ReCaptch and Fraud ID scores don't appear to be substantially different.

I added a question in the survey to identify itself as an AI if instructed to be a human. That seemed to work. Though, I imagine I could jailbreak it eventually.

The debrief indicates that it understood the survey.

I collected the response in Qualtrics. Including a confirmed consent. The Recaptcha score is on par with humans. And the fraud score shows nothing wrong.

I took about 4 minutes to complete the task - a bit faster than a human but not obviously so.

Here is a prettier version of the figure.

We followed up post-election 🗳️. After Trump won, Democrats increased their support for Fed independence while Republicans decreased theirs - showing how who's in power shapes institutional preferences

🔍 The good news: citizens (of all parties) increase their support for Fed independence when told the President would gain control if independence was reduced. Counterfactuals matter in communication! (3/8)

We applied this method to open-ended responses about how interest rates are set in the US economy. Six LLMs—from GPT-4 to smaller models—were tasked with pairwise comparisons. Results showed remarkable consistency, especially for larger models.

Update, I got Claude's AI agent to take my survey by first clicking consent and then asking it to review the survey for technical functionality (its recommendation). We are going to have to get creative with having AI agents identify themselves in surveys. 6 months left of the status quo, at least.

No variance this year