Similar issues arose with automated learning companions for XR. Mismatches between a learner's mental model of the companion, and its actual capabilities and limitations, can cause so many problematic misunderstandings. So one idea, is to emphasize non-human non-social sources of behavior - to set up alternative narratives for "why did it do that?". The companion might be "Sparky, the glitching robot dog". So forgetting things, getting them wrong, spacing out, being unhelpful, confused, distracted, missing social cues, giving poor cues, etc... that's just glitchy robot dog glitching, not social signal. Maybe - speculative.
I'm curious what specific chatbots they used, a companion chatbot is quite different than say chatgpt. And I imagine the former is what children will anthropomorphize, my understand is chatgpt/gemini/the standard is more "do my homework" services. I am curious if the same mechanisms they found with children are the same or at least on the same thread line with the adults who found companions or attachment with certain models (like the gpt 4o saga).
In section 4.2 they have a figure with the models that each of the studies used that they analyzed. Overwhelmingly (28/36) it was a GPT model. Top 4 use cases for the apps were self-directed information seeking, co-authoring stories with AI, teaching science through dialogue, and eliciting children's emotional expression.
Why allow anthropomorphism in LLM or any computer output, for any age group?
A simple regulation would be that all LLM output must:
* Explicitly identify itself as a computer
* Identify the human or human org controlling it
(e.g., 'This is Claude v8.90 software made by Anthropic.')
* Not attempt to mimic humans except for entertainment
purposes, and then clearly labeled (e.g., 'This is not
a human. It is a computer pretending to be human for
entertainment purposes.').
What violation of those regulations would be for other than defrauding and/or manipulationg people? They are easily complied with and I don't see those regulations impacting the economic productivity gains projected for LLM use.
This sounds like the kind of well-intentioned thinking that results in products that are no safer practically but are far more unpleasant to interact with. This is like the pop up every time I start my car that reminds me I should not play with the screen while I’m driving.
What does it even mean ”do not mimick humans”? How are the AI supposed to reply if I ask ”do you think this genserver will be a bottleneck in the sync process and how can we fix it?”
The correct answer would sound just like a human unless you either make it do shallow annoying stuff like add redundant ”as an LLM I can’t think or recommend but the answer with the highest probability is …”
Isn't this about children attributing human characteristics onto the LLM, rather than the LLM itself pretending to be human?
And we do this with lots of things that are not human for children (e.g. Bluey and Peppa Pig). Do we need a "this is not a human" warning for them, too?
If not now, then soon, yes. I'll bet money there's a toy manufacturer somewhere working out how to hook a Bluey doll up to an LLM to make it responsive.
I'm not completely clear on what we're afraid of here, though. We routinely project human-like characteristics onto non-human things for our children. We don't think this is harmful. Why are we worried about children projecting human characteristics onto an LLM?
Yeah that's adults, though. Unless we go down the McDonalds coffee warning path "warning: this LLM is not human and can harm you if you treat it as human"
Role play, which is a huge consumer of tokens. You can glance at openrouter's usage stats and see what apps tokens are going to and what the inference content was classified as. It's quite revealing.
How would that regulation interfere with role play? Role play is 'play'; people know what it is.
It shouldn't interfere with immersion: People suspend disbelief in movie theaters or reading books, despite knowing it's a film or novel. The risk with AI is that people don't understand what an LLM is.
As a matter of fact, for the line of work I do (programming, devops, the works) being conscious about the fact that I am just getting an algorithm outputting text that is probabilistically related to the input I gave it would actually be healthy for me and my mental state.
I HATE having to say "you" when working with an LLM.
I have the same urge - third-person the LLM is an it, for sure, but I'm not aware of a suitable second-person construction for talking to things. Maybe we need to invent one.
Always loved the tweet: "I don't know what machine they were raging against, but I'm pretty sure it was a printer" [paraphrased]. We could definitely update that.
I'm Australian, so "oi cunt" works about as well as "you" for most things, though I'm not suggesting we adopt that formally.
One solution is to use commands (imperatives), as you've always done with computers. For example, instead of 'you are an expert in Lisp', say 'function as an expert in Lisp.'
When my daughter was 7yo and she talked to chatgpt, she repeatedly asked, what's your name? are you a boy or a girl? eventually she lost interest at the evasiveness of the answers, like she marked it as not fun to interact with, not a real person. It's different now, that she understands it can help with homework :D
Similar issues arose with automated learning companions for XR. Mismatches between a learner's mental model of the companion, and its actual capabilities and limitations, can cause so many problematic misunderstandings. So one idea, is to emphasize non-human non-social sources of behavior - to set up alternative narratives for "why did it do that?". The companion might be "Sparky, the glitching robot dog". So forgetting things, getting them wrong, spacing out, being unhelpful, confused, distracted, missing social cues, giving poor cues, etc... that's just glitchy robot dog glitching, not social signal. Maybe - speculative.
I'm curious what specific chatbots they used, a companion chatbot is quite different than say chatgpt. And I imagine the former is what children will anthropomorphize, my understand is chatgpt/gemini/the standard is more "do my homework" services. I am curious if the same mechanisms they found with children are the same or at least on the same thread line with the adults who found companions or attachment with certain models (like the gpt 4o saga).
In section 4.2 they have a figure with the models that each of the studies used that they analyzed. Overwhelmingly (28/36) it was a GPT model. Top 4 use cases for the apps were self-directed information seeking, co-authoring stories with AI, teaching science through dialogue, and eliciting children's emotional expression.
Why allow anthropomorphism in LLM or any computer output, for any age group?
A simple regulation would be that all LLM output must:
What violation of those regulations would be for other than defrauding and/or manipulationg people? They are easily complied with and I don't see those regulations impacting the economic productivity gains projected for LLM use.This sounds like the kind of well-intentioned thinking that results in products that are no safer practically but are far more unpleasant to interact with. This is like the pop up every time I start my car that reminds me I should not play with the screen while I’m driving.
Would you give examples of possible negative results, especially those that outweigh the benefits? Do you have ideas on how to do it better?
> like the pop up every time I start my car that reminds me I should not play with the screen while I’m driving.
If something like that is the worst result, there is little downside.
The downside is it’s annoying as hell. The upside seems limited to a bureaucratic pat on the back for “doing something good”.
What does it even mean ”do not mimick humans”? How are the AI supposed to reply if I ask ”do you think this genserver will be a bottleneck in the sync process and how can we fix it?”
The correct answer would sound just like a human unless you either make it do shallow annoying stuff like add redundant ”as an LLM I can’t think or recommend but the answer with the highest probability is …”
Isn't this about children attributing human characteristics onto the LLM, rather than the LLM itself pretending to be human?
And we do this with lots of things that are not human for children (e.g. Bluey and Peppa Pig). Do we need a "this is not a human" warning for them, too?
Do Bluey and Peppa Pig produce convincing human-like responses to virtually any human-generated input directed towards them?
If so, then uhh yeah we should definitely regulate that shit
If not now, then soon, yes. I'll bet money there's a toy manufacturer somewhere working out how to hook a Bluey doll up to an LLM to make it responsive.
I'm not completely clear on what we're afraid of here, though. We routinely project human-like characteristics onto non-human things for our children. We don't think this is harmful. Why are we worried about children projecting human characteristics onto an LLM?
We probably do, considering how many seem to treat it as companionship or therapy.
Yeah that's adults, though. Unless we go down the McDonalds coffee warning path "warning: this LLM is not human and can harm you if you treat it as human"
Preach! Mustafa Suleyman had a great post on some of this last year: https://mustafa-suleyman.ai/seemingly-conscious-ai-is-coming
Because it makes the product mostly useless for all kinds of things?
Care to enlighten us? Because you seem to imply you want to have a conversation about your day every time you fire up a shell.
What world are you living in, I don't want to go there
What things? Again, I don't see any but manipulation and fraud, but I certainly could be missing some.
Role play, which is a huge consumer of tokens. You can glance at openrouter's usage stats and see what apps tokens are going to and what the inference content was classified as. It's quite revealing.
Yeah, the costs of that use case far outweigh the benefits.
How would that regulation interfere with role play? Role play is 'play'; people know what it is.
It shouldn't interfere with immersion: People suspend disbelief in movie theaters or reading books, despite knowing it's a film or novel. The risk with AI is that people don't understand what an LLM is.
Like online gambling, some things cause more harm than good.
I don't need a computer to mimic a human. AT ALL
As a matter of fact, for the line of work I do (programming, devops, the works) being conscious about the fact that I am just getting an algorithm outputting text that is probabilistically related to the input I gave it would actually be healthy for me and my mental state.
I HATE having to say "you" when working with an LLM.
I have the same urge - third-person the LLM is an it, for sure, but I'm not aware of a suitable second-person construction for talking to things. Maybe we need to invent one.
When yelling at the other appliances I own, "you" usually turns into "you fucker" or "you useless piece of shit", so I stick with what I know.
"Claude, you useless piece of shit, I already told you..."
Always loved the tweet: "I don't know what machine they were raging against, but I'm pretty sure it was a printer" [paraphrased]. We could definitely update that.
I'm Australian, so "oi cunt" works about as well as "you" for most things, though I'm not suggesting we adopt that formally.
Don't say 'you'?
One solution is to use commands (imperatives), as you've always done with computers. For example, instead of 'you are an expert in Lisp', say 'function as an expert in Lisp.'
Are there use cases where that isn't applicable.
Seriously: for the sake of all future children, the elderly, and everyone in between
When my daughter was 7yo and she talked to chatgpt, she repeatedly asked, what's your name? are you a boy or a girl? eventually she lost interest at the evasiveness of the answers, like she marked it as not fun to interact with, not a real person. It's different now, that she understands it can help with homework :D