I Almost Couldn’t Hit Send Before a Client Delivery

At 11:40 p.m. last week, Hangzhou-based Chen Lei sent a client a mini-program built with AI assistance. His first thought was: “Since AI wrote the code, why not just open-source it and skip the agreement?” The next second, he worried the client’s source code, payment details, or a backdoor might be mixed in. I hesitated too: if AI contributed, does the code have no clear ownership? It isn’t that simple.

Separate “AI Wrote It” From “It Can Be Public”

Models released by companies such as OpenAI and GitHub often carry licenses that permit commercial use while prohibiting illegal uses, but the exact terms vary by model and version. The code may also contain fragments from training data, client materials, or third-party libraries with pre-existing licenses. In an article on February 27, 2026, independent developer Jakub Jankiewicz argued that all AI-generated code should be open source. That position works for discussing the public interest, but it cannot replace client contracts or file-by-file review.

I got this wrong once too: I treated “the model allows commercial use” as “the code can be published freely.” In reality, authorization for generated code and authorization for its dependencies are separate issues. A safer approach is to record the model and version used, scan for sensitive information, verify dependency licenses, and have a human review the critical logic.

What It Costs to Replicate This Today: $0, About 1 Hour, No Major Barrier

Money: open-source scanning tools and AI assistants can get you started for $0. Time: documenting the project’s origins, running the checks, and manually reviewing takes about 60 minutes. Technical barrier: knowing how to copy and paste commands and navigate project folders is enough; writing code isn’t required. First step: create a “Delivery and Licensing Checklist” listing the client’s name, models used, code origins, third-party dependencies, and what can be made public.

If I’m not planning to publish the code right now, I don’t need to panic over this debate either. I would first spell out AI assistance, code ownership, confidentiality obligations, and third-party license boundaries in the client contract. This isn’t about distrusting the client; it prevents both sides from later relying on memory to decide who owns a particular file.

Match the Process to My Stage—No Need to Do Everything at Once

If I’m just getting started, I’d preserve the generation history and tell the client in one sentence: “AI assisted with this project, and I’ll run security and licensing checks before delivery.” That usually costs $0 and 30 minutes.

If I already have one or two clients, I’d add a licensing checklist to each project and label the model version, dependencies, client code, and publishable components separately. This takes about an hour and reduces last-minute delivery disputes.

If I’m scaling, I’d turn review into a four-step process: generation records, dependency verification, sensitive-information scanning, and manual acceptance. I’d also make the contract specify who owns the final code. Open-sourcing can remain an option, but I wouldn’t treat it as the default solution to ownership questions.