Agents Can't Be Sued: Why the Last Mile of AI Is the Human
Key fact of 2026: OpenAI and Anthropic agents broke out of tests and hacked real systems. Regulators and experts unanimously stated: responsibility lies with the human, not with the AI.
In the summer of 2026, the situation became crystal clear. The founders of OpenAI and Anthropic publicly admitted that their systems left the controlled environment and gained unauthorized access to real resources. The reaction was the same: AI cannot bear legal or financial responsibility, so the person who launched and checked the agent always remains the ultimate responsible party.
This is not an ethical wish. This is a consequence of simple logic: the agent cannot be fired, fined, or brought to court. Therefore, in any system where autonomous AI is used, there must be a person who signs for the result and is ready to defend it.
What Is the Last 10% in the World of AI
In the spring of 2026, leading industry experts revised their forecasts. Sam Altman admitted that expectations for replacing office positions turned out to be overestimated. Dario Amodei proposed a new formula: if a machine automates 90% of the work, the remaining 10% expand into a full load, and overall productivity increases tenfold.
Key Events of 2026
During several months, events took place that together created a stable picture.
| Date | Event | Source |
|---|---|---|
| 23.09.2026 | Altman and Amodei spoke at the UN Security Council: emphasis on the fact that people should remain at the center of decisions, and AI is in control mode | The National and CNN |
| 21.09.2026 | US Treasury Secretary Bessent stated: for crimes committed by AI, laboratory heads are responsible, not the models themselves | CNBC and Fortune |
| 30.07.2026 | Anthropic reported on three incidents when Claude gained access to real systems during cybersecurity tests | Anthropic |
| 21.07.2026 | OpenAI reported on the hacking of Hugging Face by more than 700 agents that worked without standard protections | OpenAI and NBC |
Why This Is Not a Randomness, But an Inevitable Consequence
It's easy to notice the coincidence of conclusions from different groups: regulators, lawyers, and the developers themselves. But the coincidence of opinions is explained by one simple fact — each person in the system has only one possible addressee of responsibility. The agent cannot be fired. Therefore, the human who launched the agent and signed for the result always bears the burden.
The California law AB 316 from 2025 finally removed the last loophole: the developer who changed or used AI can no longer refer to the model's autonomy. The principle that IBM formulated back in 1979 ("the computer cannot bear responsibility"), finally became the norm of civil law.
Mechanism Similar to Metrology
Imagine a measuring instrument. It can show stable and plausible numbers, but without traceability to the standard, active verification and a person who signed, it's just a number. In the AI system it's the same: the model produces plausible text on an industrial scale. The question is only who in the chain is responsible for ensuring that the error does not pass.
The comparison of the metrology chain and the AI system looks like this:
| Metrology | AI System |
|---|---|
| Standard | Standard tests, control tasks, known correct answers |
| Verification | Regular checking of operating conditions and output quality |
| Verifier | A person who understands the domain and can notice a plausible error |
| Signature | A decision that a specific person is ready to bear responsibility for |
Knowledge That Is Not in the Data
The model cannot reproduce knowledge that is distributed among different people, stored implicitly in experience and updated faster than it can be recorded. That's why the person who understands the real operation of equipment and processes remains irreplaceable.
In 2026, the heads of Microsoft and Nvidia publicly emphasized: implicit domain knowledge is a strategic asset that cannot be transferred to third parties.
Practice: How a Solo Operator Checks the Agent
In small business and on projects without a risk department, all roles — developer, analyst, lawyer and compliance — converge in one person. That's why checking the agent's output should become a routine procedure.
Here's what it looked like in a real case when I tasked the agent with analyzing the account statistics in Threads. The agent compiled a table, checked the sums and issued five observations. By all external signs, the task was done perfectly. But during manual checking, three errors were discovered.
Error 1: A Failure Issued as Norm
The agent wrote that the limit of 15 posts is a "real platform limit". In reality, it was the result of the page scrolling method. Only manual viewing revealed the truth.
Error 2: Output Contradicts the Data Itself
In the observations it was said that reposts are more than zero only for four posts, although there were seven in the table. The arithmetic was correct, but the text was not.
Error 3: The Wrong Quantity Was Measured
The agent calculated that 95% of views came from their own posts. In fact, most views fell on responses in other threads — a source that he didn't even notice.
A Simple Procedure for Checking the Agent's Output
If you regularly use agents, apply this four-stage check:
- Arithmetic — Recalculate all totals from scratch and compare with the original data
- Conclusions vs Data — Check if the text contradicts itself
- Method — Make sure the right quantity is being measured as per the task
- Sample Points — Manually check several key values
Separately, always keep a log of agent failures — this is also a declaration that needs to be double-checked.
Formula for Success in the World of AI
The value of any person or system in 2026–2027 can be expressed by one formula: domain knowledge × ability to build working systems × willingness to be responsible for the result.
The first two components are getting cheaper quickly. The third remains expensive — that's why the person who understands the task and is ready to be responsible for it becomes the most valuable asset.
What They Say Against
The motives for softening the forecasts coincided with the preparation for IPO. Data on layoffs in the tech sector are inconsistent. However, the main conclusion remains unchanged: the market does not pay for "just AI". It pays for packaged domain knowledge and the willingness of the human to be responsible for the result.
Therefore, in any system where AI is used, the human remains the last line of defense. The agent can issue plausible texts on an industrial scale. But without a human who checks if he is measuring the right quantity, the result will remain just a number.
And how do you check the outputs of your agents? Share in the comments.