It is reported that Geoffrey Hinton “the godfather of AI” is leaving Google and has voiced some serious worries about the future of humanity as AI continues to develop. I don’t have anything interesting to say about grey gloop or paperclips or AI robots waging wars, but I have been thinking a bit about the impact of AI on creative work, not limited to the production of student essays. Already we are seeing voice actors replaced by clones of their own voices and professional translators reduced to editing the output of machine translation (almost as much work, but for less money, I’m told). So what happens if AI can produce artworks (or should that be “artworks”?) such as plays, paintings, pseudo-photographs, movie scripts, novels, songs, symphonies that are indistinguishable from human productions and that people consume and enjoy? Well, one effect might be that it becomes even harder for people to earn a living producing artworks for the market than it is now. But that doesn’t mean that human production will disappear. And the reason that it won’t is because our interest in creative work isn’t just about the object of production but about its process and the exercise of our human powers (“life’s prime want”, as somebody once said.)
The invention of photography in or around 1839 may have made possible a more accurate representation of reality and in doing so may have displaced some forms of drawing whose purpose was the utilitarian representation of reality, but it hardly stopped people from painting and drawing and, indeed, gave them a new medium in which to express themselves. AI may be, even is, able to produce something that looks like a good drawing of an object, but it cannot replace the human activity of looking hard at that object and co-ordinating hand and eye to produce my (however pathetic and inadequate) represention of it. AI may be able to produce a song, but it cannot substitute for the experience of writing a song and singing it. So I suspect that even if AI gets very good and produces work indistinguishable from human work, it will not and cannot fully replace human work. It will, perhaps, somewhat devalue the artwork as the object of contemplation and consumption, except insofar as we continue to admire works as the product of specifically human intention and execution (just as we would continue to admire the moves of a talented human footballer even in a world where AI-driven robo-footballers were available). But the artwork as the product of a human process, with a renewed focus on that process as the real activity of doing and making will not cease to exist. The Milton who produced Paradise Lost “as a silkworm produces silk” will continues to write; the Leipzig literary proletarian will not. Indeed there may be more of creative labour, since if AI provides for our basic needs, we’ll have the time available to hunt in the morning and criticize after dinner, as well as drawing, painting, cooking, and writing short stories and songs, just as we have a mind. (That is, unless we are enserfed to spend our time catering to the whims of Jeff and Elon instead.)
{ 64 comments }
John Q 05.02.23 at 10:08 am
At least as regards text, there’s no need to worry about the current version of AI. Given the way it is generated, it is, by definition, mediocre. It takes a set of prompts and guesses what the average piece of text on the Internet generated by those prompts would look like. That can’t possibly be as good as the best response that already exists, let alone new and better
Kevin Lawrence 05.02.23 at 10:23 am
Unfortunately, most people are happy with mediocrity, John Q.
I can buy handmade Italian leather shoes for £400 or I can buy a mediocre pair for £40. The number of people who would prefer fine writing to the mediocre kind is approximately the same as the number of people who buy shoes of fine Italian leather. Also, there are an awful lot of people whose job is to produce mediocre writing for corporate websites and marketing brochures. Those people are now out of a job, along with the people who produce the artwork.
NomadUK 05.02.23 at 12:24 pm
As someone who works in the elementary school system in the US, I hear a lot of hopeful, positive spin similar to that of the original post and JQ’s comment. Sadly, I agree completely with Kevin Lawrence@2 that this is all whistling past the graveyard. The corporate, profit-driven imposition of ‘AI’/LLM systems on all of us is going to be an unmitigated disaster, and will dwarf even the current devastation wrought in education by the rampant (mis)use of technology (if you want to feel yourself die just a little more inside, watch a bunch of third-graders spend the first 20 minutes of class logging into their Google accounts on their Chromebooks so they can do some insipid web-based lesson plan).
Jake Gibson 05.02.23 at 1:07 pm
Machine Learning will have to get a lot better to make in inroads in commercial art. I do not doubt that it will eventually get there. I probably will not live long enough to see the AI Writer’s Guild go on strike.
Josh H-C 05.02.23 at 2:00 pm
Weird bit of trivia about the history of photography (which you may know about, Chris!): for the first few decades, a large number of photographs were really hybrids of photographs and drawings. Not only stuff like black and white photographs being coloured in with paints, but also retouching to make people look better, collages of sky and landscapes, and a florishing industry in composite pictures (you could get a phtoograph of yourself participating in various civil war battles). If there’s a model for the introduction of technology into creative processes, it’s the creation of these weird hybrid forms of art, not the automation of drawing.
Ray Vinmad 05.02.23 at 4:32 pm
These discussions are so frustrating because they continually talk about LLMs and even chatbots as ‘artificial intelligence.’ Are they? It’s unclear but there is no evidence they are in the way sci fi imagined (yet).
There’s a firehose of absurdly exaggerated descriptions of these predictive text and coding processes.
Their limitations aren’t discussed in the media in a serious way–even though it is their limitations such as their unreliability which is likely to be the most harmful. The firehose effect of the rhetoric makes it hard to talk about military, medical and other applications. These discussions are desperately needed.
The information climate is rife with propaganda. People are practically personifying LLMs as if they deserve a seat at the table. This is absurd.
There’s very little discussion about the human decisions that are going to amplify the harms if no planning or thought is applied to their use. Instead everything is described with an air of inevitability–and weirdly, about LLMs as if they can do things already they cannot do yet. It’s very similar to the self driving car discussion in a way but a bit like a car was invented and we immediately assumed the car could fly because in theory, it could.
LLMs are basically a product right now. Its creators are seeking investment capital and people to pay for it. So people should bear this in mind.
We have politicians who can’t even figure out social media and the basics of the internet and a business community that seems all in on crypto so I am not optimistic.
I would like to believe everything will settle down and the facts will become clear but this doesn’t happen the way it used to.
engels 05.02.23 at 5:07 pm
there may be more of creative labour, since if AI provides for our basic needs, we’ll have the time available
Considering the wealth of Marx quotations in this post it’s striking that Ch. 15 of Capital didn’t make the cut.
James S Daniel 05.02.23 at 5:29 pm
As someone who has been involved with computers for almost 50 years I’d like to point out that while AI is Artificial it is not Intelligent and even if it was it would not be creative. Just take a look at how AI is used to create Art, be it writing or some other art. There is always an actual human in the loop making the final decision. An just like it is the human photographer with his/her ‘eye’ who decides what photograph to keep, not the camera, so it will be with the new creative tool AI. All we have to do now is discover how to recognize what is done by humans and what is done purely by AI. Not a simple problem but not a problem we have not already solved in our long history.
mw 05.02.23 at 5:50 pm
Jake Gibson @4 Machine Learning will have to get a lot better to make in inroads in commercial art
I don’t think so. I’ve been playing around with the Dall E image creator built into Bing. It’s pretty amazing. It definitely has flaws, and does some odd things, but for a lot of run-of-the-mill illustration needs (magazine, brochure, ad-flyer, greeting card, children’s book, etc) it seems more than capable enough. And it could certainly be used for ‘artistic brainstorming’ even if an actual artist produced the final image (which would then make it copyrightable, BTW).
Anyway, give it a try — it’s free to use as long as you have a Microsoft account. I’d suggest a prompt of subject matter ‘in the style of’ well know artist/art work/style/genre.
John Q 05.02.23 at 7:00 pm
@2 and @3 I agree that LLMs are going to make big differences to work, education and so on. Whether that turns out well or badly depends on struggles that will be fought out in workplaces, including schools.
I was responding to the OP and asking whether LLMs could replace genuine creative work.
Mike Furlan 05.02.23 at 9:37 pm
#comment-823696
” Also, there are an awful lot of people whose job is to produce mediocre writing for corporate websites and marketing brochures. Those people are now out of a job, along with the people who produce the artwork.”
Agreed Kevin. And some of those “mediocre” writers, musicians and portrait painters might develop into F. Scott Fitzgerald, J. S. Bach or John Singer Sargent.
AI today, and recorded music and photography in the past will, and have removed the first rung of the ladder to a lot of careers.
Alex SL 05.02.23 at 10:26 pm
I do not doubt that this technology is and will continue to be very “disruptive” (I hate that term) in the sense of putting people out of work. But what these singularitarian tech bros generally mean when they warn about AI is indeed “grey gloop or paperclips or AI robots waging wars”. They do not understand physics, they do not understand biology. They do not understand trade-offs, they do not understand diminishing returns, they do not understand resource limits, and they clearly do not grasp that the AI stops when you simply unplug the computer running the AI.
They have read the SF novels of Iain Banks and Vernor Vinge and misunderstood them as documentaries. They see that AI can do cool things and envision it reaching out through the internet and launching all the nukes. (If the nukes are that easy to launch, we have a problem to which the cause of their launch is orthogonal, be it rogue AI, technical fault, or the US president having a bad day.)
Most importantly, they believe that a sufficiently intelligent mind can do what amounts to magic. If you point out that this or that magical ability is impossible, they respond with some variant of “the AI will be so unimaginably intelligent that it will be able to (do magic) in some way that we are too unintelligent to figure out”. In other words, if you were just smart enough, you would be able to feed yourself for two months with one grain of rice and fly by flapping your arms.
Regarding, however, the much more realistic question of AI’s impact on creativity and employment, I must admit to some skepticism about “if AI provides for our basic needs, we’ll have the time available to hunt in the morning and criticize after dinner, as well as drawing, painting, cooking, and writing short stories and songs, just as we have a mind.”
…even if we aren’t enserfed. No past increase in productivity has ever led to us working twelve hour weeks and being creative the rest of the time, so why should this one? There seems to be something fundamental to either human nature or perhaps just capitalism that forces us to either find work for people or immiserate them. Instead of cooking and writing fan fiction because AI cares for our basic needs, we would all work 50 hour weeks in open plan offices having pointless meetings and formulating and discussing procedures and policies that could be reduced to “don’t be mean to others” or “think before you touch that open wire”.
engels 05.02.23 at 10:27 pm
And some of those “mediocre” writers, musicians and portrait painters might develop into F. Scott Fitzgerald, J. S. Bach or John Singer Sargent.
Or…
Fake Dave 05.02.23 at 11:18 pm
I think we need to collectively and individually be much more skeptical and even litigious about the super-charged plagiarism machines being aggressive pushed as “creative” software with agency of its own. They aren’t really AI and they don’t have any sort of imagination or creative faculty. All they are doing is assembling rapid fire collages of anything their unscrupulous/criminal programmers chose to scrape from the web. The appropriate response is to ban most of this stuff outright until we have something approaching a legal framework for preventing mass-scale copyright violations and a culture of best practices within the industry to keep harmful and fraudulent materials from being unleashed “in the wild.”
The photography comparison partly fits and partly doesn’t. I’m not convinced this technology will actually transform the way we can capture or understand the physical environment in any way. It’s much more in tradition of collage or digital compositing. Most of the ai illustrations I’ve seen even have that glossy photoshopped offness I associate with innumerable comics and DeviantArt “tracing” controversies and the tacky 2000s trend of slapping three or four filters on someone else’s picture and calling it Pop Art. The complicated and still controversial history of sampling in hip-hop might also be a touchstone.
People fell in love with the idea of making something new out of bits and pieces of the things they loved only to find that it presented huge legal obstacles to getting paid for their work (unless they were big enough stars to get away with it.) I have heard a lot lately about how this technology can easily and cheaply enable indy publishers to add illustrations to their work without having to find and pay an artist (who might still just be some kid with photoshop). It’s a great idea in theory, but in practice, we don’t know the identity or legality of most of the “training sets” currently in use and there have already been innumerable examples of artists having their work stolen without consent or compensation. I would advise anyone currently playing around with it to leave it at that until the companies involved clean up their acts before trying to use it for anything commercial or they might wind up in the legal morass as if they had hired a human art thief.
Alex SL 05.02.23 at 11:32 pm
Ray Vinmad,
I agree with all you wrote, but I am afraid the ship has sailed as far as terminology is concerned. AI is used as a shorthand for everything software does that kind of is a bit like what a human could do. The most illustrative example IMO is how people call the computer-controlled players in 4X strategy games “AI” although in nearly all cases they are not even at the level of neural networks but merely bundles of simplistic decisions to the effect of “if I am at least 50% stronger than the other player, declare war, if not, maintain peace”.
Chris Bertram 05.03.23 at 7:45 am
I agree with engels and Alex that capitalism has a inherent bias towards output and burdensome toil over (possibly creative) leisure and, indeed, I’ve written about this myself. So I agree that making it possible for people to have more time at their own disposal depends on us changing the social and political environment in a way that overcomes this bias. But it is a bit daft to think that there have been no advances in the burdensome toil/autonomy-leisure ratio ever: people worked much longer hours in the 19th century than they do now even if Keynes’s hopes have not ben realized.
Alex SL 05.03.23 at 8:08 am
Chris Bertram,
I am not claiming that life hasn’t become easier. The introduction of washing machines, dish washers, and vacuum cleaners was the true “singularity” as far as purely technical solutions to human welfare are concerned.
But regarding only the hours worked, are we certain about that? I know quite a few fellow researchers and managers who seem to work, in some sense, nearly every waking hour, even if it is sometimes just responding to emails, reading a paper, or checking the progress of an analysis. Admittedly, they do not work these hours pushing carts in a coal mine, so I grant that they are less taxing on the body. But at the other end of the income scale are people who need to work 2-3 jobs to make ends meet, and those jobs may be quite bodily exhausting.
And to my understanding we all work longer hours than hunter-gatherers or even farm workers in antiquity.
Fake Dave,
Maybe, or maybe this is quickly becoming such a widespread practice that enforcement will be impossible. I am not saying that as if that was a good thing, just to be clear, but it is a real possibility that the Overton Window will soon be firmly planted on ‘everybody does it, and nobody is ever fined, so it must be okay’.
Chris Bertram 05.03.23 at 8:14 am
Alex, it is really quite easy to google historical statistics on working time.
engels 05.03.23 at 10:25 am
people worked much longer hours in the 19th century than they do now
According to BoE data they worked shorter hours in plague-infected C14th though (too bad they didn’t have Zoom…)
https://fredblog.stlouisfed.org/2021/10/a-short-history-of-working-hours/
mw 05.03.23 at 10:46 am
Fake Dave @14 “All they are doing is assembling rapid fire collages of anything their unscrupulous/criminal programmers chose to scrape from the web. The appropriate response is to ban most of this stuff outright until we have something approaching a legal framework for preventing mass-scale copyright violations”
Thou art the symbol of our progress, the emblem of our growth,
With thy efficiency and thy versatility, thou showest us our oath.
Thou art the embodiment of our genius, the expression of our skill,
With thy processing and thy memory, thou help us to fulfill.
That’s a stanza from ‘Ode to the Microprocessor (in the style of Keats)’ generated by ChatGPT a couple of minutes ago. Now, I’m not saying it’s good poetry but it’s clearly original and novel and would be violating no copyright even if Keat’s works were still protected. The same goes for images / art-works — they, too, are clearly not just ‘collages’ of protected works. LLMs will readily produce pastiche, though, if that’s what you ask for. But do we really want to try to legally protect against merely imitating the style of other artists’ or musicians’ work?
Are LLMs trained on works that are still under copyright protection (as well as works that are not)? Yes, but so are search engines. And so are human writers and artists — many of whom readily acknowledge their influences.
Most of the ai illustrations I’ve seen even have that glossy photoshopped offness I associate with innumerable comics
I have seen weird artifacts in some of the images I’ve generated. In some cases, it would be easy enough to clean them up. In other cases, they generate objects with Escher-like impossible geometries or weird conceptual/visual mashups. To the extent that’s true (and those problems not readily fixable — as I suspect they may not be), it will limit the usefulness of AI generated images and art. But that’s fine.
Jonathan Hallam 05.03.23 at 5:56 pm
‘(just as we would continue to admire the moves of a talented human footballer even in a world where AI-driven robo-footballers were available)’
This is a bit weird, because it’s as hypothetical as the point about artists being replaced. A more concrete example would be ‘just as handmade furniture continues to be valued, even as CNC manufactured or machine laser-cut equivalents are available at equal quality and lower price’
Sashas 05.03.23 at 9:38 pm
@mw (20)
If I challenged you to actually defend these claims, how would you do so? To be clear, I am absolutely challenging you to defend these claims. I am moderately conversant in how generative AI works behind the scenes, and I see nothing original or novel in it. (“Collage” might be a little strong for my taste, but closer to the reality as I understand it than what I think you have claimed.)
Alex SL 05.03.23 at 10:23 pm
Again, not my area of expertise, but unless I misread, a quick google suggests that available stats are pretty much last 150 years. Did specifically the industrial revolution create abhorrent working conditions that unionisation subsequently overcame, or did people in the middle ages also work six day weeks, ten hour days? Most pertinently, did middle age people work seven day weeks, fourteen hour days, because technical progress reliably reduces work times (so that the people 150 years ago had it better than them already), as would be necessary for AI to be a positive in that area?
I also wonder about accurate work hour measurement today, when people may have 40 hour week contracts on paper but can be reached by email or Zoom meeting outside of core office hours. This is tech that has actually increased work hours for many of us.
David in Tokyo 05.04.23 at 7:28 am
Someone (mw) wrote:
“‘Are LLMs trained on works that are still under copyright protection (as well as works that are not)? Yes, but so are search engines. ”
But there’s a not-so-subtle difference here. The search engine returns a link to the page that holds the text searched for, so the author get the click and the credit.
The LLM game is to encode the text, but not the link, and regurgitate that text convoluted (in the technical sense of the term) with all the other text it has ever seen, so the author is given neither the click nor the credit. This is exactly the definition of plagiarism.
This is also a total disaster for the user since the user has no way of verifying or checking the reasonableness (or lack thereof) of what the LLM spat out. Other than frantic googling to try to verify the LLM output.
IMHO, these are two really good reasons to ban LLMs.
As someone who first learned to program in 1970, audited Minsky and Papert’s AI seminar in 1972, earned an MPhil in AI in 1984, and hsa been watching the field completely fail to deal with it’s basic questions ever since, I can assure you that the whole LLM thing is nothing other than a stupid parlor trick.
(The parlor trick here is that the LLM suckers the user into interpreting its output for it. The LLM has no way of dealing with the text it produces in any way: to the LLM, it’s output text is a meaningless sequence of meaningless tokens randomly produced from input strings it has no idea what they mean. It’s a kewl trick, but the term “parlor trick” really is (a) technically correct and (b) the right way to think of these things.)
Laban 05.04.23 at 8:39 am
Chris is at Bristol Uni, isn’t he? I was chatting to a first-year history student the other night, and asked him if many people used ChatGPT. He said yes, and it’s particularly useful for “find me a reference that supports point X” type stuff. I got the impression you could write an entire essay and ask it to find and list supporting references. Of course at that point why not define the required conclusions and ask it to write and reference the essay!
I can see something similar happening in law. Who needs to have all the case law at their fingertips if ChatGPT can pull it up?
This has the drawback that (IMHO) knowledge is good, and while looking for references, in flicking through the 15 books on your desk in the library, or 15 tabs open on your screen, you may come across other things or authors of interest. Can’t do that when a bot is doing the reading for you.
engels 05.04.23 at 11:03 am
Did specifically the industrial revolution create abhorrent working conditions that unionisation subsequently overcame
<a href=https://en.m.wikipedia.org/wiki/Engels’_pause>Ahem
Neville Morley 05.04.23 at 12:31 pm
@Laban #25: someone will probably have responded to you by the time this gets approved, but if not: one critical point about ChatGPT ‘finding’ references to support a point is that it’s entirely capable of inventing them. Last month I marked a couple of student essays with some completely fictional quotations from source material and convincing but non-existent publications in the bibliography. Now, I can’t prove these were AI-generated – one student has simply ignored all attempts at communication, the other claimed that this was the consequence of poor note-taking and dyslexia – but that doesn’t matter from my point of view because they fail simply on the basis of inadequate argument and scholarly presentation. It just seems implausible that a student could produce a really plausible but imaginary journal article without having done all the reading that would mean they wouldn’t need to produce imaginary journal articles, whereas an LLM skimming the whole web for statistically-probable combinations of words can do so easily – and students trusting ChatGPT results for suggested sources are as foolish as academics putting essay text into ChatGPT and asking if it wrote it…
Neville Morley 05.04.23 at 12:50 pm
On the broader issue, I was very struck by the discussions around the fact that Bradford Literary Festival used some terrible AI-generated images for its publicity – girls with three feet and melting hijabs, books with half the pages smaller than the other half – and no one in the ‘creative agency’ responsible or the organisation noticed. It probably didn’t save the festival any money over hiring an actual artist, but it certainly improved the profit margins for the agency. And while the switch by such parasitic middlemen to AI-generated images and music may not directly deprive us of great art, it is likely to do so indirectly, as plenty of artists and musicians have relied on such work to put food on the table so they have leisure to develop their skills to the highest level, rather than the competent amateur level that most of us can manage (even if the working week gets reduced to twelve hours).
Neville Morley 05.04.23 at 12:51 pm
Sorry, forgot the link to the Bradford thing: https://twitter.com/alexdecampi/status/1653397448285356056?s=21
nastywoman 05.04.23 at 1:52 pm
But on the other hand –
if you let chatGPT write your comments on CT
they
ALL
will
pass
and you
NEVER EVER will be…
disliked?…
by and Dear Prof(s)
Right?
J, not that one 05.04.23 at 2:17 pm
About 15 years ago, I realized that some of my colleagues, rather than looking up traditional reference works, we’re doing Internet searches for snippets of example code that did what they needed to (learn to) do, and inserting those into their work product. The quality of those, for the most part, was probably no worse than what the LLMs produce. I’d guess the same is true for school essays. Where someone my age knew that Cliff Notes and Spark Notes were bad, but at least mostly reliable, it looks to me like kids find that kind of “published things with editors are better” attitude to be antiquated, and whoever puts up a site to get advertising traffic can get their attention. (In second grade, they’re told Wikipedia is bad but are taught to find definitions by googling the word and accepting the search engine’s definition; in seventh they’re told Wikipedia is actually fine.)
In other areas, where copy-paste produces how-to articles that are ten or more years out of date, most of what’s generally available, and successfully promoting itself on search engines, would seem to have the same problems as “AI.” Which is presumably incorporating all those unreliable sites, whether at first or second hand.
As the mound of seemingly authoritative garbage grows ever higher, the Leicester proletarian, lacking credentials, is already at a disadvantage with or without mechanically artificial authors stirring the pot. And a reader who doesn’t believe in editors and reference works really has only credentials to allow her to make a choice.
mw 05.04.23 at 3:15 pm
Sashas @22 If I challenged you to actually defend these claims, how would you do so?
Quite easily. They are novel and original in the sense that none of lines are plagiarized or a close paraphrase of any existing work that I can find. They produce no Google hits. Can you find where any of this text was ‘borrowed’? That is the only sense of ‘original’ that matters for copyright law.
“Collage” might be a little strong for my taste
But the difference between collage and pastiche is critical here –assuming the collage consist of reproductions of parts of protected works. That would be like sampling in music. But that is simply not what’s being done by LLM image generators. When I asked for ‘Mona Lisa in the style of Guernica’ or ‘A modern cubicle office in the style of Hieronymous Bosch’ I got visually intriguing results that were — to my knowledge — unlike anything any human artist has thought to produce. We would need entirely new, much broader copyright laws to consider these kinds images to be infringing (laws so broad that they’d surely would stifle human creativity.
Chris Bertram 05.04.23 at 3:50 pm
@mw though I do find myself entertaining the fantasy that some evil oligarch might get an AI to compose lots of melodies and then bring plagiarism cases against any subsequent composer who made use of them.
David in Tokyo 05.04.23 at 4:05 pm
Dear ChatGPT. Please write me a Crooked Timber post in the style of nastywoman.
(Hey, seriously, I sometimes try to compete in the nastywoman fast lane, and I ain’t good at it. So I’m jealous of her in-your-face creativity. And need ChatGPT’s help.)
By the way, the technology is out of the bag, and out of Google/OpenAI’s control:
https://www.semianalysis.com/p/google-we-have-no-moat-and-neither?utm_source=post-email-title&publication_id=329241&post_id=119223672&isFreemail=true&utm_medium=email
steven t johnson 05.04.23 at 4:21 pm
Perhaps I misunderstand entirely but the issues with and effects of using LLM seem to strongly overlap those found in aleatory music and action painting.
engels 05.04.23 at 6:21 pm
I’m surprised no one asked Chat-GPT to comment yet.
JimV 05.04.23 at 6:37 pm
Unless any of you have made up new words, we are all using words we learned from others. If that violates copyright, we are all guilty. It has not memorized specific phrases from its training material, in any direct form, just the correlations it has deduced among words.
ChatGPT-4 has passed actual graduate-level exams in (so far) Economics, Organic Chemistry, and Quantum Physics (A, A, and B). It is being widely used by programmers to write code routines, in several code languages, including Python and Stan.
It makes a lot of mistakes. Just not nearly as many as the average human. It will make up references unless specifically told not to do so. Many of the prompts it responds to ask it to make something up (as in the poem above), and that is how it will respond unless carefully prompted. It also helps to ask it to check what it has just produced, since it currently does not have a checking phase. (Asked to check the answers it got wrong in the Quantum Physics test calculations, after the test had been graded, it improved its score.) The programmers who use it of course test-run the code it produces and often find bugs–just not as many as if they coded the routines themselves.
Neuroscience defines intelligence, sapience, and sentience as separate qualities. According to those definitions, GPT has demonstrated intelligence (the ability to analyse and solve problems), but not the other two.
Humans are motivated to solve problems by the emotions programmed in them by biological evolution. Neural-network problem-solvers are motivated by the programming given them by humans. AlphaGo is a very intelligent Go player (having beaten the World Champion). It only plays because it has been programmed to do so. Similarly, ChatGPT responds to prompts because it has been programmed to do so. What such AI’s do is entirely up to us and how we develop them, regulate them, and share (or not) the benefits of them.
OpenAI is developing a water-marking system which will make it possible to tell (via a software checker) whether a given essay has been produced by ChatGPT. This will be also possible for image-generators.
That’s as I see it, culled from technical discussions at various science blogs. It seems like a remarkable achievement so far in an historically short time (digital computers being less than 80 years old; steam boilers were still blowing up at roughly the same time in their development).
Dave 05.04.23 at 8:25 pm
echoing JSD @8
AI as currently exists is not creative and does not generate ideas, at least as I understand those things. It’s not unimpressive, but the major obvious use case is to make spam, not art.
Fake Dave 05.04.23 at 10:16 pm
Engels @19
We’ve had that “working hours in the middle ages” conversation before and I think the consensus was that the notion that peasants had abundant leisure time comes from people applying modern assumptions about work anachronistacally. Even in the present we have a common issue of people only counting wage labor and discounting things like housework and childcare as labor. Take that capitalistic erasure of people working for themselves/their families and project it backwards to when most people were farmers and wage labor was exceptional and you get some very strange results.
mw @20
I admit I thought that was a real (bad) poem for a moment, but I wouldn’t have guessed Keats. He’s not being directly plagiarized and is in the public domain so I don’t really have a problem with it except that it doesn’t have to give any credit to the person named in the prompt and could have even lifted lines from other writers without even the prompter being aware. There’s a dangerous “man behind the curtain” effect where the people developing and marketing this software want us to be awed by what it can do, but don’t want us to really question the steps along the way, especially the steps taken by morally compromised people operating under strong financial incentives. Much of what we’ve been told about how this software works is exagerated and aspirational and in some ways reminds me of Amazon’s mechanical turk except that the people “training” the AI don’t even know they’re doing it.
It could be that language is commonly held and functionally transmissible in such a way that “microplagiarism” of a single word or phrase is an irrelevant or nonsensical concept, but these programs have also been shown to be quite capable of spitting out whole paragraphs of other people’s writing, so this isn’t just one for the philosophers. I’m much more concerned about art though because the basic components of a visual composition are much more complex and far less generic than the words in a dictionary. I’ve seen a number of examples of people getting Stable Diffusion and the like to spit out close copies of works by living artists who did not volunteer to be the little person hiding in the chess playing robot. If it can plagiarize people that blatantly with no credit, then the people using it can plagiarize without knowledge. That’s bad for everyone except the people selling the software.
I’m not saying this will be an insoluble problem forever (you could always set these things to only train off the public domain and provide appropriate attribution), but it’s certainly a problem now and will continue to be a problem as long as the industry and its boosters continue to ascribe (magical) human intelligence to mindless software.
engels 05.04.23 at 11:47 pm
Dave, idk. I don’t think the Bank of England really has an axe to grind about this and the top link I get on the debunking side (with an argument remarkably similar to yours) is our own village neoliberal, Tim Worstall:
https://www.adamsmith.org/blog/regulation-industry/medieval-peasants-really-did-not-work-only-150-days-a-year/
Alex SL 05.05.23 at 12:17 am
I am fascinated by the argument that because humans learn from each other too, everything anybody ever does is actually plagiarism, if you think about it, so nothing matters. How very convenient to anybody who wants to make off with the fruits of others’ labour.
bad Jim 05.05.23 at 2:40 am
Elsewhere I have seen it noted that ChatGPT doesn’t write poetry very well, and rhymes even when asked not to. The sample of ersatz Keats which mw supplied employs rhymed couplets, unlike any of the odes in the anthology I just consulted.
I wonder what it would do if asked to write a sestina.
mw 05.05.23 at 9:34 am
bad Jim @42 “Elsewhere I have seen it noted that ChatGPT doesn’t write poetry very well”
I’ve not been particularly impressed either (except in a ‘dog walking on its hind legs’ kind of way), but that’s not the point. The point is that if you ask it to write a Shakespearean sonnet about Chewbacca, you’re pretty much guaranteed a bad, ridiculous result but what you get will certainly not be any kind of ‘regurgitation’ of a protected work.
This is much more obvious, I think, when generating images than it is with text (or code). Below (I hope it worked) is one of the Dall-E renderings of “A modern cubicle office in the style of Bosch”. Is it actually ‘good’? Perhaps not. But would it make for a reasonable illustration for, say, a piece about the ‘anomie of the modern office? Why not? I don’t find it boring, in any case.
The common attitude here toward LLMs seems to trend toward a kind of AI Luddism, but I’d suggest folks ought to spend just a bit of time with these systems just to get a better sense of what kinds of ‘idea looms’ they’re wanting to break.
https://gcdnb.pbrd.co/images/hUSNKQc6mIB1.jpg?o=1
Laban 05.05.23 at 11:08 am
Re inventing sources – I’ve heard this elsewhere so some checking would seem necessary at the moment.
Re ChatGPT and coding – that’s what I use it for, especially if it’s a language I’m unfamiliar with. In my IT work 95% of code isn’t written from scratch, you take something vaguely like what you want and heavily edit/tweak. It can build me something like what I want, and I then hone it.
engels 05.05.23 at 12:06 pm
Chat-GPT’s take (via chat.chatgptdemo.net):
Sashas 05.05.23 at 5:07 pm
@mw
I see I misunderstood what you meant by “novel” and “original”. I was speaking of the terms in a philosophical sense. To clarify, consider: I claim that if Impressionism did not already exist, generative AI will never invent it.
As far as “novel” and “original” go with respect to copyright or patent laws, I think it’s fairly well established that the purpose of those systems is to encourage art/invention by subsidizing the hard work that goes into the creation process. i.e. We make it economically (more) viable to pursue creation because if you do you get legally exclusive use of your invention for a period of time. The relevant question then is not can a lay person tell that an AI-generated work is derivative and what it’s derived from?. The relevant question is does AI-generated work constitute an infringement on a subsidy that we still want to protect?. My own thoughts:
Should we still subsidize the creation/invention process? I don’t see why not.
Does AI-generated art interfere with an artist’s ability to profit from their creation? Yes. One of the most popular types of prompts for AI-generated art is “in the style of X”. If X’s work is not in the training set, then the AI-generated art would be somewhat less close to X’s style and would compete less with X’s own productions. I say “somewhat” because styles overlap and GPT is fully capable of interpolating between Y and Z’s styles to approximate X’s. Unless X is truly novel in which case that “somewhat” becomes “substantially”.
As far as what copyright and patent law say right now, I’m afraid I do not care in the slightest so I don’t really have anything to say in response to your points there.
engels 05.05.23 at 5:07 pm
Doubtless lacking in literary merit and rather unfair but the use of metaphor, if nothing else, seems impressive:
mw 05.05.23 at 9:09 pm
To clarify, consider: I claim that if Impressionism did not already exist, generative AI will never invent it.
But that is also true of 99.9…9% of human artists, who will never invent a new genre either. Generative AI doesn’t invent anything on it’s own, but it does not seem far-fetched to me that a human providing prompts to an LLM image generator might not invent a new genre. Recall that LLMs aren’t trained only on images produced by artists.
Note, too, that Impressionism and other genres are not invented de novo. Turner, for example, was a clear precursor and in some ways was ahead of early Impressionists, particularly in the use of abstraction (compare Norham Castle Sunrise to Impression, Sunrise and tell me Impressionism was a great leap that came out of nowhere).
As far as what copyright and patent law say right now, I’m afraid I do not care in the slightest so I don’t really have anything to say in response to your points there.
And I’m not sure what to say to that. Is this year zero during which we knock down all Chesterton fences and start anew? Regardless of whether starting from scratch would be a wise idea for intellectual property protection, the chances of it happening seem effectively nil. So why pretend we won’t be starting with the existing framework?
bad Jim 05.06.23 at 2:28 am
Once again, rhymed couplets with no discernible meter.
Fake Dave 05.06.23 at 10:50 am
Meter might be harder to fake than one would suppose because it’s the most musical part of poetry and I don’t think even a sophisticated chatbot is capable of actually “hearing” a poem as more than just text. You can plug in a rhyming dictionary and teach it to count syllables (although it seems to struggle with that), but it doesn’t have any sense of the biological rhythms humans experience and expect so it just runs on and then thuds without any sort of catharsis. Of course, plenty of human poets do that too, so maybe it’s just a matter of averaging us out into something mediocre and uninspired as one of the early comments pointed out. I agree that some of the imagery is not too bad though. I wonder where it got it from.
Phil H 05.06.23 at 4:20 pm
I’m not sure how clear my thinking is on this topic, but it seems to roll mainly along two contradictory tracks.
One is that technology reduces the amount of work we have to do. I think this is true in general. The main thing that seems to constrain it is our own desire to appear to labour and to have others appear to labour for us. To the extent that AI does anything (and it seems striking to me how little it’s done so far), it would be only to push us a little further down the curve, so human labour becomes in fact less productive, but the appearance of human labour becomes in the market more desirable.
The other track is that technology seems to be removing the incentive to produce low-level entertainment. Where singing, dancing, and toy-making may once have been local affairs, they are now largely crowded by television and games. This may have the effect of eroding creativity at the lower levels, and ultimately making intellectual production even more of an elite affair.
The first idea suggests that we’ll end up with lots more interpersonal services, because the rich will essentially make the poor dance for them. The second idea suggests that we’ll end up with less interpersonal services, because everyone will just watch a few elites remaking Succession over and over. I don’t feel like I have any of the right tools to start picking through the trends and determine which is true. If there’s anyone rich in here who can tell me whether they’ve spent more time getting a massage or watching HBO recently, that would be a first bit of anecdata.
engels 05.06.23 at 10:28 pm
You can have any verse form as long as its rhyming couplets in loose iambic tetrameter it would seem.
mw 05.07.23 at 1:19 am
The other track is that technology seems to be removing the incentive to produce low-level entertainment. Where singing, dancing, and toy-making may once have been local affairs, they are now largely crowded by television and games. This may have the effect of eroding creativity at the lower levels, and ultimately making intellectual production even more of an elite affair.
I take it you haven’t spent much time on YouTube or TikTok then (or, ahem, Only Fans I guess). The number of ordinary people producing content and consuming content by other ordinary people is mind boggling. For some the incentive is sometime just to be seen, or to communicate something they care about, while others aspire to become well-paid ‘influencers’ (and some of them actually do make some serious money). Intellectual / entertainment production seems less an elite affair now that at any time in recent history (recent as in, say, the last ~100 years that we’ve had television and film & record industries).
bad Jim 05.07.23 at 5:39 am
Perhaps the odd thing is that ChatGPT can generate original text vaguely resembling verse at all; one might suspect that it has a special rule for this case: when asked for poetry, produce rhymes.
A different AI could certainly produce nearly any sort of poetic product on demand, and something like this might even be popular with songwriters, many of whom would benefit from some sort of meter-checker or rhyme-generator.
engels 05.07.23 at 10:36 am
The future we were promised: machines do all the drudge work and humans write poetry all day
The future we’re getting: humans do all the drudge work and machines write poetry all day
mw 05.07.23 at 10:50 am
If this thread hasn’t exhausted itself, here’s another angle (this time on AI and popular music):
So we can use AI to recreate singers’ voices and also to create new songs in their styles. Scary? Well, maybe, but consider the Beatles. We’ve had Beatles tribute bands for decades who specialize in singing and playing like the fab four — some quite convincingly and there have certainly been songwriters who can knock out a Beatlesque tune. And yet there have been no money-making ‘neo faux Beatles’ hit records. Why should that change with AI?
People will still pay a lot of money for a ticket to a Paul McCartney concert, but they’re far less willing to fork over money for even fairly well-regarded new material from the man himself. So why would we expect them to suddenly spend money on AI generated songs that sound like him? Pop music is arguably as much about the personalities as about the music. Do pop fans really want to hear music from nobody? I really doubt it. I mean there may be a period where the novelty of it all will be appealing and there’ll be an AI hit or two, but in the long term, I think pop music fans want stars whose sexy photos they can ogle, who they can watch on the late-night talk shows and whose rivalries and spats and messy divorces and dysfunctional offspring they can follow in the tabloids.
engels 05.07.23 at 10:51 am
People who are saying it isn’t as good as human poetry are missing the point: under capitalism you don’t have to be better, you just have to be cheaper.
steven t johnson 05.07.23 at 1:59 pm
engels@52 God help me, I really thought that was tetrameter.
Chris Bertram 05.07.23 at 1:59 pm
“And yet there have been no money-making ‘neo faux Beatles’ hit records.”
Well, there are Oasis and Tame Impala tracks that might qualify.
David in Tokyo 05.08.23 at 5:23 am
Engels wrote
a) humans do all the drudge work and machines write poetry all day
and
b) you don’t have to be better, you just have to be cheaper.
HEY!!! CUT THAT OUT. Winning the thread (if not the whole friggin’ Internet) twice in one thread isn’t allowed.
By the way, I’m not sure that ChatGPT really is “cheaper”. It uses a LOT of computation, and that burns a lot of electricity. (When Google beat the world champion at Go, there was already a commercial PC program (that didn’t even use a GPU) that was giving professional Go players a rough time, Google just reimplemented that and ran it on a server farm (that was AlphaGo). The self-play/self-learning work came later (that was AlphaZero).) My opinion remains that the LLM idea is an inane stupidity, but it’s a seriously kewl inane stupidity: it’s essentially buiding a table of every string of words every human has ever written, cross-indexing that by the frequency of usages of the words and phrases therein (with longer and longer phrases), and then cross-indexing that by (the probabilities of the words and phrases in) what the user types. This is a seriously large amount of computational work. At some point, someone is going to need to start paying those energy bills.
And that someone is you.
(By the way, Google let the Go team use the server farms for Go only for a very limited time after that, something like 50 games of online Go (it won them all). In terms of an investment, it was incredibly effective advertising, so Google had no trouble justifying the cost. But they shut it down real quick. At some point, the LLM game is either going to see the same fate or OpenAI et. al. are going to need to figure out a revenue stream.)
mw 05.08.23 at 3:14 pm
David in Tokyo @60
This is a seriously large amount of computational work. At some point, someone is going to need to start paying those energy bills.
Yes. The building, maintaining, and querying of LLMs is computationally expensive, and it remains to be seen how many efficiency gains are possible.
And that someone is you.
Well maybe. But not against our wills. It’s really the same as search engines — the web crawling and querying use computational resources. To date, advertisers have been willing to pay enough that end-users are not charged. Will that same approach work for LLMs? Maybe. Or maybe we’ll see subscription models (as is already the case with MidJourney and ChatGPT+ and CoPilot). When/if users have to pay the full cost, they may rapidly lose interest in using it for whimsical purposes like writing bad poems about superheros. But if computational cost proves to be an issue which limits the usage, wouldn’t that be reason to freak out less rather than more?
Mike Furlan 05.08.23 at 7:52 pm
engles @55
“The future we were promised: machines do all the drudge work and humans write poetry all day
The future we’re getting: humans do all the drudge work and machines write poetry all day”
As long as they don’t take up painting landscapes we are safe, otherwise we know how that is likely to turn out.
Fake Dave 05.09.23 at 12:03 am
Over at The Guardian, <url=https://www.theguardian.com/commentisfree/2023/may/08/ai-machines-hallucinating-naomi-klein>Naomi Klein weighs in:
“)There is a world in which generative AI, as a powerful predictive research tool and a performer of tedious tasks, could indeed be marshalled to benefit humanity, other species and our shared home. But for that to happen, these technologies would need to be deployed inside a vastly different economic and social order than our own, one that had as its purpose the meeting of human needs and the protection of the planetary systems that support all life.
And as those of us who are not currently tripping well understand, our current system is nothing like that. Rather, it is built to maximize the extraction of wealth and profit – from both humans and the natural world – a reality that has brought us to what we might think of it as capitalism’s techno-necro stage. In that reality of hyper-concentrated power and wealth, AI – far from living up to all those utopian hallucinations – is much more likely to become a fearsome tool of further dispossession and despoilation.”
David in Tokyo 05.09.23 at 5:47 am
mw points out a hole in my argument: “It’s really the same as search engines — the web crawling and querying use computational resources. ”
But the web indexing trick was/is kewl comp. sci. because it uses very little computation; the LLM thing is about throwing vast amounts of computation and data at the problem (i.e. actually reading and processing all the data on every page, not just counting the words and links). My intuition is that the LLM thing is way more compute intensive than web indexing/pageranking.
But your point that you are already paying for web searches by reading the advertisements thrown up is also well taken. And brings up a pet rant here, namely that I wonder when the advertisers are going to figure out that advertising doesn’t work? At least for myself, I almost never see ads for something I’d buy. (The only counterexample is the book ads in the Japanese newspapers. Even if I don’t immediately buy the book, I may remember that the author’s publisher was plugging her, and that maybe I should pay attention to her writing.)
Whatever.
Comments on this entry are closed.