I recently posted an article laying out the speed and breadth of the AI Tsunami. In it, I quoted from an article from Business Insider India reporting that 57% of all internet content is now AI generated. They further predicted that that number will rise to 90% in this coming year. Obviously, AI generated content is going to displace human generated content and disrupt the livelihoods of a lot of Liberal Arts majors. That’s the obvious impact. But another, long-term impact scares me much more. And I think I’ve found a good metaphor to explain my anxiety.
In the 1980’s cattle farmers in England realized that they could help their cows to put on more weight, more quickly, by supplementing their diet with…cows. Yup, they turned Bessie into a cannibal so that they could boost their revenue and profits. And it worked great. Until it didn’t. You see, it turns out that cow meat is expensive. So these farmers started to branch out from cow meat to include more exotic cow-based ingredients, including ground up bones, parts of the digestive systems, and tragically, parts of the nervous system and brains.
Alas, there are always unintended consequences. In this case, cow brains can contain a protein molecule called a ‘Prion’. Prions are kind of like a proto virus. In the right environment, they can replicate themselves. Unfortunately, the ‘right environment’ in this case is inside the brains of other cows. So when farmers started feeding cow brains to their cow herds, the prions went to work replicating themselves. They turned the brains of some of those cannibal cows into hole-filled, swiss cheese-like mush. They called all of this Mad Cow Disease (officially Bovine Spongiform Encephalopathy).
When these cows trembled, stumbled and collapsed from this affliction, the farmers quickly sold them to the slaughter houses to avoid financial losses. (In hi-tech parlance, they were ‘monetizing’.) Those cow brains were then fed to new cows and so on and so on. Oh, there is one teensy little additional detail. It turns out that some human beings that ate those cows also got infected by prions (officially Variant Creutzfeld-Jakob disease for humans). People died. The public panicked. The British beef industry was devastated. It was a predictable tragedy brought on by short sightedness and greed.
Now back to the AI world. As I had explored previously, AI is trained with content. And one of the worrisome side effects that researchers are discovering is that AI models that are trained using Generative AI data tend to exhibit strange, unpredictable and unwanted side effects. Up until about 18 months ago, this wan’t really an issue because all of the content in the world was human generated. But now, as I called out above, almost everything in the environment is AI generated. In other words, the entire food chain available to AI models now has the AI equivalent of cow brains in it.
I’m lazy, so I asked ChatGPT to summarize how an AI ‘Mad Cow Disease’ might manifest itself:
1. Feedback Loops: Training models on AI-generated data can degrade quality over time, causing repetitive, less original content and amplifying existing errors.
Source: ChatGTP
2. Bias Amplification: Reusing AI-generated content can increase biases, making models less fair and potentially introducing new biases.
3. Loss of Grounding in Reality: Models may become disconnected from factual reality, leading to more hallucinations and less reliable information.
4. Legal and Ethical Risks: Using AI-generated content raises copyright concerns and complicates content authenticity, which can affect trust.
5. Increased Complexity: Filtering out AI content to maintain quality adds to computational costs and training challenges.
Sounds like a lot of fun. Perhaps you are thinking, “if ChatGPT can clearly articulate the dangers and consequences, surely the AI companies are aware and won’t feed their LLMs brains?” Perhaps. But I have two concerns. One is logical and the other cynical.
The logic concern is this: AI platforms are not currently watermarking their content, and there is no foolproof way to evaluate content after the fact to determine its origin. I have run some of my writings through these AI Checkers and they sometimes give ratings that suggest that 50% of my original work is AI generated (I try to not take that personally). I have also tested AI generated content with these checkers and they often fail to recognize its origins. So, as generative AI keeps getting better, I suspect it will become almost impossible to identify generated content simply be evaluating it after the fact.
My cynical concern is that many AI firms, under financial pressure from their investors to achieve blowout profits, will give in to the temptation to take shortcuts. Most of the Mad Cow symptoms I mention above will be subtle at first; probably easy to ignore. So just like those farmers, most companies won’t take the problem seriously. But over time, those little things will add up to a big problem. Remember, virtually all of the content we consume (and use to make economic, political, medical and interpersonal decisions) will soon be AI generated. So all of humanity will be exposed to the AI Mad Cow disease.
Treating this like a Public Health threat, the AI community should be required to identify all generative content so that we can track it and it can be quarantined from the training process. I suspect that this will need to be done at a very low level. Perhaps the answer is to indelibly encode provenance into the ASCII data of textual content and watermarked it into generative imagery? However we do it, we need to move quickly because the cows are already out of the barn, so to speak. For those of us that consume content, the best approach is to pursue verifiably human-generated content from validated sources and be aware that for everything else, we are likely ‘eating’ content that may have prions in it.
End note: both of these articles are extensions on this one that I wrote 18 months ago. I predicted there that human generated content would become increasingly valuable, like the pre-atomic steel in sunken WWII battleships. I was disappointed to learn that the concepts in my article appeared to have been lifted, whole cloth by this article in Scientific American. Apparently it isn’t just AI that repurposes content without attribution.
Leave a Reply