A common idea is gaining popularity (recently voiced by Andrey Doronichev): in the near future AI will do almost everything, and the human role will be reduced to setting goals and taking responsibility for the outcome.
The message is roughly this: “There will be no more linear employees. Every person will manage a department of at least 100 AI agents. Robots will perform the executor roles. The only role left is management: you set objectives and take responsibility for results.”
But if we look at this not as futurists but as lawyers, a key question arises: can a human realistically “take responsibility” for something they cannot fully control or even understand?
The Illusion of Control: When the Signature Exists but the Decision Doesn’t
Let’s use medicine as an analogy, although the same applies to legal practice.
Assume:
- a human doctor makes mistakes in 5% of cases
- an AI system makes mistakes in 1% of cases
The legal issue is not in the statistics, but in the mechanism of liability.
Society tolerates human errors because there is a clear accountable person: a doctor, engineer, or specialist.
But it is far less willing to accept even a smaller risk when the decision comes from a “black box” system that cannot be properly explained.
This creates a paradox:
society is willing to accept 5% of mistakes made by humans, but refuses to accept 1% made by an AI black box.
Two AI Verification Scenarios (Both Problematic)
1) The human checks everything manually
If a lawyer verifies every AI conclusion from scratch—facts, statutes, case law, argumentation—then the time spent is essentially the same as working without AI.
In that case, the technology loses its core value: no speed, no cost reduction, no efficiency.
2) The human checks selectively
This is the realistic scenario in practice: AI drafts a legal position, and the lawyer “reviews quickly” and signs off.
But then liability becomes a legal fiction.
Formally, the lawyer signed the document.
In reality, the decision was made by the AI.
And when the system makes an error (that inevitable 1%), we end up with an absurd situation: a human is held responsible for an outcome they did not consciously decide on, but merely approved based on trust in the system.
From a legal standpoint, this resembles nominal supervision rather than actual responsibility.
The Scale Problem: A “Department of 100 Agents” Is Not a Department
The issue becomes even more serious because modern AI agents can act autonomously.
One agent assigns a task to another.
The second calls a third.
The third retrieves data through a fourth.
As a result, the decision-making chain becomes:
- distributed
- hidden
- non-interpretable
- often impossible to fully reproduce
This creates a major legal conflict: humans are expected to be “managers,” but management presumes understanding how subordinates operate.
If you don’t understand how your “employees” make decisions, you are not a manager—you are simply a biological carrier of a legal signature, forced to legitimize an output produced by systems beyond your comprehension.
Hinton and the Core Issue: “We Don’t Program Behavior”
Geoffrey Hinton explains this bluntly: we do not program neural networks’ behavior directly.
We program the training method, and the internal structure emerges on its own.
Meaning the final reasoning logic:
- is not explicitly created by humans
- is shaped by data
- is not transparent or directly explainable
The key regulatory implication is obvious:
more intelligent systems cannot be fully controlled by less intelligent ones.
What Will Replace Responsibility? Most Likely — Insurance
If we accept the reality honestly, the “human is responsible for AI decisions” model does not hold up.
Because responsibility assumes:
- control
- understanding
- the ability to prevent harm
With autonomous agent systems, those conditions may simply not exist.
That is why the market may naturally shift toward another mechanism:
not liability, but insurance-based risk allocation.
Just like cars:
we do not require drivers to “guarantee no accidents,” but we insure potential damage.
In that future, clients may choose between:
- an expensive product verified by humans (“handmade legal”)
- a cheaper AI-based product with slightly higher risk, reflected in pricing
What This Means for Lawyers
In such a world, lawyers will not become “robot managers.”
They will become:
- risk managers
- process controllers
- architects of contractual liability allocation
- specialists in distributing losses
We will not “take responsibility.”
We will reassign and structure it among:
- developers
- operators
- model owners
- business clients
- end users
Can AI Become a Legal Subject?
Some propose recognizing AI as a legal subject.
But in practice, this is close to meaningless unless the entity can:
- own assets
- be punished effectively
- experience restrictions of rights
If AI has no property and no independent will, liability will still fall on humans and corporations around it.
The Key Collision of the Next Decade
The legal question of the future will not be:
“who signed the document?”
It will be:
“who should bear the risk of errors produced by systems no one can fully explain?”
And unless new regulatory regimes emerge, the final answer may be simple and unpleasant:
the consumer will pay.