Unfair use
Parrot food isn't free ... Fund Fontes' deepfakes ... And an Elon for your pocket.
You know how AI companies have been grabbing data left and right to train their systems, often without bothering to ask for permission?
Well, that Wild West vibe just slammed into a legal wall.
After years of back-and-forth in court, Thomson Reuters scored a huge win in a landmark AI copyright case against a startup that had been using Thomson Reuters' data to train its AI.
This is a pretty big deal.
It sets a clear example that creators — think writers, artists, and anyone making content — have rights that need to be respected. It's also the first step in putting the brakes on AI companies using data without the proper licenses.
Basically, it's a wake-up call: AI companies can't just take whatever they want anymore. This case is laying the groundwork for making sure data is used the right way.
About a decade ago, ROSS Intelligence thought they could shake things up in the legal research world. They launched an AI-powered tool to go head-to-head with the big players, Westlaw and LexisNexis, aiming to help lawyers find legal info faster and easier.
But in May 2020, they ran into some trouble. Thomson Reuters, the folks who own Westlaw, sued them, claiming ROSS had taken parts of Westlaw’s content — like their headnotes, which are little summaries of key legal points from court decisions — and used them to train their AI without asking.
ROSS fired back that they got the data under the “fair use” doctrine — a rule that sometimes lets you use copyrighted stuff without permission, if it’s done in a certain way.
How does a Large Language Model work?
Imagine teaching a parrot to say “hello.” You just repeat “hello” until it squawks back, not understanding, just mimicking.
AI is like a turbocharged parrot.
You feed it millions of examples — posts, pics, or legal docs — in a process called training. Eventually, it spots patterns, like “cat” near “purr” or dog shapes in images. Soon, it’s churning out sentences or tagging pics, guessing from what it’s seen.
But training it requires a ton of data – millions or billions of examples. Without that giant pile of examples, the AI wouldn’t have enough to work with -- it’d be like asking a chef to cook a gourmet meal after only flipping through a pamphlet.
When we talk about training AI, it’s not just about having data; it’s about having enough of it to make the magic happen.
But lawsuits cost a ton of money, and ROSS just couldn’t keep up. They had to shut down their business in December 2020, though they didn’t throw in the towel on the case just yet.
Then, in 2023, the Supreme Court stepped in with a ruling on a different case, Andy Warhol Foundation v. Goldsmith, that made it tougher to call something fair use.
That was bad news for ROSS. Fast forward to February 11, 2025, and Judge Stephanos Bibas made the final call: ROSS had indeed broken copyright law by using Westlaw’s content.
This wasn’t just a loss for ROSS — it set a precedent for other AI companies out there. The message? You can’t just grab copyrighted material to train your AI without getting the green light first. It’s a ruling that could change how AI folks get their hands on the data they need moving forward.
Thomson Reuters’ winning arguments
Westlaw’s headnotes, summaries, and Key Number classification system involve creative editorial judgment, making them copyrightable.
ROSS wasn’t just training AI; it was developing a direct competitor to Westlaw using Westlaw’s own work.
ROSS’s AI did not generate new creative content — it simply replicated Westlaw’s research function in a different format.
ROSS’s defense (that didn’t work)
ROSS claimed headnotes merely summarized public domain law. The court ruled that summarizing legal rulings requires editorial choices, making them protected expression.
ROSS insisted that using headnotes as AI training data was transformative. The judge wasn’t buying it, ruling that a tool built to compete directly with Westlaw was not sufficiently transformative.
ROSS never displayed the headnotes verbatim in its product, but that didn’t absolve it. Training AI on copyrighted material still constitutes copying.
Why this ruling matters for AI and copyright (and you!)
This decision has major implications for AI companies, content creators, and anyone training machine learning models. The judge noted that ROSS’s AI wasn’t creating anything new, distinguishing it from generative AI like ChatGPT.1
Global ripple effects:
Countries like Japan and Singapore have explicit AI training exceptions in their copyright laws, while the European Union allows opt-outs. The U.S. still relies on unpredictable fair use battles.
Will this case go to appeal?
ROSS (or its remaining legal team) may appeal the decision to the Third Circuit. If the ruling stands, expect it to be cited in future AI copyright cases involving books, images, music, and other copyrighted data. Meanwhile, AI companies are likely reassessing their data pipelines — because the era of free-for-all AI training just hit a legal roadblock.
What does all this mean for regular consumers?
For everyday users, this could mean that the AI tools we rely on — like chatbots, image generators, or research assistants — might come with a higher price tag. That’s not a definite outcome. Some of these corporate giants will dodge these expenses by tapping into alternative data sources or striking smart licensing deals. But if companies have to pay more to use data legally, they’re likely to pass those costs on to us.
The Thomson Reuters case wasn’t the only lawsuit out there about AI. Far from it. Similar ongoing lawsuits touch on various industries, with each one raising complex questions about fair use and intellectual property rights.
The New York Times vs. OpenAI and Microsoft: The New York Times alleges OpenAI and Microsoft used its articles to train AI models, potentially impacting subscription revenue (The New York Times vs. OpenAI and Microsoft).
Artists vs. Stability AI, Midjourney, and DeviantArt: Artists claim these companies used their artworks to train AI image generators without consent, with some claims still active (Artists vs. Stability AI, Midjourney, and DeviantArt).
Getty Images vs. Stability AI: Getty accuses Stability AI of using over 12 million photos to train its AI, a case ongoing in UK courts (Getty Images vs. Stability AI).
Authors vs. OpenAI and Meta: Authors like Sarah Silverman and Michael Chabon sue for using their books to train AI, with cases still pending after partial rulings (Authors vs. OpenAI and Meta).
Music Industry vs. Suno and Udio: Major labels sue AI music generators for training on copyrighted music without permission, with recent filings showing ongoing legal battles (Music Industry vs. Suno and Udio).
Authors vs. Nvidia: Authors accuse Nvidia of using their books to train its NeMo AI platform, a newly filed case still in progress (Authors vs. Nvidia).
Artists vs. Google: Visual artists sue Google for training its Imagen AI on their works without authorization, adding to the legal fray (Artists vs. Google).
Nearly eight months after Arizona Secretary of State Adrian Fontes first announced his Artificial Intelligence and Election Security Advisory Committee, the committee released its final report on the threats and advantages that AI can bring to the election process.
The committee included heavy hitters from companies like OpenAI and Microsoft, and from think tanks and universities. But their biggest findings weren’t exactly revelatory: Elections administrators need more money.
The report offers an overview of the benefits and challenges of AI in elections and hints at where it could be deployed: in data storage solutions, routine administrative tasks and educating the public.
It also notes where AI has fallen short – many chatbots give out inaccurate information about elections, the report states. And there’s the constant threat of AI-fueled cyberattacks, disinformation and deepfakes.
The experts suggested setting up Arizona states and “private donors” to create “AI Elections Labs,” housed at a university, to “help test or develop concepts and products for election officials to pilot, while meeting standards for use, testing, and auditing of these products.”
The latest leader in AI comes from Elon’s xAI — Grok 3.
It's been out for a week now and I’ve been enjoying talking to it — in many cases more than ChatGPT, Gemini or Perplexity.
Grok has a very unique ‘data set’ from all the interactions on X.com (formerly known as Twitter) which allows the AI to have interesting takes and real-time peer-to-peer news.
Basically, Grok gained its personality from Twitter, which makes it by far the most entertaining Large Language Model to talk to. On its phone app, you can even choose between various “AI Personas” who each have (or don’t have) a filter.
PSA: Don’t talk to the “unhinged” bot in public or at work or near kids!
Future cases may decide whether AI models that generate unique content have stronger fair use claims.