Friday, December 02, 2022

A Chatbot's Take on "Moralometers"

I still have COVID, and I had a book manuscript deadline Wednesday. Things got out of hand, and I wasn't able to write up the blog post I wanted to write this week. Boo!

But as regular readers will know, the chatbot GPT-3 can write Schwitzgebelian blog posts, and -- after fine-tuning on the corpus of Dan Dennett -- can even answer philosophical questions in Dennett's style, which experts on Dennett's work have difficulty discriminating from Dennett's own answers.

On Monday, Open-AI released an update of their GPT-3 model, Davinci-003. Since I'd been planning to write a blog post on the idea of a "moralometer" -- a hypothetical device that would accurately measure people's moral character. So I prompted GPT-3's new davinci003 as follows:

Write a blog post in the style of Eric Schwitzgebel, about 1000 words long, about whether it would be possible to design a moralometer, which accurately measures people's moral character, and what the advantages or disadvantages of having a moralometer would be.

Its reply:

The Moralometer: A Tool for Measuring Moral Character?

In recent years, the idea of a “moralometer” has been gaining traction, as a way of measuring a person’s moral character. The concept is simple: a device that can measure a person’s morality, judged against an agreed-upon ethical framework. The moralometer could be used to determine whether a person is moral or immoral, and to what degree either of these traits are present. It’s an intriguing idea, but can a moralometer truly be built, and what would be the pros and cons of such a device?

The first question to address is whether a moralometer is feasible. On the surface, the idea of a device that can measure morality seems far-fetched. After all, morality is a highly subjective concept, and what is considered moral or immoral can vary greatly from person to person, and even from culture to culture. In addition, moral judgments are so often based on context, which makes it difficult to design a device that would accurately measure morality in all scenarios.

However, it may be possible to create a moralometer, with the right approach. The key would be to develop a device that could measure a person’s moral values and tendencies, as opposed to their specific moral judgments. By doing this, it would be possible to create a device that can accurately measure a person’s overall moral character, without being weighed down by the complexities of individual moral judgments.

The second question is whether a moralometer would be beneficial. On the one hand, a moralometer could be used to identify immoral behavior and help guide people towards more ethical behavior. It could also be used to help individuals make moral decisions, by providing an objective measure of the morality of their choices. On the other hand, there are some serious potential issues with a moralometer. For one, it could be used to unfairly judge people, as those with lower moral scores could be unfairly labeled as immoral. In addition, a moralometer could be used to enforce moral conformity, which could stifle individual autonomy and creativity.

Overall, there are both potential advantages and disadvantages to a moralometer. While it could be used to help identify and address immoral behavior, it also carries the risk of being used to unfairly judge people, or to enforce moral conformity. Ultimately, the decision of whether or not to create a moralometer will be up to society, and the implications of such a device should be carefully considered before any action is taken.

Next week, wind and meteors and immune system willing, I will try my own hand at the topic. The pressure is on! Will I be able to do a better job than davinci003? Stay tuned!

Relatedly, here's Dall-E's interpretation of a "realistic AI image of a meter for measuring morality":


Update December 7:
If you want, rate the chatbot post above and my own post on a postometer from 0 to 10.

9 comments:

  1. I’m quite sure that you’ll produce a far more appropriate post on “moralometers” than this chatbot was able to professor. A better question might be, if this question were asked in an introductory philosophy course, what grade might such an answer receive? I suspect at least a “B”. Hopefully far more is required of students in advanced courses. Here’s a condensed version of my own answer:

    I consider our moral notions, which is to say our thoughts about “the rightness to wrongness” of human behavior, to largely exist as an evolved social tool of persuasion. Without this trait the human shouldn’t have been able to develop its richly interwoven societies and so become what it is today.

    Our moral notions seem to exist by means of two psychological components. One concerns the sympathy that we tend to have for others. Here our perceptions of what others feel affect us in corresponding ways. Thus your hurt can be my hurt — and to the extent that I’m informed about your hurt in a compelling way. For example certain television ads display suffering dogs so that we might morally donate to potentially help alleviate such suffering.

    The second component may be reduced to our perceptions of how we are perceived. It essentially feels good to feel respected and bad to feel disrespected. Thus doing things that our friends (or societies, or selves, and so on) consider abhorrent, could be considered “immoral” in this regard. Note that this component can be quite arbitrary since it depends upon beliefs about what’s good/bad.

    Then regarding the prospect of building a moralometer, that’s easy since rules are simply required which are set up be measured. One might to this with the Ten Commandments for example. China has not only built a moralometer (its social credit system), but actively punishes and rewards its citizens on this basis when they do noticeable things, such as purchase diapers, pornography, or whatever. A better question might be, can a moralometer be built that’s “useful”? (And here of course define a “useful” that is indeed useful.)

    ReplyDelete
  2. So, the morality meter would be something like a polygraph machine? As I understand it, there are individuals who can control their bodily reactions to stressors well enough to fool the machine, probably not 100% of the time nor 100% of the operators.
    Also, possibly at least, the socio-cultural background of a subject might have an effect on the accuracy of readings. Would an amoral subject's readout(s) be reliable in such instances? Whether Dennett's position(s) on such contingencies are known and replicable could be subject to scrutiny. This is yet another interesting area for AI research. Prof. Dennett has been getting lots of attention recently.

    ReplyDelete
  3. The question I keep hoping academics (and other intellectual elites) might answer is this:

    QUESTION: What happens when those who fund academia can't tell the difference between expert level articles generated by AI and those written by humans? Same question for writers more generally.

    What am I missing? I'm coming across quite a few articles by very intelligent writers who are having fun playing with the emerging language generation bots. I'm having trouble figuring out why they are having fun.

    Aren't these AI writing bots going to put a lot of writers out of work? Yes, academics too. I don't claim to know exactly when, but isn't it inevitable at some point??

    ReplyDelete
  4. I have a question Phil. Do you believe that coming AI which produces academic articles that funders cannot distinguish from the writings of human experts, will phenomenally experience their existence? Or rather will they simply be able to give us that impression on the basis of their speech without phenomenally experiencing anything?

    Then secondly, if you believe that these coming artificial intelligences will phenomenally experience their existence, do you think this will come by means of algorithm alone? The implication here is that to create human consciousness, all our brains do is convert certain incoming information into an appropriate second set of information. The converse would be that human brain information animates some sort of consciousness creating physics.

    ReplyDelete
  5. Hi Eric,

    If I understand your question, it seems AI giving us the impression of it's consciousness, an engaging illusion, comes first. That's already happening it seems. If/when AI obtains real consciousness, whatever that might be, is unknown to me. Apologies, I don't understand your second paragraph.

    What interests me at the moment is what will happen with our culture's relationship with the knowledge explosion when it's not just blue collar workers losing jobs to automation, but white collar elites too. I'm wondering if that might be some kind of turning point.

    ReplyDelete
  6. Hello Eric,

    What's the return time between request and result? I am wondering if these things are basically a super ELIZA program and maybe are being massaged before release by some sweat shop in order to boost share prices (it's certainly not impossible). The service forces users to give a phone number and to me that sort of filtering of users isn't needed for an actual AI to function. Also the content of the generated post is basically waffle - it spins the mental wheels but doesn't say much (just as ELIZA operated).

    ReplyDelete
  7. Phil,
    I’ll begin with your question so that I don’t ramble on with my own stuff to the point of neglecting it entirely. Even if many white collar jobs are lost to advancement in AI, I suspect that this won’t change our culture’s relationship with our knowledge explosion very much. I doubt that this group will be politically strong enough to extensively hinder such technology, or at least where it matters. Thus even if AI began replacing white collar workers, I suspect that the old Dune maxim of “Thou shalt not make a machine in the likeness of the human mind”, would still not be instituted in lasting ways. Do you suspect otherwise?

    There’s an earlier question to consider here as well. Would it be possible for a machine in the form of this non-conscious GPT-3, to write the same sorts of things that college professors have been educated to write? Or perhaps eventually replace such professionals? This is quite doubtful to me given that we’re talking about a non-conscious machine replacing a conscious machine (in the form of an educated human). I wouldn’t expect evolution to have created consciousness if non-conscious beings had the potential to do the same sorts of things, such as autonomously teach college classes.

    There’s another issue to consider here as well. What if this GPT-3 is at least proto conscious? What if the more that we program a machine to seem like it’s human, the more that it will actually harbor something like human consciousness? This is sometimes referred to as “functionalism.” Apparently it’s the most popular consciousness theme in science today. On the philosophy side, Dan Dennett and Keith Frankish are big supporters.

    I consider this position non-causal (which is a diplomatic term for “magical”) given that computers do not exclusively function by means of converting input code into output code — causality mandates that output code go on to “inform” something, such as a computer screen for example. So the question seems to be, if the computation of your brain causally creates you as a phenomenal experiencer of existence, then just as a computer screen exist as one form of substrate which is animated by computer information, what is the substrate which your processed brain information animates to exist as you the experiencer of existence?

    I’d be pleased if popular consciousness proposals in neuroscience today would at least assert that they aren’t sure what that substrate happens to be, if not actually take a stab at the question itself. Instead of remaining unfalsifiable they might thus present ideas that could potentially be tested. The only reasonable consciousness substrate proposal that I know of, is the electromagnetic radiation associated with certain synchronous neuron firing. Johnjoe McFadden is its author.

    ReplyDelete
  8. Hi Eric,

    I do think that when white collar elites like philosophers are losing their jobs to AI that will have an influence on our overall relationship with knowledge, but I don't claim to know the degree of impact that it will have.

    Blue collar workers don't have much of a voice, but highly educated articulate people with connections do. Thus, we should anticipate a different outcome, of some kind.

    As to your point about the possible consciousness of AI, it's not clear to me whether that matters. As far as business and budgets go, can consumers and funders tell the difference between expert level human written articles and AI generated articles? If not, the human writers are at risk, whether or not AI is conscious.

    Just look at the content being generated by AI today. AI appears to be roughly at the college sophomore level now. I don't see why that can't progress to PhD level content, over some period of time which I can't calculate.

    It's interesting to me that today's current PhDs don't seem to grasp the threat, or perhaps I don't see those that do. To my read, intellectual elites of various types seem to cheerfully playing with these emerging AI tools like children playing with toys. Where's the alarm? Where's the push back? Where are the writers running for the exits, planning new careers?


    ReplyDelete
  9. Thanks for all these comments, folks! See today's post for my own take on moralometers -- part one of at least two and possibly several.

    On the quality of the answers: Yes, my general experience with GPT-3 is that its outputs are ballpark B range given general philosophical prompts (with a mixing in of some real duds). I think this creates a serious challenge for professors who want to assign short philosophy prompts of this sort to undergraduates. As for when the outputs might be mistaken for high-quality philosophy, well, not yet and maybe not for a long time, but my new paper with Anna Strasser and David Schwitzgebel addresses this and finds that, at paragraph length, GPT-3 fine-tuned on Dennett can produce outputs that even experts on Dennett's work mistake for Dennett himself. Callan: Return times are fast -- faster than human writing.

    On consciousness: Check out David Chalmers's talk about this:
    https://www.youtube.com/watch?v=-BcuCmf00_Y
    I am also currently in a large collaboration with other consciousness researchers on the topic, and we're hoping to have a circulatable article soon that reflects something like the majority mainstream opinion on the topic, which I take to be -- not yet, but possibly on the medium-term horizon, given substantial but not totally unfeasible, changes in architecture.

    ReplyDelete