Skip to content

Ethics & Privacy

Values & Commitments

Equality, in hiring and in who we work with

FIRL hires on merit, without regard to race, gender, or similar protected characteristics. The same standard applies to who we work with: we do not supply Colophon, or any FIRL research output, to militaries or to organizations engaged in genocide or serious human rights abuses.

Precedent for this position

BigScience, the volunteer research collective behind the BLOOM language model, distributes it under the RAIL license, which legally bars military and weapons use and propagates that restriction to derivative models11 BigScience RAIL License (OpenRAIL-M). See the BigScience OpenRAIL-M explainer, licenses.ai, 2022, and the Hugging Face OpenRAIL overview.. It is the closest direct precedent for a research lab, not a large corporation, taking this position on an AI product.

The Future of Life Institute's 2018 Lethal Autonomous Weapons Pledge is a second, wider precedent: roughly 160 organizations and 2,400 individuals across 90 countries committed to neither developing nor supporting lethal autonomous weapons22 Future of Life Institute, Lethal Autonomous Weapons Pledge, 2018.. The Hippocratic License, published by the Organization for Ethical Source, is a third: an open-source license variant built around a human-rights-based use restriction33 Organization for Ethical Source, Hippocratic License 3.0, firstdonoharm.dev. Described here in general terms; we haven't quoted its exact clause text and recommend reading the license directly before relying on our summary.. We haven't yet adopted any of these licenses ourselves; each is precedent for an established position in open-source and AI research.

Ethically sourced AI models

Where FIRL's own work involves choosing a language model, we favor ones with real, checkable provenance over ones that only claim it. Three examples worth naming:

  • OLMo (Allen Institute for AI): releases its full training corpus, training code, and intermediate checkpoints, not just final weights, alongside its own power-consumption reporting.
  • BLOOM (BigScience): trained on the ROOTS corpus, a multilingual dataset with source documentation reviewed per language rather than blind web scraping.
  • Common Pile (EleutherAI): trained only on public-domain and openly-licensed text, legally reviewed before release.

Hugging Face's AI Energy Score rates models on energy efficiency, similar to an appliance label. It's a real, useful initiative, and industry-wide disclosure under it is still the exception rather than the norm44 Hugging Face, AI Energy Score, huggingface.github.io/AIEnergyScore. Of over 2 million model repositories on the Hugging Face Hub, only a small fraction disclose emissions data., so we cite it as a direction, not a standard every model already meets.

Looking ahead, Mozilla's llamafile project, a genuine Mozilla Builders initiative now maintained under Mozilla.ai, packages a model and its runtime into a single portable file55 Mozilla Builders, llamafile project page, builders.mozilla.org; now maintained at github.com/mozilla-ai/llamafile.. We'd like to eventually package some of these ethically-sourced models the same way, so they're easy to run locally without a cloud dependency. That's a future step, not something we offer today.

Where we fall short of this today

FIRL currently hosts Colophon on Google Cloud. That sits in tension with the stance above, since Google holds military contracts. We're a small, early-stage team, and moving hosting providers is a real undertaking; we're evaluating alternatives that better match this commitment, and we'd rather say that plainly than leave it unsaid.

Sources

What Colophon does not do Ethics & Privacy FAQ