Artists and everyday users are quietly sabotaging AI models by feeding them corrupted images, and it's working — bad news for AI companies that built their business on free data.

• Tools like Nightshade and Glaze let people alter their images so AI scrapers learn the wrong thing — a poisoned "dog" concept can turn into a cat.
• It takes very little to break a model: fewer than 100 crafted images can corrupt a single concept, and the damage spreads to related ideas like puppies and wolves.
• This started as payback for years of scraping — companies like Stability AI, OpenAI, and the LAION dataset hoovered up billions of images without asking or paying creators, sparking lawsuits from artists and Getty Images.
• The same trick now protects voices: tools like SafeSpeech scramble the audio fingerprint scammers need to clone someone, after fake voices and video helped steal $25 million in one 2024 heist.
• AI labs are fighting back by trying to filter poisoned files, but their cleanup catches only about half, is slow and costly, and poisoners adapt within weeks.

Outlook: The cost of cleaning stolen data is rising, pushing AI firms toward paid, licensed datasets like the deals Adobe and Shutterstock already use.