Who owns the data used to train AI?

Elon Musk says he owns it. Twitter's Ts & Cs suggest otherwise. James O'Malley investigates who really owns the data being used to train AI

-  James O'Malley

For decades, the fields of rocket science and brain surgery have been cited as fields of endeavour that present almost unimaginable levels of complexity. Now we might want to add another tricky job to the list: managing Twitter.

Since Elon Musk dropped $44 billion and took control of Twitter at the end of last year, it hasn't gone well. The CEO who, let's not forget, is heavily invested in both rocket and neural science - has seen the value of the social network plummet. One study found that more than half of Twitter's top 1,000 advertisers have given up on the platform since his takeover.

The stress is starting to show. When Microsoft announced that it would be pulling advertising from the platform, reportedly because it refused to pay hiked API-access fees, Musk responded with a tweeted threat: "They trained illegally using Twitter data. Lawsuit time."

His argument is that Al models such as the ones created by Microsoft and its partner OpenAI, the firm behind ChatGPT, were getting a free ride on Twitter's data. Large language models (LLMs) that power AI tools such as ChatGPT have been "trained" on text taken from across the internet. This could conceivably have included data from Twitter.

Now Musk wants his pound of flesh. But who really owns data once it's out on the internet? Does Musk have any right to lay claim to it? The answer, you'll be shocked to hear, is complicated.

Scrapes of wrath

"There are so many variables that help to answer whether a specific scraping act is legal or illegal," said Denas Grybauskas, head of legal at web intelligence collection firm Oxylabs.

His company specialises in writing scrapers - software and tools that automate the work of downloading the contents of a website or individual web page, then extracting and organising the data. It's the equivalent of saving a web page on your computer, but automated and performed at mass scale.