# Bevy AI Policy (Draft)

:::warning
This draft has been reviewed by Cart, and is in the process of moving into the offical contributors guide.
:::

## Motivation
In recent times, there have been a growing number of contributions that are fully or
partially produced by generative AI (e.g. LLMs and friends) which exhibit
characteristics that cause a lot of burden on other contributors and maintainers.
While we've seen PRs and issues with burdensome characteristics produced entirely by
humans, generative AI tools have significantly lowered the level of effort required to produce
"plausibly-worthwhile" contributions that are otherwise entirely unmergable or
incorrectly report bugs, and so have become a major source of burdensome PRs and
issues.

Whether AI generated code are copyrightable works is also a hot-button topic that is still
being openly debated and litigated. How this impacts the legal aspects of maintaining
a FOSS project is currently a unresolved question.

This policy is thus drafted as a response targeted at the problem of an
increasing frequency of burdensome PRs/issues and addressing the potential legal issues
surrounding the intersection of AI generated code and the FOSS contribution model.

The [Rust Project][rust-project-policy] is also looking at establishing a similar
policy.

[rust-project-policy]: https://github.com/rust-lang/compiler-team/issues/893

## AI Generated Communications
The unsolicited use of automated systems to communicate issues, bugs, or security vulnerabilites
about Bevy Organization projects under the guise of a human is considered unacceptable
and a Code of Conduct violation. Any individual contributor, operator of automated systems,
or company they may represent, may be barred from future contributions and banned from regular
communication channels if these communications were found to be submitted in bad faith.

This policy applies to all regular channels of communication used by members of the
Bevy Organization, including but not limited to GitHub Issues, GitHub Pull Requests, Discord,
other social media platforms, etc.

We recognize that English may not be the primarily language for all contributors and that
machine translation is an indispensable tool for proper collaboration, and thus not subject to
the above policy. However, we recommended using non-LLM machine translation
options when possible, as they tend to be less verbose while still getting the point across.

## Copyrightable AI Generated Contributions
At the current time of writing (August 11th, 2025), the US Copyright Office has
[stated publicly][us-copyright-office-response] that "human authorship is a
pre-requisite to copyright protection". A [more recent report][us-copyright-office-report]
from the same institution shows a much more contested legal space, both within the US and
internationally. In the case that AI generated works are not copyrightable, those same works
cannot be licensed to others, and thus AI-generated code and assets would not legally be
licensed under MIT or Apache.

Erring on the side of caution in light of a openly debated legal topic,
all[^1] forms of AI-generated contributions cannot be merged into repos maintained
by the Bevy Organziation. This includes both code and non-code game assets (e.g. textures,
audio, etc).

Any triage team member suspecting a pull request to be made primarily through the use of
LLMs or other generative tools should mark the PR as `S-Nominated-to-Close`, upon which a
maintainer can then review the PR for closure. To help identify these cases,
pull requests subject to this policy have characteristics such as (but not limited to):

* Needlessly or overly verbose descriptions or responses.
* Not internally coherent or even self-contradictory.
* Demonstrates misunderstanding of important aspects of what the code is doing
or the purpose of the change.

Any contributor, operator of automated systems, or company they may represent found to
have repeatedly submitted contributions with majority AI-generated code or assets may be
subject to:

* Blanket rejection of all future contributions to Bevy Organization projects.
* Retroactive removal of any potentially suspect AI-generated code and asset contributions.
* Further Code of Conduct actions if these contributions were found to be submitted in bad faith.

This policy may be revisited when the legal debate has settled.

[us-copyright-office-response]: https://www.copyright.gov/rulings-filings/review-board/docs/a-recent-entrance-to-paradise.pdf
[us-copyright-office-report]: https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf

[^1]: Trivial LLM generated content such as variable renames or autocompleted function calls, often branded "predictions" or "suggestions", that is indistinguishable from traditional methods such as a regex search/replace or an LSP autocompletion is by definition not detectable and can be treated like other regular IDE tools such as Intellisense.
This is does not include cases where the prediction generates things like entire function blocks.