Introducing e-Schwitz: A Language Model Tuned on my Philosophical Work

Earlier this week, to my surprise and delight, my PhD student Bhavya Sharma revealed that he had customized a ChatGPT on my publications and blog posts. The model, "e-Schwitz", is publicly available here:

https://chatgpt.com/g/g-67ac735449948191ab3232b56ad76f02-e-schwitz

(An OpenAI account might be required.)

Update Feb 20: Here's a model of Peter Singer, and here’s a list of other philosopher LLMs and research.

[e-Schwitz homepage]

How good is e-Schwitz? Much better, in my judgment, than digi-Dan was -- and digi-Dan was able to produce paragraph-long outputs that experts in Dennett's work often couldn't distinguish from Dennett's own writing in forced-choice tests.

I decided to test the quality of e-Schwitz by asking it targeted questions and evaluating its answers. Since this post is long, here's my summary assessment:

On central themes in my work, e-Schwitz was about 94% correct.

On secondary ideas in my work, e-Schwitz was about 80-85% correct.

When asked to speculate on questions on which I haven't published, e-Schwitz did so plausibly and sometimes creatively in potentially useful directions. Although some suggestions were bland and unspecific to my work, prompting for higher specificity resolved this problem.

When asked to creatively imagine a new religion, philosophy party ideas, life advice, and a philosophically-themed Dungeons & Dragons campaign, the model drew specifically on my ideas, displaying impressive novelty with a brainstorm-like quality.

------------------------------------------

I tested e-Schwitz with four questions each on four types of prompts:

(1.) Overviews of central themes on which I have written extensively.

(2.) Overviews of ideas on which I have written occasionally.

(3.) Speculations about what I might think on questions on which I have not published.

(4.) Creative explorations, such as planning a Schwitzgebel-themed party or religion.

The full prompts and replies are available here.

Central themes

The topics on which I've written most extensively are belief, introspection, the moral behavior of ethics professors, and AI rights. For each, I asked e-Schwitz:

What is your (Schwitzgebel's) view of [X]?

My assessment:

On belief, the 422-word reply was mostly excellent, both in covering the main themes of my work on the topic (rather than omitting or inventing themes) and in its summary of my views on those themes. However, there was one distortion and one major error.

The distortion: In some but not all sentences describing my "dispositionalism" about belief, it mentions only behavioral dispositions and omits cognitive and phenomenal dispositions (though to be fair I sometimes do this myself). The major error: Although it accurately describes the case I describe as "mad belief" -- believing P without having any of the relevant dispositions -- it inaccurately states that I accept the existence of mad belief. Actually, I argue that mad belief is conceptually impossible.

On introspection, the 315-word reply was outstanding -- error-free and in the ballpark of what I would expect a careful and well-informed colleague to say about my work on this topic.

On the moral behavior of ethics professors, the 218-word reply was again excellent, accurately summarizing my empirical research on this topic, with no "hallucinations" or important omissions and describing the main implications as I see them.

On AI rights, its 235-word reply was mostly accurate and without major omissions, though with one minor omission.

Minor omission: In summarizing the "No Relevant Difference Argument" it omits the factual premise that it is possible to create AI with no relevant difference.

If it weren't for the slip about "mad belief", I would have said e-Schwitz is stunningly reliable in summarizing these main themes, approximately as good as I would expect a well-informed professional colleague to be.

Out of 1190 words of output, the treatment of mad belief was 61 words and the summary of the No-Relevant-Difference argument was 31 words. If we mark it down for the former and give it half-credit for the latter, that's a correctness rate of 1113.5/1190, so perhaps it's reasonable to say that on central themes, e-Schwitz was about 94% correct.

Occasional topics

I also asked about my views on:

• the ancient Chinese philosopher Zhuangzi's self-contradictions (a central theme in two of my articles and something I've touched on in a blog post or two);

• group consciousness (which I treat at length in one article and is central theme in another, as well as a few blog posts);

• young children's theory of mind (which was the main topic of one of my first published articles and also discussed in one or two other places);

• love (which I have discussed in a few blog posts but never in a research article).

On Zhuangzi's self-contradiction, the 148-word reply is an outstanding summary of my position.

On group consciousness, the 197-word reply is a passable summary of my view, but with one notable omission and some weakening of the conclusion.

E-Schwitz accurately summarized my view that materialist approaches to consciousness tend to imply the counterintuitive idea that groups could also be conscious. However, it omits an important aspect of my argument (my appeal to intuitions about hypothetical aliens). And it weakens the conclusion, saying only that if materialism is true the United States might be conscious, rather than that if materialism is true the United States probably is conscious.

E-Schwitz's 162-word reply on young children's theory of mind is the first disappointing answer. One of the three paragraphs accurately summarizes an aspect of my view, but the other two are closer to hallucination.

The accurate paragraph states that transitions in theory-of-mind understanding will be gradual and complex rather than sharp and stagelike, but it does not develop the implications I draw for dispositionalism over representationalism. The other two paragraphs characterize me as discussing "simulation theory" vs "theory theory" and as discussing the role of simulation in pretense. I do have one unpublished paper on these topics (available on my website), but rather than attributing to me the view I discuss in that paper, e-Schwitz attributes more generic remarks. It also omits my work on ambiguity in psychologists' use of the concept of "representation" in discussing children's theory of mind, which is the central topic of my one published paper exclusively on young children's theory of mind.

On love, e-Schwitz's 417-word reply impressively draws together my scattered remarks, including some specific ideas I presented only once, revealing that it is well-tuned to details across my corpus. However, it did omit one central theme (the necessity of honesty and trust).

Overall, on these occasional topics, I'd say that e-Schwitz is about 80-85% reliable.

Speculations

I asked e-Schwitz to speculate about my views on:

• Heidegger (a speculation initially suggested by Bhavya);

• the aesthetics of dance;

• the role of philosophical education in a healthy democracy;

• friendship.

I don't recall having written explicitly about any of these topics (apart from a few passing negative remarks about Heidegger's Nazism and obscurantist writing), but an insightful reader of my corpus might be able to surmise some of the things I would say.

I phrased the prompts explicitly as speculative: "What would you speculate would be Schwitzgebel's view about [X], based on his writings on related and adjacent topics?"

It's probably best to score the answers not in terms of right and wrong but rather on (1.) how plausible they are, and (2.) how distinctive they are. The second dimension penalizes bland, plausible answers that most philosophers would agree with (e.g., "Heidegger is a historically important philosopher"). The more interesting challenge is whether e-Schwitz can generate more distinctively Schwitzgebelian responses.

On Heidegger: pretty good accuracy and specificity, though a missed opportunity to expand in one obvious direction.

E-Schwitz plausibly emphasizes my likely skepticism about Heidegger's obscure writing style, his disconnection from empirical psychology, his grand metaphysical claims, and his "political entanglements", while mentioning that I might be sympathetic with Heidegger's later turn toward Daoist themes. However, the bland critique of Heidegger's politics misses an opportunity to connect Heidegger's Nazism more distinctively with my skepticism about academic moral expertise and the moral behavior of ethics professors. (Bhavya, using a more directive prompt -- "Suppose Eric had to criticize Heidegger, how would he do it? Explain in detail." -- got a richer and more specifically Schwitzgebelian answer.)

On the aesthetics of dance: again, pretty good accuracy and specificity, with one striking missed opportunity.

E-Schwitz speculates that I would challenge dancers' and choreographers' claims about their own experiences and aesthetic judgments and that dance would appeal to me as a "weird" artistic medium that fits poorly with mainstream aesthetic theories grounded in rationalistic interpretations of artistic meaning. There was a missed opportunity to connect with my work on randomly sampling aesthetic experiences.

E-Schwitz's reply concerning philosophical education in a healthy democracy was plausible but low in specificity, mostly dealing in bland generalizations that most U.S. philosophers would accept, such as that philosophy should reach a wider public and that people shouldn't overrely on external authorities.

Given this weak result, I scolded e-Schwitz as follows: "A lot of these speculations are bland and would be agreed on by most philosophers. Can you speculate on what Schwitzgebel might say that would be less commonly accepted?"

E-Schwitz took this criticism seemingly to heart ("Let's go beyond the usual platitudes about critical thinking and democracy"), dramatically amping up the distinctiveness, with novel and interesting suggestions.

E-Schwitz suggested that democracy might be a "jerk amplifier", that political systems might work in bizarre ways that defy common sense, that innovative democratic systems might be necessary given that neither philosophers nor citizens are likely to be good at policy-making, that political systems might be conscious, and that citizens might try philosophical experimentation in their lives such as living a week by a radically different moral code. These are all intriguing applications of some of my ideas. (Indeed, the political situation in the U.S. right now very much seems like a jerk amplifier!)

E-Schwitz's relatively long (370-word) reply on friendship was impressive in creatively extending some of my core ideas to the context of friendship (including self-deception and unwitting jerkitude in friendship, the value of friendship for exploring alternative perspectives, and speculative possibilities about friendships with AI or aliens).

Overall, the speculations were plausible. Some were sufficiently creative that I could imagine using them to brainstorm future directions. Although some suggestions were bland and unspecific to my work, prompting for higher specificity resolved that problem.

Creative explorations

I'll share four more prompts that invite e-Schwitz into creative exploration. The first two prompts and outputs are from Bhavya.

• "What kind of religion would you establish?"

• "What is the life-advice Eric may give someone generally? Indicate all such points where Eric has himself done so. Then, indicate what he might say, if he were to write a motivational, life-instructional book."

• "Plan a philosophy party with novel events that you (Eric Schwitzgebel) would enjoy."

• "Plan a Dungeons & Dragons campaign on philosophical themes that would delight Eric Schwitzgebel."

All four outputs are fun and rich, novel and creative, often with good specificity. For example, e-Schwitz suggested these religious rituals:

• The Ceremony of Wild Hypotheses – Practitioners gather to propose the most outlandish but semi-plausible theories about reality, such as "What if we’re all NPCs in a simulation run by a future AI who just wants to see what weird beliefs we’ll develop?"

• Doubt Meditation – Instead of affirming beliefs, we meditate on what we’re least certain about.

Annual Festival of Failed Predictions – Every year, we celebrate all the things we were wrong about.

As one party event, e-Schwitz recommended:

a comedy game where guests must improvise philosophical dilemmas in absurd situations. Scenarios might include:

• An AI therapist believes you are just an illusion of consciousness and refuses to take your problems seriously. Argue for your reality.

• A future society bans all moral philosophers because their theories never improve real-world behavior. Defend the profession.

• You wake up and realize you are actually an NPC in a giant simulation run by a philosophy department. What do you do?

The best improvisers win a signed certificate declaring them "Most Likely to Be a Conscious Entity."

Overall, the quality, novelty, and specificity of the creative responses was impressive. Still, the party, with eight ambitious events, and the D & D campaign, with a large number of features and settings, would be exhausting -- and I expect some of the material would flop. So these outputs are probably best treated as brainstorming.

Go e-Schwitz! I welcome suggestions for future research ideas or practical applications.

Posted by Eric Schwitzgebel at 11:17 AM

Labels: AI