The synthetic safety net: why AI startups are turning to data they built themselves

The free-for-all where AI companies hoovered up the internet and sorted out the legal mess later is winding down, and the invoices are starting to arrive. Anthropic's $1.5 billion settlement set a price tag on pirated training data, and music publishers followed with claims running into the billions. Then, in July 2026, a Munich court handed down the first real courtroom defeat for an AI music generator: GEMA versus Suno, which found that training on and reproducing copyrighted songs was infringement rather than fair use. The parallel US cases against Suno and Udio are still working through the courts, but the message is already clear enough — training a model might survive as fair use, sourcing it illegally will not.

The bigger players have adjusted fast. Warner and Universal, instead of suing AI music companies, are now licensing them. OpenAI and Google pay for news archives and Reddit data rather than simply taking it. Regulators have their own enforcement muscle too: the FTC has previously ordered companies to destroy entire models built on unlawfully sourced data, and Anthropic's settlement required it to destroy its pirated libraries outright. That's the part that should worry founders most — once toxic data is baked into a model's weights, there's no clean way to undo it. For a young company, an order to delete the model isn't a setback. It's the end.

The mistake usually happens long before anyone trains anything. Product teams pour their energy into what the AI will do and treat where its data comes from as a detail to sort out later. Later, the path of least resistance wins — scraped web data, grey-area APIs, datasets nobody can trace. By the time the model exists, the legal risk is already baked into its foundation. And the question of provenance is no longer something founders can dodge quietly: investors now put it at the top of due diligence, the EU AI Act's transparency obligations for general-purpose models are already in force with full enforcement landing in August 2026, and California's training-data disclosure law took effect in January.

Synthetic data has been the answer waiting in plain sight. Gartner predicted back in 2021 that 60% of data used in AI and analytics development would be synthetic by 2024, up from just 1% three years earlier — and the firm now expects synthetic data to overtake real data in AI development by 2030. NVIDIA bought synthetic data specialist Gretel in a deal that exceeded the company's $320 million valuation. Microsoft trained its Phi-4 model on 400 billion synthetic tokens. The appeal for a startup is simple: generate data that behaves like the real thing statistically, without the copyright, privacy or consent baggage attached to it. A healthtech company can build millions of synthetic patient records that mirror real clinical patterns without touching an actual patient file; a fintech can simulate fraud behaviour without exporting a single real transaction.

The teams doing this properly treat it as an ongoing discipline, not a one-off fix. Every external dataset gets treated as suspect until its provenance checks out — if nobody can explain where it came from, it doesn't go into the pipeline. Where real data genuinely can't be avoided, they license it rather than gamble, and the Bartz settlement made that trade-off explicit: paying for data is expensive, but litigation can end the company. And because synthetic pipelines can leak — models sometimes memorise and spit back the very data they were trained on — the responsible teams test for that leakage continuously rather than assuming clean data stays clean.

The deeper point is less about avoiding lawsuits and more about what gets built on top of all this. Startups scraping the same internet as everyone else end up with nothing distinctive underneath their product. A synthetic data pipeline built for a specific domain, with a documented audit trail, is something a competitor can't simply copy and a claimant can't touch. Investors, regulators and enterprise buyers are all now asking the same question — where did this data come from — and the founders who can answer it aren't just protecting themselves. They're sitting on an asset the scrapers never had.