---
title: "AI Development IP Ownership and NDAs: Who Owns the Code, Model, and Data"
url: "https://www.spaceotechnologies.com/blog/ai-development-ip-ownership-and-nda/"
date: "2026-10-21T12:18:30+00:00"
modified: "2026-10-21T12:48:09+00:00"
type: "Article"
resource: "https://www.spaceotechnologies.com/blog/ai-development-ip-ownership-and-nda/"
timestamp: "2026-10-21T12:48:09+00:00"
author:
  name: "Bhaval Patel"
categories:
  - "Artificial intelligence"
word_count: 2828
reading_time: "15 min read"
summary: "Key Takeaways
An NDA keeps information confidential, but only a signed written assignment transfers IP ownership.
Under U.S. law, purely AI-generated work has no copyright owner without human autho..."
description: "An NDA keeps AI project data confidential; only a written assignment transfers IP. See who owns code, weights, and data by default, and the clauses to add."
keywords: "AI Development IP Ownership and NDAs, Artificial intelligence"
language: "en"
schema_type: "Article"
related_posts:
  - title: "AI Governance and Compliance Consulting"
    url: "https://www.spaceotechnologies.com/blog/ai-governance-and-compliance-consulting/"
  - title: "Secure AI Software Development Practices: A 2026 Checklist for AI-Assisted Coding and Production AI"
    url: "https://www.spaceotechnologies.com/blog/secure-ai-software-development-practices/"
  - title: "Data Privacy in AI Development Projects: Risks, Laws, and Controls"
    url: "https://www.spaceotechnologies.com/blog/data-privacy-in-ai-development-projects/"
---

# AI Development IP Ownership and NDAs: Who Owns the Code, Model, and Data

_Published: October 21, 2026_  
_Author: Bhaval Patel_  

![ai development ip ownership and ndas](https://www.spaceotechnologies.com/wp-content/uploads/2026/10/ai-development-ip-ownership-and-ndas-1024x541.jpg)

Key Takeaways

- An NDA keeps information confidential, but only a signed written assignment transfers IP ownership.
- Under U.S. law, purely AI-generated work has no copyright owner without human authorship.
- Contractors and development companies keep their work by default until you hold a written assignment.
- Check every layer, from code and weights to prompts and logs, before you sign.

AI development IP ownership and NDA terms decide who controls your product. An NDA keeps your code, data, and prompts confidential. Only a written IP assignment transfers ownership to you.

Purely AI-generated output lacks the human authorship U.S. copyright law requires. Developers and contractors, including any [AI development company](https://www.spaceotechnologies.com/ai-development-services/), also keep their work by default. So your contract, not the AI tool, decides who owns what.

This guide maps who owns the code, model, and data. It also shows where standard NDAs fall short on AI. Use it with your counsel before you sign anything. This page is general information about U.S. law, not legal advice. Have your own counsel review any agreement before you sign it.

## What this means when you hire an AI development partner

Any outside developer, whether a freelancer or a company, owns its work by default under U.S. law unless the contract assigns it to you. The risk is real for founders and CTOs. You can pay for a product in full and still not hold clear title to it.

Space-O Technologies has delivered client work since 2010 and runs 50+ AI systems in production. Here is how we structure AI engagements:

- **NDA-protected work:** Every engagement runs under an NDA, signed before you share code, data, or business plans.
- **Separate IP assignment:** A separate assignment transfers the code and IP we create for you. It never relies on the NDA to do that job.
- **No training on client data:** We do not train models on your data. PII masking is applied to sensitive data in AI workflows.
- **Honest scope on AI output:** We assign every right that exists in the deliverables. For AI-generated components that copyright may not cover, we assign contractual ownership and control instead.
- **Security posture:** Our [ISO 27001 certification](https://www.spaceotechnologies.com/company/awards-and-certification/) and ISO 9001 practices support confidentiality for regulated teams.

Our published terms set the handover rules. All source code, data, and IP transfer to you at project completion. Nothing is retained, reused, or shared after handover. Ownership terms are set in writing from day one. The same rules apply to the AI providers we use. PII is masked before it reaches any third-party model.

## 1. Who owns AI-generated work under U.S. law?

Under U.S. law, purely AI-generated work has no copyright owner, because copyright protects only human authorship. AI-assisted work is different. It can be protected where a person makes real creative contributions. Patents follow a separate rule, covered in the patent bullet below.

- **Human authorship requirement:** The U.S. Copyright Office registers only works with a human author. Material generated entirely by an AI system does not meet that standard.
- **Thaler v. Perlmutter:** A federal court in D.C. upheld the Copyright Office’s refusal to register an AI-authored work in 2023. The D.C. Circuit affirmed that ruling in 2025.
- **Meaningful human contribution:** AI-assisted output can be protected when a person substantially selects, arranges, or modifies it. Copyright then covers only the human-authored portions.
- **Prompts alone:** Typing a prompt is generally not enough to make someone the author. Protection comes from the creative choices a person makes in the final work.
- **Public domain risk:** Purely AI-generated material may fall into the public domain. Others may then be free to reuse it.
- **AI cannot be a patent inventor:** U.S. patents require a natural person as inventor. USPTO guidance from February 2024 requires a significant human contribution to the invention’s conception.
- **Records of human involvement:** Commit history, design documents, and code review logs show where people made the decisions. These records help defend later copyright or patent claims.
- **Trade secret status:** Code that copyright cannot protect may still qualify as a trade secret if it is kept confidential. Pasting it into a public AI tool can destroy that status.

## 2. Who owns the work: employees, contractors, or development partners?

Default ownership depends on the worker’s legal relationship to your company, not on who paid for the work. U.S. law treats employees and outside builders very differently.

![Default IP Ownership by Worker Type](https://www.spaceotechnologies.com/wp-content/uploads/2026/10/default-ownership-by-worker-type.webp)- **Employees:** Work created within the scope of employment is “work made for hire” under 17 U.S.C. § 101. The employer owns it automatically.
- **Contractors:** An independent contractor keeps the rights to their work by default. Ownership moves to you only through a signed written assignment.
- **Freelancers:** A freelancer who builds your model pipeline owns that code until a written assignment says otherwise. See [AI development company vs freelancer](https://www.spaceotechnologies.com/blog/ai-development-company-vs-freelancer/) for how the two compare.
- **Development companies:** An outside AI development company is a contractor in this framing. Its engineers’ work flows to the company, which must then assign it to you.
- **Augmented team members:** Contractors placed inside your sprints are still contractors. Using your tools and repositories does not change their default ownership.
- **AI tool vendors:** Paying for an AI tool subscription does not make you the owner of its outputs. The vendor’s terms of service set your rights to inputs and outputs.
- **Training data vs. model:** Owning the data used to train a model is legally distinct from owning the model. A contract must address each one separately.

## 3. Why doesn’t an NDA settle IP ownership?

An NDA restricts disclosure, but it does not create or transfer IP ownership. The two documents serve different legal functions. Treating one as the other is a common gap in AI projects.

- **Different jobs:** An NDA controls who can see and use confidential information. An IP assignment agreement decides who owns the resulting work.
- **Separate documents:** Keep the IP assignment separate from the NDA. Ownership terms then stay clear even if the NDA expires or is terminated.
- **Hidden IP grabs:** Some NDAs contain license or assignment language that favors the party who drafted them. Read every NDA for ownership terms before you sign.
- **Foreground IP:** Foreground IP is the new work created during your project. Your contract should assign it to you explicitly.
- **Background IP:** Background IP is what each party brought to the project, such as a developer’s reusable tooling. It usually stays with its owner under a license.
- **No default joint ownership:** Silence on ownership can leave both parties with unclear or shared claims. Allocate every deliverable in writing instead.

## 4. Where do standard NDAs fall short on AI?

Most NDAs were written before AI tools sat inside the development workflow, so they leave specific gaps. Each gap below concerns how an AI tool or its provider handles your information. [McKinsey’s State of AI research](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai) tracks how organizations manage AI-related risks.

![Eight Gaps in Standard AI NDAs](https://www.spaceotechnologies.com/wp-content/uploads/2026/10/eight-gaps-in-standard-ai-ndas.webp)- **“Third party” left undefined:** Traditional NDAs rarely say whether AI providers, cloud hosts, or subcontractors count as third parties. Your data may reach them without a technical breach of the NDA.
- **Prompts outside the definition:** Many confidential-information definitions list documents and code but not prompts or AI outputs. Prompts often contain your most sensitive business logic.
- **Training on your inputs:** AI vendors may retain inputs or use them to train models. Your proprietary code, data, and prompts could end up shaping someone else’s model.
- **Broad “business purpose” language:** Loose purpose wording may let a vendor argue that fine-tuning on your data is permitted. The permitted purpose must exclude model training.
- **Derived and de-identified data:** Vendors may treat aggregated, de-identified, or “derived” data as outside confidentiality. That can include the patterns learned from your records.
- **Return-or-destroy blind spots:** Standard deletion clauses cover files, not prompt logs, embeddings, model weights, or derived datasets. Those artifacts can outlive the contract.
- **“No AI” clauses:** Some NDAs now ban feeding confidential material into AI tools at all. Enforceability and day-to-day practicality both deserve a close read.
- **Zero data retention:** Some AI API providers offer zero-data-retention settings. Your NDA can require them for any tool that touches confidential data. Our guide to [AI agent security best practices](https://www.spaceotechnologies.com/blog/ai-agent-security-best-practices/) covers the wider controls.

## 5. AI deliverable ownership map: who owns each layer?

An AI product is a stack of layers, and each layer has its own default owner and protection. Hiring a developer does not automatically give you every layer. The map below shows what each layer needs.

![Who Owns Each Layer of Your AI](https://www.spaceotechnologies.com/wp-content/uploads/2026/10/who-owns-each-layer-of-your-ai.webp)

### Layer-by-layer ownership table

The table shows each layer’s default owner under U.S. law when no contract exists, and the clause that changes it.

| **Layer** | **Default owner with no contract** | **Strongest protection** | **Clause to require** | **Verify at handover and termination** |
|---|---|---|---|---|
| Developer-written code | The contractor or development company | Copyright | Present-tense assignment to you | Full repository and commit history transferred |
| AI-assisted code | Contractor for human-authored parts; purely AI parts may have no owner | Copyright for human parts; trade secret and contract for the rest | Assignment of all rights, plus contractual ownership of AI-generated parts | Records showing human review and edits |
| Third-party base model or API | The model provider | Third-party license | Disclosure of every model used and its license terms | Licenses held in your company’s name |
| Fine-tuned weights | Unclear; subject to the base model license | Trade secret and contract | Assignment of weights and an exclusivity commitment | Weights delivered; developer copies deleted |
| Training and fine-tuning datasets | Owner of the underlying data; compilation may belong to whoever built it | Trade secret and contract | Your data stays yours; built datasets assigned to you | Datasets returned; certified deletion of copies |
| Prompts and system instructions | Whoever wrote them | Trade secret and contract | Prompts defined as confidential information and assigned to you | Prompt library and versions delivered |
| RAG index and vector embeddings | Unclear; derived from your data | Contract only | Embeddings defined as confidential and derived from your data | Index exported; developer copies deleted |
| Generated outputs | Often no copyright owner; tool terms apply | Contract and trade secret | Output ownership with a statutory caveat | Tool terms reviewed for output rights |
| Developer’s background IP and tooling | The developer | Copyright and trade secret | Perpetual license to background IP inside your product | List of background components in the deliverable |
| Prompt and inference logs | Whoever runs the infrastructure or API | Contract only | Retention limits and deletion at termination | Written deletion certificate |

### What each layer means in practice

Each layer below is a separate ownership decision, so check every one before you sign.

- **Developer-written code:** Code a contractor writes belongs to the contractor until assigned. A present-tense assignment moves it to you as it is created.
- **AI-assisted code:** Human-edited portions can carry copyright, and pure AI portions may not. A contract gives you control over both.
- **Third-party base model:** You license a base model or API from its provider; you do not own it. Its license terms limit what you can do with derived work. Our guide to [prebuilt vs custom AI models](https://www.spaceotechnologies.com/blog/prebuilt-vs-custom-ai-model/) covers the trade-offs.
- **Fine-tuned weights:** Fine-tuned weights sit on top of a licensed base model. Contract and trade-secret practice are your main protection.
- **Training datasets:** Your raw data stays yours, but curated datasets need their own assignment. Owning data never implies owning the model trained on it. Gartner lists lineage and trusted sources among the requirements for [AI-ready data](https://www.gartner.com/en/articles/ai-ready-data), so record where each dataset came from.
- **Prompts and system instructions:** Prompts encode business logic and are rarely copyrightable alone. Treat them as trade secrets and name them in the NDA.
- **RAG index and embeddings:** Embeddings are mathematical representations derived from your documents. Without contract terms, they can survive a standard return-or-destroy clause.
- **Generated outputs:** Purely AI-generated outputs often carry no copyright. Your rights come from contract terms and the AI tool’s terms of service.
- **Background IP:** A developer’s reusable libraries usually stay with the developer. You need a license broad enough to run, change, and sell your product.
- **Prompt and inference logs:** Logs can hold sensitive inputs long after launch. Require retention limits and a deletion certificate at termination.

## 6. Which clauses should an AI development contract include?

A sound AI development contract closes each gap above with a specific, written clause. Use this checklist with your counsel when you review an MSA or statement of work.

![AI Development Contract Clause Checklist](https://www.spaceotechnologies.com/wp-content/uploads/2026/10/ai-contract-clause-checklist.webp)- **Training-use prohibition:** The developer and its AI providers may not train or fine-tune models on your code, data, or prompts. This also blocks “business purpose” workarounds.
- **Output ownership with a caveat:** You receive all rights in deliverables that the law recognizes. You also receive contractual ownership where copyright may not attach.
- **Expanded confidential-information definition:** Confidential information expressly includes prompts, embeddings, model weights, outputs, and derived data. It names AI providers and sub-processors as covered recipients.
- **Present-tense assignment:** Use “hereby assigns” language that transfers rights on creation. Do not rely on work-for-hire wording alone, because pure AI output may fail authorship tests.
- **Standalone assignment agreement:** The IP assignment lives outside the NDA, and the NDA contains no conflicting ownership terms.
- **Background license, foreground assignment:** Foreground IP is assigned to you, and background IP is licensed to you on a perpetual basis. Neither falls into default joint ownership.
- **Certified deletion at termination:** The developer deletes prompt logs, embeddings, RAG databases, and fine-tuned models it retains. It confirms this with a written certificate.
- **Sub-processor disclosure:** The developer lists every AI provider, cloud host, and subcontractor that touches your data. You can object to new ones.
- **Audit rights:** You can verify compliance with data-handling and deletion terms on reasonable notice.
- **Reverse-engineering and model-extraction ban:** No party may extract, copy, or reconstruct your model or its weights.
- **Infringement indemnification:** The developer covers infringement claims arising from its work, including from third-party models trained on scraped data.

### Get Ownership Terms Reviewed Before You Sign Anything

Our team structures every AI engagement with a separate IP assignment and a signed NDA. Share your project scope, and we will map which layers you need to own.

Scope My Project![Cta Image](/wp-content/uploads/2023/04/cta-img.png)

## Which protections matter most for your situation?

The right priorities depend on your role in the build. Use the section that matches yours.

### If you’re hiring an outside team to build an AI product

Your priority is a standalone, present-tense IP assignment covering every layer in the ownership map. Check the background IP license next, so you can maintain the product without the original team. Then confirm certified deletion for weights, embeddings, and logs.

### If you’re building on a third-party API

Your priority is controlling what the API provider keeps and learns from your traffic. Confirm the training opt-out in the provider’s terms, and turn on zero data retention where it is offered. Require a list of sub-processors behind the provider.

### If you’re adding contractors to your in-house team

Your priority is a signed assignment from every contractor before they commit code. Employee work-for-hire rules do not cover contractors, even inside your sprints. This applies regardless of the [engagement model](https://www.spaceotechnologies.com/company/engagement-models/): dedicated team, time and material, fixed cost, or staff augmentation. Compare an [AI development company vs. staff augmentation](https://www.spaceotechnologies.com/blog/ai-development-company-vs-staff-augmentation/) before you choose.

### If you’re a founder before incorporation

Your priority is assigning each founder’s pre-incorporation code and models to the new company. Investors usually check that the company, not the individual founders, holds title to the core IP. Founders own their own work until they assign it. Our [MVP development services](https://www.spaceotechnologies.com/mvp-development-services/) transfer all source code and IP to you at completion.

## Frequently Asked Questions

### Why do investors check who holds title to a startup’s AI code and models?

Investors check title because ownership gaps put the product they fund at risk. The company, not individual founders or contractors, should hold title to the core IP. Contractors keep their work by default unless a written assignment transfers it. Missing assignments show up as gaps during due diligence.

### What can an NDA realistically control when an AI tool touches my data?

An NDA controls who sees and uses your information, but it cannot create or transfer ownership. Drafted for AI, it can name AI providers and sub-processors as covered recipients. It can list prompts, outputs, embeddings, model weights, and derived data as confidential. It can also exclude model training and require zero data retention.

### Do I really need an NDA for every contractor?

Yes, sign an NDA before you share code, data, or business plans with any outside developer. The NDA only protects confidentiality. Ownership of what the contractor builds moves to you only through a separate written assignment. Have it signed before they commit code.

### Who owns code written by AI tools under U.S. law?

Purely AI-generated code has no copyright owner in the U.S., because copyright requires a human author. Human-edited portions can be protected. Your contract and the AI tool’s terms then set your rights. Keep commit history and review logs as proof of human input.

### Should the IP assignment sit inside the NDA or in a separate agreement?

Keep the IP assignment in a separate agreement from the NDA. Ownership terms then survive if the NDA expires or ends. It also avoids conflicting terms between the two documents. Use present-tense “hereby assigns” wording.

### What should happen to prompts, embeddings, and logs when a project ends?

The developer should delete them and confirm it in a written certificate. Standard return-or-destroy clauses cover files, not prompt logs, embeddings, or model weights. Name each artifact in the contract. Set retention limits before launch.


---

_View the original post at: [https://www.spaceotechnologies.com/blog/ai-development-ip-ownership-and-nda/](https://www.spaceotechnologies.com/blog/ai-development-ip-ownership-and-nda/)_  
_Served as markdown by [Third Audience](https://github.com/third-audience) v3.6.1.1_  
_Generated: 2026-10-21 12:48:11 UTC_