AMPL
The AI Model Public License. AMPL is currently at v0.9 public draft. It has an OSI-compatible base layer that does for AI what MIT or Apache 2.0 do for software. On top, four optional modules handle the obligations that software licenses…
AMPL is the shared vocabulary that the readability layer speaks. It defines four dimensions (reciprocity, commercial, provenance, ethical) that any rights-expression can be read against. Creative Commons licenses, RSL declarations, publisher terms, and custodial attestations all map into the AMPL vocabulary, which is what lets them be reconciled into a single reading by the Translator. The legal text scaffolding co-authored with AI Commons contributors is the reference implementation of what an AMPL-native instrument looks like as a working demonstration of the vocabulary's expressiveness, rather than a license AI Commons is asking the field to adopt.
The foundation.
An OSI-compatible permissive license for AI models and datasets. At v0.9, in public draft.
AMPL-base does what MIT or Apache 2.0 does for software. It provides a clean copyright grant, attribution requirements, and standard warranty disclaimers, written for AI-era artifacts (models, weights, datasets) rather than just code.
For most projects releasing an AMPL artifact, the base alone is enough. A dataset that declares "this is cleared and free to use." A model that wants a permissive container without additional obligations. A research release that prioritizes accessibility over downstream constraints.
The obligations.
Four optional modules. Stack any combination on top of the base. Each is currently being drafted with input from inner-core legal counsel.
Each module is optional. A project that does not need reciprocity does not carry it. The modules are designed to be stacked.
Software licenses weren't built for AI.
MIT, Apache, and the other established open source licenses handle code well. They were designed for a world where the thing being licensed was source code, and where the obligations were attribution, warranty, and the freedom to fork. AI introduces new questions those licenses don't answer.
What was the training data, and was its use consented to? Can a rightsholder remove their work from a future training run? What happens when the model produces outputs derived from copyrighted material? What counts as "downstream" when a fine-tuned model is itself further trained?
AMPL doesn't replace existing open source licenses. The base layer is compatible with them. The optional modules add the obligations that AI raises, in a form that can be composed, reviewed by counsel, and adopted incrementally.
What has been tried, and where AMPL fits.
Several efforts have addressed parts of this gap. RAIL licenses introduced use-case restrictions to model releases. Community licenses from major AI labs have added training data restrictions. Proprietary terms of service have added their own constraints.
Each of these has done useful work. The constraint they share is that they extend existing license structures with new obligations, which fragments compatibility downstream. A model released under one community license cannot easily be combined with data released under another.
AMPL is being designed to compose. The base layer is OSI-compatible, so existing open source projects can adopt it without disruption. The modules are independent of each other, so a project can opt into the obligations that match its values without inheriting unrelated terms.
This design is being tested in legal review and refined based on external input.
What we're looking for.
AMPL is in active development. The base layer is in legal review. The modules are being refined. Compatibility with Creative Commons and the existing open source license stack is being formalized.
We are looking for co-builders. Particularly: legal counsel and policy experts with experience in non-US, non-EU intellectual property frameworks, civil law systems, and indigenous IP traditions. AMPL will work better the more jurisdictions it is built to work within.
Public comment on the v0.9 draft is open through coordination with the AI Commons team.