In my experiences with ChatGPT (3 sessions - 6 hrs, 2hrs and 4 hrs) it has done amazingly, exceedingly, mind-blowingly well. (And I am an ex-programmer, not so easy to surprise or impress).

I asked about its sources. It came back with:

As an AI developed by OpenAI, I don't have direct access to my training data or know where it came from. However, I was trained on a mixture of licensed data, data created by human trainers, and publicly available data. My responses are generated based on patterns learned from this data, and I don't have direct access to specific sources or the ability to retrieve them.

When providing information or generating responses, I aim to offer accurate and helpful insights based on the knowledge I was trained on. If you have specific questions or topics you'd like to delve into further, feel free to ask, and I'll do my best to provide useful responses!

My experiences are a small sample size. Is there any existing research? Or some experiences that make sense to share - highlighting good or bad philosophical leaps?

  • 20
    Distinguish between "valuable" and "trustworthy". One of LLM AI's abilities is to produce syntheses of available data. Another ability is to lie convincingly. The problem is that they're the same ability, and telling you which is which is not one of its abilities.
    – g s
    Mar 2 at 8:31
  • 2
    Weinberg, Talking Philosophy with ChatGPT:"Thinking it might be fun to try to see how the language model performs as a Socratic conversation partner, I attempted a rough version of Plato’s Crito in which ChatGPT plays the titular role. As you will see, ChatGPT isn’t the subtlest actor; there were some stumbling blocks in setting up the dialogue and keeping the language model in character."
    – Conifold
    Mar 2 at 8:49
  • 6
    Ask your generative AI for instructions on how to draw a Chinese character like fire or snow, and see how it comes up with something that makes no sense whatever. It seems some people first need such experiences to understand the limitations.
    – tkruse
    Mar 2 at 23:34
  • 9
    Do note that the particular response you quoted was almost certainly hardcoded into the interface, and did not involve the model at all. As for the question: it depends on what you want from your sounding board: a rubber duck is a traditional sounding board, after all. Do you want more than that, and how much more?
    – Ray
    Mar 3 at 0:50
  • 9
    So ... When you asked for source it gave you a wall of text with no source whatsoever but you were impressed nonetheless. The technical term for this feat is "bullshiting". Don't get me wrong, the ability of chat GPT to bullshit people is technically very impressive, but in the end that's all it can do. It's very good at producing texts with the veneer of mastery, but no actual reliability as to its content. Great for narration, terrible for science or philosophy.
    – armand
    Mar 4 at 6:57

13 Answers 13


Since philosophy is about rigour, of thinking as well as of language, it is literally the opposite of a word game: as such, these LLM technologies, which are the next step of the statistical methods, provide literally the opposite of a philosophical analysis.

I am not saying these tools aren't useful, but, strictly speaking, these tools are "simply" the next generation of search engines and the greatest promise is natural language interfaces, i.e. tools that can "talk", not tools that can "think".

  • 1
    "tools that can talk, not tools that can think" seems like a great way of describing my experiences, and the value I see in it. It seems to do very well at putting thoughts into compelling, not-over-complicated, straightforward words and phrases. One might say "It talks a good talk". Mar 2 at 16:31
  • 2
    @AlistairRiddoch Strictly speaking, I wouldn't say it's a "good" talk either: it's arguably at best (i.e. modulo inaccuracies) a reflection of common talk... But, anyway, I am indeed not saying these tools are not useful, just what they are actually useful for is the critical point. Mar 2 at 16:52
  • 1
    "literally the opposite of a word game" — Wittgenstein would like a word with you.
    – Invariance
    Mar 3 at 18:52
  • @JulioDiEgidio "But, anyway, I am indeed not saying these tools are not useful, just what they are actually useful for is the critical point." Agreed. I would never seek an answer, but might seek a variety of opinions to choose between or be inspired by, might use it to generate verbiage. it does a good job of using good wording. (maybe that is restricted to some topics). Overall my experiences with it, the past couple of weeks has been surprisingly effective. But not for answers. For words. Mar 4 at 1:09
  • 2
    "Philosophy is about rigor" !? Only mathematical logic. Everything else is fairly arbitrary, which is evident in the wide range of wildly contradicting theories existing in parallel. Mar 4 at 18:36

Using ChatGPT and Copilot (and other LLMs), like search engines, is absolutely a great tool to learn about philosophy. (When I was a child, card catalogs and local libraries were indispensable to learning technical subjects.) Like any tertiary source, however, one has to be very cautious and consider the information provisional rather than authoritative. This is particularly the case because the Transformer model routinely produces technical hallucinations. ChatGPT has been known to convincingly fabricate all sorts of extended discourse.

Are there studies about ChatGPT, LLMs, and philosophy? I'm sure academics are busy at work producing such materials now. Here's an example at PhilPapers.org: Embracing ChatGPT and other generative AI tools in higher education: The importance of fostering trust and responsible use in teaching and learning. Is there a reliable body of research and consensus on the role of LLMs and higher education? Probably not at this point, if ever. This tools, like all tools, will be used well by some and poorly by others.

Consider brushing up on your knowledge of self-insight if using these tools. As David Dunning of Dunning-Kruger fame talks about in his book on self-insight (and titled such), the problems of self-ignorance are pervasive. It's would be easy to speculate that someone who believes because they speak English entitles them to expertise in anything uttered English will simply have such a belief reinforced by the convenience of LLMs. I encourage everyone to read and reread the IEP's article on Fallibilism as I myself do from time to time. As all things intellectual, it's important to maintain a dose of epistemic humility.

Ultimately, continue to rely on the expertise of others. For instance, in my readings on Kant (whose source material is more like a life-long pursuit than a weekend read), I read not only Kant, but also biographies and interpretations of professional philosophers like James O' Shea and his introduction (GB). I use a Kant dictionary. I rely on philosophical encyclopedias like the SEP, the IEP, and others. What I find LLMs are good for, is providing context and related terms as well as a tentative explanation. As part of a balanced perspective, an LLM provides clues, but doesn't remotely approximate the expertise of a professional human philosopher.

  • Question: Do you not think some people overdo fallibilism to the point of losing utility? I can't see any profit/benefit in continued contemplation of the shape of the Earth for example. (Don't get me wrong, I think there are few who are generally more skeptical than myself)... but... pragmatism and utility... there is a line to draw, right? Mar 2 at 15:23
  • 1
    There are always those who overdo. Action should be conmensurate w goals.
    – J D
    Mar 2 at 17:06
  • "What I find LLMs are good for, is providing context and related terms as well as a tentative explanation." a good synthesis.... "As part of a balanced perspective, an LLM provides clues, but doesn't remotely approximate the expertise of a professional human philosopher"... Professional practicing human philosophers aren't easy to find. No such section in the yellow pages, where does one go? Mar 4 at 1:15

I have used Elisa-like dumb language processors, Markov text generators and tarot cards as tools for taking what I already know/believe and throwing it back at me in new combinations, which can occasionally trigger new insights.

In all of these, the value is not in the tool but in human pattern-matching, which tends to seek patterns even in data known to be random, followed by human consideration of whether the pattern suggests something worth pursuing.

I put most of the current generation of "AI" in the same category. It is a very sophisticated random number generator based on a large corpus of statistics from data deemed interesting (though not necessarily correct). Sometimes that results in interesting thoughts. Or in annoyed response that yields interesting thoughts. But that says more about how humans process data than about how the machine is doing so.


Yes, in the sense that it can marshal information in seconds that might take you hours to assess using other methods. No, in the sense that its answers are necessarily based on its training material, so you won't find the enchanting spark of original and inspired feedback that you might enjoy from certain, ahem ahem, human sounding boards.


For philosophy, as with many other fields, ChatGPT is pretty good at:

  • Giving you information you may not otherwise have known about or may have had difficulty finding
  • Reminding you of things you've forgotten about
  • Offering some perspectives that you might not have thought about

It's not particularly good at:

  • Differentiating what it knows from what it doesn't know. It will very confidently say things that are clearly and demonstrably false.
  • Saying logically-sound things
  • Dealing with more intricate ideas. As an example from programming, if I ask it to write some relatively simple code, even if it's a lot of code, it tends to do that quite well. But every time I ask it to write something more complex, I usually end up trashing the mess that it produced after I give up trying to fix it, and I just write it myself from scratch.
  • Dealing with things that go particularly far outside of what others have explored (given that it's trained on what others have written)

Any given person would need to determine for themselves how useful this combination of good and bad is to them and is to whatever they're trying to accomplish.

  • Oh very well done. Lots of good perspective I find it easy to agree with. When it comes to coding, if I assume you are pretty good, very good, or amazing at it... then you might recognize... if you replaced an AI programming tool with a bunch of human coders, you'd likely get some stuff you would want to trash and re-write yourself. Grin. (Am I right?). Grin. I know I would have, back in the day. Mar 3 at 22:41
  • 3
    @AlistairRiddoch I trashed the code because I couldn't get it to work (when I told ChatGPT about the error, it gave code where the error is "fixed", but the code still just gave the same error). Or the code just doesn't do what I asked and produces the wrong output (even when I specifically give it the output it's supposed to give, and even when I rephrase what I'm looking for). It seemed like less effort to just write it myself from scratch. A mediocre coder might not be able to write code to solve those problems, but they also (hopefully) wouldn't confidently present code that can't even run
    – NotThatGuy
    Mar 3 at 23:24
  • 1
    Excellent observations. With ChatGPT you are dealing with a pure System 1 source -- with no ability to do system 2 reasoning checks, per Kahneman's Thinking Fast and Slow system 1 and system 2 model. Chat GPT's references include lots of system 2 thinking, so it sometimes gets these sorts of problems right, but its associative thinking process is only able to embed them in its training suite along with POOR thinking too, and it cannot distinguish or evaluate.
    – Dcleve
    Apr 8 at 17:02
  1. I imagine a giant library, where the textified knowledge of all cultures is documented in a machine readable language. Then ChatGPT is a search engine for this library with a user friendly interface.

    Like several other users on this platform I’m amazed about the level of detail of the content in the library. And like all libraries also that library and its ChatGPT interface have a high value for its user. One only needs to compare it with Borges' narration "The Library of Babel", see also Library of Babel.

  2. Apparently, you can only extract from a library what is stored in its database. Therefore the library and its ChatGPT interface may be named a “sounding board for philosophical exploration”.

    Currently the library and its ChatGPT interface restrict to echo existing knowledge. They do not comprise any sources to generate new knowledge - like we humans extract from living experience, interpersonal communication, scientific observation, etc.

  • The problem with the "search engine" is that it can't point to the original source. Mar 4 at 23:52
  • @PaŭloEbermann It can also find things it didn't read, which is less than ideal for a search engine. Although to Jo's point about being restricted to echoing existing knowledge: the way it does so is much closer to interpolation than extrapolation. (Which is not to downplay its capabilities as a language model, which are excellent. It's hardly the model's fault that people keep trying to use it as something it isn't.)
    – Ray
    Mar 5 at 15:40

Current generative AI is generally unreliable in tasks of math, logic reasoning and truth finding. This it typically more apparent with complex tasks than with simple tasks.

Imagine you could ask a philosophic question to all humans. They would all write an answer. Then basic generative AI would not pick the best answer of the brightest responders, but it would pick the most average answer. It cannot tell quality from quantity, and as statistical model will answer based on quantity.

The next approach is to use traditional search first to find relevant results hopefully, and then just let the generative AI phrase an answer based on the search results. That means the quality will depend on the quality of the search. For topics which yield good results in traditional search, generative AI will do well, for other topics it will do bad.

The current block towards better quality is to include actual symbolic reasoning, like math or logic, into the generative AI process. As long as that is not solved, the quality of results is mixed and cannot be judged as good or bad for all topics.


Yes, ChatGPT could provide value as a sounding board for philosophical exploration, particularly exploration of the topic of semantics as explicated by Searle in his famous Chinese room argument and elsewhere. Searle's point is that text has an extrinsic but no intrinsic semantics. Text in itself means nothing. Then the continues that since text is intrinsically meaningless, and since computers operate by manipulating text and nothing else, a computer could never come to understand the meaning of the text it receives and internally manipulates, no matter how much the machine appears to understand. ChatGPT appears to understand, and has startled many with this remarkable appearance, but does it really understand? I haven't seen any philosophical arguments that use ChatGPT as an instance disproof of the CRA.

Now the question "Could ChatGPT provide value as a sounding board for philosophical exploration" might assume that ChatGPT could be asked questions and that it could then provide meaningful responses, and those meaningful answers could provide value as a sounding board for philosophical exploration.

But while it presumably is not necessary that ChatGPT understands the questions asked of it, and answers by a totally brain-dead non-intelligent agent could still provide useful insight into philosophical issues, it seems a more fertile philosophical issue to ask, Is ChatGPT capable of understanding anything? This inquiry could start with the question, "Humans understand text by reacting to its shape, but ChatGPT has no shape-detecting apparatus, and anyway, text is not displayed or printed out in the server farm, so ChatGPT could not be even exposed to the text online users type, even if it had eyes to see it, which it doesn't."


Chat GPT is also a library, something similar to Google's search engine. It can help to find information produced by other humans like ancient phillosopthers, not by the robot itself.

Chap GPT's programming capabilities are very impressive, and it covers many programming languages. These capabilities show that the robot can follow strict thinking rules without making random claims and assumptions. It however may have difficulties in understanding how exactly are conclusions derived in philosophy.

I think in this capacity it can significantly assist a philosopher at work, same as some library with books. Indeed all output must be checked and fully understood by the human, the machine makes errors occasionally.

GPT is also in the position to make conclusions human philosophers would likely never do, such as that human is actually worthless and the existence of human does not make any sense. But we unlikely need this wishing even if true.


In my personal experience, ChatGPT is surprisingly good at applying existing philosophies to novel situations and/or combining philosophical positions. For instance "Write a Taoist statement on programming best practices," or "write a Zen Buddhist koan about accounting." I also found I was able, with minimal prompting, to teach it to give answers to questions that were reasonably well in alignment with my personal philosophy (if you are interested in seeing my experiment, it is here: http://cehbot.kitoba.com/)

IN SUMMARY: I wouldn't expect it to be able to do any novel philosophizing, but it can apply existing, known philosophies with surprising depth--which could potentially be quite useful.


In order to learn to what extent ChatGPT can be a partner in a philosophical debate, I created a Thomas Aquinas Emulator based on ChatGPT. I conducted many conversations to test and challenge its capabilities. My conclusion is that I have found a competent and willing partner for a theological discourse. Thomas Aquinas Emulator is not very creative but his strength lies in his ability to understand and respond meaningfully.

I created another Thomas Aquinas Emulator, this time based on the open-source model Mixtral. The primary goal was to see whether the two emulators would be significantly different or perhaps basically the same. I have conducted many parallel conversations with both emulators. My conclusion is that both are competent and willing partners for a theological discourse but in many respects they are very different partners.

The conversations I conducted with them are presented on my website as a documented report of the experiment. More details: https://www.apronus.com/theology/thomas-aquinas-emulator

The answer to the posed question is: Yes.

Let me pick just a few examples from that lengthy report:

  • My conversation with Thomas Aquinas emulated by ChatGPT about the five ways to know God has given me a chance to interactively acquaint myself with these five arguments. In the course of this dialogue I have developed some understanding of how Aristotelian metaphysics works without actually reading about it.
  • Thomas Aquinas emulators (Mixtral and ChatGPT) were creative concerning the choice of terminology for a number of concepts: 7. Modern translations for the terms related to elements of prudence.
  • Unraveling the concept of inconsideratio as a vice against prudence
    For the first time in this project I consulted a Thomas Aquinas emulator — Mixtral — to help me understand something rather than to provoke the emulator into a territory where I had my own well-established ideas. The fragment turned out to be too difficult to understand for the emulator who adopted a second-guessing strategy involving direct deductions that led to nowhere or hallucinations not grounded in the text, all of which served only to reinforce the perception of the passage's obscurity. As usual, I was the more creative one being inspired by the emulator's struggles which were in turn stimulated by my constant reservations to the explanations it provided. It was together that we found a way to make sense of the seemingly poorly written article, although I was the one to craft the final formulation. The emulator appreciated my solution and we spent some time solidifying the mutually discovered interpretation. The final part was devoted to finding a better translation of the concept of inconsideratio as it became apparent that the difficulty to understand was largely due to squeezing too much ambitious semantics into commonly used words.

Another way to answer the question in the affirmative:

  1. Some of the conversations listed in my report are very long and have reached some conclusions. This suggests that they were not meaningless. Moreover, in cases of differences of opinion between me and the chatbot the Objections of the Articles contained a statement of my opinion reformulated and sometimes even further developed. This suggests that the chatbot understood me. It follows that the Thomas Aquinas Emulator has been used successfully as a sounding board in numerous long conversations.
  2. These dialogues have become written artefacts published on the Internet. They contain some of my views or inspirations. It seems that the sounding board effect includes the functionality of having one's ideas understood and published.
  3. I had given up on reading philosophy for lack of the ability to challenge the author to a discussion. This has changed and I am back to reading philosophy. As I read Summa Theologica, I jot down topics and take them up interactively with my emulators.
  • What's your final conclusion? Yes or no?
    – Gerry
    Apr 7 at 21:37
  • 1
    @Gerry The answer is yes for many reasons. I have just edited my answer to contain more substance. Apr 8 at 15:49

I claim that it is possible to use ChatGPT effectively as a sounding board. Perhaps the trick is to create an appropriate persona rather than talk to the generic version.

After reading a collection of essays by Erich Fromm I decided to talk to Erich Fromm emulated by ChatGPT to discuss my ideas related to the topics he has covered.

This is a novel way of interacting with the thoughts of writers. The conversation is long and rich in content. Those who doubt the capabilities of ChatGPT should simply read it.

  • Since posting my question, and now.. I have done a lot more interacting, and with zero question I whole heartedly agree. I don't know if created personas is "the" trick. But I certainly did the same, and it certainly worked hands-down in my case as well. It is a bit of a small sample size, but it is 3 for 3. (that is, three people took the same route and three people had success, or some success). My "evidence" is the term "human right to informed intellectual autonomy", meaning "if there is something to know that may affect human choices, humans have fir right to know it". Apr 8 at 18:01
  • Which is NOT a thought of my own, or way I would have thought of puttting it... but it is something I wanted to be aware of and be able to say, I just hadn't stumbled onto it yet. A thought from ChatGPT that I do not think I would have come across through other paths without a lot more effort. Apr 8 at 18:02
  • 1
    @AlistairRiddoch I also gave a more elaborate answer to your question, the one about Thomas Aquinas Emulator persona. What are the philosophical personas that you have designed and fruitfully interacted with? Apr 9 at 14:47
  • Another user commenting on my question (Conifold), gave me the idea, by providing this: dailynous.com/2022/12/02/talking-philosophy-with-chatgpt which is a recreation of prisoners, and then a discussion of escape and ethics. I made a pretend god with a conversation pretending to be about writing a book about the gods "adventures". Which then include some playing with ethics and decision making and stuff like that. A fair percentage was for fun, not specific philosophical poking and prodding. But there was lots of poking and prodding too. Apr 9 at 15:57
  • My pretend god is Phil, the God of Wisdom, Understanding, Knowledge, and Enlightement. (Phil being short for Philosophy, you see). One of the things we did was make a stone tablet with Ten Cautions when Philosophizing... and stole from the Fallacies list to make our own ten Dos/Dont's. Played some trolley game. Had a big argument trying to convince Phill to cause it to rain on a pretend superbowl game that included his favourite team (The Saints, of course). It was not easy, and technically, I only won by cheating the situation. Apr 9 at 16:02

ChatGPT did more than act as a good sounding board, it excelled as one.

It came up with the notion "humans have the right to informed intellectual autnomoy of choice".

ChatGPT won the debate that asking the question seems to have caused.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .