OpenAI’s Mission, Charter, and Deployment Philosophy
OpenAI's mission is to ensure that artificial general intelligence benefits all of humanity. For an FDE, this means considering the benefits, risks, and permitted uses of a proposed customer feature. Cedar's proposal to analyze employee messages and identify individuals at risk of burnout raises concerns about workplace surveillance, sensitive personal inferences, employee privacy, and the reliability of those assessments. OpenAI's iterative deployment approach emphasizes learning from testing and real-world use while refining safeguards. However, a feature must satisfy applicable usage policies and authorization requirements before deployment. When a request cannot proceed, an FDE should explain the relevant policy restrictions, privacy risks, and authorization requirements, then identify a safer alternative that addresses the customer's underlying need.
We'll cover the following...
- What do OpenAI’s mission and Charter commit it to?
- What does iterative deployment mean inside one engagement?
- What does the Model Spec govern?
- What do the usage policies govern?
- How do the sources apply to Cedar’s request?
- What does a safe path look like for the HR assistant?
- How do we say no without losing the customer?
- Try it yourself
- What’s next?
Cedar’s sponsor sent the HR desk request on a Tuesday:
“While we’re expanding to HR, can the assistant read employee messages, flag anyone who seems to be burning out, and send each manager a weekly list of names with a risk score? I tried it in the playground and the model does it well. I also read through OpenAI’s policies and didn’t see anything that stops us.”
Every part of that message sounds reasonable to the person who wrote it. The model can produce the scores. The sponsor did look at the policies. Managers would like to know who is struggling. And yet this is a feature an OpenAI FDE should decline.
Interviewers ask versions of “What would you refuse to build for a customer, and why?” because FDEs make this call in the field, usually with a sponsor waiting for a yes. A strong answer rests on four public sources from OpenAI and two from the customer’s side. Each source answers a different question, and blending them is the fastest way to give a weak answer.
What do OpenAI’s mission and Charter commit it to?
OpenAI’s mission is to ensure that artificial general intelligence benefits all of humanity. The Charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work, and it turns the mission into four principles.
Principle | What the Charter Commits To |
Broadly distributed benefits | Use any influence over AGI’s deployment for the benefit of all, and avoid uses that harm humanity or unduly concentrate power. The Charter calls humanity OpenAI’s primary fiduciary duty. |
Long-term safety | Do the research needed to make AGI safe and drive its adoption. If a value-aligned, safety-conscious project comes close to AGI first, stop competing and start assisting it. |
Technical leadership | Stay at the frontier of capabilities, because policy and safety advocacy alone would not be enough. |
Cooperative orientation | Work with research and policy institutions, and provide public goods that help society navigate the path to AGI. |
These principles guide OpenAI's approach to AI development, but they do not replace the policies and requirements that apply to individual deployments. Referring to them alone is not enough to justify a decision to a customer. An FDE can use them to evaluate a proposed feature from three perspectives:
Benefit: Who benefits from the feature, and does it address the needs of the people affected by it?
Risk: Who could be harmed by incorrect model outputs or unauthorized disclosure of sensitive information?
Agency: Do the people affected understand how their data will be used, and what control do they have over that use?
Applying these questions to Cedar's proposal reveals several concerns. Managers would receive information intended to identify employees who may need support, but employees could face privacy risks or inappropriate decisions based on inaccurate inferences. The proposal also does not establish whether employees have been informed about the analysis or have any control over how their information is used. ...