Meta is using your Instagram and Facebook photos to train its AI models

🌈 Abstract

The article discusses how Meta (Facebook/Instagram) uses publicly available photos and text from its platforms to train its text-to-image AI generator model called Emu. It highlights the company's stance on not using private user data for training, and the broader issue of AI models being trained on copyrighted content scraped from the internet.

🙋 Q&A

[01] Meta's Use of Public Data for AI Training

1. What type of data does Meta use to train its text-to-image AI model?

• Meta uses publicly available photos and text from its platforms Instagram and Facebook to train its text-to-image generator model called Emu.
• The company's chief product officer, Chris Cox, stated that they "don't train on private stuff" and only use "things that are public".

2. Why does Meta's use of public Instagram and Facebook data give it an advantage in training its text-to-image AI?

• According to Cox, Instagram has many photos of "art, fashion, culture and also just images of people" which allows Meta's text-to-image model to produce "really amazing quality images".

3. How do users interact with Meta's text-to-image AI model?

• Users can create images on Meta's AI by typing a prompt starting with the word "imagine", and it will generate four images.

[02] Broader Issues with AI Training Data

1. What is the key challenge around AI models being trained on data from the internet?

• There is almost no way to prevent copyrighted content from being scraped from the internet and used to train large language models (LLMs).

2. How are companies trying to obtain data to train their AI models?

• Some companies are partnering with media outlets to license their content, like OpenAI has done.
• Meta even considered acquiring the publisher Simon & Schuster to get more data to train its models.

3. What other techniques are companies using to train their AI models?

• Companies are using "feedback loops" - data collected from past interactions and outputs that are analyzed to improve future performance.
• This includes algorithms that inform AI models when there's an error so it can learn from it.