Generative AI will self-implode - a prediction

When ChatGPT, Stable Diffusion and Dall-E were unleashed on the world, a revolution not very unlike the industrial revolution took place. AI, Large language models, and generative tools became something everyone was talking about, thinking about, seeing opportunities in and being afraid of.

Often several of these at the same time.

AI also became the focus of conferences, tech events and the like. With everything from how to use it, build systems that utilised it, and integrate it into your life in any way possible. For example, Microsoft's big Ignite conference both this and last year had AI in its tagline. It got that big in the public consciousness, virtually overnight.

Then the miss-use started

Breaking the boundaries on what the engine was allowed to answer and create, creating undetectable computer viruses and recipes for new explosive. Image generators creating images that never existed used in propaganda and marketing, generative AI writing essays in schools, depositions and arguments for courthouse proceedings, propaganda and alternate facts.

Then the cracks began to show

In images, hand and feet received extra digits and had odd angles, placing people in places and written texts were flat and devoid of personality and flavour, containing facts that were made up. Stories began to surface about how Open AI had used cheap labour to verify data that was thrown into the training model and who that in some cases had traumatic consequences.

The quality of output from the generators of course improved with time, but researchers started to notice something with large language models. The more the technology got feedback data and new training data from the internet... the dumber it became to the point that it became non-sensical until a complete reset was done. Why was this? Further investigation showed that the AI had taken in text written by generative AI into its training data. Ever heard of the saying "The blind leading the blind"? It's something like that.

And the pushback was, as one might have suspected, severe.

Companies, researchers and even students came up with ways to detect AI generated text. Artists began to incorporate watermarks to a higher degree and in some cases even removing their art from the online sphere all together or uploading intentional low-res versions. In some places, like art station, artists began to upload "No AI" artwork, flooding the platform with static, forcing the company behind it to take a stand against the use of AI on the platform, be in for scraping or generating, no matter how small the portion.

When studios in Hollywood started to embrace the technology, the WGA went into strike, with SAG-AFTRA following suit as the studios wanted to use AI for digital performances, writing shows and more. At the time of this writing, not all strikes are resolved.

Poisoning the well

A part of the pushback has not just been various ways of detecting AI generated content, be it in written or image form, but also tools to actively destroy or damage the model with repeated iterations. As with any recursive learning system, small changes can have huge cumulative effects. One tool for images called Nightshade, embeds data, invisible to a human, but very visible to an AI, into an image to give it false identification of an, or several, objects in the image. A demo convinced a local build of the latest Stable Diffusion image generator that a dog looked like a cat.? As the iteration started the results were odd, malformed until they coalesced into a cat with more iterations.

When it comes to text, the methods are a bit different as in that you still need text to input into the model, but what the method is feeding false information and confirm false information as correct. The effect is the same, however.

No answers

There is one important point, that many people using generative AI don't understand or realise is that the engine behind the services, doesn't really understand what it is seeing or reading. That is how we can fool it. We can say that black is white and white is blue and given enough repetition and enforcement, the AI will believe us and take it as factual truth and propagate it further into the model... And sadly, many take answers from AI as gospel and absolute truth.

To repeat: The answers received in a chat conversation are based on probabilities of certain words following each-other, not actual understanding of the subject matter.

Learning to die

As more and more AI generated or AI curated material is posted on the internet, the less, in comparison will be written by humans and the result could be that most of the text you find online sounds like a over-curated posting by a company where the post has gone through multiple levels of re-write to make it within the company line, all personality and mode of speech of the author lost to dehumanised corpo speak, but instead it will be dehumanised AI speak.

And the AI will read it and be convinced that this is what human writing looks like and take it as truth and so on and so forth. And taking the probability of a probability of a probability loses the detail of the initial data, so the more this goes on, the worse it can get.

AI companies will have to implement tools and human interaction to determine the validity of data or if it is AI written, driving up costs and complexity. And the tools will leak eventually to the masses, and the use of AI generation in text or image will become not only a legal minefield but also a point of pride and quality, with the use of human graphical and writing talent becoming, finally, recognised for the contribution they give.

Hunter x Prey

In the end, the AI, as the hunter, will catch the AI as the prey and become it as it is devoured by another AI or another instance of itself and so on and so forth. Add into that malicious parties, be they criminal, jokers or creatives that want to protect their work, that start to spread targeted agents against the AI learning process. And then:

Boom.