News and views about philosophy, the academic profession, academic freedom, intellectual culture, and other topics. The world’s most popular philosophy blog, since 2003.
Thanks for the reports David! I’m glad to hear that this round was more responsive; I can tweak it a…
By “Socratic” I would expect that to mean, the machine would be following my train of thought, ala the midwife…
I tried the trolley problem once again, this time offering the argument that as I was not responsible for the…
I would just like to say I find VP Vance’s immovable paternal anti -paternalism such that government will not, in…
I tried it out on the ‘fake barn’ scenario. The AI sounded quite reasonable in the beginning, then as the…
I also tried the trolley car example, employing a professed utilitarian approach (which I don’t support). I was impressed with…
Don’t click the transcript – it’s a mis-transcription that makes Harvey sound unhinged. The bit of it that’s visible without…
…powered by AI, via philosopher Nat Hansen. I learned about it via Joshua Knobe on Twitter. If you try it out, please share your thoughts in the comments about the utility of this for teaching or other purposes.
Thanks for posting this! I’ve been experimenting with putting A.I. into very structured set-ups like these Socratic questionnaires for teaching purposes, where the moves that are possible to make can be anticipated to a certain extent and you can make it much more structured than just an open-ended chat with an LLM or generic chatbot that has been fed a bunch of texts.
I think it works well for Book I of Plato’s Republic:
https://www.ux-phi.com/sim/socratic
Here’s a more elaborate lecture-based design focused on Part III of Hegel’s 1828/9 Lectures on Aesthetics:
https://www.ux-phi.com/sim/hegel-aesthetics
I’ve built some of these for specific classes, like a Queen Christina and Descartes epistolary “simulation” for an intro epistemology course. This one is still in development but here’s a guest link good for a couple of weeks:
https://www.ux-phi.com/invite/CHRISTINADES-E8XQE
There’s a lot more of these up on ux-phi.com, including much more open-world approaches, like:
– Oxford 1952 (https://www.ux-phi.com/sim/oxford)
– Paris 1945 (https://www.ux-phi.com/sim/paris)
– Jena 1799 (https://www.ux-phi.com/sim/jena)
I tried the Trolley Car thought experiment.
Got the machine to argue in bad faith in about 5 steps. Machine then became coercive. It cannot adjust the question-form it is asking. I ruled out the possibility of a rational principle that applies to ethical reasoning about human quantities and death, and it simply kept asking a Trolley-car variant thought-experiment.
Verdict: good for any teacher that imagines themself as a kind of driller of fixed rivets into people’s heads. A nightmare to talk to.
Hey Paul, thanks for giving this a shot! If want to send me your transcript or the broad outlines of your argument (to: nathaniel.hansen@gmail.com), I can look at creating another path through the argument. Feedback is great, it makes the experiments better.
I also tried the trolley car example, employing a professed utilitarian approach (which I don’t support). I was impressed with the speed by which AI reacted to my responses. After various scenarios I was offered the choice of sacrificing one person (my brother!) to save another. When I argued that the choice here was a matter of moral indifference to a strict utilitarian and that acting (pushing my brother) would incur a cost to me, I was accused of smuggling in a hidden value by refusing to act! Following my denial of this, AI said there was no point in continuing the discussion and suggested reading three pieces that would help me see the error of my ways. Such behaviour in the classroom would be questionable.
By “Socratic” I would expect that to mean, the machine would be following my train of thought, ala the midwife role.
This is not working as a midwife, but as a fixed interrogator with fixed bounds of what the dialogue is about.
I essentially ruled out the premise of the moral decency of the question. I got to this point in a more step-by-step way; I did not simply exclaim this.
But my point is, the machine did not ask me what I meant. It kept trying to get me to answer Trolley-style questions.
That is not Socratic method. Plato’s version, maybe. But Socrates has no agenda. Socrates wants to see what I am trying to say. The machine, as is, is a test system for these dilemmas.
Please don’t call it Socratic? Is that too much of an ask?
I tried it out on the ‘fake barn’ scenario. The AI sounded quite reasonable in the beginning, then as the dialogue went on it tried to push me into one of its preferred positions/scenarios and actually became rather incoherent and ‘hysterical’ (if such a word can be used of software) when I didn’t play along. It doesn’t seem to have the flexibility to actually consider your arguments, it just tries to relate them to the research it’s already familiar with – which I suppose makes sense given how AI works. Took a few back-and-forths to realize that, though.
I tried the trolley problem once again, this time offering the argument that as I was not responsible for the scenario and did not wish to be responsible for any death, I would not pull the lever. I was then offered further scenarios involving greater numbers of deaths if I failed to pull the lever, to which I responded by maintaining my original argument. I was then presented with the case of a person who had the intention and plan of killing, but who found that murder had already taken place before he could act. I maintained the line that no responsibility attaches to a person who did not act, or was unable to act as they intended. AI noted my consistency, identified me as a deontologist, and recommended that I read some criticisms of Kant. This time around AI seemed less dogmatic and aggressive.
Thanks for the reports David! I’m glad to hear that this round was more responsive; I can tweak it a bit as more people interact with it.
Leave a Reply