We’ve all seen it: you ask an AI to generate a scientific diagram, and there, buried in the pixels, is a faint, ghostly watermark.
Usually, we assume the AI is just mimicking “professional” vibes. But a recent experiment reveals something much more systemic. When the exact same prompt was given to two entirely different platforms—ChatGPT and Gemini—they produced the exact same image with the exact same watermark.
This isn’t just a coincidence. According to Jessica Nash, PhD, a software scientist and educator at Duke University’s Innovation Co-Lab, this phenomenon is a direct fingerprint of shared training data—specifically the LAION dataset.
The LAION Connection: A Shared DNA
As Dr. Nash points out, many of these models aren’t “seeing” the world independently. They are drinking from the same digital well. LAION (Large-scale Artificial Intelligence Open Network) is a massive, open-source dataset containing billions of image-text pairs scraped from the web.
If a specific watermarked image—like a 3D render of an aspirin molecule—is prominent in the LAION dataset, rival AI models will “memorize” that watermark as a fundamental part of the object’s identity. To the AI, an aspirin molecule isn’t just a chemical structure; it’s a cluster of red and white spheres that often comes with a faint logo in the bottom right corner.
Licensing vs. Scraping: The 2026 Accountability Crisis
Dr. Nash’s insights highlight a massive accountability question: Are AI companies doing their own training, or are they “laundering” data through third-party sets?
-
The “Pass-through” Problem: By using datasets like LAION, AI companies effectively outsource the “moral” part of data collection. If the dataset was scraped without consent, the AI company can claim they are just using a “research tool,” making it incredibly difficult to police.
-
The Monopoly of “Slop”: When every major AI uses the same underlying data, we lose diversity and accuracy. If the original scraped image was wrong, every AI in the world will now confidently repeat that same error.
The “Aspirin” Test
In your screenshots, the AI’s attempt to look “professional” actually reveals its scientific inaccuracy. Because it’s focused on replicating the look of a stock photo (including the watermark), it sacrifices the actual molecular geometry. It prioritizes the “aesthetic of truth” over the truth itself.
The Reality: When you see a watermark in an AI image, you aren’t looking at a creative choice. You are looking at proof of a massive, unauthorized web-scrape that has been baked into the very brain of the machine.
The Future: Transparency or Litigation?
As we move through 2026, the demand for “Clean Data” is growing. Research from experts like Dr. Nash helps pull back the curtain on these “black box” models. The goal is a future of Data Provenance: a clear trail showing that every image used to train a model was legally obtained and ethically sourced. Until then, we are all just looking at the “ghosts” of the original creators, caught in an infinite loop of AI-generated echoes.
