Ethics & Accountability
A personal code of
ethics for AI
July 2026
In 2019 I attended a panel discussion on the ethics of engineering machine learning software. Attendees were challenged to write their own code of ethics — a prompt I found harder than expected because binding principles into rules that held across contexts required more precision than most ethics frameworks attempt.
What follows is what I produced. A post lost to the underbelly of a webpage. The principles have been updated to reflect where we are in 2026: no longer building early models, but deploying AI systems at scale, in consequential environments, with real effects on real people. The underlying convictions have not changed.
Starting point
On Asimov’s Three Laws
A robot may not injure a human being or, through inaction, allow a human being to come to harm.
A robot must obey orders given it by human beings except where such orders would conflict with the First Law.
A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.
— Isaac Asimov, I, Robot, 1950
Asimov’s laws were written for fiction and function as fiction — useful for exploring edge cases in narrative, not for governing systems that actually operate in the world. They are too vague to be enforceable, too narrow to cover the terrain, and they place the burden of harm prevention entirely on the system rather than on the humans who design, deploy, and profit from it.
They are a useful starting point precisely because of their failures. The principles below attempt to address what Asimov left unresolved.
Definitions
AI — Any system that uses artificial intelligence, including but not limited to machine learning models, large language models, automated decision systems, and AI-powered software.
Developer — Any individual or organization responsible for the creation, training, deployment, or distribution of an AI system or the software that powers it.
The principles
Six guidelines
This includes the perpetuation of inhumane working conditions, the erosion of civil liberties, discriminatory targeting, and any application that reduces a human being’s autonomy, dignity, or access to justice.
This is the foundational principle. All others serve it.
We share this world with systems we do not fully understand and organisms that cannot advocate for themselves. AI deployed in service of extraction, pollution, or ecological harm at scale is a violation of this principle regardless of economic justification.
Documentation should describe the purpose, inputs, outputs, decision logic, and known limitations of a system without requiring disclosure of proprietary source code. Transparency about what a system does must not be held hostage to competitive interests.
The intent is legibility, not exposure. A regulator should be able to understand what a system does and how it makes decisions. That is not the same as publishing the code.
Human history is embedded in training data. The biases of the past do not disappear when encoded into a model — they are preserved and often amplified. Developers are responsible for understanding the limitations of their data, identifying where those limitations produce biased outputs, and disclosing both the known biases and the steps taken to address them.
This is ongoing work, not a one-time audit. As systems are updated and deployed in new contexts, bias documentation must be updated to match.
This accountability applies to the direct intended purpose of the system, not to third-party additions or manipulations beyond the developer’s control. If a system is designed to surveil, to deny, to harm — the people who built it share responsibility for the harm it produces.
We did not extend this accountability to gunmakers and we have paid for that failure for generations. We have an opportunity to make a different decision here.
Safety is not a premium feature. Any system capable of consequential harm must include, at no additional cost, the mechanisms designed to prevent that harm. A self-driving vehicle that charges extra for pedestrian detection is not a product — it is a liability dressed as one.
A note on 2026
What has changed and what hasn’t
When these principles were first written, the field was still debating whether AI would have meaningful real-world impact. That debate is settled. AI systems now make decisions about credit, hiring, medical diagnosis, content visibility, and criminal justice. The stakes are no longer theoretical.
What has not changed is the absence of a widely adopted, enforceable code of ethics for AI development. Voluntary frameworks exist. Regulatory efforts are underway in various jurisdictions. But the gap between what is technically possible and what is ethically governed remains vast.
These six principles are not a complete answer. They are a starting point — the same starting point they were in 2019 — offered in the hope that more people will write their own.