How to use this session
Everything below is sourced to a primary document - a Federal Register notice, a court opinion, a vendor's operative terms, a regulation, a specification. Where the primary source does not settle a question, this page says so rather than guessing, because a confident wrong answer here is worse than an honest gap. This is not legal advice. It is the material your counsel will ask you for, organised so the conversation takes twenty minutes instead of three meetings.
Prompts are not authorship 7 min live
The US Copyright Office settled its position early and has not moved. Its March 2023 registration guidance (88 Fed. Reg. 16190) says that when a system "receives solely a prompt from a human and produces complex written, visual, or musical works in response, the 'traditional elements of authorship' are determined and executed by the technology - not the human user", and that such prompts "function more like instructions to a commissioned artist". Its Part 2 report on copyrightability, published in final form on 29 January 2025, restates it as a bullet: "Based on the functioning of current generally available technology, prompts do not alone provide sufficient control."
LiveThe filing mechanics, and the penalty for skipping them4 min▶
If your team registers work at all, the process changes once AI is in the pipeline:
- Use the Standard Application, not the Single application.
- Describe the human contribution in the "Author Created" field.
- Exclude AI-generated material that is more than de minimis, under "Limitation of the Claim" then "Other" then "Material Excluded", described as "[content] generated by artificial intelligence".
- Do not list the tool as an author or co-author.
- Already registered something? A supplementary registration corrects it.
The cost of not disclosing is concrete: the Office says it "may take steps to cancel the registration", and separately a court may disregard a registration in an infringement action where inaccurate information was knowingly provided. In other words, the shortcut removes the very protection you were filing for, at exactly the moment you need it.
Zarya of the Dawn is the case to show your team. The registration was cancelled and reissued in February 2023 to cover the human-authored text and the human selection, coordination and arrangement - but not the Midjourney-generated images. A related Review Board decision refused registration for Théâtre D'Opéra Spatial because the applicant would not disclaim the more-than-de-minimis AI content. And in the pending case Allen v. Perlmutter, the Office's own brief records that the applicant entered prompts "at least 624 times" before post-processing, arguing that "repeatedly revising prompts does not amount to authorship" and that "the time and effort involved in creating a work is irrelevant". That case had not been decided when this session was written - do not present a predicted outcome as settled.
Self-studyWhere the human-authorship rule now stands3 min read▶
Two data points make this hard to argue with:
- Thaler v. Perlmutter - the DC Circuit held in March 2025 that human authorship is required as a matter of statutory law, affirming the refusal to register a wholly machine-generated work. The Supreme Court declined to review it in March 2026, so the holding stands.
- The Office's own volume - it states that it has registered hundreds of works incorporating AI-generated material, with the registration covering the human contribution. This is not a ban on AI in the pipeline. It is a rule about what the resulting certificate covers.
One more piece of the picture is genuinely unfinished: the Office's Part 3 report on generative AI training exists only as a pre-publication version from May 2025 and had not been published in final form when this session was written. Its stated conclusion is worth reading in full, because it is more nuanced than either camp quotes: various training uses "are likely to be transformative", but "making commercial use of vast troves of copyrighted works to produce expressive content that competes with them in existing markets, especially where this is accomplished through illegal access, goes beyond established fair use boundaries."
Read the indemnity, not the announcement 7 min live
Every major vendor announced protection for generated output. The operative terms are narrower than the announcements in the same three ways: they are gated to specific paid plans, they exclude trademark claims almost everywhere, and they are voided by modifying or compositing the output. That last exclusion deserves a moment of silence, because modifying and compositing is what a design team does with generated material every single day.
| Vendor | Who is covered | Trademark claims | Stated cap |
|---|---|---|---|
| Adobe | Creative Cloud for teams or enterprise and a plan purchased with Firefly output indemnification. Requires an eligible feature, on an eligible surface, after an export event - placing an asset on the canvas is explicitly not enough. | Included, plus publicity and privacy rights - the only one of the four that covers marks. | US$10,000 per output or per related set of claims, as the sole and exclusive remedy. |
| Microsoft | Paid Covered Products only, free previews excluded. Redirects existing volume-licensing IP defence onto output, subject to five cumulative conditions including required mitigations you must be able to hand over. | Excluded where the claim alleges the output, as used in commerce, violates a mark. | Inherits the volume-licensing cap. Not stated on the page - ask for it in writing. |
| Paid use of listed indemnified services, and only unmodified output from Google's own models. Uniquely, it also carries a separate prong covering Google's use of training data. | Excluded for trademark-related rights in trade or commerce. | Not in the clause; it lives in the underlying cloud agreement. Do not assume uncapped. | |
| OpenAI | API customers and the enterprise tier as defined in its terms. Consumer and mid-tier plans are not mentioned; beta services are excluded. | Excluded for trademark rights in trade or commerce. | Explicitly not subject to any liability cap - the most generous of the four on this point. |
LiveThe exclusions that hit design teams hardest4 min▶
Read these as a workflow audit rather than as legalese:
- Modification and combination. Adobe excludes claims arising from modifying the output or combining it with other materials, and even "the context in which the Output is used". Google covers only unmodified output. OpenAI excludes output "modified, transformed, or used in combination with products or services not provided by" it. If your process is generate-then-art-direct, assume you are outside the cover.
- Turning off the safety plumbing. All four condition cover on not disabling, evading or circumventing filters, citations or safety systems. One "let me get around this" habit voids the protection for everything downstream.
- Partner models. Adobe now hosts dozens of third-party models inside Firefly, and its own terms exclude features "denoted in the user interface as being powered or provided by non-Adobe trained models", plus anything labelled beta or trial. The same surface, the same button, a completely different risk position depending on which model is selected in a dropdown.
- Provenance metadata. Adobe's terms require that you "must not remove or alter any watermarks or Content Authenticity Initiative metadata". A pipeline step that strips metadata for file-size reasons can breach the terms you are relying on.
- One notable inversion. Midjourney's terms run the other way: the user indemnifies Midjourney, and its own words are blunt - "If You harm someone else or get into a dispute with someone else, we will not be involved." Also note its revenue rule: a company (or an employee of a company) over US$1,000,000 a year in revenue must be on a Pro or Mega plan to own assets.
Self-studyWhat the courts have actually decided about image models4 min read▶
The single most useful sentence you can carry into a risk conversation: no court has decided whether training an image generator on copyrighted works is fair use. Every image case is still pre-judgment. What does exist:
- Andersen v. Stability AI (N.D. Cal.) - on the 2024 motions, the copyright claims survived, the DMCA claims were dismissed with prejudice, and false-endorsement and trade-dress claims against one defendant survived. The court expressly deferred fair use: whether the use is "substantial enough" will "be tested on summary judgment". Watch the docket rather than the commentary here: the court's own scheduling order sets a jury trial in June 2027, and the widely repeated September 2026 trial date is stale.
- Getty Images v. Stability AI (UK) - judgment in November 2025. Getty abandoned its training and output copyright claims before closing submissions, accepting there was no evidence the training happened in the UK, so the court never decided whether training infringes. Secondary infringement was rejected on the reasoning that model weights are not themselves an infringing copy. Getty won limited trade mark findings on watermarks, and was granted permission to appeal the secondary-infringement point.
- Getty Images (US) v. Stability AI - refiled in N.D. Cal.; on the 2026 motion to dismiss, only the false copyright-management-information count was dismissed, without prejudice. Everything else, including the unchallenged copyright count and the trademark counts, proceeds.
- The decided fair-use rulings are all about text - a database case that found against fair use and is on appeal, and two 2025 book-training decisions with different outcomes on lawfully acquired versus pirated copies, one of which produced a very large class settlement. Useful context, not a rule about your imagery.
The leadership implication is not paralysis. It is that your exposure today is governed by contracts, platform terms and disclosure duties, all of which you control, rather than by case law, which you cannot.
Provenance proves a signature, not the truth 6 min live
Content Credentials are worth adopting, and they are routinely oversold internally. The specification is admirably honest about its own limits, and quoting it is the fastest way to set expectations with a marketing or comms team that thinks this solves authenticity.
LiveWhere the standard actually stands4 min▶
- Current specification version is 2.4, released in April 2026, which added an AI-disclosure assertion.
- It is not an ISO standard. A related ISO work item exists but sits at an early committee stage in ISO's own record. If someone in your organisation says "it is ISO ratified", that is not supported by the record - and getting this wrong in a client deck is an avoidable embarrassment.
- Conformance is real but young. The published conforming-products list runs to a few hundred records, all certified against an earlier 2.2 version, and includes generators from several large platforms and a stock-imagery provider. Notably, one major creative vendor appears on that list only as a validator, while documenting its generation behaviour in its own product description instead. Membership of the consortium is not the same as an implemented product - do not tell a client that a given social platform displays credentials unless you have checked.
- Durability is partial by design. The security considerations explicitly contemplate downloading an asset, stripping the manifest and reposting it. Recovery relies on watermarking and fingerprinting, which degrade with compression, cropping and format conversion. Independent research has documented both the stripping problem and demonstrations of signed but misleading content.
Adopt it anyway. It is cheap, it is the direction of travel, and being able to show your own chain of custody is worth more than being able to test someone else's.
LiveThe EU transparency duty is now live4 min▶
Article 50 of the EU AI Act applies from 2 August 2026, which for anyone reading this in the week it was written means it applies now. The Digital Omnibus regulation adopted in July 2026 delayed the high-risk obligations to 2027 and 2028 - it did not delay Article 50. The only concession is a provider-side tail: providers of generative systems already on the market before 2 August 2026 have until 2 December 2026 to meet the marking duty. Deployer duties got no extension.
| Duty | Who it binds | What it requires |
|---|---|---|
| Article 50(2) | Providers of systems generating synthetic image, audio, video or text | Outputs "marked in a machine-readable format and detectable as artificially generated or manipulated". Carve-out where the system performs an assistive function for standard editing or does not substantially alter the input. |
| Article 50(4) | Deployers - this is you, when you publish | Where the content constitutes a deep fake, disclose that it "has been artificially generated or manipulated". A deep fake is defined as content resembling existing persons, objects, places, entities or events that would falsely appear authentic. |
| Artistic carve-out | Deployers | Where content forms part of an "evidently artistic, creative, satirical, fictional or analogous work", the duty shrinks to disclosing the existence of generated content in a way that "does not hamper the display or enjoyment of the work". It shrinks the duty; it does not remove it. |
| Article 50(5) | Both | Clear and distinguishable information, at the latest at the first interaction or exposure, meeting accessibility requirements. |
Practical read for a design or marketing team publishing in the EU: the marking duty in 50(2) binds your tool vendor, not you, unless you place your own generative system on the market. What binds you is 50(4), and only when the imagery is deep-fake-shaped - photoreal work resembling real people, places or events that a viewer would take as authentic. A stylised illustration is a different case from a photoreal "photograph" of an event that never happened. The Commission has published both a voluntary code of practice on transparency of AI-generated content and implementation guidelines, and non-signatories have to demonstrate compliance by equivalently adequate means. Penalties for transparency breaches run to the higher of EUR 15,000,000 or 3% of total worldwide annual turnover, so this is not a footnote.
Self-studyBrand and character risk, and the honest state of it3 min read▶
Three findings, and the third is the one leaders find most surprising:
- There is no official trademark guidance on generated imagery. Neither the US nor the EU registry has issued guidance on whether AI-generated content infringes marks, which is structurally expected - infringement is decided by courts, not registries.
- Every filed case against an image generator is copyright-shaped. The large studio complaints plead direct and secondary copyright infringement, not trademark. The clearest trademark language appears in a pre-litigation cease-and-desist, where a studio wrote that a platform was "freeriding off the goodwill of Disney's famous marks and brands" - and the characters were removed without a suit being filed.
- No verified case exists of a brand being penalised for publishing AI imagery that reproduced someone else's mark or character. Enforcement has targeted tool providers and platforms, not downstream publishers. Treat that as a finding, not as reassurance: your realistic near-term exposure is contractual and reputational rather than litigated, and reputational damage does not need a court date. Several brands have pulled AI campaigns after public backlash; verify any specific example before naming it in a deck.
Platform terms are where your actual obligations live. Adobe's generative AI user guidelines prohibit using its features to create content that violates third-party trademark rights and state that this "may include entering text prompts designed to generate copyrighted, trademarked, or otherwise infringing content", with prompts and outputs reviewable by automated and manual methods. Midjourney prohibits using the service to try to violate others' trademark rights and, as noted, puts the indemnity on you. Neither vendor has published a brand blocklist, so do not tell your team "the tool will stop you".
Five things to write into your guidelines 5 min live
Everything above collapses into five sentences your team can actually follow. Write them once, put them where the work happens, and revisit them each quarter, because half the facts on this page carry a date.
1 · Which models are approved for client-facing work, named individually - because indemnity follows the model, not the app. Include the rule that partner models inside an approved app are not automatically approved.
2 · Never disable or circumvent filters or citations, and never strip provenance metadata. Both void the protection you are paying for.
3 · What gets registered and how - Standard Application, human contribution described, AI material disclaimed above de minimis. And the corollary: assume the generated layer itself is not yours exclusively.
4 · The disclosure rule for published work, written to the deep-fake test rather than to a vague "we label AI" promise, with a named owner for EU-facing campaigns.
5 · No brand-identifiable prompting - no marks, no protected characters, no living person's likeness without permission. Platform terms already prohibit it and no tool reliably blocks it.
Before the next session ◐ 30-40 min total
- Pull your actual contracts and fill in the four-row indemnity table for the vendors your team really uses, including the cap. If a cap is not stated, that is the email to send this week.
- List every model your team generated with last month, then mark which ones are first-party and which are partner models inside a first-party app. Most teams are surprised.
- Check one pipeline step for metadata stripping - an export preset, an image optimiser, a CMS upload. Find out whether provenance survives your own publishing path.
- Draft the five guideline sentences above in your own words and take them to counsel as a starting point rather than a question.
Official sources covered
Every fact on this page traces to a primary document read during the build: a Federal Register notice, Copyright Office reports and decisions, court opinions and orders, vendors' operative terms, the EU regulation and its implementing guidance, and the provenance specification. Where a primary source did not settle a question, the page says so. This is not legal advice. This page covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · A designer refines a prompt 600 times and ships the result. What does the US Copyright Office position imply?
The guidance treats prompts as instructions to a commissioned artist, and the Office has argued that repeatedly revising prompts does not amount to authorship. What you can register is your own expression in the output, your creative selection and arrangement, and your creative modifications - with the AI material disclaimed.
2 · Which everyday design practice most reliably voids a vendor output indemnity?
Google covers only unmodified output, OpenAI excludes output modified or used in combination with other products, and Adobe excludes claims arising from modification, combination, and even the context of use. Art-directing generated material is normal practice and it usually takes you outside the cover - which is why the cap and the conditions matter more than the announcement.
3 · A photoreal AI image of a real city square during a real event is used in an EU campaign. Which duty is most likely yours?
The marking duty in 50(2) binds providers of the generating system. As a publisher you are a deployer, and 50(4) bites where the content is deep-fake-shaped - resembling existing persons, places or events in a way that would falsely appear authentic. The artistic carve-out narrows the duty to disclosing existence; it does not delete it.