Wedding in Philly. "Dearly beloved, we are gathered here today 9 days before the opening of the season 11 days before the Sixers' first home game To witness the joining in the holy matrimony of..." True story.

An inexcusable omission, now fixed! Thanks.

Yes! Regression to the mean. Who returns for a second petition? Mostly those who had a successful first one.

This too is an interesting hypothesis; from research that I have seen (and done) so far this is not yet the case. Maybe in the future...

Yes very valid hypothesis. We don't *think* that this is what's happening here because the period we studied was not yet widespread in AI use or suspicion, but we definitely mention that as a possibility in the article. No good way (that we could think about) to find out.

Why was the AI not helpful? See the paper for more discussion, but some possible reasons (we are looking at now) include decreased specificity of petitions and (I think, more significantly) reduced effort by the petition writers, who were less committed after AI wrote for them.

Isabel came up w/ another analysis, looking at repeat petition writers who returned to write a 2nd petition. We compared people who wrote 2 petitions before the AI launch w/ those who wrote 1 before and 1 after AI feature access, confirming our results: AI access changed the text, not the outcomes.

One could ask whether these diminished outcomes are due to new authors joining the platform with the increased access that writing with AI enables (which is one of the reasons the feature was launched in the first place). We saw no evidence of a widespread increase in participation. Further...

Was the AI assistance helpful? In general, petitions were NOT more successful with access to AI -- if anything, our results show that petitions may have been less successful in hitting signature and comment thresholds (see the end of this thread for speculation as to why this was the case).

(In the supplementary materials, you can also see an increase, post-AI, in the use of words such as "implement", "urge", and "mandate" in the post-AI petitions' titles).

As you can imagine, the petitions' textual/lexical aspects shifted with access to the AI feature: they were, on average, 50 words longer with access to "write with AI"; the sentences were longer, and the words were more complex. At the same time, the petitions became more similar to each other.

If you are a data scientist/causal inference person, you know what that means: we had a natural experiment that could be leveraged using diff-in-diff analysis to estimate the causal impact of the AI feature on various aspects of the petitions.