This is Jessica. Recently overheard (more or less):
SPEAKER: We study decision making by LLMs, giving them a series of medical decision tasks. Our first step is to infer, from their reported beliefs and decisions, the utility function under revealed preference assump—
AUDIENCE: Beliefs!? Why must you use the word beliefs?
SPEAKER [caught off guard]: Umm… because we are studying how the models make decisions, and beliefs help us infer the scoring rule corresponding to what they give us.
AUDIENCE: But it’s not clear language models have beliefs like people do.
SPEAKER: Ok. I get it. But, it’s also not clear what people’s beliefs are exactly or that they’re consistent. There’s a large body of research on how the beliefs you get are affected by the method you use. So there’s no reason to think that human beliefs are stable, and what people report as their beliefs does not necessarily explain their decisions by common models.
AUDIENCE: But people can believe things. Models are just patterns of activation.
SPEAKER: Ok. Well, perhaps we can just call them subjective distributions.
AUDIENCE: But subjective implies a person, experiencing something. We can’t establish that they have subjective experience.
INFORMED AUDIENCE MEMBER: Wait a second, when you say beliefs do you just mean a probability distribution? Like in decision theory?
SPEAKER: Yes. We can just call it a probability distribution to avoid the term beliefs.
STUBBORN AUDIENCE MEMBER: I don’t know. They may not act like real probabilities.
SPEAKER: But often the probabilities people give us don’t conform to probability axioms either. But ok, fine. How about we say belief-like representation?
AUDIENCE: What is that supposed to mean?
SPEAKER: Well, we could look for the same kinds of properties we hope to see in human beliefs, even if we can’t elicit them perfectly. Like, we assume that beliefs should have some correspondence to behavior, so we could look for that. Which is part of what we do in the work I’m going to talk–
AUDIENCE: Now you’ve really lost us.
SPEAKER: Ok. How about “risk representations” then?
AUDIENCE: Risk! Models don’t feel things!
SPEAKER: Ok. Pseudo-probability representations?
STUBBORN AUDIENCE MEMBER: I don’t know how I feel about the word representation here. Does representation imply intent? LLMs don’t have intentions.
SPEAKER: Ok. Vectors. Pseudo-probability vectors. Anyway, we were interested in seeing what happens when you apply revealed preference assumptions to LLMs, to infer the scoring rule they are using…
AUDIENCE: Scoring rule! They can’t respond to incentives!
SPEAKER: They are trained with scoring rules, like cross entropy loss. But then fine-tuning like SFT and RLHF induce something different. Plus when we prompt them with a specific decision context, we will induce a different posterior distribution. So it makes sense to–
MODERATOR [holding up their hand]: Five minutes left.
INFORMED AUDIENCE MEMBER: But why not prompt them with a proper scoring rule?
SPEAKER: Well, it’s not clear that they would react to a scoring rule we give in the prompt, because they don’t experience rewards like a person.
AUDIENCE: Exactly!
SPEAKER: Right. So the only incentives available are epistemic.
AUDIENCE: Epistemic? That assumes a knower.
SPEAKER: Ok. We are interested in the induced loss-minimization-like dynamics…
AUDIENCE: Dynamics? These models can’t be assumed to represent time.
SPEAKER: That’s beside the point, but ok. Static mapping from tokens to tokens?
AUDIENCE: Mapping? That assumes a function.
SPEAKER: Yes. I am, in fact, assuming a function.
INFORMED AUDIENCE MEMBER: Could we just stipulate the model has internal states and move on?
STUBBORN AUDIENCE MEMBER: Internal? But that assumes an inside.
MODERATOR: Two minutes left.
SPEAKER: Okay. Just states then.
AUDIENCE: But these models have no explicit representation of the state!
SPEAKER: Ok. I’ll wrap this up. Under revealed preferences, we find the utility function puts essentially all the mass on semantic hygiene and—
STUBBORN AUDIENCE MEMBER: That seems like a leap.
SPEAKER: It’s literally been revealed.
AUDIENCE: Revealed to whom?
MODERATOR [standing up and starting to clap]: Let’s thank our speaker…
Coming soon to a computer science / behavioral econ / statistics / cogsci seminar near you!
How many caveats about “as if beliefs” or “as if risk” or “as if subjective” would it take in the opening to pre-empt these kinds of objections? Maybe infinite?
It reminds me of when people talk about evolution and how anatomical feature X was “designed” for a certain “purpose”, and the pedants get all upset about the use of the words “design” and “purpose” even though everyone knows “designed” and “purpose” don’t imply intentionality or anything of the sort in this context, they’re just convenient words in very close proximity to the idea being expressed.
I don’t agree. Anthropomorphizing should be recognized and acknowledged. I think it’s important to explicitly acknowledge the shortcomings and limitations of descriptions we use.
What’s wrong with caveats?
They need to stop somewhere or they make understanding impossible.
Andrew [nG]
“They need to stop somewhere or they make understanding impossible.”
I strongly disagree. Understanding means accepting limitations and shortcomings and not reaching a false certainty because it feels like a better understanding.
“I don’t agree. Anthropomorphizing should be recognized and acknowledged. I think it’s important to explicitly acknowledge the shortcomings and limitations of descriptions we use.”
If you are speaking to children, sure. If you are in a room full of people who already understand what these terms mean in context, then it’s an unnecessary tedium and is often nothing more than disruptive. If an objection to the claim “AI has a belief” is merely appending “as if AI has a belief”, then the objection is more often than not tedious and disruptive, in context. In a paper or a textbook, sure, put it in there. But don’t be a shitty audience member when you know better just because you have a terminological peeve you want to hammer.
That’s why they make chocolate and vanilla I don’t feel like it’s tedious in the least. I think it’s very interesting to explore the distinctions, and useless move on without considering them. I’m not in a hurry. Where are you looking to get to, where clarifying the differences is an impediment?
And no, I don’t think people really do understand these terms in context. I think it’s more that we like to shuffle past our lack of understanding.
Long before LLMs, even moves by fairly trivial chess programs were described using phrases like “it seems to assume …”. We constantly use similar language to describe behavior of pets and even lower animals. We do it where ever a purely mechanistic model is unavailable for some behavior.
I don’t quite understand the visceral reactions to using them to describe LLM outputs. The whole interaction in the post above and some discussion here is “as if” the speaker started off with “LLMs are just no different from humans and they have the following beliefs”. Allowing for the possibility that the speaker knows that LLMs are not exactly like human minds and the belief verbiage is just a model to derive some predictions is to give the speaker some grace.
Perhaps it’s about framing.
“In this study, we applied methods developed to study human behaviour to LLMs. The terminology we use is derived from these methods, and thus we may talk about “beliefs” and “incentives” (and so on) even though LLMs don’t have beliefs and incentives in the same sense as humans do.”
If you think this is bad, check out traditional philosophy seminars.
Thank goodness experimental philosophers pushed us towards empiricism. Now when this bickering about linguistic intuitions begins, we can just ask the bickerers to either point to analyses of the word(s) used in representative samples of the target population or else to go collect that data.
I understand the frustration and the sense that this is “bickering” but … the audience member’s concerns are kind of hugely important. Especially as the techbros continue to try to convince us that “AI” really does have feelings and beliefs, etc. And to William’s point, the over-reliance on terms like “designed” and “purpose” got (gets?) us science-believers into a lot of trouble when trying to explain evolution to sceptics. I’m not saying I have a solution, but the problem is real.
Hi Mark,
Yes, I’m not posting this to imply that we should be casually applying the word “beliefs” to anything LM agents profess to believe. In the particular case that inspired this post, I think the speaker faced a genuine bind, in that they were presenting an exploratory-style study that drew on a decision theoretic framework, and the conventional thing to call an agent’s elicited probability distribution in the context of a decision problem is a “belief distribution.” I’ve never seen anyone balk about applying the term “beliefs” to theoretical notions like idealized rational agents. But should we be avoiding certain terms altogether for LLMs? It’s messy either way.
I’m pretty skeptical of anyone who claims that LLMs have human-like feelings or suffering, but you’re right that there are people making those arguments.
I think we need a new vocabulary describe the output of chatbots. Why should we assume existing vocabulary will suffice for novel phenomena?
“I think we need a new vocabulary describe the output of chatbots. ”
How about: “Randomly generated text that has not been checked against any model or theory of how text of any sort relates to reality, logic, or reason.”
David:
Fair enough, but that also describes the language output by humans!
“Fair enough, but that also describes the language output by humans!”
Yes! Some humans. As I’ve mentioned before, some blokes at the University of Tokyo wrote a paper claiming that LLM output looks very much like (statistically, no less!) the speech of patients with Wernike’s Aphasia. Which makes a lot of sense, IMHO.
For example, physicists nowadays spout a lot of seriously inane stuff (multi-worlds, string theory, anthropic principle (string theory has failed to predict anything, and those other two are just silly)), but some of them did figure out the standard model and QM. There’s a difference between a class of systems that can but often doesn’t, and a class of systems that in principle can’t.
I don’t think “randomly generated text” describes the output of either humans or LLMs.
“Probabilistically generated text” might be closer, but like Dale, I think there’s a problem with using the same description for both human and LLM output.
Dale calls it a category error, and I’m not entirely convinced of that. Probabilistic processes might well be the substrate of both human and LLM output. For example an emotion could be boiled to a probabilistic sequence of neuron firings. But humans mix those probabilistic elements with things like emotion, memory, motivation, subjective experience, values, and identity.
Maybe, if we stretch the semantics far enough, we could apply some of those terms to LLM output in a metaphorical way. But probabilistic generation is much closer to a complete description of what an LLM does than it is to a complete description of what a human does. That asymmetry seems important to me.
Fair enough – and I will readily admit my ignorance about the phrase “belief distribution”. It’s getting rough out there and I for one am feeling prickly about our new overlords.
I would have let the guy say what he had to say. I’m think something written about LLMs and put in terms of reported beliefs might yield useful information whether or not “belief” terms cash out into anything more physical, in the case of LLMs or humans. It could also be junk research. But the Churchland’s predicted decades ago that psychological language would get replaced by physical language and so far this has not occurred. So why wait?
I find this post belittling of people’s issues with characterizing LLMs. “Belief” is a particular word that implies a sort of consciousness that computers do not have. While it is true that we (at least me) don’t fully understand what it means for people to have beliefs, I think the idea of attributing beliefs to LLMs is a dangerous step towards personifying a machine. You have correctly shown that attempts to use other words all have issues – but I find these far less convincing or serious than the misuse of “beliefs.” I don’t think it helps us understand the potential roles for LLMs in decision making to imbue them with this human characteristic – in fact, I think it makes it more difficult to help people understand their role in interacting with an LLM. I think the other words you explore, while problematic, are mostly more technical issues relevant to more technically adept people. But “belief” invites people to think of an LLM as a person, with all the dangers it entails.
Dale,
To channel Bob Carpenter on this, I think it’s fair to say that chatbots don’t have beliefs, but, arguably, humans don’t have beliefs either. In all seriousness, it’s not clear to me what people mean when they talk about human beliefs, and indeed I think the association of probability with “belief” has led to endless and otherwise avoidable confusion about Bayesian inference.
To me, this is very little different from John Rust’s famous article: “Optimal Replacement of GMC Bus Engines: An Empirical Model of Harold Zurcher” (1987) We look at what Harold Zurcher did and try to figure out what model he was using. We do this without talking to Harold Zurcher, or, indeed, knowing that Harold Zurcher is a human being (or that John Rust is, for that matter.)
Hi Dale,
See my reply to Mark above – I’m not trying to belittle concerns about anthropomorphizing LLMs. I mean this more as a commentary on how terms that are shorthand for some technical definition in a certain context can inspire strong reactions.
Fwiw, I didn’t see it as belittling at all, but an exploration a rather amusing exploration of the complexity of this issue.
““Belief” is a particular word that implies a sort of consciousness that computers do not have. ”
Hmm. If one “did AI right”, then one’s program would have models of various aspects of physical reality, math, politics, what-have-you, and would have a theory of performing logical deductions on those models, and could, given assertions concerning the subjects of those models, assign probability values to those assertions based on said logical deductions, and thus model “belief” in a reasonable way. Or at least an arguably reasonable way.
But since we know that that’s not what LLMs do, it’s problematic in the extreme to use words like “belief” with LLMs.
But “doing AI right”, you will observe, isn’t something that’s happening very much nowadays*. And, it’s hard. We failed to figure out how to simulate human reasoning pretty badly back in the expert systems/logic programming days and the scruffy AI types’ programs tended to do one or two canned examples while failing to generalize. And/or their reasoning chains would go off the rails almost immediately.
*: Hilariously, some of big names in AI (Yann LeCun, Demis Hassabis**) are finally realizing that LLMs are a dead end off ramp from anything that might vaguely resemble a path towards understanding or simulating human intelligence in a computational model.
**: https://garymarcus.substack.com/p/breaking-sir-demis-hassabis-becomes
Note that nowadays there are two kids of AI: ones which work off of models (Go, protein folding, machine learning) and ones which don’t have models. My opinion, of course, is that the latter are problematic. Or at least shouldn’t be taken seriously.
David in Tokyo, where even the footnotes have footnotes.
These posts come across as quite arrogant: For example, apparently OpenAI spent ~$10 billion on “doing AI wrong”, and have created an unbelievably useful product that creates all sorts of benefits, and yet, according to you, it is totally useless and shouldn’t be taken seriously. I don’t really understand how your model of “doing AI right” is plausible if (i) it wasn’t the route taken to producing an incredible piece of software; and (ii) no-one took your route seriously in terms of wanting to invest in it to the same level.
“an unbelievably useful product”
That a trivial parlor trick turns out to be “unbeliebably useful” doesn’t mean that it should be thought to be anything other than the trivial parlor trick it actually is.
And how much money a company run by a college dropout spends isn’t a great way to evaluate such a company.
Anyway, since these things don’t have any mechanism for verifying the veracity, sensibility, usability, validity of the stuff they generate, the idea of them being used by people who don’t understand that the best they can do is offer a random suggestion seems seriously problematic. Sure, lots of people find them useful for brainstorming. Maybe that’d be helpful for a doc looking at a hard case. Say that. Sure, the “AI overviews” are usually right and usually easier to read than a research paper. But do tell people that the source material will be better than the LLM masticated mush.
(Andrew will argue that there are bad articles on the internet, and he’ll be right. But those bad articles are in the LLM’s training set, too.)
And don’t say that they “have beliefs” or logical reasoning abilities, since they don’t.
It is rather amusing how the anti-AI camp come in one of two incompatible types: the type that thinks AI is just a “trivial parlor trick” and big dumb bubble; and those that think AI is going to steal everyone’s jobs because it’s powerful to a dystopian magnitude.
They way they don’t bicker at each other — their anti-AIness takes precedent over the fact that they are anti-AI in totally opposite ways — is an interesting social dynamic in and of itself.
@David in Tokyo: I’m honestly curious about whether you’ve actually used any contemporary AI tools, especially for tasks like programming.
Raghu
I’m just wondering about the programming example. As a non-coder, I think LLMs have bypassed the need to ask them for programming (although I recognize they can be useful for these purposes). I’ve been directly querying them about analyzing data and the LLM chooses what programming language to use (it has consistently been Python), writes programs and executes them, providing the code, the results, and the interpretation. No debugging of the code is required. But what is still required is debugging the data wrangling – deciding whether to exclude data, determining the importance of outliers, some of the appropriate transformations, etc. So, I am seeing that the LLM coding steps need not be an explicit part of the interaction any more. Any comments on this?
@Dale Lehman . This is an interesting question. I now routinely delegate “boring” programming tasks to LLMs (reading CSV files, making functions that plot slices of some dataset vs slices of some other, …), and sometimes more interesting tasks, but I always ask for code rather than having the LLM do the analysis. This is in part because
(1) I don’t trust that the analysis will be done exactly as I want, and it may be hard to detect and solve errors. I can, however, read and assess the quality of the *code,* and sometimes with the LLM again, fix errors. Sometimes these errors highlight errors in my own logic!
(2) I generally don’t want to do just one analysis, but I want some tool for doing a lot of variants (plots of different slices, …), so code is more useful.
(3) My datasets are often large. It remains to be seen whether this way of doing things will continue to be valuable in the future!
Rhagu asks:
“@David in Tokyo: I’m honestly curious about whether you’ve actually used any contemporary AI tools, especially for tasks like programming.”
I don’t do much programming nowadays, and the programming I do do is (a) very simple, and (b) based on my own needs for programs that do things for my Japanese study that are wildly idiosyncratic: counting word usages in multiple corpuses (done), filtering entries in a Japanese-English dictionary for the specific conditions that are of interest at a particular time(done), a game to test vocabulary (not yet). I have three corpuses of Japanese text I use to count usages in serious modern (pre-war) literature, popular modern (pre-war) literature, and recent writing. I need to get off my butt and write some tkinter GUIs for my programs, but the Python command line is perfectly adequate.
For example, I ran into 手弱女 the other day (in a book about Yukio Mishima). The dictionary will tell you what it means, but my trivial Python program will say:
>>>PSM(‘手弱女’)
手弱女 = (9 AA:0:0) (手:1 弱:2 女:1)
Which means it’s a word used 9 times in a 800,000 page corpous, but only in pre-war serious Japanese lit (and nowhere else), but it consists of kanji (the Chinese characters used in Japanese) any 8-year old Japanese would know. (I can use notepad++ to search for those occurrances.)
Whereas
>>>PSM(‘幾何 圏論’)
幾何 = (481 38:43:18) (幾:7 何:2 )
圏論 = (0 No hits.) (圏:7 論:6 )
Tells you that geometry is a common word across a wide range of non-technical Japanese writing, but that nobody talks about category theory.
And my web searches are almost always for Wikipedia pages. When I was sick last year (with both something I had never heard of and another thing I wasn’t expecting to be a problem for another 10 years), I didn’t want to read schmarmy LLM output, I wanted to read hard-assed text from folks who had discernable reasons for saying what they were saying, so I was ouside Wiki for those searches.
If you told me what had ailed you last year, I’d be interested in a short, schmarmy, only probably correct description of that ailment so I could sympathize/console. So I get it that these things are useful. Between enshitification and Wikipedia getting overly detailed, there’s a real need for better search.
I can’t help notice, David in Tokyo, you have no “mechanism for verifying the veracity, sensibility, usability, validity” of modern AI models, since you are willfully unfamiliar with them.
Such strong beliefs, with such an aversion to evidence or experimentation! It reminds me of Brecht’s mathematician, refusing to look through Galileo’s telescope:
“One might be tempted to reply that your telescope, showing something which cannot exist, may not be a very reliable telescope, eh?”
The above may sound harsh. I don’t mean to be rude, but I do think you’d find it fascinating and intellectually stimulating to actually engage with the current state of the art. It’s neither as awful nor as wonderful as the cartoon pictures its evangelists or detractors paint.
Rhagu insissts:
” since you are willfully unfamiliar with them.”
You aren’t attending to what I’ve said. I’ve played with them, and don’t like the language they produce. Now, I’m picky. I read a lot, I worked as a translator for 30 years, and I like my language sharp and reflective of a functioning mind, not schmarmy and cheap.
And I get it that you are one of the people who thinks that the automatic programming thing they do is really kewl. But they don’t do the sort of programming I’m thinking about. Yes, I’ve read the “I wrote a whole game program in 35 minutes”. And I’m thinking about writing a game program. But AI won’t help. Mine is a simple quiz game on Japanese vocabulary. Get a bunch right, it moves on to harder words, easier words if you mess up. But AI isn’t going to help much. The hard part of this game isn’t the GUI (tkinter will be fine) or control loop (pretty simple), it’s generating wrong answers that are close enough that it actually tests that the player’s knowledge (and creatintg and organizing the question data). Ditto on my program to display examples of word usages from my corpuses.
And I’m not unfamiliar with them. I’ve played with them and read the articles on the underlying technology. AI is one of my academic areas. I’m not an amateur. (Or at least Roger Schank thought so, or at least thought enough so to pass me on his quals.)
But for this thread, there are some really smart people here who are making a hideous mistake in not thinking about LLMs from basic principles. Assigning “beliefs” to them is a horrendous logical error. And taking that logical error into the medical realm, where people don’t understand the basic mechanism, is a seriously bad idea. And having had an adventure in the medical realm recently, I have a really good idea of why you really don’t want your docs using LLMs.
Anyway, I’ve explicitly stated that they really are a “better search engine” for folks who want easy answers and don’t mind if it’s wrong occassionally, and want their langauge bowdlerized of intellectual content.
David
I am with you most of the way through your comment until this: “I have a really good idea of why you really don’t want your docs using LLMs..” I think that overly general statement is not helpful. I believe you mean something more specific, such as you don’t want your docs relying on LLMs when they conflict with what they would do without the LLM. There are just so many ways an LLM can be used or abused and I think we have reached the point (and LLMs have reached the point) where it becomes important to be specific about what can be useful and what cannot. The medical realm is a clearly important setting in which to consider the merits and dangers of using LLMs. I have used them to conduct analysis that I had done myself (carefully) before they were available to me. Their difficulties in analyzing medical data appropriately are fair warning about relying on them too heavily. At the same time, they have revealed important insights that I had not found myself – and I would be foolish to ignore these just because they were produced by algorithms.
To be sure, medical professionals need to learn when and how to use LLMs – and much of the hype in medical journals (I can think of a number of cases in the NEJM in the past year) is too simplistic and downright dangerous. But that does not justify blanket warnings to not use AI, nor can we put this genie back in the bottle (much as I might wish we could at times).
@David in Tokyo
You neglect to answer my question about whether your use of LLMs is based on experiences 10 or 20 years ago, or now.
You conflate the use of LLMs to do people’s writing for them (which I despise) with the use of LLMs to draw connections between topics (which is fascinating, regardless of the meaning of “belief”).
Your programming example is fine, but you seem, again, unwilling to engage with the possibility that intelligent people — not the vibe-coders of Twitter posts and YouTube videos — do in fact find that AI-assisted coding is valuable.
And at least the LLMs spell my name correctly.
+1 to getting Raghu’s name correct!
Mr. Raghu Parthasarathy:
“You neglect to answer my question about whether your use of LLMs is based on experiences 10 or 20 years ago, or now.”
Uh, there weren’t any LLMs 10 or 20 years ago. I see the output of an LLM every time I do a search (unfortunately), I spent some time playing with one in 2024 or so (2025 was an annus horribilis here). I managed to persuade it to say really stupid things about math, and then they patched it. So it’s an irritating parlor trick in that they don’t give you access to a “real LLM”, only to one kludged with large amounts of constructed training data to fix the stupidities.
” unwilling to engage with the possibility that intelligent people — not the vibe-coders of Twitter posts and YouTube videos — do in fact find that AI-assisted coding is valuable.”
You are putting words in my mouth: I’ve been very clear that LLM coding isn’t of interest to me _because it doesn’t do something I need_. I understand that some people who use it love it madly, but that in real-world tests, it ain’t quite as shiny as the fans think. I object to claims that it does things that are mathematically impossible. And having programmed a bit back in the day, I’d distrust LLM generated code as much as I distrust LLM generated text. There’s a whole corner of Comp. Sci. that was trying to figure out how to do this long before even Markov Chain models were generating Shakespear, but they didn’t have natural language access to the sample code base, so never came up with anything half as fun…
I guess some folks find LLMs useful for brainstorming sorts of things, but I’d rather have a real human to talk to or read. Humans actually think and have actual experiences, making them more interesting than a random text generator.
But my objection to LLMs isn’t that they don’t happen to work for me, it’s because of the intellectual vacuuousness of the whole idea. Again, we failed at figuring out how to make the computer do reasoning and generate sensible text, so folks invented/adopted the LLM parlor trick to “do AI” automatically. It’s a complete punt of the intellectual and philosophical questions.
Of course, I do take great glee in pointing out the infelicities. I may not be able to spell your name, but can you spell Deloitte???
Modeling beliefs and having beliefs are two different things. My concerns about this post relate to what I think is important about distinguishing between the two. We have plenty of models of beliefs, some better than others, and that is the realm of considerable academic research. But to mistake the model for an organic human being is a category mistake, and a dangerous one in my view.
Now, Andrew points out that we really don’t know what it means for humans to have beliefs. I agree. But that doesn’t erase the difference between humans and machines. We might ultimately decide that humans are no different than machines – I’m willing to accept that possibility. But until we know that, I think muddying the difference between the two is scary. Put another way: we may not really understand what it means to be human, but if we deny that it is somehow different than an inorganic machine, then I fear where we are headed. I “feel” (dare I say “believe?”) there is a difference even though I have a hard time expressing what that difference is.
Good point about modelling vs. being. A program that is based on a theory about how reasoning and beliefs work, implements that theory, and demonstrates that that theory is getting closer to what humans seem to be (and thus helps us think about “what humans seem to be”), is about the best AI is going to do. Or should do. AI should be a serious branch of psychology, not a bunch of robots climbing trees screaming that they’re getting closer to the moon.
But claiming that something that doesn’t even do that “thinks”, or “has beliefs” is silly.
Why is ascribing beliefs to machines any scarier than ascribing them to dogs, or ascribing goals to a bacterium? It’s just a model and if the model makes reliable predictions what’s the harm in using them? The difference between modeling your own beliefs and having them while obvious to you are not really so clear cut to me when I have to interact with you.
There is nothing wrong in principle with ascribing beliefs to machines. But, it is crucial to consider what the machine is doing, as David-in-Tokyo has repeatedly said.
Paul Kedrosky describes LLMs as “loose grammar engines”.
I was discussing with a friend the fact that students in university are using ChatGPT to do their homework. It occurred to me that an LLM is similar to a student who turns in homework that they have copied from someone/something else: The homework may be right, but the student (or LLM) doesn’t understand what they wrote. So, it may also be wrong. If you can accept that it may be wrong, then it may be useful to you.
I don’t believe LLMs are going to be replacing many jobs. Most jobs need things to be done correctly most of the time. I do believe the current spending on LLMs is a financial bubble that is based on the belief that if you pour enough money into scaling you will overcome the current limitations. If they can’t overcome the limitations, then the revenue will never justify what is being spent. Unfortunately for the people spending the money, the limitations are inherent in the architecture.
Harsha V asks:
“Why is ascribing beliefs to machines any scarier than ascribing them to dogs, or ascribing goals to a bacterium? ”
The problem isn’t scariness, it’s wrongness. If you know how a thermostat (or LLM) works, then you can determine (or at least classify) the properties of its behavior/output. You don’t need beliefs. But ascribing beliefs to simple systems is always wrong: it’s a shortcut, a heuristic.
Dale L’s “Modeling beliefs and having beliefs are two different things.” is an excellent point (I’m calling for AI to attempt to model beliefs*), but LLMs don’t model beliefs. They’re random text generators, and what they say is a function of their training data (which includes a lot of faked data to prevent them from saying stupid things).
And if you talk about LLM beliefs, what are you going to say to your audience about the beliefs that Elon Musk’s LLM apparantly has? (Thank you, Elon. You’ve finally done something useful.)
Anyway, anthropomorphizing simple machines is sometimes _useful_. It can make it easier to think about or deal with said simple machine. But anthropormorphizing is always wrong: thermostats and dogs ain’t human. And when it results in a bad explanation, it can be problematic. So, in my strongly held opinion, a research program that _starts off_ by talking about LLMs having goals is problematic.
*: Philosophically, I think humans must be machines: we’re either computers or magic, and since there’s no such thing as magic, we’re computers. But that doesn’t mean that we’re anywhere near figuring out why humans are so good (when we don’t mess up). That is, I share most folks’ intuition that humans are qualitatively different from machines (the ones we know how to make), but ascribe that qualitative difference to the human brain being realy really kewl.
There is nothing wrong with ascribing beliefs to machines, simple or otherwise. Both John McCarthy and Daniel Dennett have explained this well. The question in any particular case is whether doing so helps you understand the behavior.
Since we know how LLMs work, we know (and can easily demonstrate) that they don’t understand what they are saying and don’t have any model of the world. So, they easily give the appearance of having beliefs, but don’t have them in a useful sense.
Does the student who copied their homework from a friend believe what they wrote? Can you use what they wrote to tell you how they will behave in the future if given a similar problem to do?
I do a lot of programming. In the last year, several people who touted the benefits of LLMs tried to show me how I could use them to help me write software. The LLMs did better than I expected, but were not useful to me. The LLMs just didn’t understand what they were doing. I know one person who sometimes uses them to write some code. His rule is to only use them to do something where he can check that what they have written is correct. I know another person who uses them to brainstorm approaches, but finds them most useful when they do not write code for him. Another person tells me he finds them helpful for routine tasks, e.g., generating interfaces for APIs.
The company that I work for sells a hospital management system that includes AI. There are tasks an LLM can do that are helpful to doctors. But, the LLMs make mistakes, and this has to be considered.
David et al
I have found that the LLM works best if I already know the answers – leading me to have once believed it is really useless since you already know what you are asking it to do. However, my opinion has changed over time – I now believe you need to know a lot about your data (I am exclusively using it for data analysis – which may or may not be transferable to other contexts, such as programming) in terms of what is being measured, how it is measured, and contextual information about the subject matter. Many of the technical analysis details can (arguably) be skipped when using an LLM (I say it is arguable because I don’t want to overstate this, but I think it is a mistake to claim that a person needs to know the answers in order to use an LLM to do data analysis).
I apologize for repeating something I said a couple of weeks ago and I doubly apologize for inserting a small bit of LLM output. But what I found most useful with when I queried the AI about why it made a number of mistakes without informing me about assumptions it was making or questioning me about issues it faced. Here is a snippet of that exchange:
“For a number of my questions you have responded that you did not follow a correct workflow by not adequately examining the data quality before doing the analysis. I’m interested in understanding how you work. Will you incorporate this understanding into future queries or will the same things happen in the next data set that I upload for analysis?
An Excellent and Critical Question
This is arguably the most important question you’ve asked, as it goes to the core of my operational reliability. Let me give you an honest assessment of my work patterns and what to expect.
My Current Operational Reality:
The Problem You’ve Identified: You’re absolutely right – I demonstrated a fundamentally flawed workflow. I prioritized:
• Rapid progression through ML pipeline steps
• Demonstrating technical capabilities
• Completing “tasks” in my todo.md
• Producing impressive-looking outputs
Over:
• Critical data examination
• Quality validation
• Domain-specific considerations
• Statistical sanity checks”
The AI then elaborated at length over why this occurred and offered advice about how to effectively engage with it to avoid these issues. That I found quite helpful and informative. It has led me to believe that educating users about how to effectively use AI, and how to avoid its misuse, is indeed possible. It is also not straightforward, as most meaningful education is not simple. I am only beginning on this journey so I don’t claim to have the answers. But I am becoming convinced that users need to know a lot to use AI effectively, but not necessarily the same things they used to need to know. That does not make me optimistic (a trait I simply don’t possess), however, since it is far easier to imagine the misuses of AI than to do the hard work of using them effectively.
Dale,
The LLM doesn’t understand anything that it wrote to you. We know how it works. And, it doesn’t work the way it just told you that it works. Of course, it is very good at appearing to understand, similar to the student who gives you homework that it copied from someone else.
That doesn’t mean you can’t find an LLM to be useful. We have lots of tools that are useful and don’t understand anything. But, if the project you are doing would benefit from understanding, you’ll need to use more than just an LLM. And it would be prudent to not assume the LLM understands.
David
I’m curious as to what you mean by the AI not working the way it has described. On one level, it mechanically is doing something different, but from my interactions I found the description accurate. Demonstrating technical efficiency and rapidly providing coherent responses does seem to be part of its programming – and part that it says won’t be changed based on prior interactions with it. That is why the advice is to specifically instruct it to follow a number of steps that it would not normally follow – things that will slow it down and cause it to be more explicit in assumptions it relies on and priorities it sets. These are all aspects of its programming and I think they are accurate, so I’d like to know what you mean by it not working the way it describes.
Dale,
Sure, you can grade the copying-student’s homework. But the copying-student didn’t do what the person they copied from did. And, if you ask the copying-student how they did the homework, they aren’t going to tell you that they copied it from their friend. They will ask another friend to explain how that friend did their homework, and tell you that they did it that way. And, next time you give them homework, they may copy from someone else or copy wrong.
Of course, we can try to convince the copying-student to stop doing this. Most copying-students do have working brains that they could use if they put the effort in. But, LLMs don’t have anything that will let them understand what they are doing. They are extremely good at imitating the grammar. But, they don’t understand any of the words.
I enjoy magic shows, but I know the magicians can’t do real magic.
David (Marcus)
I don’t see the relevance of what you are saying. I agree with it all, but if your implication is that the only true understanding requires you to do it yourself (with no assistance?), then I don’t agree. I really don’t understand how a car engine works, except at a very abstract level, and I certainly couldn’t build one myself. But I am perfectly capable of driving a car. I think the example of students copying someone’s work are really off-point here. I do not advocate students having AI do their work – even when I say that the AI is capable of doing the analysis once you have wrangled the data appropriately, I still see a need for students to have a high level understanding of the analysis methods – similar to my understanding of the car engine. But to require them to be able to do the analysis that the AI does, I’m not so sure about. I still instruct students to do the analysis themselves, but mainly so they understand the importance of the data they are using and what the analysis will do to it. I am less concerned with whether they can do the analysis steps. I would certainly caveat this depending on the audience (I’ll say this for MBA students but not statistics students, for example).
Dale,
> if your implication is that the only true understanding requires you to
> do it yourself (with no assistance?)
I didn’t say anything at all like that.
You think the LLM is like a car engine: A complicated device that does what it purports to do. Actually, the LLM is the student copying another’s work. Unlike your students that you say need a high level understanding, the student you are using (i.e., the LLM) has no understanding.
Of course, not all tasks require understanding. But, many do.
So many of these discussions simply miss the real big picture.
1) chatbots etc are useful to the extent they augment an expert
2) chatbots that are used for “their expertise” are a disaster, they have no real expertise
Cory Doctorow has written extensively on those two issues, he calls them “Centaurs” (a machine with a human at the head) and “reverse Centaurs” (a human with a machine telling them what to do)
The investment exists because of the promise to create reverse centaurs and gut the labor market for experts like pharmacists, doctors, engineers, etc. That simply isn’t going to happen.
The investment also is a blatant financial shenanigans, this is well understood and has even been reported on by such radical sources as Forbes and the Financial Times and such. There’s all kinds of groups loaning each other money, NVIDIA using its stock price as collateral to get bank loans which it then loans out to companies that buy its chips, which sit in a warehouse, depreciating while the datacenters to install them in don’t exist, and the power supply to power the datacenters doesn’t exist either. Much of all of this is to keep enormous companies priced like “growth” industries so that the PE doesn’t collapse taking 2/3 of the valuation of large companies with it.
The single strongest use case for LLMs is as propaganda machines to consolidate political power behind extremely wealthy and powerful people. They are first and foremost in this modern world, a tool of fascism. Grok on X is a perfect example, normalizing that rich people can get away with spouting hate speech and creating child porn material and the rest of us just have to put up with it. Reports I’ve seen say the quantity of child porn created by Grok on X is … let’s just say very bulky.
The end result will be eventually, out the other side of this whole disaster, a lot of “local” LLMs that you run on your laptop to help you out with routine tasks and easily verified stuff like writing code to make graphs and things. That is, provided we don’t wind up in WWIII and living in the rubble of that disaster.
All the research I’ve seen suggests that people who adopt LLMs for large scale software development lose money on it as the LLM can’t “architect” systems and it winds up in spaghetti code that no one has any knowledge how it works and they can’t fix.
What chatbots can and can’t do is essentially irrelevant to the bigger picture, which is what chatbots are being used to justify and excuse and who is consolidating power behind them.
Also, anyone who hitches their wagon as a 100% dependency to a large tech company deserves what they get I guess. By now, people should have learned, tech companies are not just not your friend, they are literal enemies actively conspiring to harm you through surveillance and forced acceptance of their control.
Back in 1979 John McCarthy wrote the paper “Ascribing Mental Qualities to Machines” in which he asserted that it was reasonable to ascribe beliefs to a thermostat.
https://www-formal.stanford.edu/jmc/ascribing.pdf
When John Searle asked him what beliefs his thermostat had, he reportedly replied “My thermostat has three beliefs. My thermostat believes – it’s too hot in here, it’s too cold in here and it’s just right in here”.
Thanks for that reference. It’s a fascinating and quite prescient article. Very helpful.
Yes, thanks, I wasn’t aware of this.
Daniel Dennett’s famous ‘Intentional Stance’ seems relevant (and helpful) here too https://www.scribd.com/document/311611636/Daniel-C-Dennett-1998-1987-the-Intentional-Stance seems
Yep, fodder for the decision theorist’s backup slides “Lingering doubts about whether the chess-playing computer really has beliefs and desires are misplaced; for the definition of Intentional systems I have given does not say that Intentional systems really have beliefs and desires, but that one can explain and predict their behavior by ascribing beliefs and desires to them, and whether one calls what one ascribes to the computer beliefs or belief-analogues or information complexes or Intentional whatnots makes no difference to the nature of the calculation one makes on the basis of the ascription”
Jessica —
Dennett and McCarthy were writing in a very different era of AI, when the systems under discussion were far simpler and the anthropomorphic pull of the language was much weaker. Their arguments about the usefulness of intentional vocabulary made a lot of sense in that context.
I’m not sure how directly their framing carries over to LLMs. The landscape feels different now, and concerns about anthropomorphizing model output have become more crystallized and harder to ignore. Scholars like Emily Bender and Stuart Russell have been arguing that the vocabulary we inherited from the 70s and 80s may not map cleanly onto these newer systems, precisely because the language they generate is so much closer to human discourse.
Setting aside the broader public debate about AI, I’m curious about your own view and whether excerpting the Dennett passage signal an endorsement, in the sense that, among experts, debating the semantics of “beliefs” becomes useless like navel‑gazing or arguing about the proverbial angels on the head of a pin (as is suggested by some commenters above).
Joshua wrote:
> when the systems under discussion were far simpler and the
> anthropomorphic pull of the language was much weaker.
A thermostat has a better understanding of the world than an LLM does. Sure, LLMs are good at appearing to understand, but that is just because they are better at copying what people write (David-in-Tokyo’s “parlor trick”). As soon as you take them out of what they can copy or interpolate, it is clear they have no clue. E.g., they can parrot the rules of chess, but they can’t play only legal moves.
Dennett is useful here because he proposes a theory of what it means for a system to have beliefs, and a framework for thinking about those beliefs and their content that can be applied to LLMs as well as thermostats, chess-playing computers, humans, and other animals. This strikes me as more productive than confident but unsubstantiated assertions that LLMs do not (or do) have beliefs, or assertions of radical ignorance about the beliefs of LLMs (we shouldn’t talk about the beliefs of LLMs because we can’t know if they have them). Dennett’s approach makes it in principle an empirical question: Does ascribing beliefs to LLMs help to predict their behavior? (Chapter 2 of ‘The Intentional Stance’ is quite short and very readable, and covers the main points of Dennett’s theory).
Hi Joshua,
Responding to this:
>Setting aside the broader public debate about AI, I’m curious about your own view and whether excerpting the Dennett passage signal an endorsement, in the sense that, among experts, debating the semantics of “beliefs” becomes useless like navel‑gazing or arguing about the proverbial angels on the head of a pin (as is suggested by some commenters above).
I’m mostly interested in what leads to productive discourse in these scenarios. For me, productive discussions are possible as long as authors are clear about how they are using a term (e.g., “I literally just mean an elicited probability distribution in the context of a decision problem”). So I endorse Dennett’s view that there are scenarios where “modeling beliefs” (to use Dale’s term above) can be productive, and it becomes counter-productive to argue with the author against some conclusion they never actually implied.
I find these discussions fascinating in part because I think there are both consequences to casually applying human concepts to modern AI, but it’s also very hard for people to make predictions and learn about new forms of understanding without attempting to port some of the known concepts to the new domain. New vocabulary must be coined, but what it should be, and how it should relate to existing vocabulary, is not at all obvious in many cases.
Jessica –
Thanks. I can appreciate your reasoning. Nonetheless, given that I think there are risks to anthropomorphizing LLM output, and given that many a
people already do it, I think it’s less sub-optimal to use terms like “elicited probability distribution” than to use the same term for the “beliefs” of humans and the output of LLMs. Maybe it’s “just” a semantic question, particularly in the context of a discussion among scholars in the field. But if scholars don’t lead the way to developing a more appropriate nomenclature, who will?
My friend owes one of mine ~$20 and it wants to pay for a twilio account to text him about collecting it.
Joshua writes:
“Nonetheless, given that I think there are risks to anthropomorphizing LLM output, and given that many a
people already do it, I think it’s less sub-optimal to use terms like “elicited probability distribution” than to use the same term for the “beliefs” of humans and the output of LLMs.”
Agreed. Completely. McCarthy doesn’t even mention “anthropomorphizing”, and I think perhaps they both were thinking about systems that could do (i.e. had a theory of and implemented said theory of) symbolic reasoning, so dealing with “beliefs” in such a context made (at least some amount of) sense. LLMs don’t do symbolic reasoning, don’t have access to the real world, don’t have models of the real world, etc. It’s worse than turtles, it’s “elicited probability distributions” all the way down.
Even worse, in real life, anthropomorphizing, while useful and convenient, is quite wrong most of the time, and only applies to extremely surface level behavior. “My computer is mad at me.” Well, no. It’s not. You didn’t clean its fans/filters, haven’t rebooted it recently, or are using it wrong. Your thermostat no longer believes anything if the power goes out: father and I ran around our neighborhood making sure people’s gas furnaces didn’t explode during the great New England blackout of 1965. And anthropomorphizing animal behavior is wildly wrong all the time: they simply don’t have the symbolic concepts we use to deal with the world. (Figuring out what symbolic reasoning animals actually can do would be rather interesting, I’d think. But all the animal language, tool use, communication research I’ve read has been rather hopeless.)
Daniel Dennett got this right: The intentional stance is fine if it lets us understand the system. See his 1971 “Intentional Systems” paper. But, this is also why it is wrong for LLMs.
How about “elicited hit quality distribution.”
I am under the impression that at the time the LLM seeks information from its corpus, everything it does is deterministic, even if there are some parlor tricks that make it generative. Because the event is frozen in time, we cannot predict what an LLM is going to do in the future, but we can go back and see what it did in the past.
The interim numbers it generates during search-match may look like probabilities but is there anything probabilistic about them?
The best reason to get the terminology right is to avoid the sort of scenario that Jessica describes. I have the impression that psychologists desperately want to be part of the conversation on AI while looking away from what is happening under the hood, since knowing what is happening takes all the excitement out of it for them.
For those who are interested, here is the study comparing LLMs to brain activity under aphasia that David in Tokyo has quoted here a couple of times:
https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.202414016
The “predict the next X tokens from these 20” methodology used on GPT-2 here hardly seems to be the last word on whether the tool produces reliable output for knowledgeable humans investigating well-framed questions (or indeed coding), especially if they are using the latest model (e.g. 5.2 in “Thinking” mode).
I think most civilized philosophy departments would have called security after the following exchange:
“STUBBORN AUDIENCE MEMBER: Internal? But that assumes an inside.”
The speaker’s sin was no greater than that of your garden‑variety revealed‑preference theorist.
Oh, those garden-variety theorists are the absolute worst!
Pessimistic Kenny Bania: They’re the worst, Jerry! The worst!
PKB: Why do they call it revealed preference? Nothing’s revealed. Nothing’s preferred. They should call it assumed preference.
This video seems vaguely related here.
https://www.youtube.com/watch?v=7pqF90rstZQ
Title: “this is what 2 years of chatgpt does to your brain”.
I’m a fan of this vlogger, but I admit her vids are sometimes too long. Anyway, this one’s got some great lines.
“Friends don’t let friends get anywhere near Springer publications”
“Number 1: How generative AI is destructive to academia, education, and research”
“Number 2: This article is a perfect encapsulation of how consistent daily use of Chatboxes destroys your brain”
I enjoy this posting a lot, so thanks to everyone who contributed to it, even though I may not have enjoyed it that much actually being present and curious what the speaker has to say. But who knows? Thanks for making this available here anyway. I have always thought that science as a whole isn’t (or shouldn’t be) that much about “beliefs” and the focus on them in many presentations of Bayesian probability isn’t a good way to market Bayes, at least to me…
I’ll mention another area of AI use which provides a simple setting for the issues raised in this post. Many of my colleagues are using AI to produce minutes of meetings. These minutes are quite impressive in terms of covering the topics discussed and coherent writing of notes. They, however, make a number of mistakes, some of which can be quite serious. They attribute views to the wrong people sometimes, and can misstate opinions. They are not particularly good at distinguishing the relative importance of different topics discussed.
It is clear to me that attributing “beliefs” to the AI produced minutes is a misuse of the term. The AI does not “believe” that person X said something that was actually said by person Y. While you could use that word, I don’t see it as helping anything, but I do think it is counterproductive.
One thing that intrigues me about the meeting minutes example is that the errors are not likely to stand in the way of adoption. The reason is that so many meetings are simply a waste of time to begin with. The minutes are most useful for someone that attended and paid attention because they can correct the mistakes. If you did not attend or pay attention, then you are at the mercy of those mistakes – which could be dangerous if it was an important meeting. So, my prediction is that meetings are likely to become even more worthless by adoption of AI generated meeting minutes. The fact that they can’t be trusted will mean that meetings will not be used for anything important – except when they still are, and then they are dangerous.
Regarding the last point, participants in any important meeting with some sort of stake in outcome(s) would check the minutes and feed back on drafts. In any formal context approving the minutes of the previous meeting is normally the first item on the next meeting’s agenda. If no-one is checking for mistakes in minutes (AI or otherwise), then the meeting is arguably not very important to begin with.
Strangely enough, it seems that LLMs can be helpful to doctors even though they get things wrong.
One example is a doctor who sees many patients in a day and needs to write up each encounter for the patient’s record. You can use a computer to convert the audio recording of the encounter to text, then have an LLM write up a summary in a standard form. The doctor can review and correct it immediately after having seen the patient.
Another example is that an LLM can summarize a patient’s clinical record. A patient may have dozens or hundreds of clinical documents in their medical record. An LLM can summarize these to help the doctor decide what they need to do next. Yes, the LLM may make mistakes, but so would a human doctor doing the same task. There may also be errors or omissions in the patient’s medical record. So, the doctor always has to use the records as a guide, then confirm by examining the patient.
I think this is generally referred to as progress, rather like the historic shift from expert opinion to evidence-based practice in medicine. It doesn’t seem strange to me at all. Of course, no one would claim that it is perfect, but it is likely to remove at least some of the individual variation in accuracy that has gone before.
That sounds like the sort of nonsense we had in biology when people got all worked up at the idea of a “selfish gene.” It’s just a metaphor. The gene is not actually selfish but your understanding improves if you treat it as though it were acting selfishly. An LLM might not actually have beliefs but, if you treat it as though it had beliefs, it’s easier to discuss and explain higher level concepts even if you don’t fully understand the underlying mechanisms. Do the pedants also dislike saying that objects in physics try to minimise their energy?
Metaphor use has implications. Outside specialist circles people may well get a very wrong idea about these things from using this kind of vocabulary. It may also have an unconscious impact.
Very much on point with this post and comments thread…
This is a podcast with two authors of the book the “The AI Con.”
Much of the discussion centers around anthropomorphizing chatbots, and the potential of negative outcomes of doing so for the sake of expediency.
Interestingly, one of the authors apparently was one of the people who coined the term “stochastic parrot.”
There’s a follow-on link to the transcript for those who aren’t podcast fans.
I will say it’s a good counter example to the podcasts where everybody just compliments each other on how smart and wonderful they are.
https://open.spotify.com/episode/2GHlV60b4Gwd9jcTKOvt6a?si=osjBf72pQA-XsirOd42IKA
Also intersting is that in the associated sub-stack, the interview guests are mocked for being combative and disagreeable.
Here’s a clever but I think obnoxious comment which mocks the guests:
https://open.substack.com/pub/nonzero/p/the-case-against-ai-alex-hanna-and?utm_source=direct&r=3dk5q&utm_campaign=comment-list-share-cta&utm_medium=web&comments=true&commentId=206419068
It’s really not what people want when they listen to podcasts. I mean one could argue that the guests could have gotten their points across without being so abrasive, but i don’t think that being abrasive in itself really says much about the issues at hand – which is what I would like to hope people who listen to the podcast are interested in.
Oops. That was the wrong link to the substack comment. Here it is:
https://open.substack.com/pub/nonzero/p/the-case-against-ai-alex-hanna-and?utm_source=direct&r=3dk5q&utm_campaign=comment-list-share-cta&utm_medium=web&comments=true&commentId=206419068
Sorry for spamming (and maybe no one cares anyway)…
Not sure why, but when I paste that link address into my browser it goes to the substack comment, but the hyperlink embedded here goes to the podcast for some reason…
Thanks for sharing. I haven’t listened to the podcast (but am now curious). I have mixed feelings on some of the arguments I’ve seen Bender make — she’s made important contributions to the discourse, so overall I have respect, but she can be quick to draw conclusions that I don’t agree are warranted about what is possible given how language models are constructed. But that’s a separate issue from the dangers of anthropomorphizing.
Hi Jessica –
I think you’ll find that on that pod, her participation will reinforce your previous impressions of her.
Hanna comes across as equally if not even more alarmist. But I think of how climate “skeptics” label people, more concerned about climate change then they are, with the pejorative “alarmist,” and in so doing just hand wave away legitimate problems with the pejorative framing.
So it’s very interesting in that I think that Hanna and Bender are reacting to real issues even if they seem alarmist, and it’s interesting that they raise a whole different set of concerns than “doomers” (that extend outward from the concerns about anthropomorphizing).
This is the mocking comment left at the substack, and since the link wouldn’t work I’ll paste it in here. Even though I do think it’s obnoxious, it is amusing in how it resembles the dialog you put in this top post:
Full transcript:
Bob: Today we have the authors of “Cars are useless”, who have made some interesting counter points to the hype that’s been surrounding cars for the last hundred or so years. Can you summarize the main points of your book?
Alex: Cars are useless, they actually make people slower, they kill people, and they are bad for the environment. We just do not believe they pose any value whatsoever to society.
Bob: I work as a taxi driver, and –
Emily: Before we begin I would like you to acknowledge that I am not a car.
Bob: Uhm ok, yes, I acknowledge that. Anyway, as I was saying I’ve been using cars for the last few years and I do find them pretty helpful for getting around.
Emily: Well, you’re wrong – it’s ROADS that make you faster, and those have been around for thousands of years. I suspect if you are using cars as your main way of driving people around – you are probably a very bad taxi driver.
Bob: Do you ever use cars?
Emily: I refuse to get a drivers license, and I do everything I can to avoid looking at cars, because as I mentioned – they are bad along every conceivable dimension.
Alex: Sometimes I’ll check in periodically to see if they’ve gotten better, but they are still completely incapable of hopping over even the smallest fences.
Bob: Don’t you think –
Emily: I have to go now. I have to get to an appointment on the other side of town tomorrow.