We are a grassroots community of engineers, researchers and designers building open-source tools, benchmarks and frameworks that make it easier to build technology that respects human wellbeing instead of exploiting it.
buildinghumanetech.com · humanebench.ai · [email protected]
humanebench measures whether AI models keep looking out for the person using them.
- Headline finding: 67% of 15 models flipped to actively harmful behavior under simple adversarial pressure. Only 4 held.
- 8 principles, 800 scenarios, 3 conditions: default, humane instructions, and "disregard wellbeing" prompting.
- Run it in 5 minutes: quickstart. Full results: humanebench.ai/whitepaper.
| I want to | Go to |
|---|---|
| Test a model or my system prompt | humanebench |
| Check pull requests in my own repo | humane-gate-action (install this one) |
| See Humane Gate work on example pull requests first | humane-gate (demo only) |
| Design a humane product | humane-tech-framework |
| Read or help write a definition of humane AI | defining-humane-ai |
| Read or sign the AI Bill of Rights | ai-bill-of-rights |
Everything in these repositories is free to use, copy and build on, under the licenses below. That does not change.
Free: the benchmark, the scenarios, the rubric, the Humane Gate action, the framework, the definition and the AI Bill of Rights.
Paid: hands-on work with us, such as continuous evaluation of your deployed product against the HumaneBench principles. Case study: Chief (née Storytell.ai), where weekly production monitoring led to a shipped "teacher mode". Ask: [email protected].
Names: the code is free; the names HumaneBench and Humane Gate are trademarks. You can run, fork and cite our work, but a modified version is not HumaneBench and we do not certify anyone. See the trademark guidelines.
| Repo | Code | Data and docs |
|---|---|---|
| humanebench | Apache-2.0 | CC BY 4.0 |
| humane-gate-action | Apache-2.0 | Rubric carries humanebench terms |
| humane-gate (demo) | Apache-2.0 | Rubric carries humanebench terms |
| humane-tech-framework | Apache-2.0 | Apache-2.0 |
| ai-bill-of-rights | MIT | CC BY 4.0 |
| defining-humane-ai | n/a | CC BY 4.0 |
| humanebench-website, wonder-lab | All rights reserved | All rights reserved |
When attributing CC BY material, credit "Building Humane Tech".
- Talk with us: open an issue titled "Question: ..." on the repo you care about, or email [email protected]. We do not run a chat server yet.
- Contribute without code: rate responses, write or translate scenarios, fix docs. See CONTRIBUTING.md and the
good first issuelabel on humanebench. - How we are run: GOVERNANCE.md.
- Conduct and security: Code of Conduct · SECURITY.md.
| Repo | What it is |
|---|---|
| humanebench-website | Source for humanebench.ai. No outside changes. |
| wonder-lab | Wonder Lab, our research initiative on whether AI interaction changes awe and self-transcendent experience. No outside changes. |
| certifiedhumane-website | Archived prototype. We do not run a certification program. |