Was this app vibe-coded?
An app built quickly with AI tools can look finished and still be hard to change. A buyer wants to know two things: how much of the code an AI tool wrote, and whether anyone engineered it. The repository answers the second question well. It answers the first only when the team said so: nothing in the code itself proves who, or what, wrote it. This guide says what you can check, what you cannot, and what to ask. It is not legal or investment advice.
What the repository can tell you
| Question | Can the repository answer it? | From what |
|---|---|---|
| Did the team use AI tools? | Only where they declared it | Commits that name an AI agent as co-author, say they were AI-assisted, or were made by an agent's own account; a written policy on AI contributions |
| How much of the code? | A lower bound | The share of lines and commits that declare an agent, all time and over the last 12 months |
| Undeclared AI code? | No | Nothing reliable: code reads the same whoever wrote it |
| Was it engineered? | Yes | Whether code was revisited after it was written, how fast it arrived, how commits were written, and repeated blocks |
| Is it tested? | Yes | Tests and their proportion, whether they grew with the code, whether fixes came with tests, whether CI runs them |
| Was it reviewed? | Partly | How changes reached the main branch |
Declared AI help, and why it is a floor
Many AI coding tools sign the commits they help with: a
Co-Authored-By line naming the agent, or a note that the
change was AI-assisted. Some agents commit under their own accounts.
Those declarations are what a repository can show.
They are a lower bound, not a measurement. A team that turned the signature off, squashed its commits, or rebased its history leaves no trace, and a single rebase removes the lines this rests on.
Absence of a declaration is absence of disclosure, not absence of AI authorship.
Tools that estimate AI authorship from the code alone are estimating, and can be wrong in both directions.
As of 19 November 2026, across Borehole's library of 1,657 open-source projects, 1,222 of 1,657 have commits that declare an AI agent as co-author: 5 in 2023, 1138 in 2026. 357 state a policy on AI-generated contributions. The full count, year by year.
Engineered, or merely generated
A high AI share is not a defect by itself. What a buyer inherits is a team's ability to change the code safely, and the history shows that whoever typed it:
- Revision. Was code revisited after it was written, or does most of it sit exactly as it first landed?
- Pace. Did large amounts of code arrive faster than a person reviews it, or in one first commit carrying most of the project?
- Commit messages. Written by people about their change, or templated?
- Repetition. Identical blocks of code repeated across files, the cheapest thing a generator does.
- Tests. Do tests exist in proportion to the code, did they grow with it, and do bug fixes come with the test that would have caught them?
- Review. Did changes reach the main branch through review, or straight in?
None of these proves how the code was written. Together they say whether anyone can maintain it, which is the question a buyer pays for.
What to ask the seller
- Which AI tools does the team use, for what, and does it declare them in commits?
- Which parts of the product were generated, and who reviewed them?
- Can the team change what a tool produced, or only regenerate it?
- How do the tests show the generated code behaves as intended, rather than that it passes?
- What dependencies did the tools pull in, and who understands them now?
- Do the people who built it come with the deal?
The questions to ask the CTO cover the rest of the conversation.
How Borehole checks it
A Borehole survey reads the code and its history with 107 checks across seven dimensions. It states the declared AI share as a disclosure, with a confidence level and the reasons for it, and that share never changes a rating. The checks above (revision, pace, commit messages, repetition, tests and review) carry the ratings. Every finding cites its evidence.
- A public repository. Paste its address. The survey is free.
- A private repository. The collector reads it on your machine. What it sends: never whole files; facts and short excerpts of up to 120 characters where a check matched. 3 private surveys in any 30 days are free once you sign in. What happens to your code.
- A deal that needs a person. Hire us for what a repository cannot show.
The check on one page
- Look for declared AI co-authorship, and treat it as a floor
- Read any policy on AI contributions
- Check that code was revised after it landed
- Look for code arriving faster than review, and repeated blocks
- Check that tests exist, grew with the code, and come with fixes
- Check how changes reached the main branch
- Ask who can change the generated parts, and whether they stay