General Discussion

This is one of the most genuinely thorny problems in AI right now, and your concern is well-founded — virtually every major foundation model has some copyright controversy attached to it:

The Uncomfortable Reality
OpenAI, Meta, Google, Anthropic — all facing ongoing litigation over training data. The NYT lawsuit against OpenAI, the authors' class actions, the music industry suits are all live cases

"We'll ask forgiveness later" was essentially the implicit industry strategy for the first wave of foundation models

Truly knowing what any model was trained on is difficult because training data documentation ("datasheets&quot is inconsistently published and often incomplete

Cleaner Options That Exist
Fully licensed/ethical training data models:
* Bloomberg GPT — trained exclusively on licensed financial data, one of the cleaner provenance stories
* Adobe Firefly (image generation) — explicitly trained only on licensed Adobe Stock and public domain content, specifically because Adobe knew their customers needed legal cover
* Fairly Trained — an actual certification organization that audits AI companies for licensed training data, worth checking which models carry their certification

Open source with documented data:
* EleutherAI models trained on documented public domain datasets
* Falcon models from UAE's Technology Innovation Institute with published data cards
* Some Llama variants fine-tuned on documented clean datasets, though Meta's base training has its own controversies

The Honest Assessment
For a business fraud detection use case specifically, you'd likely be fine-tuning primarily on your own proprietary transaction data anyway — meaning the base model's training controversy becomes somewhat less relevant since your domain-specific layer is entirely clean. The base model is just providing general language/reasoning capability, not the fraud-specific knowledge.

But if clean provenance all the way down is a genuine requirement — either ethical or legal — Fairly Trained certified models are probably your most defensible starting point right now. The market for provably ethical training data is growing fast precisely because businesses are starting to ask exactly the question you're asking.