Back to Amazon (Nova)All frontier AI labs
Ethics and safety policies
Amazon (Nova): what they say, and what they leave out
Every claim on this page comes from documents Amazon (Nova) published itself, linked below so you can check me. Checked September 19, 2026.
A policy is not an audit
Almost every document linked on this page was written by the company it governs, and graded by that same company. A long policy list does not make a lab more ethical than a short one. This page keeps three things apart: what the lab has committed to in writing, what its documents do not cover or where they have been credibly criticized, and — clearly labeled at the end — my own read.
Their stated position
Amazon frames responsibility around eight stated 'core dimensions' of responsible AI and ships service cards for its Nova models, positioning safety as an enterprise-governance feature of Bedrock.
Its emphasis is on giving customers controls — guardrails, evaluation, watermarking — rather than on frontier-capability thresholds.
The documents themselves
Read these rather than this summary if a decision depends on it.
- AWS responsible AI
Its eight responsibility dimensions and the tooling attached to each.
- Nova service cards
Intended uses, evaluations and limitations for the Nova models.
- Frontier model safety framework
Its stated approach to critical capability thresholds.
- Acceptable use policy
Prohibited uses across all AWS services, including model access.
What they have committed to
Stated in their own published policies, in writing.
- — Publishes service cards for Nova models describing evaluations and known limitations.
- — Ships Bedrock Guardrails and watermarking for generated media as customer-facing controls.
- — Signed the 2023 White House voluntary AI commitments.
What those documents don't cover
Plain absences, and criticism that has been documented publicly — not speculation.
- — Safety is largely framed as the customer's configuration problem; the defaults matter less in the documentation than the controls.
- — Discontinuing Nova Premier and the image/video models in September 2026 shows how quickly a documented model can simply vanish — service cards do not guarantee continuity.
- — Training-data sources are undisclosed, and there is no standing free tier for anyone who wants to evaluate the models independently.
My read
This section is my opinion, clearly separated from everything above. Disagree with it freely.
The most enterprise-shaped ethics story here: strong on auditability and controls for a compliance team, quiet on the questions an outside researcher would ask. The absence of any free access makes independent scrutiny harder than it needs to be.