Data Provenance in AI? Sitting on a Legal Landmine

Alright, AI folks, let’s talk about something no one likes to admit but everyone needs to hear the data behind your models might be more sketchy than a dodgy used car deal. You’ve got your shiny new language models and all that, but how often have you actually checked where the data came from? Yeah, thought so.

An Audit That Should Make Us Sweat

Some folks did the hard work and dug into over 1,800 datasets, and what they found isn’t pretty. Turns out, 70% of them don’t even have proper licensing info. It’s like building a skyscraper with bricks that “fell off the back of a lorry.” Now, you might think, “Not my problem,” but it absolutely is when the lawsuits start piling up. Mislabelled licenses on platforms like Hugging Face, GitHub, and others mean your models could be swimming in a sea of copyright infringement without you even knowing. And these platforms, bless their hearts, get it wrong up to 66% of the time.

Commercial and Non-Commercial Data Are Heading for a Messy Divorce

Here’s the kicker: the best, most diverse data that could really push your models to the next level? Locked down under non-commercial or academic licenses. It’s like being at an all-you-can-eat buffet, but you’re only allowed to eat the soggy salad. Commercially open datasets are mostly boring, while the fun stuff creative texts, rare languages, juicy synthetic data is kept in a vault. If you're a startup or a small player, good luck breaking in.

“More than 70% of licenses for popular datasets on GitHub and Hugging Face are ‘unspecified’, leaving a substantial information gap that is difficult to navigate in terms of legal responsibility. Our rigorous re-annotation finds that 66% of analysed Hugging Face licences were in a different use category, often labelled as more permissive than the author’s original licence.”

The Legal Chaos Is Only Just Getting Started

You think the rules are clear? Think again. We’re seeing big names like Stability AI and OpenAI already duking it out in court over who owns what. Some folks are banking on the idea of "fair use" to cover their tracks, but that's a bit like trying to use a colander as a life raft—good luck. U.S. copyright law might give you a bit of wiggle room, but it’s mostly just lawyers' paradise. And don’t get me started on platform-specific rules. OpenAI says you can’t use their outputs to make something that competes with them. Bold, right?

How to Not End Up in Court Over Data Usage

If you're not in the mood for a lawsuit, there’s a tool that might save your hide—DPExplorer. Think of it like a detective for your datasets. It helps you track where data comes from, who owns it, and whether you can legally use it without waking up to a cease-and-desist letter. It doesn’t solve all your problems, but it’s better than crossing your fingers and hoping for the best.

The Bottom Line Is Simple, Wake Up and Smell the Legal Coffee

AI’s future isn’t just about making models smarter; it’s about getting smarter about how we handle data. If you’re not paying attention to where your data comes from, you’re building on a foundation that could crumble faster than you can say “copyright infringement.” Get your act together, and maybe, just maybe, we can keep the AI hype train rolling without getting derailed by a lawsuit.

Original study here - https://www.nature.com/articles/s42256-024-00878-8

Make sure you subscribe to "What's Up With AI" for fresh insights, captivating stories, and intriguing reads - https://lnkd.in/gbme5JMt

Connect with me on GitHub ?? - https://github.com/mgks

Stay tuned for the next issue of What's Up With AI. Until then, keep questioning, exploring, and enjoying the journey!