Responsible AI
· 5 min czytania

Can AI be responsible? The three conditions behind a promise

Autor: Amine Gaaliche · szkic przygotowany z pomocą Claude'a, sprawdzony i poprawiony ręcznie

No. To be responsible is to make a promise that other people can rely on. That trust rests on three conditions: you are competent, you have something to lose, and you cannot escape the consequences. An AI can meet the first. It cannot meet the other two, however capable it becomes.

Responsibility is a promise

When you take responsibility for a piece of work, you make a promise. You tell other people that the work is right, and that if it is not, you will answer for it. A lawyer who signs a brief makes that promise. So does an accountant who certifies a set of accounts, or an engineer who stamps a drawing.

A promise is useful only if people believe it. A client relies on the signed brief because they trust the lawyer to keep the promise. So the real question is what makes that trust reasonable. Three conditions do, and all three have to hold.

The three conditions

Competence. You must be able to do what you promise. That includes being informed. A skilled tax adviser who has not read the client's file cannot promise that the return is right.

Something to lose. You must have something you would forfeit if the promise fails: a licence, money, clients, a reputation built over years. The stake is what makes keeping the promise worth the effort.

No escape from the consequences. A court, a professional body or a regulator must be able to take that stake when you fail. You must be identifiable, and you must not be able to walk away. This reaches beyond criminal law. Being struck off is not a criminal sentence, and it still costs a lawyer their livelihood.

Remove one and the promise fails

Competence makes it likely that the work is right. The stake makes failure costly. The third condition makes sure the cost is actually paid. Take any one away and the other two cannot carry the promise.

Lawyers call someone with nothing to lose judgment-proof. You can win your case against them and still collect nothing. They may be skilled and easy to find. Their promise gives you no protection.

Now picture someone wealthy and skilled who lives where no court can reach them. They have plenty to lose, but nobody can take it. Their promise is worth no more.

Finally, picture a well-meaning amateur with savings at risk and a known address. You could not rely on their work in the first place, and suing them afterwards does not undo a bad filing.

Notice what the list leaves out: good intentions. A stake and the certainty of paying it make keeping the promise the sensible choice, whatever the person feels. That is why strangers can trust each other at all.

Where AI stands

An AI model can meet the first condition, and often does. You can even give it the file.

It fails the second. A model owns nothing, so it has nothing to forfeit. Switching it off or retraining it costs the model nothing it values. If a company posts a bond to cover its model's mistakes, the money at risk is the company's.

It fails the third. You cannot sue a model, fine it or strike it off. When an AI tool gets something wrong, the claim goes to a person or a company: the one who sold it, or the one who used it.

Better models will not change this. More capability improves the first condition. It does nothing for the other two.

Competence is only one condition

People who say an AI can be responsible have usually reduced responsibility to competence. They see output that is usually right and conclude the system can carry the promise. Being usually right is what we ask of a tool. Responsibility is what we ask of the person who uses it.

The same test explains why a company can be responsible when a model cannot. A company has no body either. It does have assets to lose, courts can reach them, and people inside it answer for its decisions.

What this asks of the person who signs

If the AI cannot hold the promise, a person has to. A professional who signs AI-assisted work brings the stake and can be held to account. The second and third conditions are covered.

The first is where it breaks down in practice. A professional who signs AI output without reading it may be very skilled. They are still not competent for this document, because they do not know what is in it. They are making a promise they have no basis for. Reviewing the work is how they meet the first condition.

Anonymisation shows this clearly. When detection misses a name, nothing breaks and no error appears. Only someone who looked can vouch that the file is clean. In promptShield, detection proposes and a person confirms each region before the file is produced. By default, a file with marked items cannot be downloaded until review has verified them. We explain what that review looks at in human-in-the-loop anonymization.

Common questions

Can an AI be held legally responsible?

No. Courts and regulators deal with the people and companies that build, sell and use AI systems. A model cannot be sued, fined or struck off, and it owns nothing a judgment could take.

Isn't the AI vendor responsible?

Yes, for what its contract with you promises. A company has assets and can be sued, so it can hold that promise. Your client, though, relies on your name, and the vendor's contract does not change that.

What would it take for an AI to be responsible?

A stake of its own, and a way for others to take it. Giving a model legal status would not supply either. Any money at risk would come from the people behind it, so they would still be the ones answering.

Czy ten tekst był przydatny? Prosimy dać nam znać.

Udostępnij