I wrote this after a long summer break, and it captures a good part of what I believe and strive for. This blog is my AI engineering learning scratchpad, so it felt like the right first post: a way for you to get to know me, and a starting point for tracking, in public, our progress toward AI superintelligence that benefits humanity.

The Case for AI Stewardship

Beneficial in purpose. Trustworthy in operation. Governable by humans.

From early August through mid-September, I spent six weeks largely unplugged from the web. When I returned, my news agent caught me up on the “pacing the frontier towards safe AI” saga. I loved that the safety topic got so much attention, but I was disheartened by how divided the mainstream reactions to the proposal were.

That reception echoed the fear, uncertainty, and doubt I heard from non-technical friends and family during my time away, across Mexico, Denmark, and Iceland. I spent a lot of that time discussing “safe AI” with them. My first lesson was to stop framing the conversation with the words “safe AI.” That phrase can encompass privacy, misinformation, discrimination, cybersecurity, economic disruption, and catastrophic loss of control. Without explanation, people hear the same words and enter very different conversations.

I think that, as advocates, we must frame AI differently. The framework that worked for me this summer, and one that summarizes my current mindset, is AI Stewardship.

A steward accepts responsibility for something consequential and an obligation to those affected by how it is used. Applied to AI, stewardship means pursuing meaningful human benefit while understanding and managing risk, acknowledging uncertainty, and remaining answerable for the result.

The responsibility of AI stewardship belongs to the people, corporations, and institutions developing, deploying, or overseeing AI. It continues as systems change and as their consequences become clearer. It includes (at least) three core commitments:

Beneficial in purpose. Advanced AI has the potential to improve people’s lives through scientific discoveries and through gains in knowledge, productivity, creativity, and automation. Beyond expanding what we can do, it should also make those capabilities more affordable.

Our shared responsibility, as stewards, is to ensure that this potential is realized and that it benefits as many people as possible. If you build AI, maximize its capabilities while keeping it safe. If you host or deploy it, point it at problems that matter and widen access to it. If you are a potential user, embrace it, learn how to use it responsibly, and help us deliver on its potential.
Trustworthy in operation. For any system to be trustworthy, it needs to demonstrate reliability, robustness, security, transparency, and observability. All these attributes must be backed by evaluation and continuing scrutiny; trust is earned with evidence, not assurances. The evidence required for each system should be proportional to its capabilities, what it is authorized to do, and the consequences of getting it wrong.

The “system” here is a very broad continuum. It can be an individual consuming AI through an assistant offered by a frontier lab, a cloud provider hosting a model and exposing it to enterprises, or an enterprise hosting models in-house. Every participant on that continuum has a responsibility. The frontier labs creating the models need to align them and contain them. The hyperscalers and enterprises also need to secure them, govern them, and contain (or sandbox) them.

Reading that the frontier labs are committing to giving independent evaluators access to their upcoming models is encouraging. That is the kind of evidence trust is built on. There is a lot more that the labs need to do. Safety research has not progressed at the same speed as capabilities. It is time to tilt the research scales back towards mechanistic interpretability, unlearning, scalable oversight, and many other techniques that help the labs understand and control what the models are doing.
Governable by humans. AI must remain subject to human authority. People must be able to set its objectives and boundaries, oversee its actions, intervene, correct its behavior, and stop its operation when necessary. People need not approve every individual action, but they must remain able to revise the boundaries within which AI acts and withdraw its delegated authority. The people exercising that authority must themselves remain accountable, and they need the model host to be accountable too and to offer appropriate governance.

The term “governable” here includes more than a legal or standards compliance metric. A system can satisfy documented requirements while remaining difficult to redirect or stop. We need the ability to exercise human authority at a moment’s notice. Builders should design the oversight and stop mechanisms into the system from the start. Companies and hosts need tested controls and the willingness to use them. Everyday users need to stay in the loop: set the boundaries and remember that delegating a task does not delegate the responsibility.

Governments must be stewards too. They are how a society exercises its authority, and no company can stand in for that. Treating AI mainly as a race to be won quickly is a trap: speed without safety adds risk, and the lead worth having is the one the world can trust. There is a lot governments can do today. They can explain in plain language which laws and policies already apply to AI, to the public as well as to the labs. Policymakers must become better educated about AI and consult third-party experts. We must hold every widely-used model to the same independent safety and alignment tests. I am all for importing models, provided they pass the safety tests. Finally, governments should open the conversation about what abundant intelligence and automation will mean for society in the long term. Today those questions are largely unanswered, and the unknown is what leads many people to expect the worst.

Let’s all do our part

We do not need to wait for a perfect policy or a single agreed definition of “safe.” Stewardship starts with whatever part of AI is in your hands today. If you build it, show your evidence. If you host or deploy it, secure it, govern it, and stay answerable for it. If you govern it, set clear rules, test independently, and keep the conversation open. If you use it, be curious, be responsible, and bring someone along with you.

My conversations from this summer told me that people care deeply about getting this right. AI must ultimately be a force for good, amplifying human creativity, agency, and productivity; we can all shape that future by becoming stewards today.