Creative Control in the AI Copyright Age - The AI Optimist Podcast
Ask about getting content back out of AI, I keep hearing,
"AI is a black box. NO one understands."
And...you accept that. After all, I’m not an engineer. Nor are many creators whose work are now inside AI. Fair use in their eyes, unfair in creator’s eyes.
Stuck in this AI Black Box myth - and told I wouldn't ever understand. Until a listener shared a comment that made me triple take. Dang.
Now with the AI tools to detect and protect copyright, it's time to turn around and stop looking in the black box.
Wow, crack 1 – the AI black box hype may just be that, hype.
Some companies are finding their content inside an AI system with just a few tweaks - content tech companies say is "transformed beyond recognition" and impossible to track.
I've been repeating what the big AI companies told us - once your content turns into "tokens," it's like mixing paint.
You can't unmix it, you can't identify it, you can't remove it.
I bought it. All of it.
And why wouldn't I?
These companies kept saying the same thing - it's all math and statistics, it's all transformed, it's fair use. The black box defense.
Then reality hit me.
The New York Times, in their legal filing, shows how ChatGPT spit out their articles almost word-for-word, possibly by tweaking a "randomizer" setting.
It's like finding a knob on your blender that suddenly unblends your smoothie back into whole fruits.
I found out about Patronus AI and their CopyrightCatcher. They ran a test with ChatGPT and found that 44% of what it produced was copyable content.
Not "inspired by" or "in the style of" - actual copyable content that could land someone in legal hot water.
If AI isn't really a black box after all, all those company valuations built on the idea of "we transformed it, so we own it" come crashing down.
If your copyrighted content (or something close enough) can be found inside these systems, the whole black box story changes.
Now creators gain control and protect their work if they take a few steps.
No engineering degree required. Often for free.
Recreating Copyright Content: Does AI Remember?
The idea that AI as an impenetrable black box? The evidence is hard to ignore:
Patronus AI's Discovery: 44% of ChatGPT outputs were copyable content in their early test (2024)
The NY Times Prompt Hack: How adjusting a simple "randomizer" setting revealed much of the original articles.
The Visual Evidence: Researchers Gary Marcus and Reid Southen showed Midjourney could create images displaying copyrighted characters from Marvel, Star Wars, Sonic, The Simpsons, and other major franchises—proving the same issue exists across text AND visual AI
The Key Insight: Thomson Reuters, via its Westlaw legal business, won against a Fair Use Defense by Ross Intelligence, who legally purchased data from a 3rd party to train their AI.
Ross are finding it and removing it from their AI. Copyright fair use defeat.
Turn down the randomness, and out comes content that looks an awful lot like what went in.
This matters because it may mean your work isn't being "transformed beyond recognition" as we've been told—it's being stored in patterns that can be pulled back out. At least in these examples.
Patronus AI's approach worked because they focused on outputs rather than trying to peer inside the box.
Instead of trying to catalog all copyrighted material in training data (impossible), these tools focus on clear ownership of specific works and detecting when those works appear in AI outputs.
It's a bit like how Shazam works. Shazam doesn't need to understand music theory or know every song ever recorded.
It creates unique fingerprints of songs and matches them against what it hears. The same principle applies here - create a unique fingerprint of your creative work, then search for matches.
So if these systems remember and reproduce specific content, shouldn't creators have some say in how their work is used?
That's what the next generation of tools is helping to address.
The AI Optimist shares tools, insights, and approaches to AI.
Take advantage of AI, before it takes advantage of you.
