I don't think it's quite right. AI is much better as a helper than as something that does the whole job. It also doesn't actually think, so replacing doctors with it would be a terrible, terrible idea.
Now, as an assistant, that's amazing. As an actual finished result, it's not good. It needs tweaking and fixing, to make everything work right. Same goes for AI images.
So here's what I think AI will actually endanger: stock photography. If you're writing a blog and need to stick a random illustration on an entry for extra appeal, the AI is perfect. You don't care if it fits in a theme, or if it's quite right. You need a picture of a cute cat, so most any cute cat will do.
Doing actual, specific illustrations with an AI is hard work and you'll find beating your head against it quite frequently, especially if you need more than one, and if you need something that's not quite well covered in the dataset. AI works much better to compliment a human artist that can guide it to generate what's needed, and fix the problems in post.
Obviously, I agree replacing doctors is a terrible idea.
That said, in the last 5 years I have broken an ankle in a rock climbing accident, and gotten a herniated disk in my neck from wrestling. Both times, medical professionals misdiagnosed me for months (both times until I eventually found a doctor who would refer me for an MRI which proved the issue in both cases). This resulted in a lot of undue suffering and increased damage.
In both cases, I explained the situation and symptoms to GPT3 (this is before ChatGPT was released), and it correctly diagnosed the issue in both circumstances, down to the exact bone I most likely fractured in my ankle (talus), and down to the exact vertebrae I was likely herniated in my neck (C6-C7).
I think this may end up being the most powerful application of GPT. While I've been using it for coding, most of it is high level "what are the trade-offs?" kind of thinking and very little actual code.
A person with some background and sufficient general knowledge, but who is a non-expert on a topic, can dive into a topic and allow the AI to handle most of the detail work. What do I need to know to understand this? It falls out naturally from the exploration of the topic, in a faster and more integrated way compared to chasing down articles and definitions.
It's like having a research assistant, that is an expert that has read everything ever published about the topic, and that is also infinitely patient. Of course, it's also prone to bouts of being confidently incorrect, and it can't actually reason, or think up anything new. That may seem like a serious limitation, but it also describes many of my college professors, and they still taught me plenty. ;)
I think ChatGPT is most similar to text-davinci-003 which does have gaps in it's knowledge which it can be stubborn about but code-davinci-002 has fewer gaps and can be more reasonable about requests. It's sometimes gets stuck in a loop though and is in beta with very limited capacity.
Point being ChatGPT is the not quite best programming model.
I watched a show in the 80s about expert systems essentially playing 20 questions with a patient (via a human typing assistant) and the computer was already more accurate than doctors in a study. The fucking 80s.
Doctors have a very powerful Union. They don’t call it a union, but it’s a union.
Perhaps it's not going to be about replacing doctors, it's about remembering everything they know in aggregate and distributing it wholesale to augment their work along with healthcare provision in general, while increasing access inclusiveness by empowering patients. Has the potential to be most transformative in parts of the world and socioeconomic stratums/zones where internet connected smartphones are widespread, but timely, inclusive and affordable healthcare is not.
I'm an ex editorial illustrator and midjourney can do the whole job often enough. With about 20% of ideas I try out I've been able to get submission ready work. Maybe another 30% can be cleaned up without too much hassle in Photoshop. I don't expect these odds to get worse in the future.
A photographer friend of mine had a gig to do some festive food arrangements. He needed a full day and a half, a studio, $30k in gear and years of professional experience. All I needed was 5 minutes and some clever prompt poetry. I did not have the heart to show him why he might have a hard time getting a similar gig next year.
As Clayton Christensen's theory of disruption proposes: the disruptive product does not need to be better. It needs to be good enough and far cheaper. So cheap now that there is zero marginal cost for bespoke art.
That's what people have been telling themselves for over a decade, I think. At least articles on news sites about who should be wary of AI and how much had been quite fashionable back then. And then people would respond with "yeah, it will probably automate away those other jobs and it will be able to do the simple and boring part of my work that unskilled/low-end people in my field do, but experienced/knowledgable dudes like me will always be in demand. It's just that our work will be more interesting." And you would hear lawyers and graphic designers and accountants respond with this same theme. (For some reason, software developers were always thought to be untouchable anyway.)
And then AI came a long way, improved to levels hard to imagine back then, it's already threatening programmers, but people just keep telling the same story. That somehow it will run out of steam below their level. If it can talk to a professional it can just as much talk to whoever talks to the professional now. (I mean, unless it gets stuck at a lower level.)
However, when people talk about the future dangers, they still keep looking at the current status quo and not at how fast it evolves. Of course, it's hard to predict, but it seems like some people don't even try and just judge based on what we have today. While what we have been seeing so far is that it's evolving faster than what most experts have expected.
Replacing doctors is probably further away, because being a doctor is like being an airline pilot: it's about trust big time. And everybody likes to think their doctor (or pilot) is especially good. Oh, and also doctors do have to interact with the real world, like examine the patient. (Programmers don't ;) )
I agree 100% with you. I have a free digital monthly magazine in PDF. It has articles, etc. Finding images was a hard task: I had to find good images, creative commons, etc.
For the last month I paid MidJourney and it was amazing. I enjoy doing the magazine much more now. As you said, the image doesnt need to be exact, they must have some objects / surface a topic or idea. And I can modify them with Photoshop if I want to change something (or add text).
I think that's where the IA will prevail: scenarios where the requirements are not really important.
To be fair though, that's the poem it produced when the entire prompt were the words "ode to a space lizard".
It's like with describing a commission: If you just give basic info, then there is a lot left for the artist to guess, and they might be inclined to err on the safe side.
I often got better results when I included more info in the prompt.
E.g., here that could be: What is a space lizard? What's so amazing about it? What should the general emotion of the poem be? Should it be of some particular style? etc etc.
True, I was intentionally trying to make it do something weird to see how it'd fail. But it's not just a lack of information. A tail might help with swimming but won't propel you through the void. And "anything deployed" is awkward. And what's an "intergalactic norm"? And so on.
I think more information would result in something more interesting, but I don't think it'd fix the problem that some of the concepts used just don't add up, because it's not actually a thinking machine, just very fancy statistics that work surprisingly well.
>A tail might help with swimming but won't propel you through the void.
No offense, but I feel like if you're criticizing ai-written poetry for a lack of absolute realism, maybe you're the one that doesn't quite grok poetry. Next you'll be telling Poe that ravens can't actually enunciate human speech.
It's 3 am in the morning, there's a medical condition that came up suddenly, I'm on hold forever - you can be sure I will jump on an AI first. I'll get clear answers first that probably compete with or exceed over the phone diagnosis accuracy and the virtual diagnostician wont disappear after I hang up.
That will cut costs, so it will be used no matter the outcome. Humans make mistakes too, right?
US is ripe for multi-tier medicine. Limiting diagnostics to AI will be preferable to huge medical bills for so many.
Ignoring the mountains of regulation the issue with doctors is that bedside manner plays a huge role in effective treatment. Patients are often hiding symptoms in order to avoid a scary diagnosis or unaware that something they've been experiencing for years is worth mentioning. There's a very human element involved in extracting the data you need and convincing people to follow treatments.
As people hook themselves up to more AI assistants and health monitoring applications I'm not sure if that will always remain true. Our telemetry devices may start to pick up on these things before we even do.
Is OCR that good yet? From what I've seen, it's good if you have uniform text in a single standard-looking font. When layout or font varies widely, or there's extraneous stuff on the page (for instance: headers, footers, page numbers, marginal notes, same-page footnotes, low quality source material with marks on it, or scan artifacts), quality degrades. Good OCR engines I've seen can still OCR all the things and present them in a somewhat readable text format, but general intelligence (human or AGI) allows quick, automatic recognition of different sections of text that a narrow OCR AI struggles with. A human or AGI knows this text and that text are both blockquotes, or marginal notes, and instinctively attaches semantic meaning to each area, font, style, color encountered. An OCR engine struggles to get beyond blocks of text each with their own margins and no semantic meaning attached, leading to markup hell.
To highlight the limitations, look at an OCR'd version of a technical book with code samples and different fonts and styles that have different meanings, and that has both footnotes and endnotes. The text will be readable, but disorganized, probably inconsistent styling, and even if some footnotes and endnotes are linked by a good engine, I suspect that's less than fully reliable. For the purposes of reading the book, I'd rather have the scanned pdf with page images for reading, with the OCR'd text as the text layer for searching.
Lower-quality source images seem to cause major problems for tesseract, and even ABBYY judging from archive.org text conversions. Those engines confuse more ambiguous letter or punctuation combinations, while humans can still read the images without much trouble.
Thanks! It's definitely a short list at this point.
I had a few more items in mind but was too eager to share it.
I will add some of your suggestions in the coming days.
> Spotify knows me better than anyone and every song on my Discover Weekly hits the spot!
I'd wish. If there is one area where AI had utterly failed it's ads. Product recommendations are always off for me. I love a joke that goes like "meatbag bought a sofa recently, must like sofas, let show them all the sofas we have" because it's silly but very true. And checking out those items leads you down to a rabbit hole where some smartass decided that if I look at something it means I'm potentially interested in more of that, so a machine must feed me more of "related products".
Spotify recommendations are 90% of time provide no value to me - I stopped bothering with their recommendations and "discover" playlists. I actually hate that they start playing music of their choice when my playlists are over. The only reason I keep a subscription is because a) they have a fairly decent collection of things I like and b) it's easy to share a playlist with my wife.
Same with Kindle book recommendations. Same with all those streaming services. Same with Steam games. Even though I spend some time there trying to tell all those platforms what I like and what I don't.
I always have to browse catalog, checking every single thing individually. Like a quest looking for a good apartment on Airbnb among the sea of places that weren't ever good for living, only crashing for the night. Hopefully there is some user-curated collection to start with - because "related" stuff automated classifiers generate are typically echo chambers full of things I'm not looking for.
The problem is, machines recommendation systems fail to understand the reasons for picking one thing over another, because they don't know much about the product they're recommending, only its relationship with other products. They don't read the book (or watch a movie), so they have no clue about its language (or actors play), tropes, and how interesting (or stupid) the plot is. They don't listen to the song so they have no clue about how it sounds like, the vocals, and certainly don't care about lyrics. And so on.
Maybe you're just an outlier and AI doesn't show you good ads because it doesn't have a set of good ads that meet your needs that pay a decent amount?
Unfortunately with bulk data over populations you are going to have some members that are in the long tail. Instead the operators of the system look at the efficiency of the system over larger parts of the population, even if it has the potential to lose sales for a small part of the population.
Yeah, well, I could imagine being an outlier with some service or two - but everywhere? I'm a grumpy ass who loves to nitpick, but I don't think I'm that unique.
Even the Google Ads (which are supposed to have some magic pixie dust algorithms done by the fanciest experts in the field because that's Google's bread and butter) are typically showing me some products I couldn't care less about. It's always some random post somewhere (which could be an ad) that actually makes me interested and drives me to making a purchase, not some bullshit video banner.
> the operators of the system look at the efficiency of the system over larger parts of the population
My guess is that they look at engagement metrics, not consumer happiness or satisfaction. And even though I'm pretty much disappointed, I still use those services and still click on "related products" and even check some book samples etc. - in hope that maybe I'm wrong, and maybe this time it's a good recommendations. Very infrequently that happens, but mostly it's not what I'm interested in. So I go find some non-machine recommendations, search for those and thus remain an active user operators want on their platform.
I agree most recommendation engines simply suck. (I can't imagine how they spent millions hiring Machine Learning PhDs and end up with that crap)
That said, youtube recommendation algorithms are usually sane. I suspect the depth and breadth of available videos (which are actually potentially interesting) helps, whereas for ads, the platform's primary goal is to shove unwanted stuff down viewer's throats, and if 99.9% of the people don't actually want to see the ads, they still need to pick the ones "most likely" wanting to see them, even if that probability is like 1%.
I feel like the specific case of "sofas forever" is because of a clear lack of feedback channels.
If you buy a sofa, especially at a different vendor, it doesn't send a signal back to the ad network which could be used to deprioritized the sofa ads.
There's also no trust in the advertiser relationship, so you couldn't just have an "already bought/don't show anymore" button because people would either ignore it or insist they bought everything in the hopes of poisoning data or suppressing ads.
Ad algorithms are never going to do what you want because companies which make great products don't need to advertise. Advertising is a mechanism for inducing artificial demand: the products you actually want don't need to be advertised to you because you'll seek it out at no expense to them - the market for your attention is in manipulating you into buying inferior and overpriced alternatives.
I'm generally bearish on AI, but I do feel that there are going to be a surprising number of applications where 'good enough' results are actually perfectly fine.
In particular, a couple of industries that I think are likely to get disrupted in the next few years are stock art, and the art-by-commission scene. The ability to use a Stable Diffusion + Dreambooth style setup to easily generate new artwork in the style of your favourite artist that's "good enough" is incredibly powerful.
I personally would appreciate a more expansive and less opinionated (forgoing the value judgments about the ultimate good or bad about AI being applied to something) version of this, as a way of tracking AI progress month to month. Maybe crowdsourcing a rating system from 1-10 in each category for how well AI can complete each category of task.
I miss programmers from the list. For some reason we tend to think we'll be among the last ones to be automated away, while programming is probably the best documented job on the internet, the job with the largest training set. We also see that the machine is coming for us, we just keep telling that nonsense story that it only does the boring/low level stuff and imply that it will be like this for a long-long time which in my interpretation would be several decades at least. (Otherwise I don't get the dismissive tone/remarks.)
OTOH recognizing songs can still make you look cool. Like being good at chess, poker or being able to lift heavy weights are still cool even though machines are better than people (or just most people, in case of poker). The only downside is that now you can be caught being wrong ;)
If programming is automated but driving trucks still takes a human, someone will tell GPT-5 "Write a program that automates driving trucks". If that doesn't work, programming isn't automated. If it does, it doesn't matter that programming was the first of the jobs automated in the last 48 hours of the age of man.
It's not all or nothing. At least it doesn't make sense to talk about it as if it was. It doesn't have to do away with all programming jobs to make it to the list. (Especially, since the list itself defines differing levels of dangers.)
And, of course, it doesn't have to be at the level of producing a perfect solution for the input "Write a program that automates driving trucks". Programmers can't do that either. It's enough if it can produce usable and testable code based on user input and then it can refine the solution based on the feedback and then get to a complete solution with this method. If it can start asking questions, all the better.
BTW, ChatGPT can already do some of this without being tailored to do so. You can tell it to create a todo app (I know, that's just copypasta), defining the language and the framework to use. (I went with TS and Vue.) And then tell it to create a backend and to connect the two. Then you can tell to add persistance to the backend (because it does not, by default) and it will update the backend code. Before ChatGPT, I thought that back references like these would be the tricky thing. Sure, in my experiment I didn't talk like a non-tech person wanting an app, so you might say I was programming, but I really wasn't. And that's where things are now.
> it doesn't matter that programming was the first of the jobs automated in the last 48 hours of the age of man.
Now that's the thing I have always been unsure about. Whether automating programming is far enough from (super strong) AGI for programmers to feel bad about their job :)
You're confusing "automating programmers" and "automating super-human programmers". Automating programming in this context means replacing 99% of the normal types of programmers. We don't know how to program driving trucks with humans yet, so there's no driving-truck-programmers to automate.
The fact is given the state of AI development today, all our jobs could be in risk in a couple years. I'm not betting on that for now since I personally think the chances are still low (due to other reasons). But being in denial through creative wordplay doesn't change the fact a single bit.
Reminder that all argumentation about the impact of AI on society, industry, etc., should be made with great care with respect to prediction.
It is a very common trope to see assertions about what AI does and does not do, and how this naturally limits impact in various ways.
Every assertion about such limits should be bracketed, always, by, "today."
Not because every such limit will fall within some specific timeframe; but because we continue to see advances which make prior arguments of this type appear woefully shortsighted and naive.
IMO we as of yet are literally unaware of what the natural limits, should they exist, are. We are aware of a great many problems, always have been; but we are also finding that a surprising number of them fall before "scaling" in ways not many predicted.
And computation continues to get cheaper. And resources continue to amass.
A useful question for me is what is the relative cost and timeframe to increase the resolution of indistinguishable replication behavior of a system from that of a human agent, in various domains.
That is a moving, fractal, shoreline.
How fast is the tide coming in? How high shall it get? Those are the questions.
Try to frame reasoning about impact in terms of such dynamicism.
I thought it's going to take decades to dethrone Google. Lo and behold a chatbot funded by Microsoft is giving some serious competition to Google. Bing might finally beat Google in case they don't rename Bing or launch completely different product!
Google is toast. It has been in continuous regression for the last 10 years. On the other hand, language models are progressing by leaps and bounds.
It is likely that within the next year, a new open chatGPT model will emerge that is comparable to "ClosedAI", to be used as the next stage after search. This model, along with others in its family such as Stable Diffusion and Wisper may become integral components of browsers and operating systems and could potentially be used as the main interface for accessing the internet.
As people become tired of ads and spam, they may turn to language models as a way to shield themselves from these distractions. These local models may also serve as personal creative spaces, allowing individuals to work and explore ideas without outside interference.
Google can trivially duplicate the research and compute needed for their own version of GPT-3 if they want to. Given the size of the company, I wouldn't be surprised if several people or teams have done so without them all knowing about each other.
Google also has at least two generative models of their own that they, like OpenAI is criticised for, don't publish "just in case"; one for duplicating voices, and the LaMDA chatbot that got in the news before ChatGPT.
People will get tired of ads, but I bet someone will use a text agent to rewrite stories and scripts so heroic characters always enjoy the great taste of $beverage while saving the day, the lead romantic opportunity is $consumer_gender_preference and likes wearing $fashion_brand.
Can duplicate, yes, but it's not going to increase their revenues. The chat interface is less profitable than the search box + ads. Search has reached a paradigm change moment.
The search concept might not be that disruptable in a lot of niches, because chat tends to provide an answer without context.
The traditional search-box-and-list-of-URLs at least provides attribution, which can be of huge value in some search scenarios.
If I have a health-related query, I want to see the links from research papers from legitimate institutions, or trusted authorities like the NHS. A human-language answer doesn't necessarily say "based on the groundbreaking research of wificausesherpes.com" at the end.
Any sort of comparison and recommendation searches are similarly iffy. That's why there are a trillion "best 8k webcam under $30 December 2022" affiliate-link-farm sites, but The Wirecutter and Consumer Reports still have some minor level of clout.
google already has this technology and have submitted their own research. this is just hard to deploy at google scale. chatgpt is millions, google is billions and many companies rely on that entire ecosystem/ads
> organizationally, google is a hot mess that can’t even figure out how to launch a product
Maybe, but for Google LLMs wouldn’t be a product, they’d be an algorithm behind the universal query box and conversational assistants, which are established products. Adding additional backing algorithms and changing output UI to prioritize their output isn’t something Google has a problem with, even if it arguably does with new product launches.
Google have literally invented LLM using Transformer. LLMs are trained on large corpus of text. Google have bigger corpus than anyone. So this prediction of google's demise don't make any sense. ChatGPT literally have no advantage in terms of technique and dataset that Google don't have.
Reason companies like google are exposing their internal LLMs are they are not production ready to handle 50 billion queries a day
> Reason companies like google are exposing their internal LLMs are they are not production ready to handle 50 billion queries a day
If anyone can do that, Google can. Even at that scale, I'd bet Amazon and Microsoft can also do that right now, possibly also Apple and Twitter, and the only reason I'm not listing Tesla is that (to my surprise) SOTA image AI uses way less RAM than SOTA text AI, and Tesla's probably all about image data.
Nah, all of them have got a lot of potential reasons why they might not make them public, but I don't see that being one of them, and especially not Google given how often they jump in with half-baked launches that they try to make work later.
I think sites with user generated content are accurately flagged. There is a danger that infinite, free, low quality content will lower the value of their platform unless they figure something out.
I disagree with occupations getting flagged. For many professions, AI will become human augmentation. Illustrators, writers, and other forms of design will require people to adopt new tools but they will become more creative and more efficient.
This list significantly affects journalists, writers, illustrators and digital artists and especially StackOverflow which programmers both senior and junior will be impacted and less roles going around for both of them.
As for paramedics, doctors, lorry drivers and Google, etc the impact of AI is hardly a threat to them and this website has greatly exaggerated their own predictions.
Great backstory! Been playing around with a very similar tool (for a specific, niche context: context for archives of government documents) and then was excited for GPT-3 to help solve it versus other options and have run into the same experience: Good, very expensive way to get what feels like a great solution and then you start noticing the many cracks.
I feel like people talking about "Killed by AI" don't understand how Machine Learning works. There are a lot of fields that can be "killed" by current AIs if trained correctly - almost anything menial, repetitive and something that people could answer on their own, but "skill issue" prevented them from.
Art? Yes, you can get it to draw something in style of, but it won't be able to create new style.
StackOverflow? Well, what are you going to train it all? Current AI is good at answering questions that been already answered by SO.
Spotify's recommendations? Absolutely not. Probably the worst recommendations out of all streaming services. Sharing playlists EW not because of AI, but because of music accessibility at unprecedented levels. People don't even make playlists for themselves anymore, let alone share them. Sidenote: 4/5 of my Discover Weekly is insta-dislike because Spotify doesn't get _why_ I liked the song similar to it.
Art? Yes, you can get it to draw something in style of, but it won't be able to create new style.
I don't think that's true. Because DALL-E knows how to create art in various styles, it implicitly knows what are the parameters which separate one style from another -- for example, thickness or shape of brush strokes, degree of realism vs. surrealism, typical subject matter, etc.. I imagine if you asked some variant of DALL-E to create a new style, it could arbitrarily alter the parameters that define "style" and come up with something new, and even alter them in a way such that the new style could reasonably be expected to be pleasing/well-received.
It knows how to create in styles it been trained on. Yeah, you can tune and tweak parameters, but it's not the same. I will for sure kill most of digital artists, though.
It is not true at all. I have already seen more new styles since summer from Stable Diffusion than in the last 20 years of going to art galleries.
Every artist has their influences. If they come up with something new it is from a unique combination of their influences. Stable Diffusion makes this literally trivial.
A big difference is that a good artist does this in a meaningful way, creating novelty in a way that triggers a specific emotional impact or conveys a message. With generative AI, it’s all statistical sampling with any “meaning” inferred after the fact. But I do believe AI will be invaluable in helping artists explore idea space and refine their vision.
I think you just discovered that art galleries either show art by dead artists (can't be a new style) or zombie formalism. I haven't seen anything groundbreaking new by Stable Diffusion or DALL-E. Well, sometimes it produced funny images because it doesn't understand what it's actually "painting".
So what's really going to be killed by AI is originality. The moat to cross to get to anything original will get wider as AI makes it easier to churn out rehash.
The only real "threat" of AI is it lulling people into apathy and shutting their brains off. When the majority defer their thinking to AI because it's "good enough," civilizational progress will plateau. You can't invent the future using a machine that was only trained on the past (especially if you've atrophied your own ability to think about/interpret what the machine spits out).
I don't think it will get killed, but I think those models will eventually replace people that work on non-original stuff. Like, I don't doubt AI can "direct" the next 100 CoD games and no one will notice anything.
If the obnoxious ads I keep seeing on tiktok are any indication, replika is trying to take over the "unlicensed therapist" sector (though based on recent ads, they seem to be pivoting towards "sexting"? Unclear.)
You say that as a joke, but I actually think it's a very relevant point. People keep pointing to mistakes the AI makes in order to make an argument that there will always be a place for people doing that kind of work. The website in this post also does this. But often they fail to take into account that people make mistakes all the time as well. And the success rate of AI is only going to get better.
Me: Can you write a short title for a forum post about a humorous website that lists things that people think AI is threatening to replace. Bonus points if you can give credits to "Killed by Google", a website that inspired it.
ChatGPT: Here is a possible forum post title that combines humor with a reference to the website "Killed by Google":
"AI Taking Over the World? Here are the jobs (and hobbies) it's gunning for (with a nod to 'Killed by Google')" [...]
Me: The same but not trying to be too cute.
ChatGPT: Here is a more straightforward forum post title that gives credit to the website "Killed by Google":
"AI Threats: What is Artificial Intelligence Endangering? (Inspired by 'Killed by Google')"
The "RIP-off" is a terrible pun that someone would utter and say "Get it, do you get it, because it involves death, and it's a copy, a rip-off!".
Don't try to be too cute, OP (submitter's username is the same as the github username linked at the bottom of the webpage, and the RIP-off (har har har) pun is also at the bottom)...
Thank you! It's my first post and I have to admit the rules confused me a bit.
In Show HN rules: "Off topic: blog posts, sign-up pages, newsletters, lists, and other reading material. Those can't be tried out, so can't be Show HNs. Make a regular submission instead" => felt like this would be categorized as a list but I'm glad you promoted the post as Show HN.
I asked it to write a poem and got this: https://imgur.com/DBHjki5
Now, as an assistant, that's amazing. As an actual finished result, it's not good. It needs tweaking and fixing, to make everything work right. Same goes for AI images.
So here's what I think AI will actually endanger: stock photography. If you're writing a blog and need to stick a random illustration on an entry for extra appeal, the AI is perfect. You don't care if it fits in a theme, or if it's quite right. You need a picture of a cute cat, so most any cute cat will do.
Doing actual, specific illustrations with an AI is hard work and you'll find beating your head against it quite frequently, especially if you need more than one, and if you need something that's not quite well covered in the dataset. AI works much better to compliment a human artist that can guide it to generate what's needed, and fix the problems in post.