In March 2025, legal teams across the US publishing industry did a double take at a newly unsealed court filing. It was already known that Anthropic had settled its copyright lawsuit with authors and publishers. But this new filing wasn't about the settlement amount itself — it revealed that authors, publishers, and literary agencies were now locked in a three-way fight over who gets how much of that money. The payout an AI company had agreed to hand over had become a fresh battleground within the very camp receiving it. It marked a shift: copyright litigation was moving from an external fight against AI companies to an internal dispute over how content creators divide the spoils.

Around the same time, a different alarm was sounding in the financial markets. A technical paper accompanying LG AI Research's EXAONE Finance release laid out data on why general-purpose models can be risky when deployed as AI agents for investment decisions. The numbers traced a path in which better performance actually makes the market as a whole more dangerous.

A New Fight Lands on the Settlement Table

Until now, AI copyright litigation has followed a fairly simple script: AI companies stand accused of feeding copyrighted work into their training data without permission, and content owners sue for damages. The lawsuits The New York Times, the Chicago Tribune, and The Atlantic filed against OpenAI and Microsoft are the textbook example.

Anthropic's case adds a new layer to that script. In late 2024, Anthropic reached a confidential settlement in the copyright suit brought by US author groups and publishers. The trouble started after that. Publishers, author guilds, and literary agencies each showed up with a different claim on who was entitled to the settlement money. Publishers argue their contracts give them the underlying rights to the works; author groups counter that the writers themselves are the creators; agencies maintain that their contract structure entitles them to a commission.

As of March 2025, the dispute remains in court-supervised mediation. The total settlement amount hasn't been disclosed, but the trade publication Publishers Weekly noted that "the structure of this dispute could become the template for AI copyright settlements going forward." Around the same time, on the news-media side, Ziff Davis filed a new lawsuit against OpenAI, explicitly citing indirect traffic diversion through article-summarization features as grounds for the suit. It's a case where a dispute that began over training-data use has stretched all the way into how AI-generated output gets monetized.

LG AI Research published a technical report alongside its EXAONE Finance model in March 2025. According to the report, a model specially trained on financial time-series data behaved very differently from general-purpose language models in market-shock simulations. When general-purpose models were deployed across many independent agents, their trading decisions converged in a statistically significant way — a byproduct of their shared training background.

As the Performance Curve Rises, So Does the Herding

"A better AI produces better results" mostly holds true at the level of an individual user. But the picture changes once the same model — or models built on the same architecture — is used simultaneously by hundreds or thousands of investors.

That's exactly what the EXAONE Finance report flags. As more AI systems that share the same training data and architecture populate the market, their behavior converges even when each agent appears to be reasoning independently. Diversification doesn't just become less effective — it structurally degenerates into correlated trading. The report confirms a path in which the better individual agents get, the stronger the collective herding becomes.

This phenomenon is known as systemic risk. Its structure resembles how algorithmic trading amplified the 2010 Flash Crash and the March 2020 COVID market shock. But where those earlier episodes involved a handful of institutions running systems built on identical rule sets, the AI-agent era is harder to track: countless agents, run by different companies for different purposes, now share similar training backgrounds.

That context explains why LG AI Research chose to build a finance-specific model instead of relying on a general-purpose one. General-purpose large language models are plenty useful for understanding financial terminology or summarizing reports. But when the judgment call requires finance-specific patterns — abnormal trading volume in a given stock, interest-rate cycles, sector-level return time series — the reasoning demands a model built on fundamentally different training data.

A similar pattern shows up on the copyright side. AI companies have fed more, and higher-quality, data into training in pursuit of better performance. The logic of performance improvement has driven the logic of data collection, and friction with content creators has accumulated along the way. The fiercer the performance race gets, the wider the surface area for copyright conflict becomes.

A similar tension runs through how people design their careers and businesses. Moving in the direction of growth and moving safely don't automatically point the same way. What the AI ecosystem is showing right now, simultaneously on the legal and financial fronts, is that a technology's performance curve and its systemic-stability curve can move in opposite directions.

What Solo Entrepreneurs and Planners Should Check Right Now

This moment leaves Korean solo entrepreneurs, content directors, and solo PMs who use AI in their day-to-day work with three practical things to review.

It's time to reread the contract clauses governing your content sources. What the Anthropic settlement dispute exposed is that there's still no established standard for how money paid by AI companies should be divided within the content ecosystem. South Korea, too, could see AI training-data copyright guidelines take concrete shape sometime in 2025. Reviewing the contract structure behind the content you provide or use — especially the licensing clauses and the scope of secondary use — now can help you avoid unnecessary disputes later. If you work with publishers, media outlets, or agencies, it's worth checking whether your contracts even include a clause addressing how AI-related revenue gets attributed.

Get in the habit of factoring model diversity into your choice of AI tools. The EXAONE Finance case is about financial markets, but the same principle applies to marketing, content planning, customer analysis, and other fields. If every competitor is using the same AI tool, the same prompt patterns, and the same data sources, the room for differentiated output shrinks. Even when using the same model, it becomes important to build your own context — through the data you feed it or how you design your prompts. It's also worth considering running a domain-specific small model or a fine-tuned version alongside a general-purpose one.

Change how you document your copyright-related work. Few people currently keep records of the source and usage method when they include AI-generated text, images, or code in commercial deliverables. But as US court rulings begin to address the copyright status of AI-generated output in concrete terms, Korean companies have picked up the pace of reviewing their own contract clauses since 2025. Deciding now how you'll flag AI use in client deliverables, and how you'll record which tools and versions you used, puts you ahead of potential disputes later.

Slow down before connecting financial AI tools directly to your investment decisions. The systemic herding risk the EXAONE Finance report warns about applies to individual investors too. The more often you act directly on a stock pick or trade timing an AI recommends, the more your actions are likely to overlap with those of other investors using similar models. When the market wobbles and you feel the urge to sell, it means countless other people using the same AI are likely receiving the same signal at the same moment.

Choosing an AI tool is partly a matter of work efficiency, but it has also become a decision about what kind of risk structure you're stepping into. Alongside "how much faster does this tool make my work," you now need to ask "how many people in a position similar to mine are using this tool in an essentially identical way."

While Anthropic's settlement money sits undivided on the mediation table between three parties, the data piling up elsewhere shows that the more accurate AI gets, the more synchronized markets become. Closing the gap between how fast performance improves and how fast we understand the structures that use that performance — that's the homework left for anyone putting AI tools to work today.