I. Why This Guide

A. Purpose: To Inform, Not Alarm

The rise of generative AI has made almost every creator—artist, writer, musician, developer—a contributor, often by default, to large-scale AI training efforts with limited visibility into how that use occurs or how it might be constrained. Works that once lived quietly on websites, social-media feeds, and portfolio platforms are now copied, parsed, and recombined to train the next generation of “creative” machines. Much of this occurs without clear notice, express permission, or compensation.

This Guide proceeds from a simple premise reflected in the Human Baseline Principle: that meaningful consent requires a clear, intelligible opportunity to choose, rather than participation by default.

One goal is to help restore visibility into how authorship is treated within current systems. This is not a manifesto against technology; it is a practical guide to navigating how AI training affects creator rights. Its main purpose is simple: to equip creators with practical and procedural tools—and with clear-eyed understanding of the limited legal structures that currently exist—to limit or object to the use of their work in AI training, where such choices are still possible.

Platforms often argue that using uploaded content for AI training is a reasonable cost of providing free or low-cost services. But when such use is imposed by default, described in broad or ambiguous terms, and difficult to refuse in practice, the central issue is no longer justification—it is whether creators have any meaningful ability to say no.

Opting Out of AI Training: Protecting Creator-Uploaded Content on Major Platforms—Guide No. 1 of The AIRights Guide™ Series—focuses specifically on platforms where creators actively upload their works of authorship, and where the terms governing AI training are defined at the point of upload, including creative platforms, social platforms, and developer platforms. It evaluates how those platforms disclose AI training, what choices they offer creators, and whether exercised choices meaningfully limit future use.

This Guide does not:

• evaluate AI model developers or AI services that train or deploy models using data acquired through licensing, partnerships, scraping, or other downstream means; or
• provide guidance for protecting self-hosted content from web scraping or AI training.

These omissions are intentional. The consent mechanics, power dynamics, and avenues for recourse in those contexts are fundamentally different from those that arise when creators hand their work directly to a platform. Those issues are addressed separately in companion publications: Protecting Self-Hosted Content from AI Scraping (Guide No. 2) and Using AI Systems and Using AI Systems: Protecting Your Inputs, Outputs, and Rights (Guide No. 3).

Some rights apply and remain enforceable, others operate only in part, and many more remain unsettled. Knowing which is which can make the difference between feeling powerless and acting strategically. This Guide therefore focuses on actionable literacy: how to recognize the points where consent, notice, and control still operate. It identifies opt-out mechanisms where they exist and highlights blind spots where they do not, and translates the fine print of “AI data use” policies into plain language.

1. Three Kinds of Protection, None Built for AI

a. Privacy law protects who you are.

It governs the use of personally identifiable information—names, images, voices, locations, or digital traces. When an AI system trains on human faces or health data, privacy rules apply. But they are not applicable for creative works of expression. A painting or a poem is not “personal data,” even when the artist’s name appears in the metadata.

b. Trade-secret law protects what you know privately.

Businesses guard confidential data, algorithms, and internal documents under the doctrine of secrecy: protection lasts only while the information stays out of public view. Once a dataset or piece of code is published—or scraped from the open web—that protection evaporates. But creators, whose livelihoods depend on sharing their work, cannot rely on secrecy as a shield.

c. Copyright law protects your expressive works—the things you make public.

It attaches automatically to original works of authorship but presumes copying will be visible, traceable, and human. AI training defies those assumptions. Developers claim that ingesting millions of works to “analyze patterns” is fair use because no literal copy survives in the finished model. Courts are still deciding whether that metaphysical distinction holds water.

2. Mismatched Shields

Each framework guards a different doorway; none guards the hallway between them.

A creator can file copyright notices, submit takedown requests, or toggle an opt-out form and still see the same work reappear elsewhere.

Where is consent recorded—and who recognizes it?

There is no central registry of consent, no universal “do not train” tag with legal force.¹

Fair-use, consent, and secrecy each operate under different logics—and all vary wildly in practice. Copyright asks whether copying is “transformative.” Privacy asks whether permission was given and personally identifiable information removed. Trade-secret law asks whether the information was ever public to begin with. For creators, the outcome is the same: uncertainty about who can use their work, and for what purpose.

[cont’d ↗]

3. Judicial Drift and the Myth of Implied Consent

Early lawsuits challenging AI training practices have resulted in a couple of judicial decisions that hint at a trend: courts are showing deference to the idea that training on publicly available material may be permissible under fair-use principles. The issue remains unsettled, but the first summary-judgment decisions—in Bartz v. Anthropic and Kadrey v. Meta (N.D. Cal. 2025)—have not favored creators seeking to block AI training on copyright grounds.²

Why should “public posting”—the very act by which artists share their work with the world—constitute permission to use to train AI?

This raises a fundamental interpretive question: whether public posting—traditionally understood as sharing with an audience—can and should be treated, at machine scale, as permission for use in AI training.

The answer to that question has significant implications for how consent is understood in AI-mediated environments.

4. Platform Fine Print

The big platforms—social networks, code repositories, creative-hosting sites—frame their AI policies in the language of “innovation” and “personalization.” The Terms of Service, however, include permissions broad enough to encompass nearly any use of uploaded content. A single clause authorizing “improvement of services” can justify extensive model training. Platforms often emphasize that data is not “sold,” but those same provisions typically allow sharing, licensing, or affiliate use that renders opt-outs ineffective—shifting the real issue from compensation to control.

Opt-outs, when they exist, often resemble gym memberships: easy to start, hard to cancel, and quietly renewable. Some platforms offer region-specific forms, others route users through generic “data rights” portals. Few confirm what actually happens to data already ingested into models. “We’ll remove it to the extent technically possible” is the standard policy statement

5. Net Result

The typical creator faces two intertwined paradoxes: legally recognized ownership of their work, yet almost no practical control over its algorithmic afterlife; and a supposed remedy that would require never making the work public at all—a difficult proposition when visibility is, for many, both the purpose of authorship and the prerequisite for livelihood.

Enforcement options—copyright claims, privacy requests, even contractual complaints—remain costly and uncertain. But clarity itself is a form of power. Knowing where the system fails allows for tactical adaptation: choosing safer platforms, limiting uploads, watermarking, filing timely objections before data enters model pipelines, and understanding which legal reforms are worth advocating for.

Awareness is not symbolic; it is preparation.

B. What This Guide Will and Won’t Do

It Will Do Four Things

1. Map the terrain. Explain in plain language how data from creative platforms flows into AI training systems—what’s collected, where it goes, and which rights attach along the way.
2. Compare the rules. Show what each major platform must offer by law (if anything), what it chooses to offer voluntarily, and what happens when neither exists.
3. Offer workable actions. Step-by-step summaries for opting out, requesting data deletion, and setting privacy controls where they actually matter.
4. Provide context without panic. The goal is not to scare creators offline but to help them publish on their own terms—aware of exposure, not paralyzed by it.

It Won’t Do Two Things

1. It won’t pretend to solve the problem. The imbalance between individual creators and AI companies is structural, not procedural. A few forms and privacy toggles will not rebalance it.
2. It won’t offer legal representation or guarantees. Every jurisdiction is in flux. What holds today may vanish with the next ruling or executive order. This Guide informs; it does not litigate.

In short: this is a field guide, not a philosophy text. If you reach the end knowing what you can do this week to limit AI exposure of your work, it has succeeded. And if it helps clarify why the system works as it does—and where future reform might matter—that, too, is progress.

[cont’d ↗]

II. Copyright Law: The Framework That Fails to Scale

The Copyright Conundrum: How do we let inventors and artists control and profit from what they make, yet still allow others to build upon it?

To understand why these constraints persist, it is necessary to examine how existing legal frameworks—particularly copyright—interact with AI training.

A. How Copyright and Fair Use Worked, Pre-AI

Copyright was designed for a world where copying was visible, traceable, and usually profitable. The system rested on three assumptions:

1. Identifiable authors. A human created something new.
2. Detectable copying. Someone else reproduced it.
3. Commercial context. The copying replaced or diluted the market for the original.

Within that framework, fair use operated as a pressure valve. Quoting, parodying, or referencing a work could qualify as “transformative,” meaning the copier added new meaning rather than merely reusing content. The balance worked—imperfectly but with some predictability—so long as copying remained human-scale, visible, and interpretable.

Then came automation. The doctrine never contemplated training sets or neural networks. It was written for photocopiers, not planet-sized data centers.

B. How Data Scraping Rewired the Internet

When the web matured, search engines and aggregators began scraping publicly available pages—copying text and images to index, summarize, or rank them. Courts and regulators generally tolerated this: it served a public function, it displayed only snippets, and it didn’t compete directly with the underlying works.

AI training operates at a fundamentally different scale. Instead of indexing, it ingests; instead of reproducing snippets, it generates new content that can directly compete with the original. In that earlier indexing model, consent typically attached to specific, identifiable uses—such as displaying, indexing, or storing content. AI systems instead operate through ongoing processes of training and reuse, where the full scope of use cannot be fully specified in advance. As a result, consent is asked to govern not just discrete acts, but evolving uses over time. In effect, the process has shifted the open web into a form of industrial data extraction—something closer to a ‘data quarry’ than an index.

Consent was never revoked because it was never formally given. Creative platforms’ terms of service, written long before AI, permitted broad “service improvement” uses. That innocuous phrase now justifies the ingestion of entire archives.

In effect, the legal system accommodated early forms of scraping as a “transformative” utility, without fully accounting for how those practices would scale into large-scale data extraction.

C. How AI Slipped Through the Cracks

AI training practices now occupy a legal gray zone with three major gaps:

1. The Doctrinal Gap. Training involves the mass reproduction of works—the first and most fundamental element of any copyright infringement claim—but defenders argue that no “copy” remains once the model internalizes patterns. Courts are still deciding whether that metaphysical distinction holds water.
2. The Temporal Gap. Laws presume discrete acts of copying. Models, by contrast, retrain continuously, blurring the line between past and ongoing use.
3. The Accountability Gap. A model that outputs a style “inspired by” a million artists implicates everyone and no one. Copyright requires an identifiable infringer; machine learning produces none.

Other jurisdictions, such as the EU and U.K., now require opt-out mechanisms under their text-and-data-mining exceptions. The U.S., by contrast, still relies on broad fair-use doctrines that leave such decisions largely to corporate discretion.

Creative works thus occupy a no-man’s-land: too expressive to count as personal data, too public to claim privacy, and too scattered to police through individual enforcement. The system built to protect art now protects abstraction instead.

D. Why This Matters Now

For creators, these legal ambiguities translate into practical exposure. Every image uploaded, lyric posted, or snippet of code shared can become training material by default. The tools to contest it—copyright takedowns, privacy requests, platform forms—address individual instances, not industrial pipelines.

The system built to protect art now operates more effectively at the level of abstraction than attribution.