Model collapse might break the internet

What happens when AI models cannibalize their own data? You get "model collapse."

Not actually a peacock (left), despite what your AI search tells you.

Have you run an internet search lately? Chatted with a customer service ‘bot? Asked a search engine to find you a picture of something? You’re using an AI, whether you know it or not — which begs the question, can you trust the answer?

There’s a really interesting ShortWave podcast about what happens when AI cannibalizes its own data.¹

Most AI you’ll run into today falls into the large language model (LLM) category. When you ask Claude or ChatGPT a question, you’re using an LLM. You’re also using an LLM when you ask, for instance, “what does a baby peacock look like?”

How do you know if the answer you get is correct? Or, on the other hand, if it’s been totally made up by the AI?

That’s a problem AI engineers have been struggling with — and the problem could become even more challenging to solve.

AI models are trained on huge data sets. Think about it this way: a brand new LLM needs data — so, where’s the biggest source of …