Description du poste
I followed my own instructions for exploring
Massachusetts 2026 ballot question arguments. I figured out how I want to vote.
I also figured out some improvements to make to
HowTruthful and
its associated agent skills.
• Since users are rating statements' truthfulness on a 1-5 scale, it's not very
useful to have a hedged statement like "Massachusetts legislators who run
unopposed can take voters for granted and ignore important issues." If the
rest of the statement is even a possibility, the word "can" makes the
statement as a whole true. The AI skill should change such statements into a
definitive form like "Massachusetts legislators who run unopposed will
take voters for granted and ignore important issues." Then the user can rate
that statement false, dubious, debatable, plausible, or true.
• There was a contest where I was among
the winners, even though I
disagreed with the organizers'
vision of having AI assess
truthfulness. In my contest entry I opined that humans should do the final
assessment; AI's role would stop at ingesting and structuring arguments.
(More on the contest in a future post.) Having just gone through 400 or so
statements in order to figure out how to vote, my position has
shifted. I think it would make sense, in the course of a human's dialogue
with an AI, for the AI to offer to put a set of truthfulness ratings onto a
set of statements in response to that dialogue. A human could accept such an
offer and save work.
• Related to the previous point, the one-way process I've set up isn't good
enough. An AI ingests and structures arguments. You have a dialogue with it
to refine the arguments. Then you import them into HowTruthful where you
assess truthfulness. While I was assessing ballot questions, there were
several points where I wanted to go back and use AI assistance to improve the
argument tree. Having rated a lot of statements, I didn't want to start over
with everything marked debatable. I wanted to start from the partially-rated
argumen