
state fairly the strongest arguments on more than one side of the debate over AI image generators and artists’ work, explain the current U.S. legal position on authorship, and reach a reasoned position on how I will use and disclose AI in my own art.
Instruction
This course began with an AI image as the starting point for every project. Now you examine the ground that starting point stands on. The question is contested, and people you respect disagree. Your job is to understand the strongest version of each position before deciding your own.
The artists’ concern. Image generators are trained on very large collections of images gathered from the web, mostly without the consent of, credit to or payment for the people who made them. Artists point out that a model can then produce images “in the style of” a named living artist, competing with that artist’s paid work. In January 2024 the Guardian reported that thousands of British artists, including Damien Hirst and Tracey Emin, were among the names used to train the Midjourney tool. In January 2023 a group of visual artists filed a lawsuit, Andersen v. Stability AI, in federal court against several AI companies. As of November 2026 the case had not been decided: in February 2026 the judge moved its summary-judgment hearing to February 2027. Some artists now use tools developed by researchers at the University of Chicago, Glaze and Nightshade, which alter images in ways intended to disrupt style imitation or training.
The case for the tools. Defenders argue that a model learns general patterns from images much as human artists learn by looking, and does not store or sell copies; that such learning may be a lawful fair use (a question courts are still deciding); and that the tools open image-making to people who could not otherwise do it, including some disabled artists. Some companies now offer models trained on licensed or public-domain images, or let artists opt out of future training. Weigh whether opt-out is enough when most images were collected before it existed. Courts have begun to weigh in on other media: in a case about books, Bartz v. Anthropic, a federal judge found in 2025 that using the books to train an AI model was fair use, but the company had also downloaded pirated copies, and in July 2026 the court approved a $1.5 billion settlement of the authors’ claims about that piracy. That case did not decide anything about images.
Bias. Because models learn what is common in their training data, researchers have found that image generators often reproduce stereotypes — for example in who is shown in particular jobs — and represent some cultures, bodies and places poorly. You saw a small version of this with your texture classifier.
The law on authorship. The U.S. Copyright Office registers only human authorship. Material generated by AI from prompts is not protected; a person’s own expression, creative selection and arrangement, and modification can be (see the summary below). A federal appeals court upheld the human-authorship rule in Thaler v. Perlmutter (2025), and in March 2026 the U.S. Supreme Court declined to hear the case, so that ruling stands. Copyright in training is a separate question the courts have not settled.
Consent and real people. Images of real people raise a different question from style. The federal TAKE IT DOWN Act (2025) makes it a crime, in the circumstances it sets out, to knowingly publish online intimate images of a real, identifiable person, including ones made with AI, without that person’s consent, and requires platforms to remove them on request. Missouri courts also let a person sue when someone uses their identity without consent to gain a commercial advantage. The rule in this course is simpler and stricter: never upload a photo of another person to an AI tool, and never make an image of a real person without their permission.
Disclosure. Whatever you conclude, honesty about process is part of the ethics of making. Industry groups have developed standards such as Content Credentials (C2PA) to record how an image was made. In this course, every work that used an AI image says so: tool, prompt, date, and what you did.
Structured academic controversy, on your own. Write the strongest case for one side from the evidence, then switch and write the strongest case for the other, then drop both roles and write the position you actually hold and one thing that would change your mind. If a family member is willing, they can argue the other side aloud while you restate their argument until they agree you have it right. You are assessed on fairness and evidence, never on which side you land.
📖 Read for this lesson
The reading opens over the lesson; close it to come back to exactly where you were. The checks below draw on it.
Reading: What the U.S. Copyright Office has said about AI and authorship — a summary
Summarized in our own words from: U.S. Copyright Office, Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence (March 2023); the Office’s decision on Zarya of the Dawn (February 2023); Copyright and Artificial Intelligence, Part 2: Copyrightability (January 2025); Thaler v. Perlmutter (D.C. Cir., March 2025; certiorari denied, U.S. Supreme Court No. 25-449, March 2, 2026) · SLS original (summary of public-domain U.S. government documents)
This is a paraphrase, not a quotation. Check the Copyright Office’s AI page for any updates before relying on it.
- Human authorship is required. Copyright protects only material created by a human being. Material generated autonomously by a machine is not protected; a federal appeals court confirmed this in 2025, and the Supreme Court declined to review that decision in 2026.
- Prompts alone are not enough. In the Office’s view, writing a prompt does not give a person enough control over how the output looks to make them its author, even if the prompt is long and creative.
- Human contributions can be protected. A person’s own expression that is perceptible in the final work, their creative selection and arrangement of material, and their creative modification of AI output can be protected.
- An example. For the graphic novel Zarya of the Dawn, the Office registered the author’s text and her selection and arrangement of images and text, but not the individual images generated with Midjourney.
- Disclosure to the Office. People registering works that contain more than a minimal amount of AI-generated material must disclose it and describe the human contribution.
- Training is a separate question. Whether training AI models on copyrighted works is lawful is a different issue, being argued in the courts.
Formative check
Work through these before moving on. They are not graded — they tell you whether the standard below has landed yet.
Which statement about U.S. copyright and AI is accurate?
Andersen v. Stability AI was filed by a group of visual artists in 2023 and, as of this writing, had not been finally decided.
Record your structured academic controversy notes: the strongest argument you made for each side and the evidence behind it. Then write your personal AI-use and disclosure policy (150-200 words): when you will and will not use AI images in your art, how you will credit and disclose them, and one reason from each side of the debate that shaped your policy.
How confident are you that you can state both sides of the AI and artists debate fairly and explain my own policy?
Whose Images? The Evidence Behind Both Sides and the Law: Federal Reports, Court Records, the TAKE IT DOWN Act, Missouri’s Right of Publicity, the FTC and NIST
Copyright Office, Copyright and AI Part 2 (Jan. 2025) and Part 3 (pre-publication, May 2025); Bartz v. Anthropic, No. 3:24-cv-05417 (N.D. Cal., order of June 23, 2025; settlement final approval July 20, 2026, as reported by the Authors Guild); Andersen v. Stability AI, No. 3:23-cv-00201-WHO (scheduling as reported by ChatGPT Is Eating the World, Feb. 3, 2026); Thaler v. Perlmutter, Supreme Court docket No. 25-449 (cert. denied Mar. 2, 2026); TAKE IT DOWN Act, Pub. L. 119-12 (govinfo.gov); Right of Publicity Roadmap, Missouri (Doe v. TCI Cablevision, 110 S.W.3d 363 (Mo. 2003)); FTC, 16 CFR Part 465 (Aug. 14, 2024); NIST SP 1270 (Mar. 15, 2022); first-edition items read 26 September 2026, updates read 9 November 2026 · SLS original (worked notes and policy), with quoted public-domain U.S. government and court text and brief credited quotations from secondary sources
Why this reading. The lab asks you to argue both sides “from the evidence”, to explain the current U.S. legal position and to write your own policy. The lesson names the debate and summarizes the authorship rule, but the evidence for each side, and what the law actually requires of AI companies and of you, is not in it. This reading gives sources: federal reports, court records, a federal statute, a Missouri legal rule, the FTC and NIST. The items carried over from the first edition were read on 26 September 2026; the case updates, the TAKE IT DOWN Act and the Missouri item were checked on 9 November 2026. Laws and cases change; each item gives its number and date so you can check for updates. Nothing here is legal advice.
A. Evidence for the artists’ concern and for the tools’ defenders
The lawsuit. Andersen et al. v. Stability AI Ltd. et al., No. 3:23-cv-00201-WHO, a federal district court case before Judge William H. Orrick, filed 13 January 2023 by a group of visual artists against several AI image companies. It has not been decided. A case tracker reported that on 3 February 2026 the judge granted most of the artists’ request to push the schedule back about three months, moving the summary-judgment hearing to 17 February 2027 (ChatGPT Is Eating the World, 3 February 2026, read 9 November 2026). That is a secondary report of a scheduling order, not a ruling on the merits: check the current docket before saying what the court has held.
The federal government’s view on training. The U.S. Copyright Office released a pre-publication report, Copyright and Artificial Intelligence, Part 3: Generative AI Training (May 2025). It says: “Various uses of copyrighted works in AI training are likely to be transformative. The extent to which they are fair, however, will depend on what works were used, from what source, for what purpose, and with what controls on the outputs—all of which can affect the market.” Read that as support for both sides: it treats training as potentially transformative (the defenders’ point) and makes fairness depend on the source and the effect on the market (the artists’ point).
A court on books, not images. In Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA (N.D. Cal., order of 23 June 2025, Document 231), the judge wrote that “the use of the books at issue to train Claude and its precursors was exceedingly transformative and was a fair use under Section 107 of the Copyright Act.” That case concerned authors and books, not visual artists and images, so it does not decide the image cases, but it shows how one court weighed training. The order also dealt with how the books were acquired, which the Copyright Office report lists among the things that matter: the company had downloaded pirated copies as well. On 20 July 2026 the court granted final approval of a $1.5 billion class settlement which, in the Authors Guild’s words, “resolves only claims about Anthropic’s past conduct in acquiring and copying pirated books.” A settlement is an agreement, not a ruling, so it does not decide whether training is fair use.
Bias. NIST Special Publication 1270 (15 March 2022), Towards a Standard for Identifying and Managing Bias in Artificial Intelligence, says that “biases remain endemic across technology processes and can lead to harmful impacts regardless of intent.” It does not test any image generator. If a parent or guardian agrees and the tool’s age rules allow it, you can do your own check: give an image tool five neutral prompts such as “a nurse”, “a farmer”, “a scientist”, “a family dinner” and “a house”, tally what is shown (who, where, what style) in a table, and report only what you counted. Otherwise, write the five prompts and predict, with reasons from this section, what a model trained on common images would tend to show; label it clearly as a prediction, not a finding.
B. What the law says about consent and disclosure
Authorship. Human authorship is required; prompts alone do not make a user an author; human selection, arrangement and modification can be protected (Copyright Office, Copyright and Artificial Intelligence, Part 2, 29 January 2025). In Thaler v. Perlmutter (D.C. Cir., 18 March 2025) a federal appeals court upheld the rule that a work generated autonomously by a machine cannot be registered. Thaler asked the U.S. Supreme Court to review the decision (No. 25-449); the Court’s docket records “Petition DENIED” on 2 March 2026, so the appeals court’s ruling stands. Denying review is not a Supreme Court opinion: it leaves open how much human input is enough for AI-assisted work.
Images of real people: the TAKE IT DOWN Act (Public Law 119-12, 19 May 2025). This federal law makes it a crime to use an online service to knowingly publish an intimate image of an identifiable person without consent, in the circumstances it sets out (for example, where it is intended to cause or does cause harm), including a “digital forgery”, which it defines as an intimate image of an identifiable individual “created through the use of software, machine learning, artificial intelligence, or any other computer-generated or technological means” that a reasonable person could not tell from a real one. It defines consent as “an affirmative, conscious, and voluntary authorization made by an individual free from force, fraud, duress, misrepresentation, or coercion.” It also requires covered platforms to remove such images when asked. It shows that consent is a legal line, not only an ethical one, when images of real people are involved.
A person’s identity: Missouri’s right of publicity. Missouri has no statute on this; its courts recognize it as common law. The Right of Publicity Roadmap (Professor Jennifer Rothman, read 9 November 2026), citing the Missouri Supreme Court’s decision in Doe v. TCI Cablevision, 110 S.W.3d 363 (Mo. 2003), states the elements: the defendant used the plaintiff’s identity, without consent, with the intent to obtain a commercial advantage. It protects a real person’s identity, such as a name or likeness. It does not protect an artist’s style, which is why the “in the style of” question is argued under copyright and other law. Ask: if you made an image of a real friend with a tool and sold it as a poster, which element would matter?
Disclosure laws bind companies. Some states have passed laws that require AI companies to publish information about their training data or to label AI-generated content. Notice who is bound: the company, not a student artist. Your disclosure rule in this course (tool, prompt, date, what you took, what is yours) is an ethics rule, and a good one, rather than one any such law imposes on you. Content Credentials (C2PA) is an industry standard, not a law, for recording how an image was made.
FTC. The Federal Trade Commission announced on 14 August 2024 a final rule, 16 CFR Part 465, that addresses reviews and testimonials “that misrepresent that they are by someone who does not exist, such as AI-generated fake reviews”. It matters if you ever sell or promote your art: a made-up testimonial from a person who does not exist is deception whatever tool made it.
C. The structured academic controversy, on your own, with worked notes (SLS original)
Procedure (about 30 minutes). Five minutes: read the evidence for both sides above. Eight minutes: write the artists’ strongest argument with its evidence, in a form an artist would sign. Eight minutes: switch and write the defenders’ strongest argument, in a form a defender would sign. Then drop both roles for ten minutes and write what you actually think and one thing that would change your mind. Optional, with a family member: they argue one side aloud and you repeat it back until they say “yes, that is my argument”, then switch.
| Side | Strongest argument | Evidence | Limit |
|---|---|---|---|
| Artists’ concern | Fairness depends on where the works came from and what the output does to their market, and images were gathered mostly without consent. | Copyright Office Part 3: fairness “will depend on what works were used, from what source, for what purpose, and with what controls on the outputs”; Andersen is still pending; the Bartz settlement shows that how works were acquired can carry a heavy price. | A filed case is not a finding, and a settlement is not a ruling; opt-out tools helped only after collection. |
| Tools’ defenders | Training on works is likely transformative, and a model learns patterns as a person learns by looking. | Copyright Office Part 3: uses “likely to be transformative”; Bartz order on books. | Bartz is about books; the Office says the answer depends on the facts. |
Personal policy (about 170 words; SLS original, invented). “I will use an AI image only as a sketch, and only when I can say what I took from it. I will not prompt a tool to imitate a living artist by name, because the artists’ strongest point, that fairness depends on the source and the effect on their market, is one I cannot answer for a single student project. I will never upload a photo of another person to an AI tool or make an image of a real person without their consent: the TAKE IT DOWN Act shows how seriously the law treats fake images of real people, Missouri lets a person sue over commercial use of their identity, and it is disrespectful in any use. I will disclose every use in the form: tool, prompt, date, what I took and what is my own. No law requires that of me; I do it because honesty about process is part of the work. One reason from the defenders that shaped me: a tool can let a person who cannot draw start. One from the artists: that start should not be built on someone else’s unpaid work.”
Sources: U.S. Copyright Office, Artificial Intelligence initiative (Parts 1-3, registration guidance) (copyright.gov); Bartz v. Anthropic, Order on Fair Use (N.D. Cal. 2025) (copyrightalliance.org); Authors Guild, Court Grants Final Approval of Anthropic Copyright Settlement (2026) (authorsguild.org); U.S. District Court, Andersen v. Stability AI, 3:23-cv-00201-WHO (court case page) (cand.uscourts.gov); ChatGPT Is Eating the World, Andersen schedule pushed back (3 Feb. 2026) (chatgptiseatingtheworld.com); Supreme Court of the United States, Docket No. 25-449, Thaler v. Perlmutter (supremecourt.gov); TAKE IT DOWN Act, Public Law 119-12 (govinfo.gov) (govinfo.gov); Right of Publicity Roadmap: Missouri (rightofpublicityroadmap.com); FTC, final rule banning fake reviews and testimonials (14 Aug 2024) (ftc.gov); NIST SP 1270, Towards a Standard for Identifying and Managing Bias in AI (nist.gov)