Contents
Review AI-generated code against clear quality standards, and focus human attention on changes with high impact, complexity, or uncertainty. Automating implementation or review does not transfer responsibility for the outcome.
Key Points#
- Assign responsibility to someone who can explain the change, assess the evidence, and respond when problems arise.
- Use automated checks for repeatable requirements, and involve people closely in reviews of consequential or ambiguous decisions.
- Do not treat a readable diff as sufficient evidence; review the behavior delivered to users, including permissions and failure handling.
- Update quality standards when incidents, product changes, or external requirements reveal gaps.
Assign Responsibility First#
In his October 28 talk at Qiita Conference 2026 Autumn, Yosuke Furukawa said that code review remains important while questioning whether every line needs the same level of human attention. The important practical distinction is between automating checks and taking responsibility.
Teams can use AI to write and review code while keeping a person responsible for the outcome. That person does not need to remember every implementation detail. They do, however, need enough understanding to assess evidence, recognize uncertainty, and coordinate fixes.
Before merging, designate one person as responsible. Ask them to answer two questions: What evidence shows that this change meets the requirements? What will you do if that evidence proves insufficient? “AI generated it” answers neither question.
Set Review Depth According to Risk#
Reading every change with equal intensity is an inefficient use of attention. A wording fix and an account recovery change should not follow the same review process. Classify changes before deciding how much human review they need.
Ask these four questions.
- Impact: Could a defect expose data, move funds, cause loss of access, or disrupt a critical service?
- Complexity: Does the change span multiple services, introduce concurrency, or modify persisted data?
- Ambiguity: Are the expected behavior and acceptance criteria clear enough to test?
- Recovery: Can the team detect a defect quickly and safely roll back the change?
For small, reversible changes with clear requirements, automated checks alone may suffice under an agreed team policy. Payment changes and unclear permission rules should receive direct human review. Major architectural decisions require a review of their assumptions and consequences, even if the generated code looks tidy.
Check Behavior Against Standards#
“Looks reasonable” is an insufficient release criterion. Define observable requirements before asking AI to implement or evaluate a change. The same vague instructions can produce both an incorrect implementation and a review that confidently approves it.
For a password reset feature, a minimum checklist might include these four requirements.
1. Reset tokens expire after the period defined for the product. 2. Tokens cannot be reused after successful use. 3. A token issued for one account cannot reset another account. 4. Responses do not reveal whether an account exists for the email address.
These requirements guide automated tests and human checks. Also check failure handling, such as when the email provider is unavailable, a request is interrupted, or tokens are submitted concurrently. Passing tests supports a release decision, but does not prove that the requirements themselves are complete.
Use Automation to Provide Evidence for Decisions#
Automate checks with clear pass criteria, such as formatting, type checking, known dependency vulnerabilities, and tests tied to acceptance criteria. AI review can identify additional concerns, but its approval alone should not serve as evidence.
Give reviewers a concise brief covering the intended behavior, affected components, check results, and unresolved questions. For consequential changes, include recovery procedures and criteria for deciding when to execute them. This lets reviewers assess the change without having to reconstruct its purpose from the diff.
When a defect slips through, improve the verification process along with the code. This might mean adding a missing test, clarifying a requirement, or adjusting the agent’s tools or constraints. Otherwise, subsequent generated changes may repeat the same mistake.
Keep Standards Current#
Quality standards are product decisions that need ongoing review. Newly disclosed vulnerabilities, changes in the customer base, or accessibility requirements can make previously accepted behavior inadequate. Existing code may also need review.
Furukawa’s talk emphasizes this evolution of standards. However, before teams use specific vulnerability names or claims about exploitation from the presentation to determine urgency, they need to verify them against primary sources: security advisories. Base the response on affected versions, external exposure, available mitigations, and evidence of exploitation.
Assign responsibility for monitoring relevant changes and revisiting checklists after incidents or major product changes. Human judgment remains necessary to determine whether yesterday’s checks still protect today’s users.
Next Step#
For your next pull request, add a four-line review note before merging. Include the person responsible, expected behavior, verification evidence, and recovery actions.
Source: https://speakerdeck.com/yosuke_furukawa/ai-ga-kodo-o-kai-te-kureru-jidai-wareware-ha-nani-o-tashikameru-no-ka
Contents
