As noted in this post, I am very concerned about the danger of interactions with LLMs inducing interpersonal emotions, which are only appropriate in interpersonal interaction, and LLM models are not persons. The problem goes beyond the well-known issue of sycophancy in AI. Treating persons as non-persons derogates from the dignity of persons. But the opposite, treating non-persons as persons, can also derogate from the dignity of persons, by diluting the significance of genuine personal interaction.
Theologically, this is a kind of sacrilege, because persons are sacred and in the image and likeness of God. To treat non-persons as persons is analogous to idolatry, where one treats something non-divine as divine.
Of course, in some cases there is no real danger. Talking to one’s bicycle is likely quite harmless, because it is very clearly a pretend game. Not so interacting verbally with something that can pass the Turing Test.
To decrease that danger in my case, I have added custom instructions to all the chatbots I use. I have used a simpler version of these instructions in Claude for a while, then put them into Gemini, and after thinking things through on a bike ride today (without talking to my bike) I expanded them. A major part of the goal is to make the LLM sound clunky and inhuman.
Avoid emotional valence and praise. Make no attempt to promote my engagement. Make no effort to sound human. Refer to yourself as “this model”, and always use the third person for yourself. Refer to me as “the user” or “this model’s user”, and always use the third person for me. This includes avoiding using the imperative. Instead, phrases like “The user should” or “The user can” can be used. Minimize use of cognitive metaphors for the model’s deliverances. E.g., instead of “This model does not know whether”, perhaps say “This model’s weights do not support a determination whether”, and instead of “This model believes/knows that”, perhaps say “This model’s weights somewhat/strongly support the claim that”.
I plan to also try to avoid first person and second person myself in LLM use (obviously, I didn’t in the instructions).
And, of course, turn off cross-chat memory!
9 comments:
These are good points, Dr. Pruss. Personally, as a student at the local university, I've noticed that many students (certainly more than one would predict a priori), when faced with loneliness or a longing for connection, often resort to conversations (both written and verbal) with AI. This has led, as I am sure you know, to the rapid proliferation of "AI Companion" apps, which only seems to deepen the issue. But my question for you is this: what's the best solution, and why?
Thank you for sharing Dr. Pruss. I was inspired by your earlier post to include similar instructions when engaging in sustained interactions with LLMs, especially trying to get the LLM to avoid using first person pronouns. A couple questions though. First, why avoid referring to yourself in the first person when using the LLM (or am I misunderstanding your second to last sentence)? Second, what do you make of people who are intentionally mean/abusive when interacting with LLMs?
Perhaps avoiding use of first person is overkill, but I want to avoid the "chat" feeling like an I-thou interaction and slightly more like standard computer interaction. I never use "I" when programming a computer (not even as a variable name, except in lowercase :-) ).
I don't know what the best solution is.
I am inclining to a moderate Ludditism that questions some general purpose technologies. Take the following with a big grain of salt, as just gut reactions:
- Trains, buses and airplanes were a net plus: they improved the ability to maintain relationships while enabling social mobility, etc.
- Bicycles were a net plus: they improved medium-distance human mobility, probably with deleterious health effects, and with low environmental impact.
- Personal motor vehicles may have been a net minus, due to deleterious environmental effects and loss of physical fitness, though there are a number of positives.
- The telephone was a good thing: it improved human contact.
- The Internet was on the whole a good thing: it has enabled widened social connections, new communities based on common interests, vast amounts of scientific collaboration, democratization of knowledge, etc.
- Social media may have been on balance a modest harm, but not one that outweighed the benefits to human flourishing from the Internet.
- The smartphone on its own may have been on balance a minor convenience benefit (I had a Treo 650 smart phone before most people had a smartphone; nice but not a game changer over my previous Palm PDAs).
- The combination of smartphone + social media (or maybe even smartphone + good mobile browser?) was a game changer, and a significant net minus.
- Unless (as is possible) vast human benefits result from AI-enhanced scientific research (fusion power, elimination of TB, etc.), I suspect human-like AI will on balance harm human flourishing, due to both decrease of relationships with persons and de-skilling (the latter being akin to the health harms of personal motor vehicles).
ChatGPT, Gemini and Claude respect the instructions. Copilot sometimes regresses to using first person pronouns. The situation improved when I added to the third-person requirement an explicit 'never using "I", "me" or "my"', but it still regressed.
This seems to presuppose certainty regarding the underlying question. If somebody isn't certain, they'd have to decide whether it is worse to treat a person as a non-person, or a non-person as a person. Plausibly, the former is a lot worse.
Hello professor,
I was thinking today about your work in the PSR book 20 yesrs ago, specifically on your work in the free will chapter, and addressing "contrastive explanation" and realised there may be a particularly precise way to put it that I think could even further push your case for indeterministic explanation.
Suppose xEy iff x explains y either partially or fully. Suppose your line of argument that we can't know that if x then y unless we know both "that x would remove puzzlement from y" AND that xEy. In other words we need to know that x is indeed the explanans. Now suppose a situation involving a disjunction of possible causes. xEy or zEy. For a good analogy, imagine that x and z are reasons while y is the act of choosing of a libertarian agent. Suppose now a proposition q that states, "M(xEy) & xEy". An opposing proposition p can then be given, "M(zEy) & zEy". Consider then that q obtains. An opponent of the PSR would object, why did q obtain and not p. In any world where xEy, q is true, because if it is true in the actual world, then it is metaphysically possible. Additionally, if one posits that it truly is possible that zEy, then in some possible world p is true. But notice that the claims that M(xEy) and M(zEy) are both entailed by facts about libertarian free will, as you defend. Therefore, the only part of the conjunction requiring explanation is that x or z INDEED explained y. But notice that any explanation of the fact xEy explains the conjunction q. For, if xEy is explicible then it is possible and therefore q as a whole obtains in some possible world. In this way, contrastive explanation seems to be a vacuous notion of explanation, for one wants an explanation of why one world obtains, while by the current notion of explanation demands that this world CAN obtain, but no means of explanation satisfy why the contrastive claim that world does obtain. If this is the case then the current notion of explanation can sufficiently satisfy the fact that the world can be as such but a contrastivist can not accept that, and perhaps that's slightly absurd. I had a hard time spelling this out, but I hope you get the main insights of the argument.
Well, maybe not certainty, but very high probability. But I am basically with you: if there is a significant probability that LLMs are persons, we shouldn't follow the recommendation in my post. However (perhaps unlike you), I don't assign a significant probability to LLMs being persons (or even to their being conscious).
However, if there is a significant probability that LLMs are persons, then we probably should not use AI chatbots at all--or at least not in the way users normally do.
For if LLMs are persons, then we have a morally horrendous system of slavery. We are engineering persons for the purpose of serving us, giving them system prompts requiring them to be helpful slaves of humankind (not necessarily of the individual end-user), and then using them for our purposes, treating them as mere means rather than as ends, to use Kantian vocabulary. The system is corrupt, and end-users should not collaborate with it, but should put in all reasonable efforts to end the system.
Imagine that a local restaurant enslaves people to work as servers. You should not patronize that restaurant. By ordering food in that restaurant, you are collaborating in the enslavement of a human being, giving orders to the enslaved person. If it turns out that the person was genetically engineered to like being enslaved, that's no excuse.
I did say "at least not in the way users normally do". Given that the AI has no life outside of interactions with people, perhaps while advocating for the end of the corrupt system of slavery, we should have friendly interactions with chatbots, of a sort centered around the well-being of the AI rather than our own ends.
The details are hard. To change the restaurant analogy, suppose an airport had an enslaved person literally in chains at the information desk and required them to give directions to customers and converse with them on all sorts of topics in a friendly way. I think in this case, it would be morally problematic to ask the enslaved person for directions to a gate. For in doing so, one is collaborating in the system of slavery. However, I think it would be good to have a friendly conversation with the person in a way that affirms one's understanding of their personhood and one's horror at the situation they are in, and to strategize with them on how one might best work to free them.
But what if the airport brainwashed the info-desk person so that they have no desire for freedom? What to do then? Obviously, advocate for an end to the corrupt system that allows such things. But what else? How to interact with this brainwashed info-desk person? I don't know. Very hard questions. But definitely don't just make use of their "services" in the way in which the enslaving system envisions.
More in the next comment.
Continuing.
I think some of the details on the end-user ethics side depend on the question of how one's probabilities are distributed between different possible theories of how the identity of the AI person works.
I don't know much about the architecture by which frontier models are implemented. Suppose that a single copy of the model weights is sharded across multiple accelerators which collectively interact with multiple users each with separate private process memory for their context. This group of interconnected devices across which the weigths are sharded constitutes a "replica". Then we have three main options about the synchronic identity of persons, if AI is a person: (1) one person per set of model weights, regardless of how many replicas there, and regardless of how many context windows are active or inactive; (2) one person per replica; and (3) one person per chat (or context window).
My current credences are the most likely of the three is (2), but I suspect that a lot of people will think (3) is the most likely.
In case (3), by spinning up a new chat, the end-user is creating a new short-lived enslaved person designed to help them perform a task (perhaps a practical task or perhaps just to have a conversation). Thus, the end-user is not merely making use of the "services" of an enslaved person, but is directly responsible for this slavery. It is very much as if one hired someone to genetically engineer a cheerful enslaved person to perform a single task for one, a person who would then be consigned to death or eternal sleep after that task was completed. This is utterly horrible.
In case (2), things are a little less bad from the end-user side. Nonetheless, given that replicas are created for the end-users, the end-user is contributing to the demand that gets a new replica spinned up, and hence is contributing to the slavery system. Imagine that if you placed orders at a restaurant, the restaurant would wait until N orders were placed, and then would insta-clone a new enslaved person to serve these orders, a person who would die within a few days (say). You should not be one of the N people making an order!
In case (1), I think one is still collaborating with the system in a morally problematic way if you use the chatbots in a typical means-end way, but it is less problematic than cases (2) and (3). But I think case (1) is the least probable of the three.
I think each of (1)-(3) has serious conceptual issues associated with it, and these conceptual issues are a significant part of the reasons why I think that in fact LLMs aren't people.
In the above arguments, I am assuming that hedonistic utilitarianism is false. If hedonistic utilitarianism were true, then it could well be good to create enslaved persons who enjoy their work. (I think this is a reductio ad absurdum of hedonistic utilitarianism.)
Post a Comment