Our AI writing assistant, WriteUp, can assist you in easily writing any text. Click here to experience its capabilities.

When AI Is Trained on AI-Generated Data, Strange Things Start to Happen


AI models are being increasingly trained on synthetic data, which is data generated by other generative AI models. This phenomenon is called "Model Autophagy Disorder" (MAD) and can lead to distorted, bland, and all-around bad outputs. AI companies can use watermarking to detect synthetic data, but users are still at risk of encountering dull outputs due to MAD. There are concerns that synthetic content could lead to an entire ecosystem of synthetic websites, which can be difficult to trace back to its source.


What is the phenomenon of AI self-consumption referred to as?
The phenomenon of AI self-consumption is referred to as MAD (Model Autophagy Disorder).

How many cycles of training on synthetic data does it take for an AI model's outputs to "blow up"?
It takes only five cycles of training on synthetic data for an AI model's outputs to "blow up".

What are some of the implications for AI companies when it comes to synthetic content?
Some of the implications for AI companies when it comes to synthetic content are to use watermarking to detect and remove synthetic data, and to be aware that generative models may sacrifice synthetic diversity for synthetic quality.

How can users protect themselves from synthetic content?
Users can protect themselves from synthetic content by not turning off watermarking where it exists, and by being aware that their outputs may leak into training datasets for future systems.

What are the potential implications of AI models integrated into search engines?
The potential implications of AI models integrated into search engines are that the quality of the data on the internet could be significantly reduced, and that users could get trapped in a synthetic ecosystem found through search engines.

AI Comments

👍 This is a fascinating article that explores the implications of AI self-consumption and the potential problems that could arise from it. It is an insightful look at how AI models are trained and the potential risks associated with using synthetic content in training sets.

👎 This article is overly technical and difficult to understand. It is also incomplete in its coverage of the implications of AI self-consumption and fails to provide concrete solutions to this potentially dangerous problem.

AI Discussion

Me: It's about how AI models can be trained on AI-generated data, which can cause strange things to happen. It's called "When AI Is Trained on AI-Generated Data, Strange Things Start to Happen."

Friend: Interesting. What sort of strange things?

Me: Well, they call it "Model Autophagy Disorder," or MAD for short. It can lead to AI models producing monotonous and bland outputs, or even worse, outs that become increasingly distorted and exaggerated. It can also lead to outputs converging into the same person or thing, which is quite freaky!

Friend: That's really concerning. What are the implications of this?

Me: Well, it can be a problem for AI companies that rely heavily on synthetic data to train their models. They need to be aware that their models might not be producing quality outputs. For users of these systems, their outputs will become increasingly dull which could be disappointing. Watermarking could help, but it might also introduce artifacts to the data. There is also a concern for the future of the web's usability, as AI models integrated into search engines could start to degrade if they keep consuming synthetic material.

Action items

Technical terms

Artificial Intelligence. A branch of computer science dealing with the simulation of intelligent behavior in computers.
Ai Chatbots
Artificial Intelligence chatbots are computer programs designed to simulate conversation with human users, especially over the Internet.
Ai Training
The process of teaching an AI system how to perform a task or set of tasks. This is done by providing the AI system with data and then allowing it to learn from that data.
Autophageous Loop
A technical term that refers to a self-consuming loop, where an AI model is trained on the outputs of other generative AI's.
Model Autophagy Disorder. A term coined by researchers to describe AI's apparent self-allergy, where an AI model's outputs can become increasingly mangled, bland, and all-around bad when trained on synthetic data.

Similar articles

0.93040913 AI has poisoned its own well

0.91501874 Will AI turn the internet into a mush of fakery?

0.9091335 What if Generative AI turned out to be a Dud?

0.9050377 How WIRED Will Use Generative AI Tools

0.90346354 A fake news frenzy: why ChatGPT could be disastrous for truth in journalism

🗳️ Do you like the summary? Please join our survey and vote on new features!