Tech AI News - 테카이

13. Can you put AI-written code straight into your company's product?

· 6 min read

This post was translated from the Korean original by AI.한국어 원문 읽기 →

What happens if you copy and commit a Copilot suggestion as-is, only to discover it's nearly identical — down to the variable names — to open-source code under the GPL license? From that moment, the problem is no longer about code quality; it moves into the territory of copyright law and license compliance.


Background — why this question matters now

AI coding tools like GitHub Copilot and Cursor are already part of the default workflow for many development teams. The problem is that the practice of committing "AI-written code" without review has spread alongside them. It's easy to assume that if the code works, quality has been verified — but because AI coding tools learn from public repositories and work by "predicting" the next code to come, there's a structural possibility that a prediction turns out nearly identical to a specific piece of code in the training data.

That this isn't just an abstract worry is confirmed by an actual lawsuit. In a class action brought by developers against GitHub, Microsoft, and OpenAI (Doe v. GitHub), the open-source license violation claims partially survived at the district court level in 2023, and an appellate court ruling followed in September 2026.

Key data & current state

IssueStatusSource
Code recurrence rateAccording to GitHub's own documentation, suggestions matching public code account for roughly 1% of all suggestions, concentrated in low-context situations like new projects and new filesGitHub official documentation
Litigation progressIn the Doe v. GitHub lawsuit, the Ninth Circuit Court of Appeals finalized the dismissal of the copyright-notice-removal (DMCA) claims on September 16, 2026, but the open-source license violation claims are still being heard separately at the district courtLeagle ruling, AFS Law analysis
Copyright attributionGitHub's terms state that users own the Copilot suggestions they receive, but the U.S. Copyright Office maintains that purely AI-generated output with little human creative contribution is not eligible for copyright protectionterms.law community legal analysis
Scope of vendor liabilityIn its Copyright Commitment announced in September 2023, Microsoft said it would take on legal liability for infringement lawsuits, but only for paid Copilot customers who have guardrails and content filters enabledFOSSA blog

The core point of this piece, though, isn't "what percentage of code recurs." What matters more than the fact that matches occasionally happen is the structure of how liability spreads once even a single match makes its way into a product.

In-depth analysis

1) How code can come out "nearly identical"

AI coding tools are trained on vast public repositories, which mix permissively licensed code like MIT alongside code with conditions attached, such as GPL and LGPL. In 2022, it was reported that Tim Davis, a professor at Arizona State University, disclosed that Copilot suggested his LGPL-licensed sparse matrix code verbatim — down to variable names, helper functions, and comments (ReversingLabs blog, October 2022). GitHub itself acknowledges in its official documentation that such matches can occur, explaining that they appear especially often when there's little existing code to serve as context, such as when writing a new project from scratch.

2) What "license contamination" actually means

GPL-family licenses carry "copyleft." That means any work built using this code (a derivative work) is also required to carry the same license. If a snippet of GPL code reproduced by AI ends up mixed into a commercial product, that could, in theory, create an obligation to release the entire product — or at least that module — under the GPL. The Free Software Foundation (FSF) already declared Copilot "unacceptable and unjust" back in 2021, raising unresolved legal questions about whether training on public code counts as fair use, and whether reproduced code infringes the original author's rights (InfoWorld report). There are also reports that code reproduced by Copilot often loses the original license and copyright notices along the way, making it easy for a user to pass it along without even realizing they picked up conditional code.

3) How far has the litigation gotten?

In the Doe v. GitHub lawsuit, the presiding judge in 2023 let both the copyright-notice-removal (DMCA) claims and the open-source license violation claims partially proceed. On September 16, 2026, the Ninth Circuit Court of Appeals finalized the district court's dismissal of the DMCA-related claims, but that ruling concerned the allegation that copyright notices were stripped from reproduced code — the claims addressing the open-source license violation itself appear to still be proceeding separately at the district court. In other words, the core question — "can using AI-written code as-is create license problems?" — still hasn't been resolved by the courts. While the law catches up, the risk of that gap falls on the development teams actually committing the code.

4) The practical test — when it's risky and when it's fine

  • Getting a short piece of helper code within a context-rich existing codebase carries relatively low matching risk. GitHub itself states that matches rarely appear when existing application code is provided as context.
  • Generating a new project from scratch, or requesting code that has "only one right answer" — like sorting functions, math routines, or well-known algorithms — raises the relative likelihood of overlapping with specific public code.
  • Putting that code into a product you commercially distribute or sell carries a different weight of risk than a personal project. If a license obligation is triggered, it becomes a question of the disclosure scope for the entire product.
  • Whether you've turned on GitHub Copilot's "Suggestions matching public code" setting to Block is also a key fork in the road. This setting compares a suggestion and roughly 150 surrounding characters against public code, and suppresses the suggestion itself on a match or near-match. If it's left off (Allow), this first line of defense doesn't exist, and Microsoft's Copyright Commitment only applies when guardrails are enabled.

In practice — four principles to check before putting AI-generated code into a product

  1. Don't leave duplicate-detection filters off. Setting Copilot's "Suggestions matching public code" to Block keeps suspicious suggestions from ever appearing on screen — a blunt but reliable first line of defense.
  2. Run a license scanner right after generation. Scan AI-generated code the same way you'd scan an external library, using a software composition analysis (SCA) tool like FOSSA or Snyk, to check whether a conditional license is mixed in.
  3. The less context a request has, the more carefully a human should review it. If you generated a new project from scratch or asked for a well-known algorithm wholesale, search for similar code once more before using the result as-is.
  4. Mark the sections AI generated. Tagging AI-generated sections in commit messages or code comments lets you swap out only the problematic piece later, rather than rewriting everything.

References

  • #ai
  • #artificial intelligence
  • #ai coding
  • #github copilot
  • #open source license
  • #copyright
  • #software license
How It Works

Transformer — The Ancestor of Every AI Today

From GPT to Gemini to Claude, every major AI model in 2026 runs on the Transformer architecture Google published in 2017. This article covers how self-attention works and the line of variations from decoder-only models to MoE and Mamba hybrids.

· 5 min