You uploaded a photo, a song, a painting. Somewhere, a model is training on it, and nobody asked you first.
This site is a plain-English guide to what's happening to your data online, who it hurts, and the tools you can actually use to push back.
Who's affected
Generative AI is trained on the open web. That sounds abstract until you realise the open web means your selfies, your mother's voice note, the sketch you posted in eighth grade, the reel you filmed last Tuesday. Here are five ways that plays out.
01 · Ordinary users
Every public photo is a training sample. Face-swap apps, undressing tools, and non-consensual deepfake porn all rely on models that learned from ordinary selfies, most of which were never meant for that purpose. Once an image leaves your phone, it can be reconstituted into something you never posed for.
e.g., a global surge in AI deepfakes driving new cases of child sexual violence.[1]
02 · Voters
Cloned political voices and AI-generated rally footage have already circulated during election seasons. Detection tools lag behind generation tools by months, sometimes years. The cost of making a convincing fake dropped from a studio to a laptop and ten minutes.
e.g., AI-generated deepfake videos and cloned voices of political leaders flagged as threats across Indian elections.[2]
03 · Artists
A diffusion model that has seen a few dozen of an illustrator's pieces can generate new work in their style on demand, for free, in seconds. Commissions dry up. Clients pick the $0 version. The artist's own search results get buried under imitations. Class-action lawsuits are slow; by the time courts rule, the career damage is done.
e.g., the Glaze user study: 1,156 artists, 88% want a protection tool.[3]
04 · Creators
Platforms now auto-dub creators into languages they don't speak, matching lip movements and voice timbre. Nuance, slang, and regional identity get flattened. More importantly: the AI-modified version often gets more reach than the original, so the original effectively disappears.
e.g., YouTube's auto-dubbing feature automatically translating creators' videos into other languages.[4][5]
05 · Everyone
Most platforms bury AI-training clauses inside 40-page Terms of Service that the average person will never read: just 9% of adults say they always read privacy policies before agreeing.[7] Opt out of AI training is either missing, hidden three menus deep, or only available to users in specific jurisdictions.[6]
Consent, in practice, is manufactured by exhaustion.
Consent, in practice, is manufactured by exhaustion. The pattern across every major platform's ToS, 2023–25
The pushback from publishers isn't aimed only at OpenAI or Anthropic. When a
newspaper updates its robots.txt to keep AI crawlers out, the same
rule frequently locks out the Internet Archive too: the public-memory service
that has preserved over a trillion web pages since 1996. By late 2024, a
large share of the world's top news sites were blocking at least
one major AI crawler;[8] many major
news sites have also moved to block the Wayback Machine.[9]
Journalism and the public record used to share one infrastructure. AI-training
anxiety is pulling them apart.
The bind is brutal: the same tools publishers reach for to keep AI models out also keep librarians, researchers, and ordinary readers from preserving what was published. The remedy we build for one problem is quietly dismantling the piece of the web most of us assumed would always be there.
India in focus
Most global AI-consent discourse is framed around GDPR and the EU AI Act. Indian users rarely see themselves in it.
India is one of the largest sources of training data on the planet: photos, voices, languages, faces, handwritten scripts. It is also, for now, one of the jurisdictions where AI companies face the least friction. The EU can fine a company billions under the AI Act. A user in Berlin can demand their data be deleted under GDPR. A user in Bengaluru, in most cases, cannot.
| Jurisdiction | Key law | What it gives users |
|---|---|---|
| European Union | GDPR (2018)[11] + AI Act (2024)[12] | Right to erasure, right to object to automated processing; AI Act penalties reach up to €35 million or 7% of global turnover for prohibited practices. |
| California, USA | CCPA / CPRA[13] | Right to know, delete, opt out of sale or sharing, including for AI training. |
| India | DPDP Act, 2023[14] + DPDP Rules, 2025[15] | Consent-based framework; rules notified in 2025 and Data Protection Board still being operationalised. No AI-specific provisions. |
| United Kingdom | Data Protection Act + emerging AI regulation | Similar to GDPR; still evolving. |
Several patterns recur in reporting on AI deployment in India:
The pattern isn't conspiracy: it's economic gravity. Where the regulatory cost of a mistake is low and the labour cost of data work is low, companies go. India is currently both.
The Act gives you the right to consent, purpose limitation, and erasure. It is narrow, and it is still rolling out, but it is what you have.
Glaze, Nightshade, and C2PA work regardless of jurisdiction. Laws can wait; your images can't.
If your work or likeness appears in a scraped dataset, keep a paper trail: it's the evidence base that legal and advocacy follow-up is built on.
The Toolbox
None of them are silver bullets. Each addresses a different layer of the problem: proving what something is, hiding what a model can learn from it, and poisoning what a model tries to steal. Used together, they shift leverage back to the creator.
C2PA
A tamper-evident receipt for every piece of media.
An open standard, backed by Adobe, Microsoft, BBC, Nikon, Leica and ~300 others, that attaches a cryptographically signed manifest to your file. It records who made it, when, with what tool, and whether AI was involved.
Deep diveGlaze
A cloak for your art style.
From the University of Chicago's SAND Lab. Nearly-invisible perturbations that change nothing for a human eye but make a diffusion model learn the wrong style. Defensive. Over 93% effective in the authors' study of 1,156 artists.
Deep diveNightshade
If you scrape me, you poison your model.
From the same lab, but on offence. It modifies an image so that, to a human, it still looks like a dog, but to a model training on it, "dog" starts to mean "cat." As few as 50–100 samples measurably corrupts a concept.
Deep dive| C2PA | Glaze | Nightshade | |
|---|---|---|---|
| What it does | Labels content with signed provenance. | Hides your style from learning. | Poisons what a scraper learns. |
| Who it's for | Everyone. | Artists with a style at risk. | Artists willing to escalate. |
| Works if already scraped? | Partially (for future edits). | No. | No. |
| Requires cooperation? | Yes, AI companies must honour it. | No. | No. |
| Stance | Defensive / legal. | Defensive / technical. | Offensive / technical. |
| Cost to user | Free (Adobe CAI). | Free. | Free. |
The Agency Simulator
Below is a side-by-side sandbox. On the left: a regular image, as a scraper sees it, featureless, learnable, repeatable. On the right: the same image with C2PA credentials attached and a Glaze-style perturbation applied. Watch what the simulated scraper picks up in each case.
Your image never leaves your browser. Everything above is processed locally via the Canvas API: no upload, no server.
What artists actually think
Researchers at UChicago's SAND Lab surveyed 1,156 professional artists — illustrators, concept artists, animators — for the Glaze paper (USENIX Security 2023). It is, to date, the largest structured record of what working creators actually feel about generative AI training on their work. The numbers are not ambivalent.
77% say models mimic their specific style — what a client would recognise as theirs, not a generic look. 24% have already lost paid work to it. The harm is current, not hypothetical.
88% wanted a dedicated protection tool before one existed. After trying Glaze, 93% rated it successful at disrupting mimicry. The appetite for a technical opt-out isn't fringe — it's nearly the whole profession.
89% have taken or plan defensive measures: watermarking, style-cloaking, locking portfolios, leaving platforms. Over half are considering pulling their work off the public internet entirely.
51% said they would personally join a class-action against an AI company training on their work without consent. In a profession historically reluctant to litigate, a clean majority willing to go to court is a structural shift.
Source: Shan et al., Glaze: Protecting Artists from Style Mimicry by Text-to-Image Models, USENIX Security 2023.[3] 1,156 professional artists surveyed, recruited through artist community networks: predominantly full-time (46%) or part-time / freelance (50%), across illustration, concept art, animation, and related fields.
The survey is not a random sample of all artists; it skews towards creators already aware of AI-training concerns and engaged enough to respond. Read the figures as the view from the affected, attentive core of the profession: the people who will define what the rest eventually demand.
Bibliography
Sources behind the claims, statistics, and examples throughout this site.
Get involved
Spot a missing tool, a better explanation, or a case study worth adding? Open a PR or an issue on GitHub.
The Agency Simulator can grow, and the site itself has rough edges.
Good first issues are tagged good-first-issue on GitHub.
Pass it along to anyone whose work, voice, or likeness lives online. The point of an educational resource is to be read.
Content licensed CC-BY-SA 4.0. Code licensed MIT. Attribution, not permission, is the default.