Skip to content
Kosplora

Routes What does an AI company's safety policy actually commit it to?

What does an AI company's safety policy actually commit it to?

understand + decide · 4h25m alone, 45m together · 8 stations · 4 sources

Careful reading. No legal or technical background needed.

kosplora-route/0.2 · by Bentley Moon

What would change this route

If the company you chose revised its policy while you were reading, the route would have you grade both versions and record what changed. If a law began requiring these policies to say specific things, the reading in station 3 would add that law's list.

The whole route, before you start

  1. 1 · orient · 30mRead METR's comparison of the companies' published safety policies.
  2. 2 · source · 1hChoose one of the twelve companies METR lists and read its current policy in full, from the company's own site.
  3. 3 · practice · 1hFor each shared element, copy the sentence in the policy that commits the company to it, or write "none found".
  4. 4 · source · 30mRead the rubric's seven criteria and their indicators, then grade the policy on each from A to F, citing the sentence each grade rests on.
  5. 5 · counterpoint · 30mRead the summary of an independent study that scored all twelve policies against risk-management practice from aviation and nuclear power, and found most commitments too vague to check.
  6. 6 · make · 30mMake a one-page table: each shared element, the policy's sentence or "none found", whether keeping it can be checked from outside, and your seven grades with their reasons.
  7. 7 · person · 15mGive the policy and the rubric, without your grades, to one other person and ask them to grade it.
  8. 8 · reflect · 10mRecord the one commitment you would most want made checkable, and how.

Optional study preferences. These stay in this browser and do not change the route’s steps.

Station 1 · orient · 30m

Read METR's comparison of the companies' published safety policies. Write down the elements it finds they share, such as capability thresholds, securing model weights, deployment safeguards and conditions for halting development.

Common Elements of Frontier AI Safety Policies (December 2025 update) · METR · 2025-12-09

Station 2 · source · 1h

Choose one of the twelve companies METR lists and read its current policy in full, from the company's own site. Note the version and date of the document you read.

Station 3 · practice · 1h

For each shared element, copy the sentence in the policy that commits the company to it, or write "none found". Beside each, say whether an outside reader could ever tell if the commitment had been kept.

Station 4 · source · 30m

Read the rubric's seven criteria and their indicators, then grade the policy on each from A to F, citing the sentence each grade rests on.

A Grading Rubric for AI Safety Frameworks · Jide Alaga, Jonas Schuett and Markus Anderljung · 2024-09-13

Station 5 · counterpoint · 30m

Read the summary of an independent study that scored all twelve policies against risk-management practice from aviation and nuclear power, and found most commitments too vague to check. Find its score for your company and write down where your grades and its score disagree.

Evaluating AI Providers' Frontier Safety Frameworks · Lily Stelling, Malcolm Murray, Bruno Galizzi, Max Schaffelder, Siméon Campos, Henry Papadatos and others · 2025-12-01

Station 6 · make · 30m

Make a one-page table: each shared element, the policy's sentence or "none found", whether keeping it can be checked from outside, and your seven grades with their reasons.

Station 7 · person · 15m

Give the policy and the rubric, without your grades, to one other person and ask them to grade it. Compare the two sets of grades and talk through the largest gap.

Station 8 · reflect · 10m

Record the one commitment you would most want made checkable, and how. If you want to go further, the free course below covers how frontier AI is governed.

Frontier AI Governance · BlueDot Impact · undated

What you leave with

A one-page reading of one AI company's safety policy: what it commits to, whether each commitment could be checked, and a grade on each of a published rubric's seven criteria, with a second person's grades beside yours.

The standard it is checked against

A Grading Rubric for AI Safety Frameworks

Jide Alaga, Jonas Schuett and Markus Anderljung · 2024-09-13

The policy is graded on each of the rubric's seven criteria, from A (gold standard) to F (substandard), and every grade cites the sentence in the policy it rests on.

Who checks it: You, criterion by criterion with the policy's own words, and then one other person who grades the same policy without seeing your grades.

Kosplora does not issue a certificate. A partner who recognises this standard would assess the report; no partner has agreed to yet.

When you are done