Skip to content

ETHICS.md

Latest commit

 

History

History
162 lines (121 loc) · 5.96 KB

ETHICS.md

File metadata and controls

162 lines (121 loc) · 5.96 KB

PAI Ethics

PAI is a tool. Tools are not neutral. This document states what we will build, what we will not, and how we think about the line between them.

See also: SECURITY.md · docs/src/PRIVACY.md · CODE_OF_CONDUCT.md

Our position

PAI exists because people have a right to compute privately. A bootable USB that runs a capable AI model without network dependency, telemetry, or account login restores a default that commercial AI has removed — the ability to use software without being logged, fingerprinted, or monetised.

We believe this default is worth defending for everyone: journalists, dissidents, researchers, patients, lawyers, students, and ordinary people who simply want their private thoughts to remain private.

Privacy is not a shield for wrongdoing. It is a precondition for autonomy, dignity, and free thought. PAI is built on that premise.

What PAI is designed for

  • Private research, writing, and thinking — using AI without a third party recording your prompts or a subscription keeping a log.
  • Journalism and source protection — drafting and analysis on a machine that leaves no trace on the host.
  • Operating under surveillance or censorship — for people who cannot use cloud services safely in their jurisdiction.
  • Personal data that should never leave your device — medical, legal, financial, or emotional material that cloud AI would ingest.
  • Offline environments — field research, travel, air-gapped labs, classrooms without reliable connectivity.
  • Learning and experimentation — running, auditing, and modifying AI without platform gatekeepers.

What PAI is not designed for

PAI is general-purpose software under GPLv3. We cannot and will not police what users do with it. But we can be clear about what the project is for, and what we will not help build.

We will not build or merge features intended for:

  • Targeted harassment, stalking, or non-consensual surveillance of any person.
  • Generation or facilitation of child sexual abuse material (CSAM). We will remove and report any such use of project infrastructure.
  • Non-consensual intimate imagery (deepfake pornography, undressing apps, etc.).
  • Large-scale disinformation operations or automated impersonation campaigns intended to deceive at scale.
  • Coordinated fraud, scam automation, or phishing kits.
  • Credential stuffing, botnets, or other infrastructure for unauthorised access to systems the user does not own.
  • Weapons design — chemical, biological, radiological, nuclear, or autonomous lethal systems.

This is not a list of things the law forbids. It is a list of things the maintainers will refuse to accept contributions for, and will remove if they appear.

Dual-use honesty

Most privacy tools are dual-use. Encryption protects both dissidents and criminals. So does a live-USB operating system. So do local language models.

We accept this trade-off consciously. The question we ask is not "can this be misused?" — everything can — but "does the benefit to legitimate users outweigh the marginal uplift to bad actors, given that equivalent tools already exist?"

For a private AI USB, the answer is clearly yes. Cloud LLMs are freely available to anyone with a credit card and a fake email. PAI does not make bad actors meaningfully more capable; it makes private, legitimate use possible for people who currently have no option.

Models and weights

PAI ships with, and makes it easy to run, open-weight models from third parties (currently the Ollama library). We do not vet every weight. Users should be aware:

  • Model outputs can be wrong, biased, or harmful. A local model is no more trustworthy than a cloud model by virtue of being local.
  • Fine-tuned models published by third parties may encode values or capabilities their authors did not disclose.
  • PAI provides a runtime, not an endorsement. The choice of model is the user's.

We will not ship, and will not make it easier to ship, models whose explicit purpose is to generate content in the forbidden categories above.

User responsibility

PAI gives you a capable, offline AI on hardware you control. With that comes responsibility:

  • You are accountable for what you produce. "The AI wrote it" is not a defence.
  • Local ≠ safe. Sensitive material on a lost or stolen USB is at risk. Use persistence encryption and strong passphrases.
  • Anonymity ≠ invisibility. PAI raises the cost of casual surveillance; it does not make a determined adversary powerless.
  • Your jurisdiction applies. Some of what PAI makes possible is illegal in some countries. We do not advise on that; consult a local lawyer.

Maintainer commitments

The PAI maintainers commit to:

  1. Not adding telemetry, analytics, phone-home, or "anonymous usage stats" — ever, regardless of pressure from funders, partners, or governments.
  2. Not accepting contributions that backdoor, weaken, or circumvent the privacy guarantees documented in SECURITY.md.
  3. Publishing transparency reports if we ever receive a legal request that would compromise users, to the extent allowed by law.
  4. Staying free and open — PAI will remain GPLv3, with no "enterprise edition" that extracts privacy features behind a paywall.
  5. Documenting the threat model honestly, including where PAI fails. See SECURITY.md § Known weaknesses.

Enforcement

If you believe a PAI contribution, release, or maintainer action violates this document, raise it:

  • Community issue: via CODE_OF_CONDUCT.md enforcement channels.
  • Security issue: via SECURITY.md reporting.
  • Ethics issue: email the maintainers at contact@pai.direct.

We will respond, investigate, and publish a resolution.

Last reviewed: 2026-04-20. This document evolves; pull requests welcome.