ChatGPT

From language models and AI-assisted coding to art generation and incredible deepfake technology, 2022 has opened some big AI-related cans of worms. I have worries of many different natures about what this will mean in the coming years, and I wanted to share them.

Code generation

I'm starting with this since this is the one thing I actually know a thing or two about. First, I wanted to showcase some of the capabilities and use cases I found for ChatGPT in particular, so without any particular order, here they are:

It still has many limitations, the main ones being limited context (it cannot consider an entire project, understand every functionality and make sure a change doesn't break anything), and the fact that it still can get things very wrong, introducing bugs. Copy pasting code you don't understand from it can be rather risky, since it may introduce hard-to-find errors/vulnerabilities. It won't be replacing developers in the immediate future.

However, I have no doubt that's the end goal. Microsoft (major OpenAI investor) employs 221k people worldwide, and 112k just in the US ¹, with over 100k of those being developers ² That's one big payroll. Reducing that number could easily save billions of dollars. We've already seen a large number of layoffs this year ³, I'm hoping this doesn't cause this trend to continue.

Furthermore, it can be a great service to offer to other companies. We're already seeing paid tiers for both Github Copilot and ChatGPT, but in the long term I can see a product that can be used with zero programming knowledge that, for example, builds a website to your specifications, with any custom functionality you may need. Something like Squarespace, but orders of magnitude more powerful. That kind of product can quickly remove the need for non-software companies to have any devs.

So the incentives are very much there to replace as many developers as possible. I don't believe every dev will be fired yet, bugfixing and other things that require a more complete understanding of the product are still out of reach for AI, but the demand for tech workers may decrease enough for it to stop being such a well remunerated profession.

One last though I have is that it will make getting into programming (on a professional level) much more difficult. Trainee devs are not that different from these AIs: they both require a rather detailed description of the tasks at hand, and help along the way; a more senior programmer must review their code more thoroughly to catch mistakes; neither of them know the full context of how the product works. But AI is pretty much instant in its code generation, already knows the basics of whatever language/framework you're using, and it's probably way cheaper. This may either significantly raise the bar for entry level positions; or maybe unpaid internships will become even more common in this industry. Either way, it doesn't look great for inexperienced devs.

Content moderation/Spam

Content moderation has always been a complex endeavour in the internet, due to de sheer amount of content that is generated every second. But until not too long ago, at least is was viable to flag bot generated content somewhat reliably. Recently, an absurd amount of legitimate-looking content has flooded every platform.

Stackoverflow had to ban AI-generated responses, mainly because the number of mistakes it made ⁴. It is unknown how they enforce this rule ⁵, or whether they can even do that effectively.

Clarkesworld had to close submissions altogether ⁶. As they explain ^{7 8 9}, most solutions don't really work for them, due to being too costly or too restrictive.

Since these kind of practices can be quite profitable (from scams to karma farming/account selling), this won't stop any time soon. This sadly has the potential of completely ruining many sites with user-generated content, especially the smaller ones that cannot afford the necessary tooling for detection (assuming this tooling exists and is actually any good). Even worse, this is one of the few areas where a completely open source solution probably wouldn't be too helpful, since by its very nature it would reveal potential attackers how to defeat it.

Art generation

I see artists being the ones that suffer the most in the short term. Obviously this technology is not ready for replacing top-tier profesional artists, but it is already amazing for quick prototyping and brainstorming. As usual, smaller creators who are already struggling to make a living will be the most affected. Why buy your furry fanart from some dude on Instagram when you can get a custom drawing for free, from an AI that was trained with this guy's art.

As good as these can often look, I do fear that originality will be lost if AI starts being used in mass. There's a finite amount of human created art to train the models¹⁰ and (supposing human-made art is dead¹¹ ) at some point we will run out of new styles and ideas.

And I'm not only talking about images, but also video, music¹² , and any other form of art. AI already exists for all of them, and it'll only get better. Considering that most media companies already have a tendency to prioritize quantity over quality, I do not believe it will be that long before we start seeing some of these things used in mass¹³. But I'm not sure I want to sit and watch a soulless 6 season Netflix show that feels exactly like all the others except that it has targeted advertising baked into it.

Apart from directly supporting independent artists, I think that one of the most important things in this area will be potential legislation to regulate the use of copyrighted material for training.

Another thing that worries me in a much longer term is a disappearance of shared experiences. What happens when everyone can have an AI-generated movie/song perfectly tailored to their tastes? Will the idea of watching the same movie as other people and then discuss it die? Will you be unable to talk to strangers about the music you listen to, because they have literally never heard of it? Even if the creations are in fact perfect, this detracts from the experience of consuming it in a way that I do not believe can be made up for.

Confirmation bias

An interesting pattern I noticed in my own usage is what happened when I asked it a question I already thought I knew the answer to: When it gave me another answer, I tended to asume it was making it up, and just clicked on "Regenerate response" until it gave me something that aligned with my previous beliefs

When I noticed this, it got me thinking about this new type of confirmation bias. Now, you are not limited to trying to find people that agree with you; you can try to make an AI agree with you. This isn't limited to making generate as many different responses as possible, another more subtle way is by having some (probably unintentional) bias in the question itself.

We've already seen what widespread cognitive biases can do (*gestures broadly at every single conspiracy theory*), so I don't need to explain the risks; but it is apparent to me that this can be considerably worse than existing search engine related biases.

Licensing

I'm really not a lawyer nor do I have enough knowledge to add much to the conversation here, so all that I can say is that:

1. Licensing in this area right now is a nightmare, and I expect some big lawsuits in the following years. Even if I wasn't concerned in any other way, this would be enough to turn me away from using any of these tools in a profesional setting for the time being.
2. Consent from creators of any kind (artists, programmers, or even the writer of some Reddit post) should be required if their creation is to be used for training, especially if it's done with the intention of profiting. This should come with an appropriate compensation.

Skynet?

Despite it sounding a bit exaggerated, it may be worth considering that creating a super-human intelligence with goals misaligned from ours is not a great idea, especially without having a good system in place to control something like that¹⁴.

This has already been quite long, and I know for a fact I couldn't explain this better than Rob Miles, so I recommend checking out some of his videos. If you don't know where to start, I recommend the one on instrumental convergence (i.e. why it would want to make bad things even if it has the right end goals), or the one on reward hacking (i.e. why it's so hard to define the right goal in the first place).