How LLMs Are Disrupting Patent Strategy — Mark Cuban's Warning on Patents as AI Training Data

Overview

Mark Cuban made a thought-provoking observation on X (formerly Twitter): “When you publish a patent, it becomes training material for LLMs.” The patent system is built on a social contract — disclose your technology in exchange for a monopoly. But as LLMs consume published patent documents at scale, this fundamental premise is beginning to crumble.

This article analyzes how patent strategy must evolve in the LLM era, starting from Cuban’s insight.

Mark Cuban’s Core Argument

Cuban’s argument can be summarized as follows:

1. Patents are public documents: Filing with a patent office means detailed technical disclosure
2. LLMs learn from public data: Patent documents are included in training datasets
3. AI ends up “knowing” patented technology: The monopoly right exists, but the technical knowledge itself is absorbed by AI

This goes beyond simple patent infringement — it means the fundamental value exchange of the patent system is breaking down.

graph LR
    A[Company Files Patent] --> B[Patent Office Publishes Technology]
    B --> C[20-Year Monopoly Granted]
    B --> D[LLM Learns Patent Documents]
    D --> E[AI Absorbs Technical Knowledge]
    E --> F[Competitors Use AI to<br/>Develop Similar Technology]
    C -.->|Hollowed Out| F

Why the Patent System’s Premise Is Shaking

The Traditional Social Contract

The patent system has operated on the following premise for over 200 years:

Inventor Side	Society Side
Disclose technology in detail	Grant 20-year monopoly
Describe at an enabling level	Free practice after expiry
Contribute to technological progress	Provide foundation for follow-on invention

How LLMs Change the Rules

In the LLM era, this contract’s balance tips dramatically:

• Learning speed: AI learns in seconds what human engineers take years to read
• Abstraction ability: AI extracts core ideas from patents and applies them in modified forms
• Scale problem: Millions of patents learned simultaneously, discovering connections between technologies
• Legal gray zone: Unclear whether development based on AI-learned knowledge constitutes patent infringement

What’s Already Happening

This problem is already materializing in several areas:

1. Code generation AI: GitHub Copilot generating code similar to patented algorithms
2. Drug discovery AI: AI trained on published pharmaceutical patents designing similar compounds
3. Hardware design: AI trained on semiconductor patents assisting with circuit design

How Companies Should Rethink Patent Strategy

1. Revival of Trade Secret Strategy

Protecting through trade secrets instead of patents is gaining renewed attention.

Advantages:

• LLMs cannot learn it (it’s not public)
• No time limit (20 years vs. perpetual)
• No filing costs

Disadvantages:

• Vulnerable to reverse engineering
• Cannot prevent independent invention
• Risk of leakage through employee turnover

graph TD
    A{Choose Protection Strategy} -->|Can Disclose| B[File Patent]
    A -->|Core Know-How| C[Trade Secret]
    A -->|Hybrid| D[Core as Secret +<br/>Peripheral as Patent]
    B --> E[LLM Learning Risk]
    C --> F[Leakage Risk]
    D --> G[Balanced Protection]

2. Strengthening Defensive Patent Strategy

Using patents as defensive tools rather than offensive weapons:

• Patent pools: Industry-wide patent sharing for mutual deterrence
• Defensive publication: Publishing technology as prior art instead of patenting, preventing competitors from patenting
• Cross-licensing: Mutual technology exchange through reciprocal licenses

3. AI-Era Patent Drafting

Approaches to writing patents that are harder for LLMs to fully learn:

• Separate core know-how: Include only minimal information in patents, protect implementation details as trade secrets
• Control abstraction levels: Write broad claims but strategically craft specifications
• Multi-layer protection: Protect a single technology with a combination of patents and trade secrets

4. Legal Responses to Restrict AI Learning

Legal and policy-level responses are also needed:

• Legislation restricting AI training on patent data: Under discussion in some countries
• robots.txt-style patent protection: Adding learning-restriction metadata to patent databases
• Patentability of AI-generated inventions: Whether patents can be granted for AI-created inventions

Industry Impact Analysis

Industry	Impact Level	Key Risk	Recommended Strategy
Pharma/Biotech	Very High	Compound patents used for AI drug discovery	Trade secret + patent hybrid
Semiconductors	High	Circuit design patents used in AI-assisted design	Keep core processes as trade secrets
Software	Medium	Algorithm patents affecting code generation	Open source + service model pivot
Manufacturing	Medium	Structural patents used in CAD automation	Maintain manufacturing know-how secrecy

Conclusion

Mark Cuban’s observation isn’t mere concern — it’s a call for fundamental re-examination of the patent system. In an era where LLMs absorb all public knowledge, the 200-year-old social contract of “disclosure equals monopoly” may no longer serve its intended purpose.

Companies should immediately consider three things:

1. Assess LLM exposure of their current patent portfolio
2. Redesign the optimal mix of trade secrets and patents
3. Develop an IP strategy roadmap suited for the AI era

The paradigm of patent strategy is shifting. Only companies that adapt quickly will maintain their technological edge.

References

• Mark Cuban’s X Post — Original post on patent disclosure and LLM learning
• WIPO — AI and Intellectual Property — World Intellectual Property Organization’s AI policy discussions
• USPTO — AI-Related Patent Guidelines — U.S. Patent Office guidance on AI inventions