Overheard an engineer at a coffee shop talking about AI training data collapsing

I was sitting near a table at a Starbucks in Denver last week and this guy was explaining to his friend how AI models are starting to train on content other AIs created. He called it "model collapse" and said the quality keeps getting worse with each generation. That got me thinking about how much of what I see from these tools now feels recycled and samey. I work in marketing and use AI for drafting social posts, but I wonder if I'm just getting bad output because the training data is polluted. Has anyone else noticed their results getting dumber lately or know of any tools that filter out AI-generated content from their training sets?

2 comments

2 Comments

daniel55221d ago

Used to think the whole "AI eating itself" thing was just tech bros being dramatic, but now I'm not so sure. Tried using ChatGPT for some SEO descriptions last week and kept getting these weird circular sentences that sounded like they were written by a bot that had only read other bot writing. Looked into it a bit and apparently there's already research showing models slump hard after like 5 rounds of training on synthetic data. Kicker is we're probably all contributing to the problem by posting AI slop everywhere without realizing it lol.

the_uma20d ago

Man, that part about "written by a bot that had only read other bot writing" hit Different for me. I used to roll my eyes at the doom and gloom stuff too, thought people were just scared of progress. But last month I tried to get a recipe from one of those AI-powered sites and it read like a robot trying to explain how to boil water after only reading other robots explaining it. Just this weird, hollow loop of words that went nowhere. Now I'm actually paranoid every time I see a paragraph online that sounds too clean or perfect. We're definitely filling the internet up with dead air and calling it smart.