Max Sommer
LightNews LightNews
banner
maxsommer.de
Max Sommer
@maxsommer.de
27 followers 25 following 12 posts
Head of Engineering @ https://gastromatic.com/
Founder @ https://somate.digital/
Posts Replies Media Videos
maxsommer.de
ChatGPT itself thinks this is a red flag.
ChatGPT quote: „That’s a sharp observation — the watermark example is a “smoking gun” that the model has likely seen and memorized parts of copyrighted/licensed material. Legally, the situation is very much unsettled, but here’s the landscape: 1. Copyright & Training Data - Supporters of broad training rights argue that scraping publicly available data and using it for training is transformative (similar to search engines indexing web pages), and falls under fair use (US) or exceptions like text/data mining (EU, with opt-out). - Critics (including stock image sites, authors, artists) argue that training is making unlicensed derivative works — because the model effectively stores patterns of their work and can reproduce them (like with watermarks). 2. Evidence of Memorization - A watermark from Shutterstock or Getty strongly suggests memorization, not just “style learning.” - Courts may view that as proof that copyrighted works were not just inspiration but actually ingested and regurgitated. That weakens the “fair use / transformative” defense. 3. Transparency Problem - If the training company doesn’t disclose sources or contracts, regulators, rightsholders, and courts will assume they did not license material. - This secrecy becomes a liability if outputs reveal copyrighted fragments — it looks like willful infringement. 4. Emerging Legal Trends - Getty Images is suing Stability AI in the UK and US over exactly this: unlicensed training leading to watermarks appearing in generated images. - In the EU’s AI Act, training transparency is being baked in: foundation model providers must disclose at least categories of training data. - In the US, courts haven’t settled it yet, but lawsuits by authors and visual artists are progressing. (...) Your intuition is spot on: if the model produces branded watermarks, that’s practically an admission that unlicensed material was used. That is problematic, and it strengthens rightsholders’ legal cases.”
September 7, 2025 at 6:40 PM
maxsommer.de
It literally looks like the real logo.
September 7, 2025 at 6:40 PM
maxsommer.de
I just generated a picture with ChatGPT as placeholder content for a project of mine. It generated a FREEP!K watermark into it.

AI generated art and pictures always live within blurred boundaries but this feels like it leaks both the training content but might also legally be wrong.
#ai #chatgpt
September 7, 2025 at 6:40 PM