I was recently asked to speak on balancing AI and Copyright for creators at the Understanding Trust and Safety in AI seminar by SFLC.IN. It was then that I had to sit down and clarify my messy thoughts on AI for creators, if it is copying, and what copyright even is in the age of AI.

Following is me working through where that line is. These are my raw personal thoughts and not meant to be spoken for others. I’m not a policy expert or a lawyer. Just a creator trying to figure out what feels right and what doesn’t to him personally.

When AI Learning vs Copying Became Real

I have been using AI, not really bothered about imitation. But then, recently, I saw the viral deepfake video of the actress Rashmika Mandanna. It wasn’t that deepfakes and AI-generated fake videos were new. We’ve all seen them at some point by now. It was then that I finally started seeing the difference between an AI learning my style and an AI cloning my identity.

You see, I’d been comfortable with AI for months by this point. I use these as tools daily. According to ChatGPT’s year-end recap, I’m literally in the top 1% of users by messages sent. I had concluded that AI training on my writing was fine.

But watching someone’s face get stolen like that? That didn’t feel right.

The ‘Powered Screwdriver’ – AI for Creators Explained

I’ve been a paid ChatGPT and Perplexity user for about a year and a half. I experiment constantly. Some of it is serious writing stuff. Some of them are ridiculous side projects like helping me swipe through dating app profiles.

From all that, my view on AI for creators is that of a powered screwdriver. It doesn’t build the furniture for you, but it lets you tighten screws faster so you can focus on what you care about. Perhaps the design. As a writer, it’s like having a powered pen attached to a master’s-level research assistant.

Recently, I saw a video of a fellow author who needed a pink jewel with a specific mythical history for a story scene she was working on. If you’re doing this manually on Google, you’re looking at maybe 15 minutes of searching and cross-referencing. She asked Claude and got three solid options in two minutes.

Or yesterday. I needed Kerala Police Housing specs for a scene. This is boring information, like room sizes, pay grade-wise assignment, and so on. I could have spent an hour going through architectural diagrams and boring Government policies. Instead, I asked Perplexity Pro to research and got what I needed to write the scene in minutes.

You still need skill, though. You need to know how to talk to these models, what words unlock useful outputs, and you learn by doing. And the models keep changing, so nothing stays fixed. But the acceleration is real.

The Ethics of Training AI for Creators

In preparing for the roundtable, I read a paper (more appropriately, I read an AI summary of that paper), “A Scanner Darkly: Copyright Liability and Exceptions in Artificial Intelligence Inputs and Outputs” by Andres Guadamuz, published in GRUR International. It gave me a new language on how to think about AI training.

The argument was simple. Models don’t store the original files. They store patterns, like vectors, mathematical representations of patterns.

A simple analogy I liked is asking a human to draw a dog. You don’t retrieve a stored photograph from your brain. You have a pattern in your mind. Fur, four legs, a tail. And you create something new from that understanding. That’s what these models do. That is what the model is “trained”.

Take, for example, how we learn art. Art students imitate Picasso to understand brushwork. Musicians learn ragas composed by masters who died centuries ago. We’ve always learned by absorbing what came before us and building something new from those patterns.

Sometimes models do spit out verbatim content. Like AI images with the Getty watermark embedded. So, verbatim reproduction is possible, but less likely. At least that is what I understand.

That’s why I don’t mind AI training on my writing. It’s learning patterns. As long as it does not photocopy my words.

The Gap I Can’t Cross

But here’s where my comfort ends. Right now, as you read this, this itself was written by an AI that transcribed my English voice locally on my machine. But I can’t do the same thing for Malayalam, my mother tongue. There isn’t enough training data, or local language models aren’t as easily accessible, or both.

I can’t transcribe my own thoughts in my own mother tongue. I can’t efficiently translate regional books. And I’m not alone. This is a huge gap in regionally trained AI for creators in India. There is a wall that English speakers don’t face.

This is where the licensing debate stops being theoretical. If Indian companies have to pay massive fees to access training data, be it regional or otherwise, will they ever build Malayalam tools? Or will we stay locked out while English speakers get more powerful models every quarter?

Challenge is part of growth. But what if the Government’s idea, Working Paper on Generative AI and Copyright, if implemented as it is, accidentally makes the challenge insurmountable?

Where I Draw the Line on AI for Creators

Back to that deepfake. And the Anil Kapoor case. And the government proposal about mandatory training on all our data, with some vague “compensation” attached.

I have a YouTube channel. My face, my voice. They’re public. According to the new policy, at least as I understand it, if implemented, AI can be trained on all online resources. What if a model is trained on those videos, and a deepfake video of mine is made where I say things I would not?

The line then became clear. For me, at least. Imitating writing style and imitating identity are not the same thing, for me at least, right now.

If someone trains an AI on my writing and asks it to produce something “in the style of Ashik Satheesh,” I’m okay with that. Of course, as long as readers know it’s AI-generated. Style isn’t something you can lock down. A “Van Gogh painting of a cat” isn’t copyright infringement. Style is learned, imitated, remixed. It’s been that way for centuries.

But Mickey Mouse? That’s an identity, and reproducing it is theft.

Similarly, my face and my voice are my identity. At least what people identify legally as me. My voice saying things I never said isn’t “style learning”. It’s identity cloning. And what if the compensation for that, according to the proposed Working Paper, is five rupees? It also gives it total legal legitimacy, if my understanding is correct, that is.

So here’s my line. 

Learn from patterns and style, yes. 

Clone identity or copy verbatim, no. 

Live and let live.

What I Still Don’t Know

I wish I could wrap this up with a neat policy proposal, but I can’t. That is not my area of expertise.

I don’t know if it’s technically feasible to calculate fair compensation for individual contributions. Say my face, voice, my indie book, my health data, anyone’s creative work. At least in a meaningful way. I don’t know how opt-in and opt-out would work across text, style, face, and voice without becoming a legal mess.

Maybe this takes decades of court battles to sort out. Maybe we never get it perfect. I don’t know.

But I know what my line is now. Let me use the powered screwdriver. Make it easy for people to make those powered screwdrivers. Let it learn the style, patterns. But don’t steal identity.

What about you? Where do you draw yours?