI needs the same rules. The question is what’s at stake for a real person. We sort AI into four levels.

How to find the level

Ask two questions:

  1. Does it make a decision about someone’s life?
  2. Does it use their personal data without them knowing?

🟢 Low risk: it doesn’t make decisions about people or use their data secretly.

🟡 Use with care: it talks with people or creates content that could pass as real.

🟠 High risk: it makes decisions about people’s lives or uses their data without them knowing.

🔴 Off limits: it’s on the banned list below.

Not sure? Choose the higher level.

🟢 Level 1: Low risk, so let it be

Examples: spam filters, movie picks, autocorrect, maps.
Protocols:

  • No special rules beyond normal privacy and consumer law.
  • Let people turn off personalization where it’s practical.
  • Collect only the data the feature needs.

🟡 Level 2: Use with care

Examples: chatbots, AI-written content, deepfakes and AI images or voices.
Protocols:

  • Mark AI chatbots, AI-generated images, video and voice clearly as AI.
  • Never use AI to impersonate a real person without that person’s written consent.
  • Have a human review AI-written content before it’s published or sent to a client.
  • Don’t put personal, financial or health information into public AI tools.
  • Give people an easy way to reach a human.

🟠 Level 3: High risk, so it needs real rules

Examples: AI that decides or recommends about a job, home, loan, insurance, healthcare or schooling.
Protocols:

  1. Human makes the final call. A named, trained person reviews and approves every decision, and can override the AI.
  2. Bias testing. Test for unfair results before launch and at least once a year, and keep the results on file.
  3. Notice. Tell people when AI is part of a decision about them.
  4. Explanation. Give the main reasons for a decision in plain language when asked.
  5. Appeal. Offer a way to challenge a decision, reviewed by a human, with an answer in a reasonable time (we propose 30 days).
  6. Data care. Use only the data the decision needs, protect it, and never sell it without clear permission.
  7. Records. Keep a log of AI-assisted decisions and human reviews.
  8. Accountability. Name a person responsible for the AI system.

🔴 Level 4: Off limits

  • Secretly scoring or ranking people, such as hidden social scores.
  • Mass facial recognition or tracking of people in public without legal authority.
  • Manipulating vulnerable people, including children, older adults and people in crisis.
  • Letting AI decide to use lethal force without human control.
  • Buying or selling people’s personal data in bulk without their knowledge and consent.

For businesses that adopt these standards

  • Keep a written AI policy that lists the tools used and the level of each one.
  • Train staff on it once a year.
  • Have a simple way to report AI mistakes and complaints, and fix them promptly.
  • Review the policy every year.

Our principle

AI isn’t the enemy. A human needs to stay in charge, and people’s privacy and safety come first.

Have feedback? Share it when you take the pledge on our homepage.

Take the free pledge