Deepfakes and the AI Ethics Gap in Indian Classrooms

Two students at a school in Guwahati allegedly spent part of their week in June 2026 making deepfake images of their classmates and teachers, then circulating them on Telegram. The Assam Police Crime Branch opened an investigation this month. Read that again. Not "researchers warn this could happen." It already happened, in a real school, days ago, with kids who probably have a coding class on their timetable too.

This is the part of the "AI in education" conversation nobody puts in the syllabus. Everyone talks about generative AI in classrooms as a cheating problem: ChatGPT writing essays, an app solving the calculus homework. The actual urgent ethics problem walking around Indian campuses right now is teenagers using free tools to generate sexual images of people who never agreed to it, then calling it a joke when they get caught.

The tool got too easy, and nobody updated the warning label

A few years ago, making a convincing fake image of someone took real skill. GAN training, careful frame alignment, a GPU that could handle the workload, patience. Sergio Alexander, a researcher at Texas Christian University who studies this space, makes the obvious but underrated point: you can now do it on an app, download it straight off social media, and need zero technical expertise to get a realistic result.

That gap between "this requires expertise" and "this requires a phone" is the entire problem. The old ethics conversation assumed friction: a person with enough skill to build a deepfake also had enough judgment to think twice before doing it. That friction is gone. The decision to make one now takes less effort than picking what to order for lunch.

I build developer tools for a living. I get genuinely excited about how fast generative models have improved. But being excited about the tech and being honest about what it enables aren't in conflict. Pretending the second doesn't exist because you like the first is exactly how a school in Guwahati ends up in a cyber crime probe.

What "it was just a joke" actually triggers under Indian law

If you're a student in India and you think the law hasn't caught up to AI-generated content, that assumption is wrong, and it's an expensive one to hold.

Section 66E of the IT Act covers capturing or sharing images that violate someone's privacy. It carries up to three years in prison and a minimum fine of two lakh rupees. Sections 67 and 67A cover obscene and sexually explicit material, going up to five years and a ten lakh rupee fine when the content is sexual in nature. Sections 66C and 66D cover identity theft and cheating by impersonation through a computer resource, three years each. None of these sections care whether the image is "real." A deepfake is still a depiction of a real, identifiable person, generated without consent, and Indian courts have been treating it that way.

The Bharatiya Nyaya Sanhita adds more layers. Section 356 covers defamation. Section 351 covers criminal intimidation, which becomes directly relevant the moment a deepfake is used to threaten or pressure someone. Section 77 punishes capturing or distributing images of a woman in circumstances where she expected privacy, and a fabricated intimate image falls squarely under that even though nothing was technically "captured."

In October 2025, the Ministry of Electronics and IT amended the Intermediary Guidelines specifically to deal with synthetic media. Platforms now have two hours to take down a highly invasive, morphed deepfake after a verified complaint, 36 hours for other non-consensual intimate images, and seven days for anything else flagged as synthetically generated. That timeline exists because the old standard, review it whenever, let images outrun any takedown request through WhatsApp groups and Telegram channels.

None of this requires you to have posted anything publicly. Sending a fabricated image to a private group chat, one friend, a class WhatsApp group, all of it counts as distribution under these sections. "I only sent it to three people" is not a defense in front of a magistrate. It's an admission.

The case that should actually scare you

In Faridabad in October 2025, a 19-year-old college student named Rahul Bharti was blackmailed using AI-generated deepfakes of his sisters. Two people hacked into his phone, reached him over WhatsApp, and threatened to release the fabricated images unless he paid them. He died by suicide roughly two weeks into the blackmail. Police registered the case under the Bharatiya Nyaya Sanhita and are still working to identify the people behind it.

Look closely at what's missing from that case. Rahul didn't make the deepfakes. He didn't share them. He was the target, and the damage still reached him hard enough to end his life. This is the part of the ethics conversation that gets skipped when people frame deepfakes as a victimless prank or "just editing a photo." The harm doesn't stay contained to the person whose face got used. It spreads to siblings, friends, classmates, anyone connected to the target, through psychological pressure that has nothing to do with whether the image was technically real.

Where the actual line is, and it isn't where people think

Consent. That's the whole line. Not whether it was funny, not whether you meant any harm, not whether the group chat thought it was hilarious. If the person in the image didn't agree to be depicted that way, you've crossed it, and intent doesn't move the line back.

This is different from satire or parody, which Indian courts have generally protected as free expression when it's clearly understood as commentary on a public figure and not designed to deceive or sexualize. A meme mocking a politician's speech sits in a different category from generating an explicit image of a classmate. If you're a CS student building anything that touches image or video generation, that distinction needs to live in your product design, not just buried in your terms of service. Refusing certain prompt categories involving real identifiable people, watermarking outputs, logging generation requests against an account: these aren't checkbox compliance. They're the difference between shipping a tool and shipping a weapon with a UI.

What I'd actually tell a student about this

If someone sends you a deepfake of a classmate, the worst move is forwarding it "just to show someone," or keeping it because deleting feels like admitting you saw it. Report it to a teacher or parent and get yourself out of the distribution chain immediately. Possession and forwarding both carry legal exposure under the same sections that apply to the person who made it.

If you're experimenting with these models for a college project, build refusal behavior for prompts involving real identifiable people without consent, from the start, not as a patch after something goes wrong. "I was just testing what the model could do" won't hold up any better than "it was just a joke." The students in Guwahati weren't malicious masterminds. They were teenagers handed a tool more capable than their understanding of consequences, and that gap is closing slower than the tools are improving.

Here's the take most school assemblies won't say out loud: the technology isn't the ethical failure. Treating a sixteen-year-old's judgment as the only safeguard standing between a tool this capable and a classmate's life is the failure. Until schools, parents, and the platforms building these tools actually act on that, the next case is already loading.