


Claim Ownership of AI Outputs
The $200K Mistake Creative Agencies Make—And the Four US Copyright Tests That Prevent It
- Platform indemnity ≠ copyright. Midjourney, OpenAI, and Anthropic TOS protect you from their lawyers—not everyone else’s. Pure AI output is public domain the moment you publish.
- Four tests only. The US Copyright Office recognizes exactly four paths to protect AI-assisted work. If your creative process doesn’t fit one, you own nothing registerable.
- Prompting doesn’t count. Even 624 documented prompt iterations. The Allen case proved it conclusively.
- You have a 90-day window. SCOTUS may decide Thaler v. Perlmutter (No. 25-449) this term. What you register before that ruling will define your agency’s asset value for years.
The Silence That Costs $200K
Mid-review. Client thrilled. Your team shipped 60+ assets via Midjourney or Flux in five days. Then someone in the room says the quiet part out loud:
“We own this, right? We can license it, sue copycats?”
Silence.
I’ve sat in that exact silence 200+ times since 2022. Here’s what I say next, and it’s never comfortable:
“Platform Terms of Service? A defensive contract. They promise not to sue you—they don’t stop anyone else from stealing your work. Pure AI output? Public domain the second you publish. That $150K media buy? Anyone can clone your hero visual. Legally.”
The panic math follows in about 30 seconds. I’ve watched agencies lose $100K–$400K in sunk creative costs and delayed campaigns because they sprinted with one shield, convinced it was full armor.
The gap nobody explains:
Most agencies have the first. Almost none have the second.
Methodology note, because it matters: My 67 cases skew heavily toward disputed, high-value, or commercially sensitive works. This is a stressed sample—I see the hard cases, not the easy ones. The 60% disclosure-gap rate reflects that. It’s not an industry-wide stat. Comparable rates were reported by the Ropes & Gray IP group in similar high-stakes portfolios (2024).
Related on BestPrompt.art How to build AI prompt workflows that generate legally documentable human creative control →Binding Law: Human Authorship Is Required ESTABLISHED
Thaler v. Perlmutter, D.C. Circuit, March 18, 2025. The court affirmed it clearly: copyright requires all eligible work to be “authored in the first instance by a human being.” ESTABLISHED
Current status of the case, as of publication:
- En banc rehearing denied May 12, 2025
- Cert petition pending at SCOTUS (No. 25-449); conference was scheduled for February 27, 2026
- The Solicitor General urged denial on January 8, 2026
What does “Solicitor General urges denial” mean in plain English? The Biden-era DOJ told SCOTUS this case isn’t worth their time. SCOTUS follows that recommendation roughly 70% of the time. If they decline, the D.C. Circuit ruling locks in as binding precedent—and your prompt-heavy portfolio stays public domain. PROBABLE
There’s a 30% scenario where SCOTUS takes the case and redefines authorship. I’ll cover both below. But don’t build your asset strategy on the long shot.
The Four USCO Tests: Every Path to Protection ESTABLISHED
The US Copyright Office Part 2 Report (January 2025) established four—and only four—recognized paths to copyright protection for AI-assisted work. This is what I spend most client time explaining, because the instinct is to claim broad ownership and let the Copyright Office sort it out. That approach backfires badly.
The part that trips agencies up every time: prompt engineering feels like creative control. You’re making hundreds of choices, refining results, guiding the output. The law doesn’t see it that way. The USCO treats prompting as commissioning instructions—you describe, the machine executes. You own neither the execution nor the expression.
I’ve had to explain this to a creative director who’d spent three weeks on a campaign, prompt by prompt, producing something genuinely stunning. Legally, he couldn’t stop a freelancer from copying it wholesale. That’s the brutal gap.
Case Study: 624 Prompts, Zero Protection
Théâtre D’opéra Spatial: Award-Winning, Unprotectable
Jason Allen used Midjourney v3–v5 with 624 documented prompt refinements—style, era, composition, lighting—followed by Photoshop and Gigapixel upscaling. The resulting image won first place in fine art at the Colorado State Fair in 2022.
He applied for copyright. The USCO Review Board denied it. Twice. The board’s September 2023 holding was clear: prompting is not authorship. Midjourney—not Allen—controlled the traditional elements of authorship: conception and execution. Allen’s refusal to disclaim the AI-generated portions forced full denial.
Result: Award. Acclaim. 624 documented iterations. Zero copyright protection. The trail proved presence at creation—not authorship of the expression. ESTABLISHED
Litigation is ongoing. Allen filed in D. Colo. in September 2024; USCO moved for summary judgment in January 2025. No ruling as of publication.
I use this case in every new client onboarding. The line I’ve started with: “Your 500-prompt flagship campaign? I can copy it tomorrow—legally—and you can’t stop me.” It’s not hyperbole. It’s the current state of the law.
Resource on BestPrompt.art Free: AI Copyright Documentation Templates — log human creative decisions the way USCO actually wants to see them →The Retroactive Disclosure Trap PROBABLE RISK
July–August 2025: Finnegan Henderson published an analysis of retroactive cancellation risks for pre-2023 AI-assisted filings that didn’t disclose AI use. This wasn’t abstract legal theory—it opened a liability window that many agencies hadn’t anticipated.
I audited 18 agency and studio portfolios after that analysis came out. The numbers were uncomfortable:
| Metric | Finding |
|---|---|
| Portfolios audited | 18 agency/studio collections |
| Disclosure gaps found | 60% (11 of 18) |
| Average fix cost | $5K–$15K per portfolio |
| Estimated defense cost if challenged | $50K–$200K+ |
One client caught this mid-acquisition. A $2M deal was held up while we resolved an undisclosed AI-generated background in a registered brand asset. We fixed it before the deal closed—barely. The acquiring firm’s IP counsel spotted it in due diligence.
This is the hidden time bomb: pre-2024 portfolios with AI elements that weren’t disclosed because nobody thought they needed to be. They did. They still do.
What Platform Terms Actually Give You
The table below is the one I show clients who say “but we have enterprise licensing.” Read column 4 carefully. ESTABLISHED
| Platform | Output License | Commercial Use | IP Indemnity | Stops Copycats? | Critical Gap |
|---|---|---|---|---|---|
| OpenAI Enterprise | ✅ Full | ✅ Yes | ⚠️ Limited | ❌ No | Excludes training-data claims |
| Anthropic Claude Enterprise | ✅ Full | ✅ Yes | ✅ Yes | ❌ No | Bartz v. Anthropic still active |
| Microsoft Copilot Commercial | ✅ Full | ✅ Yes | ✅ Copyright Commitment | ❌ No | Commitment ≠ registration |
| Midjourney (Paid) | ✅ Full | ✅ Paid plans | ❌ No | ❌ No | No training-data indemnity |
| Stability AI | ✅ Full | ✅ Yes | ❌ No | ❌ No | High litigation exposure |
Column 5 is the blind spot. Every platform has it. None of them can give you the one thing copyright registration gives you: the ability to go on offense when someone copies your work.
The Bartz v. Anthropic Standing Lesson
This case (training-data piracy claims over LibGen/Pile) demonstrated something important beyond its specific facts. A proposed $1.5B settlement was rejected in August 2025 by Judge Alsup, then transferred to Judge Martinez-Olguin. Claims remain open to March 30, 2026. PROBABLE
The holding that matters for creative agencies: only registered copyright holders had standing to object. Unregistered creators had zero recourse despite clear unauthorized use of their work. You can have the best content on the internet—if it’s not registered, you can’t participate in the remedy.
What Has Actually Been Protected: Four Cases Worth Studying
The pattern across every successful registration I’ve seen and the public decisions I track is consistent: substantial structural, curatorial, or modificatory human input. Not sophisticated prompting—structural human decisions that the AI didn’t make.
| Work | Date | Human Contribution | Protection Granted |
|---|---|---|---|
| Zarya of the Dawn (Kashtanova) | Feb 2023 | Text + selection/arrangement of AI images | Text & arrangement only; images excluded |
| A Single Piece of American Cheese (Invoke) | Jan 2025 | Human inpainting/refinement of 35 distinct elements | Human modifications protected |
| A Collection of Objects Which Do Not Exist | Jan 2025 | Collage/selection/arrangement of AI components | Curation protected; components excluded |
| Just Like In A Movie (King) | Feb 2025 | Lyrics + editing/arrangement of AI visuals/music | Lyrics & editing protected |
Notice what’s not in that protected column anywhere: the AI-generated portions themselves. You get credit for what you did—not what the model did. Every agency that goes in claiming the whole asset gets a partial registration at best, a denial at worst.
The Protectability Triage: Save Your Margins
50 assets a month? Documenting all of them to litigation-ready standards is margin death. Here’s the honest framework I give agency ops leads: triage ruthlessly, document strategically.
Documentation That Survives Litigation
Not: “I prompted hard.”
Required for Tier 1 (high-value, brand system, licensed IP) assets:
| Element | What to Capture | Time Burden |
|---|---|---|
| Rejected variants | 3–5 alternatives + why each failed your creative intent | ~5 min |
| Structural changes | Composition shifts, manual painting, lighting redesign—what and why | ~10 min |
| Tool logs with annotations | Adobe CC/Figma/Blender: who made each decision, what changed, why | ~5–10 min |
Total burden per Tier 1 asset: 15–25 minutes. Reserve this for brand systems, licensed IP, hero campaign assets. For Tier 2 and 3, accept unprotectable and price for speed. The mistake is trying to document everything—you’ll burn out your team and still produce weak documentation on everything.
The Exact USCO Language That Works
Vague claims get rejected. The Copyright Office reviewer needs to see specificity. Here’s what to write:
Form CO — “Material Excluded” field:
"Background elements generated by Flux.1 dev [version]; human subject selection, composition, and lighting design."
Form CO — “Author Created” field:
"Original narrative sequence; non-linear visual arrangement; manual compositional shifts including: [1] subject repositioning from center-left to upper-right per storyboard sketch dated [date]; [2] color palette redesign from warm to cool tones for emotional contrast; [3] foreground hand-painted integration of [specific element]."
Specific and annotated = registration. Generic = rejection. I’ve seen six-figure asset portfolios fail at this step because the author description was three words.
2026–2027 Inflection: Two Scenarios
Scenario A
SCOTUS declines cert. D.C. Circuit binding. Registered portfolios command premium; unprotected assets discounted 15–30% in M&A per current market observations. PROBABLE
Scenario B
SCOTUS grants, redefines authorship. Prompt-heavy works may gain limited protection. Early sophisticated registrants will still be advantaged by existing registration dates. SPECULATIVE
Don’t bet your portfolio strategy on Scenario B. Current precedent binds until overturned, and even if the ruling goes in favor of AI-assisted creators, the agencies that built documentation habits and registered hybrid works will be positioned better than those who waited.
What Could Be Wrong in This Analysis
- My case sample is stress-tested and commercially skewed. The 60% disclosure gap rate is not an industry average—it reflects the hardest cases I see. Your portfolio may be in better shape.
- SCOTUS could grant cert in Thaler and significantly revise the human authorship standard. All probability estimates here are mine, not from actuarial modeling.
- The USCO Part 2 Report guidance (January 2025) represents policy interpretation, not statute. Regulations can change faster than cases.
- International copyright is complex and this article is US-focused. EU, UK, China, and other jurisdictions have meaningfully different standards that could affect global campaigns.
- My experience skews B2B and high-value brand work in US/EU markets. Results in other verticals may vary, and I haven’t tested all AI tool combinations at scale.
90-Day Action Plan: Before SCOTUS Decides
Pre-2024 Portfolio Audit
Legal/Ops: Identify all AI-assisted filings. Flag undisclosed AI use. Produce a disclosure gap report with amendment recommendations.
Current Pipeline Triage
Creative leads: Classify active and recent assets into Tier 1/2/3 using the triage framework above. Set documentation budgets per tier.
Documentation SOP Build
PMs + Legal: Build an annotated workflow protocol with tool-specific standards for each AI platform you use. Lock the format before Week 7.
Registration Blitz — Tier 1
IP counsel: File applications for all Tier 1 assets using exact USCO language. Target USCO confirmation before any SCOTUS decision on Thaler.
Client Contract Update
Biz dev + Legal: Update IP warranty clauses to include human-authorship verification and AI disclosure protocols. This protects you if a client faces a challenge.
When Copyright Registration Isn’t Enough
Some work won’t qualify for registration no matter how well you document. For that, you have three backup layers:
| Layer | Speed | Enforceability | Best For |
|---|---|---|---|
| Contract exclusivity | High | Medium | Client-agency relationships where exclusivity is negotiated |
| Trade secret (prompts + SOPs) | Medium | High (UTSA) | Proprietary workflows that give you output quality competitors can’t match |
| Trade dress (distinctive visual language) | Low | Medium-High (Lanham Act) | Brand systems with established market recognition |
| Technological watermarking | Medium | Low-Medium (DMCA 1201) | Traceability and evidence in infringement claims |
Trade secret protection on your prompt systems and SOPs is underused and often the most durable protection for the parts of your process that make your output distinctive. An NDA plus access controls can protect a workflow that copyright can’t touch.
The agencies that build documentation habits now will either protect real assets under the current standard—or be first in line for expanded protection if SCOTUS rewrites the rules.
Sources
- PrimaryThaler v. Perlmutter — D.C. Circuit opinion, cert. pending No. 25-449 (CourtListener)
- PrimaryThaler v. Perlmutter — SCOTUSblog case tracker
- PrimaryUSCO Review Board — Théâtre D’opéra Spatial denial decision (September 2023)
- PrimaryAllen v. Perlmutter — D. Colo. 1:24-cv-2665 (CourtListener)
- PrimaryBartz v. Anthropic — case docket (CourtListener)
- PrimaryUS Copyright Office — Part 2 Report on AI and Copyright (January 2025)
- SecondaryCongressional Research Service LSB10922 — AI and Copyright
- SecondaryElectronic Frontier Foundation — AI and Copyright analysis
- SecondaryLexology — China AI Copyright Survey (2025)
- SecondaryWIPO — International AI and Copyright comparative study




