Under human control: this is Principle 5 from the CEPEJ Charter. When we wrote it, we were laughed at. How can you have human control over something like AI? Even the developers don’t know what the AI is going to do. Eight years on, the consensus is that human control is the norm. But how? We now need to figure out how to shape this human control. International organizations develop principles and standards, the EU legislated the AI Act, courts have guidance, they provide training. All good, or maybe not.
Recently, I was part of a panel on AI in the Courtroom at the International Bar Association Annual Conference. It provided lots of food for thought, but also more material for my diatribe on how we talk about AI. Even if we do not consider AI to be human-like, we tend not to specify what we mean when we say AI.
The panel members all had very different angles to what they had to say about AI. For the lawyers, a big problem right now is that clients who have unspecified forms of AI put together witness statements and even expert reports. Is a witness statement made by AI still a statement from that human witness? And is the expertise presented in the report the expertise of AI, or of the expert? Courts in the People’s Republic of China use Ai to triage incoming cases and do various other tasks that can be regarded as ancillary to the judicial function. This is not just any unspecified AI, but AI developed based on policy choices and trained to consistently follow the policy. But what are “ancillary tasks”? The EU AI Act’s says its high-risk qualification for justice systems should not extend to AI systems intended for purely ancillary administrative activities that do not affect the actual administration of justice in individual cases, such as anonymization or pseudonymization of judicial decisions, documents or data, communication between personnel, administrative tasks. But the general agreement is clearly that the final judgment is with the human judges.
An anecdote: someone, let me call them Bee, recently told me this story. Bee had submitted their profile to be published on a web site. Publishing the profile was essential for the next step in Bee’s career. Once the profile was up, Bee checked whether it was correct. To their horror, the text was quite different from the submitted one. Of course, Bee asked what had happened to their profile, and requested that the original profile was put in place of the rewritten one. The web site management told Bee AI had made the changes. It works like this: If something is uploaded to the web site, AI automatically makes adjustments and changes. That is how uploading to this web site works. Fortunately for Bee, it was possible to upload the original text without interference from AI. Clearly, there was no human control involved, and the web site management felt that was OK.
