The Ownership Trap: Why Paying for AI Development Doesn't Mean You Own It

You've signed the check, the developer has delivered the build, and the software works. There's just one problem: under U.S. copyright law, you may not own a single line of it.

This is the ownership trap that catches business owners hiring developers for custom software — and AI development raises the stakes considerably. According to WIPO, the software developer automatically retains IP ownership unless a software development agreement explicitly assigns IP to the client. Many ventures mistakenly assume that paying for custom software grants them full ownership. It doesn't.

Under U.S. copyright law, ownership vests initially in the author of the work. In an outsourced development arrangement, the author is the developer — not the person footing the bill. WIPO identifies four ownership models available in development agreements: IP assignment, IP licensing, work for hire, and joint ownership. If your contract is silent or vague, the default lands with the developer.

The problem is compounded by contract language that looks strong but fails in practice. A legal analysis by Willcox Savage notes that many template agreements contain IP clauses that do not actually transfer code ownership. Two common failure modes stand out:

• "Work made for hire" designations are likely insufficient for software, since commissioned works must fall within nine enumerated categories under 17 U.S.C. § 101 — and many software programs fit none of them.
• "Agree to assign" language creates only a future promise. Federal Circuit decisions in 2018 and 2021 held that such language did not effect a present assignment of the IP.
• Many form agreements were compiled by non-attorneys, making experienced counsel review before signature critical.

AI development adds a layer of complexity that traditional software contracts never faced. Purely AI-generated code cannot be copyrighted in the U.S. — human authorship is required, a position affirmed by the D.C. Court of Appeals in Thaler v. Perlmutter in March 2025. Code produced solely by AI may be freely used by anyone unless protected by other means, such as trade secret.

The scale of the issue is growing fast. Microsoft's CEO estimates that 30% of the company's production code is AI-generated, and Meta's CEO projects roughly half of its codebase will be AI-generated this year. Ownership strength now scales with human involvement: when developers meaningfully edit, refine, and integrate AI output, copyright can attach to those human contributions.

For business owners hiring AI developers, the practical takeaway is simple but urgent. Ownership must be secured by contract language, not assumed from an invoice. At Agents by AIQ, we believe the client should own everything they've paid for — which is why ownership terms deserve as much attention as the build itself. Before you sign any AI development agreement, ask one question first: who owns what when this project ends?

When Standard Contract Language Fails: Work for Hire and 'Agree to Assign'

You paid for the code, the contract says you own it — and a court may disagree. Two of the most common clause types in development agreements quietly fail to transfer ownership at all, leaving clients with a working product and no legal title.

The first failure mode is the "work made for hire" designation. It sounds airtight, but for independent contractors, the work must fall within one of nine enumerated categories under 17 U.S.C. § 101 — and many software programs fit none of them. As a result, a Willcox Savage analysis concludes these designations are likely insufficient for software, meaning the developer — not the client — remains the legal author by default.

The second trap is "agree to assign" language. Courts have treated this phrasing as a mere promise to transfer rights in the future, not a transfer itself. In Advanced Video Techs. LLC v. HTC Corp. (2018) and Omni MedSci, Inc. v. Apple Inc. (2021), the Federal Circuit held that such language did not effect a present assignment. The developer retains ownership until a separate, later document actually conveys it — which rarely gets signed.

The fix is deceptively simple: present-tense assignment language, sometimes called the "magic words." A clause stating the developer "agrees to assign, and hereby does assign" all right, title, and interest in the code and deliverables effects the transfer at signing. Attorney Joseph B. Allen recommends pairing this with:

• A broadly defined "Work Product" term covering deliverables, inventions, and work product
• A perpetual license to the developer's pre-existing or background IP
• A waiver of moral rights
• A commitment that the developer will assist with IP enforcement if needed

This matters even more in AI development, where outputs may carry weak or no copyright protection at all. Since purely AI-generated code is ineligible for copyright under the human-authorship requirement, contract language can't create protection where the law doesn't — but it can assign whatever rights do exist and preserve your commercialization rights regardless, a point emphasized in contract drafting guidance from Saul Ewing partner Matt Kohel.

At Agents by AIQ, we build every agent engagement on month-to-month terms where the client owns everything — the code, the configurations, the integrations — because ownership language should never be the thing that surprises you later. Before you sign any development agreement, check the assignment clause for those two words: "hereby does." If they're missing, ask for a redline.

The AI Twist: Human Authorship and the Three-Part IP Split

Paying an AI developer no longer guarantees you own what they build — because in some cases, no one owns it at all. That's the uncomfortable reality at the center of modern development contracts.

Under U.S. law, purely AI-generated code cannot be copyrighted. The D.C. Court of Appeals confirmed this in Thaler v. Perlmutter (March 2025, rehearing denied May 2025), affirming that copyright requires a human author — a position shared by both the courts and the U.S. Copyright Office, as detailed in this legal analysis of AI copyright and work made for hire. Code produced solely by AI may be freely used by anyone unless protected another way, such as a trade secret.

Human involvement changes the picture. When developers iteratively prompt, edit, refine, and integrate AI output, copyright can attach to the human contributions — and flow to their employer through the contract. But as Judge Millett noted, only the human modifications are protected, not the underlying AI-generated material. Simply accepting generated code weakens your position; meaningful human authorship strengthens it, according to IP attorneys analyzing AI code ownership.

This matters more every year. Microsoft's CEO estimates 30% of the company's production code is AI-generated, and Meta's CEO projects roughly half of its codebase will be AI-generated this year. The proportion of machine-authored work in your project will only grow.

That's why AI development agreements can't treat "the IP" as one lump. Contract structuring guidance for AI deals identifies three components that must be allocated separately:

• Core models and background IP — the developer's pre-existing algorithms versus new models created during your project
• Training data — your proprietary datasets and RAG knowledge-base documents (almost universally customer IP) versus the provider's data
• Outputs — explicitly assigned to you, since assignment language preserves commercialization rights even where copyright may not attach

The central negotiation point is the fine-tuned model itself — the version trained on your data. Deals typically land on customer ownership, joint ownership, or a license, and which one you get shapes what you can do long after the engagement ends. For businesses hiring outside AI developers — whether a full build or an agent that answers calls and follows up on leads — clarity here prevents the worst outcome: a system that runs your operations but that you can't freely use, modify, or take with you.

Contracts can't manufacture copyright protection where the law doesn't provide it. But they can assign every right that does exist, lock down your data, and make the fine-tuned model yours in writing. That's the standard Agents by AIQ holds its own engagements to, and the standard any client should demand.

Your Contract Checklist: Clauses to Demand Before You Sign

Before you sign anything, remember this: paying for custom software does not grant you ownership. Under U.S. copyright law, ownership vests initially in the author — the developer — and the developer automatically retains IP ownership unless your agreement explicitly assigns it to you, as WIPO guidance makes clear.

That gap between what you pay and what you own is closed with specific contract language. Attorney Joseph B. Allen calls these the "magic words," and they matter because common clauses routinely fail. "Work made for hire" designations are likely insufficient for software, since independent-contractor works must fall within nine enumerated categories under 17 U.S.C. § 101 — and many software programs fit none of them. Meanwhile, "agree to assign" language creates only a future promise; in Advanced Video Techs. LLC v. HTC Corp. and Omni MedSci, Inc. v. Apple Inc., such wording did not effect a present assignment.

Demand these clauses before you sign:

• Present-tense assignment: the developer "hereby does assign" all right, title, and interest in code, deliverables, and inventions developed under the agreement.
• Broadly defined "Work Product" — so nothing you paid for slips outside the definition.
• A perpetual license to background IP — the developer's pre-existing tools and models.
• A moral-rights waiver, plus the developer's assistance if you ever need to enforce the IP.
• Open-source license restrictions, since some open-source licenses carry restrictive commercial consequences.

For AI projects specifically, the assignment should also cover outputs. Purely AI-generated code cannot be copyrighted in the U.S. — human authorship is required — so contract language preserves your commercialization rights regardless of copyrightability. AI development contracts should also allocate three IP components separately: background models, training data, and outputs, with customer-supplied data almost universally treated as customer IP.

One more caveat: don't assume you need to own every line. Most software is not 100% original, and WIPO notes ventures often need clear permissions to use, modify, and commercialize rather than full ownership of the codebase. Pair those permissions with warranties and indemnification for non-infringement — though recognize these are only as strong as the developer's ability to verify the provenance of what they ship.

This is the standard we build to at Agents by AIQ: month-to-month engagements where the client owns everything we deliver. If you're scoping an AI agent for your business, bring these questions to the table — whoever builds it should welcome them.

After the Handoff: Documentation, Provenance, and End-of-Contract Protections

Delivery day feels like the finish line, but for AI projects it's when your ownership position is either locked in or quietly lost. The code, the fine-tuned model, the RAG knowledge base, the credentials — everything your business paid for needs to actually end up in your hands, and the record of how it was built needs to end up in your files.

Start with provenance. Because U.S. copyright law requires human authorship — purely AI-generated code cannot be copyrighted, as affirmed in Thaler v. Perlmutter — your strongest claims come from documented human involvement. Legal guidance from MBHB LLP recommends logging which prompts produced which parts of the codebase, and recording human edits, refinements, and review decisions throughout the project. If a developer meaningfully modifies and integrates AI output, only those human modifications are protected — so the log is what separates a defensible claim from a vague one.

This matters more as AI adoption accelerates. An estimated 30% of Microsoft's production code is AI-generated, and Meta's CEO has projected roughly half of its codebase will be this year. As one AI risk strategist puts it, the market divide will be between companies that can prove what they built and those that cannot — in diligence, disputes, and enterprise sales. Keep records of tools used, what was entered into prompts, and whether outputs were scanned for licensed or open-source code.

At handoff, secure the full package. WIPO advises always requesting source code, documentation, and passwords as part of termination provisions, since access is what makes ownership usable in practice. Your end-of-contract provisions should also run in the other direction — protecting your data from the developer's systems. Per contract guidance from a Saul Ewing partner, agreements should require secure destruction of fine-tuned models and RAG databases upon termination, specify deletion methods such as cryptographic erasure or physical destruction of storage media, and demand verifiable certification that customer data and derivatives are irrevocably erased — including from backups.

• Source code, technical documentation, and all access credentials delivered at handoff
• A provenance log covering prompts, tools, human edits, and output-scanning decisions
• Secure deletion of fine-tuned models and RAG databases built from your data
• Verifiable certification that your data and derivatives are erased from all systems and backups

The same principle applies to how you engage a build partner. We structure our agent work at Agents by AIQ on a month-to-month basis where the client owns everything — the code, the configurations, the knowledge base — so a change in direction never means losing what you paid for. If you're scoping an AI receptionist, follow-up, or workflow agent and want the ownership terms settled before the first line is built, book a call and we'll walk through it.