A common refrain I’m seeing in this post is that if there’s something wrong with the model you can just retrain it. There’s a couple problems with that assumption.

The state of the technology actually makes training a model somewhere between difficult or opaque. And what I mean by this is that in order to train a model you need to give it data. A lot of data. An amount of data that a single person frankly either doesn’t have access to or has no simple way to generate. And even then, there’s no way to be sure how the model performs until after the training completes, so even if you’ve collected all that data you won’t know it’s an improvement.

But for the sake of a hypothetical let’s ignore the current state of the technology and imagine that wasn’t a problem.

If an AI representative votes for me, and it gets that vote wrong, I won’t know about it until after it has voted for me. And by then it’s too late - I’ve already voted against my interest.

Also it seems that your position is that these AI reps are for people who care enough about politics to care, but don’t care enough to do. I don’t know that those people would ever confirm that their model is actually voting in their favour. If they don’t care enough to vote, then they don’t care enough to confirm their votes either.

The most damning thing about using AI for policy though - AI is NOT a decision making tool. Ask anybody who actually works on AI. It might fool the people who use it, and the people who sell it to you will tell you anything to make an extra dollar. AI is just a formula that spits out words instead of numbers. Sometimes it strings together a cohesive sentence and sometimes it hallucinates. There isn’t any Intelligence happening under the machine, it’s all Artificial.

AI is essentially autocomplete on steroids. It has no capacity to reason or argue, it just says what it’s trained for you to expect. It’s not a thinking machine and I sincerely doubt it ever will be