Session 7 · 27 October 2026
Should AI companies be allowed to train on copyrighted material without permission?
4 min read · 2 positions · 5 cited sources · Session recording
Both teams agreed on more than the motion suggests. Nobody argued for exact copying, and the For team’s own proposals — a visible opt-out symbol, Terms of Service check-boxes, government oversight — were all ways of asking permission. The disagreement was over the default: should AI companies be free to train unless a creator says no, or barred unless a creator says yes?
The debate turned on Novak’s comparison with an artist drawing from a reference. Aayan, moderating, pressed the For team on whether that meant one rule for people and another for AI, and Angelo argued that what makes the comparison break down is scale: a company doing it to the whole internet is not a person looking at a picture. The For team’s strongest answer was practical — nobody can negotiate with every creator, so clear opt-outs are the realistic route — and the Against team’s was that the opt-outs which exist today are buried where users never see them.
Shashank was debating for the first time as an honorary member. Every argument above was made in Room 213; the sources were found afterwards to check the claims the debaters made.
Sides in this debate were assigned, not chosen. Neither position below should be read as any debater's personal view. Our standards.
For the motion
AI learns from references the way artists do, and the answer is clear opt-outs, not a ban.
Novak and Shashank argued that training is not copying. Novak's comparison was an artist drawing: before you draw something, you look at a picture of it first. An AI model does the same, he argued — it looks at existing work as a reference and makes something new inspired by it, rather than reproducing the original file. Exact reproduction would still be a straight copyright breach. Being inspired by something is not.
Their second argument was about what AI is for. If AI is going to be used in fields like medicine, or to help people build new businesses in areas like fashion and technology, it needs to learn from information about the real world, and much of that information is copyrighted. Shashank argued that AI is now a core part of how businesses and startups work, so the question has to be looked at from a business point of view, not only from the creator's.
Novak also argued that training does not have to raid companies' most valuable work: there is a huge amount of free stock imagery and public content to learn from, and companies can limit how much of any one source a model uses. He pointed out that chatbots already refuse some requests involving famous logos, which shows guardrails exist. Shashank added that AI can be programmed not to take part in criminal activity, where a person who gets hold of copyrighted material can't be controlled in the same way.
Shashank's main proposal was a practical one. Instead of every AI company negotiating with every creator, which the room agreed would be impossible, there should be a visible symbol — like the © and ™ marks — that tells AI companies a work is off-limits, backed by clear Terms of Service check-boxes so users know when their content is being used. Rules like this already exist in part: in the EU, AI developers may mine content that is publicly available online unless the rightsholder opts out in a machine-readable way. And where companies want more, they can strike deals, as OpenAI did with Reddit. In closing, the For team argued that the government should set and enforce these boundaries, rather than banning training outright.
Sources (2)
- Defining best practices for opting out of ML training — Open Future (Paul Keller and Zuzanna Warso), 2023
- Reddit and OpenAI Team Up for AI Integration and Advertising — MediaNama, 2024
Against the motion
Taking creators' work without asking is infringement at industrial scale, and nobody is stopping it.
Angelo and Sonal argued that scraping, downloading and storing other people's text, code and art without consent is copyright infringement, whatever the purpose. Angelo's example was a logo: ask an AI for something that looks like Spotify's logo and it will base its answer on Spotify's logo. To the company, that is not "just a logo" — it is intellectual property it created and paid to protect.
Against Novak's comparison with an artist drawing from a reference, Angelo answered with scale. One person looking at a picture is an individual; an AI model is a huge company taking other companies' and creators' work across the whole internet. Sonal added the money: when a person downloads a stock photo, the creator gets paid, and when an AI company takes it, they get nothing. There is a legal way to do this — a licensing contract — and skipping it, she argued, takes away creators' livelihoods. And because developers rarely publish exactly which websites, books and images they trained on, creators often cannot even find out their work was used. As Angelo put it, AI has grown fast over the past five years and the major rules have not caught up.
The Against team argued that the For team's safeguards do not work in practice. Sonal raised Meta using people's photos to train its AI, and Angelo explained how: users were opted in by default and had to find the setting to object. Opt-outs only work if people know they exist. Angelo then argued that the race between ChatGPT, Gemini, Grok and Claude means companies can put safeguards in place but don't, because they just want to be on top — a point Novak agreed was true. Sonal also warned that AI-generated content is now a large share of what is published online, so models increasingly learn from other models' output, copying copyrighted styles at one remove.
Finally, they asked where all this material is stored. Scraped work sits in the companies' own data centres, and those can be breached: Sonal pointed to the hacking of OpenAI's internal systems. Angelo also argued that the more of humanity's creative work AI absorbs, the more people will hand creative tasks to it, and the less original people themselves become. Their conclusion: copyrighted material should belong to the people and companies who made it, and right now the law does not protect them.
Sources (3)
- Hey, UK! Here's how to 'opt out' of Meta using your Facebook and Instagram data to train its AI — TechCrunch, 2024
- More than half of new articles on the internet are being written by AI — The Conversation, republished by AIhub, 2025 (reporting a study by Graphite)
- Hacker Stole OpenAI Internal Documents – Report — Silicon UK, 2024
Who argued which side
For the motion
Novak, Shashank
Against the motion
Angelo, Sonal
Most Valuable Debater, scored on the rubric from the recording: Sonal
Credits
| Name | Role |
|---|---|
| Aayan | Moderator |
CAS learning outcomes
This piece is evidence toward:
- LO5 Demonstrate the skills and recognise the benefits of working collaboratively.
- LO6 Demonstrate engagement with issues of global significance.