TecLeads TecLeads Blog
2026-11-16 · 6 min read

The AI Tool Your Team Needs Is a Provenance Gate

Abstract streaks of blue light forming a network
ai newsopen sourcedevsecopssoftware supply chainapplied ai

A pull request lands late in the afternoon. The tests pass, the diff looks plausible, and the author says an AI assistant wrote most of it. Then someone asks a question the repository has no field for: what are we allowed to do with this code?

That is not a theoretical licensing seminar. It is a release problem, and release problems have a habit of appearing after everyone has already agreed to ship.

Most AI news rewards the visible event, an LLM release, an AI launch, another coding assistant with a polished demo. A short article published on 15 November makes a less comfortable proposal: all AI-generated code should be open source. It is not announcing another model. It is asking who gets to close the source after models learned from an enormous body of code written and shared by other people.

That argument deserves attention. More importantly, engineering teams need a way to handle it before policy, procurement, security, and legal all arrive in the same meeting with different assumptions.

The argument is sharper than the headline

Jakub T. Jankiewicz makes two linked claims. First, AI models were trained on open source code and material from communities such as Stack Overflow, often without the permission or attribution expected by the original licenses. Turning that accumulated work into proprietary code, he argues, resembles license laundering.

Second, the article argues that code produced entirely by a non-human system should not receive ordinary copyright protection in the US. It also acknowledges that copyright rules differ between countries, including the treatment of computer-generated works in the UK. The proposed answer is explicit: release AI-generated code under an open source license, or dedicate it to the public domain with CC0.

That is the author's position, not a settled global rule. The distinction matters. Human contribution varies from accepting a whole generated module to using one completion and rewriting every meaningful line. Copyright, license obligations, tool terms, and company policy can point in different directions. A checkbox marked “AI used” does not resolve any of them.

Still, dismissing the article because the legal edges are disputed would miss the engineering problem. Teams are accepting generated code without recording enough provenance to discuss those edges later.

Your repository already has a supply chain

Generated code does not arrive alone. It can introduce packages, copy familiar-looking implementations, embed insecure defaults, expose secrets from prompt context, or reproduce an API pattern that stopped being sensible three versions ago. Reviewers see the resulting text, but usually not the prompt, model, extension, context sources, or amount of human revision behind it.

Call this an AI exclusive if the CMS insists. In the repository, it is a software supply chain question.

We already know how to reason about supply chains. Record origins. Minimise unreviewed inputs. Scan the output. Require named ownership. Keep enough evidence to investigate later. The same hardened cloud foundations used for an LLM application, RAG service, agent, or ML pipeline should extend to the code and configuration used to build it.

This is where many new AI tools get an easier ride than they should. Teams approve the editor extension, add a paragraph to the acceptable-use policy, and assume the code review process will absorb the rest. It will not. Review cannot account for provenance that nobody records.

Put a gate where the code enters

The practical takeaway is simple: add an AI provenance gate to every pull request.

Do not wait for perfect detection. Generated-code detectors are not a dependable source of truth, and trying to infer authorship from coding style turns review into theatre. Ask the contributor directly, preserve the answer with the change, and define what happens when the answer is yes.

A pull request template can start like this:

## AI assistance

- [ ] No generative AI tool materially contributed to this change.
- [ ] A generative AI tool materially contributed to this change.

Tool and model, if known:

Human review performed:

Generated or heavily assisted files:

New dependencies or copied external material:

Intended licensing treatment:

“Materially contributed” needs a local definition. Autocomplete for a variable name is noise. Accepting a generated function, test suite, infrastructure module, migration, or policy file is not. Pick a boundary engineers can apply without consulting a philosopher every time they open a pull request.

For a yes answer, route the change through the controls you already trust. Run secret scanning, dependency analysis, static analysis, license checks, and the real test suite. Require an accountable human reviewer for security-sensitive paths. If generated code is intended for public release, make the license choice explicit. If it remains proprietary, record that decision and send genuinely uncertain cases to whoever owns legal and licensing risk.

Do not make one engineer reconstruct this six months later from an editor subscription and a vague memory.

The checkbox is not the control

A provenance declaration only works when it changes behaviour. If every answer follows the same path, contributors will correctly treat it as paperwork.

The useful control is the decision attached to the declaration. A generated Terraform change might need review from the platform owner. Agent code with tool access deserves threat modelling around permissions, prompt injection, and data exposure. RAG ingestion code needs checks on source rights and sensitive data. An ML pipeline change still needs reproducibility, artifact tracking, and deployment controls even when the generated Python looks ordinary.

This also gives product teams a better conversation. The question stops being “do we allow AI?” That question is far too broad to produce a useful answer. Ask which tools may touch which repositories, what context they may receive, what evidence a contributor must retain, and when generated work can be distributed under the repository's existing license.

Those decisions will survive the next round of AI updates. A policy tied to one vendor's current interface will not.

If you only do one thing this week

Add the provenance block to your pull request template and assign an owner to every yes answer.

Do not begin with a company-wide argument about whether all generated code must be open source. Begin by making generated contributions visible. After a few real pull requests, you will know where the unresolved questions actually sit: licensing, sensitive context, weak review, dependency sprawl, or unclear ownership.

The article's headline is deliberately absolute. Your first control does not need to be. It needs to make the decision explicit before generated code crosses the release boundary.


If this is on your plate, TecLeads does exactly this as part of our AI & Applied ML work. If you'd like a second pair of eyes on your setup, book a 30-minute call or explore what we do.

📍 Tech Pulse · today's quick question 🟡 Level: Intermediate Cybersecurity

What is a zero-day?

Pick an answer to see how other engineers voted.

Want a hand with this?

TecLeads helps engineering teams ship faster and more securely.

Book a 30-minute call

← All posts