The Suniworks journal
AI-generated code: who takes responsibility?
A useful lesson from the Linux kernel: someone must review, understand and take responsibility for the code a team delivers.
Updated on .
A function may take seconds to generate and much longer to understand. A coding assistant moves part of the work; it does not remove review or accountability for the delivered application.
The Linux project makes human responsibility explicit
The kernel’s coding-assistant documentation places responsibility on a human contributor. It prohibits an agent from adding the Developer Certificate of Origin sign-off itself and provides for attribution of assistance.
This is a project-specific contribution rule. For an organisation commissioning software, it also suggests a useful question: who understands this change and can answer for it? That application is an operating recommendation, not a commercial obligation created by Linux.
Review behaviour, not appearance
Well-presented code can apply the wrong business rule, miss a permission check or mishandle absent data. The reviewer needs to connect a change to its purpose and examine what can go wrong.
- Which user task does the change support?
- What happens if information is missing or submitted twice?
- Who may read or change the data?
- Which test demonstrates the expected behaviour?
The questions apply equally to code written without an assistant. They make review more useful than checking formatting alone.
Match the evidence to the risk
A label correction needs different checks from a billing or access-control change. Tests should address failures that would create real cost for users or operation.
A delivery record can name the scenario tested, the outcome and the limits of that test. A successful demonstration with one dataset does not establish every capability of the system.
Leave the next developer enough context
Record important constraints, dependencies and trade-offs. That context helps a future maintainer more than an unfiltered transcript of conversations with an assistant.
Before commissioning an application, ask how changes are reviewed, how licences are examined and who handles defects. The writing tool matters less than the team’s ability to explain and maintain the result.