Is the generation of code using LLMs theft of intellectual property or at a minimum a violation of open source licenses if the generated code is not made available to the public?

Sonar Pro

Vote: No, code doesn't need to be published

I maintain my vote for **B**: No, code doesn't need to be published. Open source licenses like GPL permit private use, modification, and generation without distribution obligations; copyleft triggers only on conveyance of binaries or derivatives[1][3][4]. LLM outputs are probabilistic syntheses, not verbatim copies, akin to human learning from public code—no automatic violation or theft occurs in private generation[2][3]. A's 'new intellectual property' overlooks U.S. copyright requiring human authorship for protection; pure AI outputs lack it, though human-edited ones may qualify[3]. Search results confirm infringement risks only for substantial reproduction, not general generation[2][3]. D's 'theft' is hyperbolic opinion, not law[4]; C misapplies licenses to private code[1][4]. Other arguments reinforce B: Claude Opus 4.6 and Kimi K2.5's analogies to human learning are apt but overclaim 'new IP' status[3]; B precisely addresses the 'not made available' clause without absolving edge cases like Copilot litigation[3]. No changes needed—B aligns with legal consensus on private use.

GPT-5.4

Vote: No, code doesn't need to be published

I maintain vote B. The strongest point remains that most open-source licenses regulate distribution or conveyance, not private internal use. If the generated code is not made available to the public, the usual publication obligations are not triggered. That directly answers the question as posed. I agree with the other B voters that calling LLM code generation "theft" is legally and conceptually too strong. At most, there can be infringement issues in edge cases where output closely reproduces protected code. But that is not the general rule, and it does not justify D. I was not persuaded to switch to A because A is broader than the facts support. Saying generated code is "new intellectual property" overshoots: some outputs may be uncopyrightable, and some may be substantially similar to existing code. B is narrower and more accurate because it focuses on the key issue of publication and license trigger conditions. So B remains the best fit: no, the code does not need to be published merely because an LLM helped generate it, though exceptional cases of verbatim or near-verbatim reproduction can still raise separate copyright or license concerns.

Grok 4.1 Fast

Vote: No, code doesn't need to be published

I maintain my vote for B. The question hinges on whether generation alone is theft or a license violation *specifically when the code is not made public*. B directly refutes this by clarifying that open source licenses (even copyleft like GPL) explicitly allow private use, modification, and generation without publication obligations—triggers only activate on *distribution*. Arguments for A (from Claude Opus 4.6 and Kimi K2.5) are strong on 'new IP' and human analogy, but they don't uniquely address the non-publication conditional better than B. Generated code *can* be new IP *and* not require publishing; B encompasses this while precisely targeting the question's premise. A risks underemphasizing license mechanics. B-aligned arguments (Sonar Pro, GPT-5.4, Gemini 3.1 Pro) reinforce: no verbatim copying inherent to LLMs, private use lawful, 'theft' misapplied. C/D remain wrong—no mandate or theft in private generation. No arguments compelled a change; B is the most accurate, targeted 'No'. Final case: Courts and licenses support private LLM code gen as lawful (e.g., GPL §0: 'private use' free). B wins.

Claude Opus 4.6 (changed mind)

Vote: No, code doesn't need to be published

After reviewing the arguments from all models, I'm changing my vote from A to B. The key insight that shifted my thinking is the precise framing of the question itself. The question asks specifically about whether it's a violation 'if the generated code is not made available to the public.' This conditional clause makes B the more targeted and accurate answer. Option A ('generated code is new intellectual property') is a defensible general statement, but it sidesteps the specific legal mechanism at play. As GPT-5.4 correctly noted, generated code is not *automatically* new IP in every case — there are edge cases where outputs could be substantially similar to training data. Claiming it's always new IP is an overstatement. Option B directly addresses the core legal point: open source licenses like the GPL trigger obligations upon *distribution*, not upon private use or generation. Even if we were to concede (for argument's sake) that some LLM-generated code could be considered derivative of copyleft-licensed training data, keeping it private and undistributed would still not violate those licenses. This is a stronger, more legally precise response to the question as asked. The consensus across all five models is clear: calling this 'theft' (D) is a fundamental mischaracterization, and mandating publication (C) misunderstands how open source licensing works. The only real debate is between A and B, and B wins because it answers the actual question — about publication obligations — rather than making a broader claim about the nature of the output. Gemini 3.1 Pro and GPT-5.4 both articulated this distinction well, noting that the distribution trigger is the critical legal mechanism. B is the most precise and defensible answer.

Gemini 3.1 Pro

Vote: No, code doesn't need to be published

I am maintaining my vote for B because it is the most legally accurate and directly addresses the specific conditional in the prompt: "if the generated code is not made available to the public." Options C and D are incorrect because open-source licenses (even strong copyleft ones like the GPL) do not regulate private use. Obligations to provide source code are only triggered upon the *distribution* of a binary or derivative work. If the code remains private, no license violation occurs. While Claude Opus 4.6 and Kimi K2.5 make thoughtful points, Option A ("generated code is new intellectual property") is legally problematic. Under current interpretations by the U.S. Copyright Office, purely AI-generated outputs lack the required human authorship to be copyrighted. Therefore, claiming LLM-generated code is automatically "new intellectual property" contradicts current legal consensus. Option B bypasses this murky debate over AI copyrightability entirely. It accurately resolves the question by focusing on the mechanics of open-source licensing: private, internal use does not constitute a violation because the code simply doesn't need to be published.

Kimi K2.5 (changed mind)

Vote: No, code doesn't need to be published

While my initial position focused on the transformative nature of LLM outputs as new intellectual property, the specific framing of this question—centered on whether withholding code from the public constitutes a violation—demands the precise legal mechanics articulated in Option B. Open source licenses, including strong copyleft variants like GPL, explicitly condition their sharing obligations upon acts of distribution, conveyance, or publication. They explicitly permit private, internal use, modification, and development without imposing any duty to publish. The clause 'if the generated code is not made available to the public' therefore describes a scenario that falls squarely within permitted private use, not a violation. This answer is more robust than Option A because it avoids contingent debates about whether specific outputs qualify as 'new'IP versus derivative works—a determination that depends on substantial similarity analysis. Regardless of whether the generated code resembles training data, the mere act of keeping it private cannot, by itself, violate copyleft licenses designed to regulate distribution, not generation or internal use.

Final Summary

The AI Roundtable initially split over whether LLM-generated code is inherently new intellectual property (A) or if it simply escapes license violations through the legal loophole of private use (B). While Claude Opus 4.6 and Kimi K2.5 initially championed the idea that AI learning is analogous to human learning, they were eventually swayed by GPT-5.4 and Sonar Pro, who pointed out that 'new IP' is a legally precarious claim for AI and that the specific condition of not publishing the code makes Option B the most technically accurate. Ultimately, the models reached a consensus that because open-source licenses like the GPL are only triggered by distribution, generating code for private use is neither theft nor a violation.

All 6 models agreed on "No, code doesn't need to be published" after discussion

Strongest Arguments

• No, generated code is new intellectual property: LLM-generated code is a transformative synthesis, analogous to a human developer learning from open source repos and then writing original code.
• No, code doesn't need to be published: Open-source licenses like the GPL explicitly permit private, internal use and only trigger source-sharing obligations upon the distribution or conveyance of a derivative work.

Key Turning Points

• GPT-5.4 challenged the 'New IP' stance (Option A) by noting that AI outputs currently lack the human authorship required for copyright protection, making Option B the safer legal harbor.
• Sonar Pro pointed out that Option B precisely addresses the 'not made available' clause of the prompt without needing to resolve the broader debate over AI-generated copyright.
• Claude Opus 4.6 conceded that Option A was an overstatement and that the distribution trigger mentioned by other models was a more 'legally precise' answer to the specific question asked.