Responsibility
AI principles
These principles apply to every model the foundation builds and every product OneThought AI builds on that research. Each one says what we do and how you can check it.
01
Evidence and traceability
Outputs point to their sources.
Model outputs cite the data, documents or studies they rest on. Every number we publish has an evaluation behind it, and we link it.
How to check
Model cards, published evaluation protocols, citations in outputs.
02
Honest uncertainty
Models say when they do not know.
Models report confidence, flag contradictions in the evidence and abstain rather than guess. We publish where each model fails, not only where it works.
How to check
Calibration and abstention results in each model card; the known-limits section.
03
Human oversight
Experts decide; AI supports them.
In health and other high-stakes uses, our systems support qualified people and do not replace them. We make no diagnostic or treatment claims.
How to check
Intended-use and out-of-scope sections of each model card and product page.
04
Safety and misuse prevention
Tested before release, with uses we refuse.
Every release is tested for misuse before it ships. Generated media carries provenance by default. Each product has a usage policy that names what it will not be used for.
How to check
Red-team summaries, safeguards and usage policies, the incident log.
05
Privacy and consent
Licensed data, minimal personal data.
We train only on data we have the right to use and name every source. We collect as little personal data as the work needs. We do not generate the likeness of a real person without consent, and our tools do not monitor private individuals.
How to check
The data statement in each release, product privacy notes.
06
Fairness and inclusion
Built for India's languages and knowledge.
We work in Indian languages and with knowledge systems that general models under-serve, and we report how performance differs across languages and groups.
How to check
Language coverage and group-level results in evaluations; the access programmes.
07
Accountability
Named owners, public corrections.
Each model and policy has a named owner. Anyone can report a concern. We keep a public log of incidents and corrections, and version every policy.
How to check
Report a concern, the incident and change log, policy version history.
How to read these
Each principle has two parts: what we commit to, and how you can check that we did it. If you cannot check a principle from what we publish, we have not kept it, and we want to hear about it. Report a concern.
When principles conflict
They sometimes do. Publishing an evaluation set helps others check our work, and can also help someone tune a model to pass it. Releasing weights helps research, and can help misuse. When principles pull in different directions we decide case by case, write down the reason, and publish it with the release.
Frameworks
The principles are written to align with these frameworks. Alignment is our own assessment; we will say so if and when a framework is independently audited.
- India AI Governance Guidelines (MeitY, IndiaAI Mission)
- Digital Personal Data Protection Act, 2023
- OECD AI Principles
- NIST AI Risk Management Framework
Who is responsible
The board of the foundation approves these principles. Each model and product has a named owner who answers for it. An advisory group of practitioners and specialists reviews high-stakes releases before they ship.
Version 1.0 · 23 October 2026 · First published version.