A Computer Cannot Be Held Accountable

There is one ethical boundary in my work that is absolute and non-negotiable.

I will not work for, contribute to, or support the development of autonomous AI systems that can make decisions to target, engage, injure, or kill a human being without meaningful human judgment, authorization, and accountability.

That includes autonomous weapons and targeting systems, and it can include guidance systems when the system itself is empowered to decide who or what becomes a target.

More broadly, I would prefer not to work for a weapons manufacturer. But autonomous lethal decision-making is the line I simply will not cross.

It is 2026, and I find it extraordinary that those of us working in technology have reached a point where we actually need to articulate this boundary.

We can now build machines capable of observing their environment, interpreting enormous amounts of information, identifying people and objects, recommending actions, and increasingly acting upon those recommendations themselves.

The technical question is rapidly becoming: Can we?

I think technologists have an obligation to keep asking another question:

Should we?

Authority is not the same thing as capability

My concern is not simply with weapons.

It is with delegating consequential human judgment to systems that cannot themselves accept responsibility for the consequences.

AI can be extraordinarily useful in assisting human decisions. It can identify patterns a person might miss. It can summarize thousands of documents. It can model possible outcomes, detect anomalies, make recommendations, and automate enormous amounts of routine work.

I build systems that do exactly those things.

But there is a difference between allowing a machine to inform a decision and allowing it to become the decision-maker.

That distinction matters to me.

An AI system I design may assist someone making decisions about hiring, compensation, performance, or other matters affecting another person. It may organize information and make recommendations.

But the consequential decision belongs to a human being.

The same principle applies with enormously greater stakes when the decision involves taking a human life.

IBM understood the problem almost half a century ago

A line from a 1979 IBM internal training manual expressed the problem remarkably well:

“A computer can never be held accountable. Therefore a computer must never make a management decision.”

IBM Training Manual, 1979

The computer in 1979 and the computer in 2026 are almost unrecognizable to one another.

The accountability problem is not.

An AI model cannot stand before a family and explain why their loved one died.

It cannot accept moral responsibility for firing an employee.

It cannot experience remorse.

It cannot be imprisoned for criminal negligence.

It cannot resign because its judgment caused harm.

And it cannot meaningfully accept responsibility for a decision, because responsibility ultimately belongs to the people and institutions that designed, deployed, authorized, and operated it.

Giving a machine greater intelligence does not give it greater accountability.

Asimov was asking the right question

I have always had an affinity for Isaac Asimov’s Three Laws of Robotics.

First Law: A robot may not injure a human being or allow them to come to harm through inaction.
Second Law: A robot must obey human orders, unless they conflict with the First Law.
Third Law: A robot must protect its own existence, provided this does not conflict with the First or Second Law.

– Isaac Asimov

They were fictional, of course, and Asimov spent a substantial part of his writing demonstrating just how complicated and contradictory apparently simple rules for machine ethics could become.

But the fundamental idea remains important:

When we give machines agency, protecting human beings must be a foundational constraint rather than an optimization target.

Modern AI safety obviously cannot be implemented by putting three sentences into a system prompt. Real systems require engineering controls, auditing, human authorization, governance, and clearly defined responsibility.

But the philosophy underneath those controls matters.

Hence the additional law:

Zeroth Law: A robot may not harm humanity, or allow humanity to come to harm through inaction

– Isaac Asimov

Human in the loop cannot become a euphemism

There is also a danger in saying that a system has “human oversight” when the human exists only to rubber-stamp decisions generated by a machine.

If a system presents an operator with a recommendation every few seconds and experience teaches that operator to simply click Approve, we have not meaningfully preserved human judgment.

We have created the appearance of it.

Meaningful human oversight requires that the person making the decision has sufficient information, sufficient time, actual authority to disagree with the system, and responsibility for the decision they ultimately make.

Otherwise, “human in the loop” becomes an interface feature rather than an ethical safeguard.

Some decisions should remain human

I am enthusiastic about AI.

I have spent years building intelligent systems precisely because I believe computation can dramatically expand what human beings are capable of doing.

But enthusiasm for a technology does not require believing that every possible application of that technology should exist.

There should be lines.

One of mine is simple:

But when a consequential decision must be made about another human being, accountability must remain human.

And when that decision is whether another human being lives or dies, I believe that principle must be absolute.

The machines have changed enormously since 1979.

The principle has not.

A computer still cannot be held accountable.

We should be extraordinarily careful about giving it decisions for which someone must be.

---

Historical note: IBM itself now identifies the quotation as coming from a 1979 IBM training manual. The surviving image’s archival provenance has also been investigated independently; IBM’s corporate archive was unable to locate its own copy of the original presentation.