Who Owns What AI Built?
Updated: 6 hours ago
AI is the only co‑founder that works nights, weekends, and public holidays — and still refuses to sign an IP assignment.
It drafts your product copy. It proposes UX. It writes code. It suggests claims for a patent. It even “helps” with a contract.
And then, at the worst possible moment — fundraising, an acquisition, a dispute, or a takedown — someone asks the question that always sounds simple… until it isn’t:
Who owns what you built with AI?
Because “we used AI” is not a legal strategy.
And in 2026, it’s no longer a cute footnote — it’s a due diligence category.
Below is the landscape I’m seeing most often: where ownership actually lives, where risks hide, and how to turn AI output into something a serious investor (or buyer) can trust.
1) Ownership has three layers — and teams keep mixing them up
When people say “we own the output,” they often mean three different things at once:
Layer A: Contractual rights (what the AI provider’s terms give you)
Some providers assign rights in output to the user. Others say they won’t claim ownership, but also tell you outputs may be similar for everyone.
That’s not academic. In a dispute, your starting point is often: what did you agree to when you clicked “Accept”?
How we handle this in practice: AI Terms & Output Ownership Review (tool-by-tool), mapping:
what you own (if anything),
what you can commercialise,
whether the provider can use your inputs for training,
whether you have indemnity, and under what conditions.
Layer B: IP law (whether copyright/patent protection exists at all)
Even if a provider “assigns” output to you, copyright may not attach if the law says there’s no human authorship.
Meaning: you can “own” something in a contractual sense — and still have nothing enforceable against copycats.
How we handle this in practice: We help teams design copyright‑resilient workflows: documentation, human contribution, versioning, and “chain‑of‑creation” evidence that stands up when a counterparty asks uncomfortable questions.
Layer C: Chain of title (does your company own it, or an employee/contractor does?)
Even if the output is protectable, the next question is:
Was it created by an employee in scope of employment?
By a contractor without an assignment?
By a founder “on their own” before incorporation?
By an agency using AI tools under their own accounts?
How we handle this in practice: We structure clean IP assignment packages (employees, contractors, founders) and fix gaps before they become negotiation leverage.
2) Copyright: the law still expects a human to show up
Here’s the blunt truth: in many jurisdictions, purely AI‑generated output can be legally “ownerless” in the copyright sense.
Two US anchors that investors still cite (even outside the US) because they’re clear and early:
Thaler v. Perlmutter (U.S. District Court, D.D.C., No. 1:22‑cv‑01564, Aug. 18, 2023) — the court held that US copyright law requires human authorship.
U.S. Copyright Office policy guidance (88 Fed. Reg. 16190, Mar. 16, 2023) — works with AI‑generated material can be registered only to the extent there is protectable human authorship, and AI material must be disclosed/disclaimed.
And a practical example that became a reference point:
Zarya of the Dawn (U.S. Copyright Office; registration VAu001480196; Feb. 21, 2023 letter) — a comic where the text/selection/arrangement could be protected, but the Midjourney‑generated images themselves were not treated as human-authored.
What this means for founders & product teams
If your “AI co‑founder” writes:
your marketing copy,
your brand language,
your UI text,
your illustrations,
your documentation,
your product photography,
your “icon pack”…
…you may have output that is commercially valuable, but not automatically copyright‑enforceable the way people assume.
How we handle this in practice: We build AI‑Assisted Creation Playbooks: practical rules that keep you on the right side of human authorship and create evidence of contribution (without killing speed).
3) “But the terms say I own it” — read that sentence again
Most mainstream AI tools now include some form of: you own the output / we won’t claim ownership.
Good. But there are two catch‑phrases that matter just as much:
Catch #1: “Outputs may not be unique”
Meaning: you might “own” your output — and someone else may receive something extremely similar, lawfully, from the same model.
Example: Google’s Gemini API Additional Terms say Google won’t claim ownership, but may generate the same or similar content for others.
This is where teams confuse ownership with exclusivity.
Catch #2: “Free tiers can be training tiers”
A lot of teams use a free web tool for a serious commercial project — and don’t realise the default is: your inputs and outputs can be used to improve the service, sometimes with human review.
Google is unusually explicit here in its Gemini API terms: unpaid services may use submitted content and generated responses to improve products, and warns not to submit confidential information.
Anthropic has also publicly described how consumer settings can affect model training usage, and that commercial products are treated differently.
How we handle this in practice: We do AI Toolchain Risk Mapping, including:
which tools are used,
on what plan (free / paid / enterprise),
whether training is enabled,
what categories of data are shared,
whether a DPA exists,
and how that maps to NDAs, GDPR, trade secret status, and deal warranties.
4) “Training data” litigation: your output can become someone else’s lawsuit
A growing part of the risk isn’t what your team does — it’s what the model was trained on, and what it might reproduce.
Key cases worth knowing (and preparing for, even if you’ll never be in a courtroom):
Andersen v. Stability AI Ltd. (N.D. Cal., Case No. 3:23‑cv‑00201‑WHO, Order signed Aug. 12, 2024) — artists alleged infringement related to training and outputs; early‑stage rulings show courts are taking some theories seriously, while trimming others.
The New York Times Co. v. Microsoft Corp. & OpenAI (S.D.N.Y., Case No. 1:23‑cv‑11195, filed Dec. 27, 2023) — high‑profile claims around training and output similarity; still developing.
Getty Images (US), Inc. v. Stability AI, Inc. (D. Del., Case No. 1:23‑cv‑00135, filed 2023) — among other things, Getty alleges copying of protected images/metadata and points to outputs that appeared to contain Getty‑style watermarks.
Whether these plaintiffs ultimately win or lose isn’t the point for a founder.
The point is this:
If your product ships AI‑generated content at scale — you need a liability position.
Not vibes. A position.
How we handle this in practice: We help product teams and boards build an AI Output Risk Framework, including:
“allowed / restricted” use cases,
vendor selection and contract leverage (including indemnity),
customer‑facing terms (who owns what, and what you disclaim),
internal review processes for high‑risk outputs (e.g., marketing visuals, celebrity voice, brand‑sensitive content).
5) Code is the quietest place to lose the most money
If you’re using AI for code (Copilot, ChatGPT, Claude, code‑gen tools), your biggest risk isn’t “bad code”.
It’s unclear provenance.
There’s a reason the GitHub Copilot litigation exists:
Doe 1 v. GitHub, Inc. (N.D. Cal., Case No. 4:22‑cv‑06823‑JST, Order on motions to dismiss dated Jan. 3, 2024 / publicly filed Jan. 22, 2024) — plaintiffs alleged, among other things, open‑source licensing and attribution issues.
Now, here’s where founders get blindsided:
Even if your AI provider assigns outputs, the provider may still warn that generated code can implicate third‑party rights — including open‑source licenses.
How we handle this in practice: We run Open‑Source & AI‑Assisted Code Scans and produce a due‑diligence‑ready memo:
what tools were used,
what settings were enabled,
what the risk profile is,
and what remediation looks like (including attribution, replacements, or refactors where necessary).
6) Patents & inventions: AI can accelerate R&D — but it can’t be the inventor
AI can help you:
explore design spaces,
generate hypotheses,
propose structures,
optimise configurations,
draft claims.
But in patent law, naming the inventor is not a branding exercise.
A few landmark decisions make the direction clear:
EPO Legal Board of Appeal — J 8/20 and J 9/20 (Dec. 21, 2021): an inventor must be a natural person, so an AI system cannot be named as inventor.
Thaler v. Comptroller‑General of Patents — UK Supreme Court [2023] UKSC 49 (Dec. 20, 2023): same conclusion in the UK.
Thaler v. Vidal — U.S. Court of Appeals for the Federal Circuit, 43 F.4th 1207 (Fed. Cir. 2022): same in the US.
So yes: AI can be a tool in invention.
But you still need human inventor(s), with evidence of contribution.
How we handle this in practice: We do AI‑Assisted Invention Harvesting: capturing what humans actually contributed (the part that matters), and building documentation that survives scrutiny.
7) Trade secrets: the easiest IP to destroy with one paste
Trade secrets don’t die in court.
They die in Slack threads, “quick prompts”, and convenience.
If your team pastes:
customer data,
pricing logic,
product roadmaps,
proprietary datasets,
unreleased designs,
confidential code…
…into the wrong tool under the wrong plan, you can end up with:
NDA breaches,
loss of “secrecy” status,
GDPR exposure,
and a due diligence headache you don’t want.
This is where “AI as a co‑founder” becomes “AI as an uncontrolled subcontractor”.
How we handle this in practice: We build AI Usage Policies + Contract Controls, including:
what can and can’t be used in prompts,
which tools are approved,
training settings and retention,
enterprise configurations,
and customer/partner clauses to keep confidentiality intact.
8) Nano Banana, Claude, Open Source: a practical reading of the fine print
Let’s translate the licensing world into founder language.
Claude (Anthropic) — commercial terms in one sentence
In Anthropic’s Commercial Terms (for business use), the structure is clear: you own your inputs and outputs, and Anthropic states it won’t use customer content to train/improve the products (subject to the agreement).
That’s exactly the kind of paper a serious investor likes to see: not because it eliminates all risk, but because it shows discipline.
How we handle this in practice: we align the commercial terms with your actual workflow (who has access, what gets uploaded, what your team assumes vs what the contract says).
Nano Banana / Gemini (Google) — ownership, similarity, and data settings
Google’s Gemini API Additional Terms are explicit on two points that matter for founders:
Google won’t claim ownership of generated content, but may generate similar outputs for others.
For unpaid usage, submitted content and generated responses may be used to improve products, with a warning not to submit sensitive/confidential information (and special rules for EEA/UK/CH contexts).
If you build on Gemini‑powered generation, your legal posture depends on:
whether you’re on unpaid vs paid routes,
whether you’re in an enterprise configuration,
and what you promise your users.
How we handle this in practice: we build your AI vendor compliance pack (tool terms + product terms + internal policy), so you’re not explaining it mid‑deal.
Open source models: “open” is not “anything goes”
Open‑weight AI is powerful — and dangerous for teams that treat it like public domain.
For example:
Meta’s Llama 3 community license includes obligations and restrictions (including attribution/notice requirements and limitations that matter at scale).
Some image model ecosystems use “OpenRAIL”‑style licenses (designed to permit broad use while restricting harmful use).
Others are genuinely permissive (Apache 2.0‑style), but you still need to track what’s inside your distribution.
How we handle this in practice: we run Open‑Source + Model License Compliance Reviews, so your product doesn’t inherit obligations you didn’t budget for — or can’t live with later.
9) The “investment‑grade” way to build with AI
If you want AI in your product and peace in your data room, aim for three outcomes:
A documented toolchain
A defensible rights story
A clean chain of title
Because when the money gets serious, the questions get boring — and boring questions decide outcomes.
LAWWISE IP Scan: the fastest way to de‑risk “AI as a co‑founder”
If you’re building with AI (internally or in your product), we offer an IP Scan designed for exactly this reality — a structured review that covers, in one integrated picture:
AI toolchain & terms (Claude / Gemini / Nano Banana‑type tools / OpenAI / others),
training and retention settings, confidentiality exposure,
open‑source and AI‑assisted code risk,
IP ownership & chain of title (founders, employees, contractors, agencies),
patent/inventorship readiness for AI‑assisted R&D,
and a due‑diligence‑ready action plan you can actually execute.
If you’re raising, scaling, licensing, or simply tired of guessing where the risk sits — this is the most efficient place to start.
Closing thought
AI is not your co‑founder in law.
It’s your co‑author in practice.
And that gap — between practice and law — is where companies either build real value… or accidentally donate it.
If you want to build with AI and sleep at night, let’s talk.
Comments