ai compass tests
I recently stumbled upon this Hacker News submission which linked to Jordan's Protect your consciousness from AI. My interest piqued after reading these statements from simonw's most recent reply to a discussion on AI sentiment:
Maybe there's room here for one of those political compass style quizzes.
Note: If you know what a political compass is, skip this paragraph. If you don't, The Political Compass is a website with a test of 62 propositions. They rate your political ideology along economic (authoritarian or libertarian) and social (left or right) policy axes. Those who have finished get their overall position along these axes indicated by numeric values and also a dot somewhere on the compass. If you're interested in taking the test, feel free to take it there or at this automatically updating, one-page version: Interactive Political Compass.
My first thought was that maybe quizzes like that already exist for mapping personal sentiment toward AI whether positive, negative, or neutral. A quick search for AI political compasses returned a hit that instead pertained to political biases among leading AI chatbots: Tracking AI's AI Political Compass Scores. Searching for AI compasses returned a few hits that fit my bill.
The closest one I found was Keaton's AI Future Compass Test. Its 20 questions seek out your current answer to "What do you think about the future of AI?" Once finished, your position gets plotted on the compass based on these axes: dangerousness (whether AI will be existentially beneficial or dangerous) and governance (whether human institutions will handle AI safely or not). Those that don't mind spoilers may want to check out the overall Community Results. I find them interesting to say the least. With almost 600 anonymous responses so far, I think we can pump those numbers up.
The AI Safety Compass test has double the questions as Keaton's test. It uses different axes: alignment (being skeptical of AI risks to prioritizing AI safety research) and access (closed source or open source). A nice addition is the AI Model Responses page which helps "compare how different AI models respond to the same questions". However, I find it unfortunate that there's no overall comparison among AI models off the bat like Tracking AI's compass above.
Some questions come to mind:
- What tests do you know of that work similarly to the above?
- What are your results and findings from the above tests or others?
- In what ways were your test results similar to or different than your expectations?
We can even riff on simonw's project idea from earlier in his reply:
Honestly at this point you could build a full periodic table of AI hesitancy/resistance/criticism and have it be a useful document!
How much is still unknown when it comes to our beliefs toward AI? What specific visualization types and specific cohorts can be combined to further elaborate on that?