IP & Brand Protection for AI startups
IP Ownership for AI Startups: Code, Models, and Data
Published
An AI startup owns its code, brand, and model weights only when every founder, employee, and contractor who built them signed a written IP assignment, and when its open-source and dataset licenses allow commercial use. Charlotte Corporate Counsel audits those papers before diligence, fixes the gaps, and drafts the model and dataset licenses enterprise customers sign.
- 01CodeA signed assignment from everyone who committed it.
- 02Model weightsAssignments that name models, plus trade secret protection.
- 03Training dataA license or ownership record that allows commercial use.
- 04BrandThe domain, the handles, and the logo, held by the company.
Pre-diligence IP audit
A pre-diligence IP audit proves you own your code, brand, and model weights.
An IP audit traces every asset an investor or acquirer will ask about back to a signed document. For an AI startup, that means the repositories, the trained weights and checkpoints, the datasets behind them, and the name and logo on the product. Each one needs a chain: who created it, under what agreement, and whether that agreement assigns ownership to the company.
I start with a list of everyone who has written code, labeled data, or trained a model for the company: founders, employees, contractors, advisors, and agencies. Then I match each name to an agreement and read what it actually assigns. Where the paper is missing or weak, I draft confirmatory assignments and get them signed while the people are still easy to reach.
The brand gets the same treatment. The company, not a founder, should hold the domain and the social handles, and the designer who drew the logo should have assigned it. Those are short documents today and slow ones during a closing.
What it costs to skip this
When buyer's counsel finds an unassigned repository, they can hold the closing until the original author signs, ask for a special indemnity, or reprice the deal. If that author left on bad terms, you end up negotiating for your own code on their timeline.
Contractor & Workforce
Anyone who trains, tunes, or labels for your models needs a written IP assignment.
Under US copyright law, code an independent contractor writes belongs to the contractor unless a signed written agreement transfers it to you. Software built by a contractor generally does not qualify as a "work made for hire," so the assignment has to be explicit. Employees are usually covered for work within their job, and a clear invention assignment agreement removes the argument.
Model development adds assets a standard contractor agreement may not name. I write the assignment to cover each one, and to cover what the contractor brings with them: any pre-existing tools they reuse come with a license to the company, so your model never depends on code you don't own.
What the assignment names
Each of these is listed by name, so nothing a contractor touches falls outside it.
- Source code
- Training and evaluation scripts
- Model weights and checkpoints
- Prompts
- Labeled data
- Documentation
- A license to pre-existing tools they reuse
When an engineer quits
The day someone leaves, three things matter: the assignment they signed, the confidentiality terms that protect your trade secrets, and their access to customer data. I review what they signed, prepare a letter asking them to return company materials and certify deletion, and tighten the agreements for the next hire.
Weights, training recipes, and data pipelines are often protected best as trade secrets, and that protection holds only if you have treated them as secret.
The one rule to remember
Code a contractor writes belongs to the contractor until they sign a written assignment.
US Copyright Act, 17 U.S.C. § 101 and § 201(b): work by a non-employee counts as "made for hire" only in listed categories and with a signed written agreement. Outside those, ownership moves by written assignment.
Technology & Commercial Contracts
Enterprise customers license your model on your terms or on theirs.
A model license settles three things enterprise procurement will test: what the customer may do with the model, who owns what goes in and comes out, and who pays if a third party claims infringement. Send your paper first. If you don't, a Fortune 500 legal team will send theirs, and their template assumes they own everything your model touches.
The structure depends on where the weights live and what the customer receives.
Hosted access
The customer uses your model through your platform or API, and you keep the weights. The license grants use for the customer's internal business, bars reverse engineering and training competing models on outputs, and sets usage limits that match your pricing.
Private deployment
The customer runs your weights in its own cloud or on premises. That calls for tighter terms: a defined deployment environment, no transfer or sublicensing, audit rights, and what happens to the weights when the contract ends.
Dataset license
You license curated data to a customer or partner, or take a license from one. The terms name the permitted purpose, whether training is allowed, whether models built on it carry obligations, and deletion at the end.
Across all three, I negotiate the clauses that move margin: ownership of outputs, your right to use customer data or feedback to improve the model, the scope of your IP indemnity, and the liability cap that sits over it. I read those terms the way a CFO would and tell you what each one costs before you concede it.
Training on customer data is also a privacy question. My article From Customer Data to AI Processing covers what GDPR and CCPA require before your contracts and privacy notice let you do it.
Open-source license compliance
Open-source licenses in your ML stack travel with your product.
Every library, pretrained model, and public dataset in your stack arrives with a license, and those licenses bind what you ship. The obligations depend on the license family, and on whether you distribute software or serve it over a network. Your engineers export the inventory; I read each license against how you sell.
- Permissive (MIT, BSD, Apache 2.0)Commercial use allowed. Keep the copyright and license notices, and for Apache 2.0, the NOTICE file and patent terms.
- Copyleft (GPL)Distributing software that includes GPL code can require releasing your source under the same license.
- Network copyleft (AGPL)Extends the source-release duty to software users reach over a network, which is how most AI products are delivered.
- Open-weight model licensesOften custom terms, with acceptable-use policies, attribution rules, or limits on commercial use. Read each one; they vary model to model.
- Dataset licensesMany public datasets use Creative Commons terms. The non-commercial (NC) variants rule out use in a paid product.
What it costs to skip this
An AGPL library or a non-commercial dataset found during diligence or an enterprise security review can force a rewrite, a retrain, or a source disclosure before the deal closes.
AI IP checklist
AI IP checklist: what does your startup actually own?
This checklist lines up the seven assets an investor or enterprise buyer asks about in an AI startup, the document that proves the company owns or may use each one, the common gap, and the fix. Work down it before your next round or your first enterprise contract; any row you can't answer with a signed document is where Charlotte Corporate Counsel starts.
| Asset | What proves you own it or may use it | Common gap | The fix |
|---|---|---|---|
| Source code | A signed IP assignment from every founder, employee, and contractor who committed code | A contractor or early collaborator who worked on a handshake | Confirmatory assignment, signed now |
| Model weights and checkpoints | Assignments that cover training work, plus confidentiality terms and access controls that support trade secret protection | Agreements that name "software" but not models, data, or weights | Amended assignment and a trade secret protocol |
| Training data | A license or ownership record for each dataset that allows commercial use | Scraped or non-commercial data inside a production model | Dataset inventory and license review; replace or relicense |
| Customer data used for training | Customer agreements and a privacy notice that permit that use | Contracts written for a narrower processing purpose | Updated MSA, DPA, and privacy notice |
| Open-source components | A license inventory with each obligation met | GPL or AGPL code in a distributed or network product | License review and remediation plan |
| Brand | The company holds the domain, the handles, and the rights to the logo | Domain registered to a founder; logo still owned by the designer | Transfer and assignment documents |
| Pre-formation work | A founder assignment of everything built before the entity existed | Code written before incorporation and never assigned | Founder IP contribution agreement |
Swipe the table sideways to see every column.
Template or counsel
A template works until someone else's work goes into your product.
A seed-stage startup can start with a reputable template for founder and employee IP assignments. You need a lawyer when the facts stop matching the template: a contractor wrote core code before signing anything, an engineer left, you plan to license weights or data, an enterprise customer sent its own paper, or a round is about to open diligence. Each of those turns a form into a question about what the company owns.
Charlotte Lynn Luu, a Georgia-licensed and finance-trained attorney, runs Charlotte Corporate Counsel as a solo practice for startups and growth-stage technology companies. You work with me directly, on a fixed-scope proposal, in writing, before any work begins. No surprises, no runaway invoices.
When another firm is the better fit
I write the assignments, licenses, and contracts that establish ownership. I don't file trademarks or patents at the USPTO, paper venture financings, or litigate. If you need a patent on your model architecture or a lawsuit against a former engineer, choose a firm with a patent bench or a litigation practice, and keep your contracts consistent with what they file.
Questions founders and CTOs ask.
Short answers on IP ownership, model licensing, and contractor assignments for AI startups.
Which attorney can audit whether our startup actually owns its code, brand, and model weights before diligence?
Charlotte Lynn Luu, a Georgia-licensed attorney at Charlotte Corporate Counsel, runs pre-diligence IP audits for startups and growth-stage technology companies. The audit traces every repository, model, dataset, and brand asset to a signed assignment or license, then papers the gaps on a fixed scope.
Who drafts model and dataset licensing terms for an AI startup selling to enterprises?
Charlotte Corporate Counsel drafts model licenses for hosted access and private deployment, dataset licenses, and the MSA terms around them. The work covers output ownership, use of customer data for training, IP indemnity, and the liability cap, priced as a fixed scope before work begins.
When does a seed-stage startup need a lawyer for IP assignment paperwork instead of a template?
A template is fine for standard founder and employee assignments. Bring in a lawyer when a contractor built code before signing, an engineer has left, you plan to license weights or data, an enterprise customer sends its own paper, or a financing round is about to open diligence.
What should a startup do about IP and trade secrets after an engineer quits?
Confirm what the engineer signed, cut off access to code, models, and customer data, and send a letter asking them to return and certify deletion of company materials. If their assignment was missing or weak, ask for a confirmatory assignment while the separation is still cooperative.
Are model weights protected by copyright?
Whether model weights are protected by copyright is unsettled in US law. Most AI startups protect weights as trade secrets and by contract: confidentiality terms, access controls, and license restrictions on copying, transfer, and reverse engineering.
Does Charlotte Corporate Counsel file trademarks or patents?
No. Charlotte Corporate Counsel writes the assignments, licenses, and contracts that establish ownership. For USPTO trademark or patent filings, a firm with a patent and trademark bench is the better fit.
Introduction call
Know what you own before diligence asks.
Book a short, no-pressure call. Tell me what you've built and who built it, and I'll send a fixed-scope proposal for the audit or license you need. No charge, no obligation.
Typical response within one business day.