On #robot judges: part 9.

In part 8,¹ we explored why an #AI #LLM can neither (a) make reasoned decisions; nor (b) give the reasons why certain output was generated in response to a prompt (notwithstanding appearances).

In this post, I'll address other propositions in the article from The Business Times,² in which the author argues there are "good reasons" to "embrace legally binding and enforceable AI-generated reasoned determinations".

---

Proposition 2.

The author suggests that using robot judges would lead to time and cost savings, as opposed to conventional dispute resolution mechanisms such as Court proceedings.

I agree.

But this not the be-all and end-all.

Suppose you and I enter into a contract for the sale and purchase of widgets. The contract stipulates that in the event of a dispute, we will flip a coin, and the winner will get their way.³

Cheap and fast. But fair?

I will go out on a limb to suggest that while speed and cost are important, they are secondary to procedural fairness.⁴ So while speed and cost may be a good reason to use robot judges, it does not mean that robot judges should be used in all cases.

---

Proposition 3.

The author suggests that we should not be worried about LLMs generating hallucinations or exhibiting bias and prejudice, because human judges also make errors, and have individual beliefs, experiences, and idiosyncrasies.

This is... a strange argument.

Indeed, human judges aren't perfect. And that's why (a) judges write grounds of decisions; (b) there are appeal and review mechanisms; and (c) we have recusal procedures.

I am not saying that these mechanisms are perfect. But their existence acknowledges the fallibility of human judges, and reflects society's attempt to mitigate these imperfections.

The author is, however, silent as to what can be done to mitigate the problem of hallucinations, bias, and prejudice on the part of LLMs.

Why switch from one imperfect system, which has checks and balances, to another imperfect system, for which the checks and balances remain unknown?

I am unconvinced that the downsides of robot judges have been sufficiently mitigated.

---

Proposition 4.

The author suggests that confidentiality is not a concern when using LLMs as robot judges, since conventional Court cases are generally a matter of public record.

I don't disagree. But this only indicates that a potential problem is not really a problem. It's not an argument as to the merits of robot judges, and is, at best, a neutral factor.

Also, what about arbitration proceedings, which are generally confidential? Until LLMs can address confidentiality concerns, parties who wish to keep proceedings confidential via arbitration would not opt for robot judges.

I don't see how this argument demonstrates the merits of robot judges.

---

In part 10, I'll address the author's final proposition, and proffer some concluding thoughts.