Bertrand Meyer Opens Q&A Follow-Up on AI and Software Verification
Bertrand Meyer published a July 21 follow-up to his May 7 ACM Tech Talk on software verification in the age of AI, posting unanswered attendee questions and saying he plans to address those he can over the next two or three weeks. The questions focus on safety-critical code, testing, formal methods, and the ethics of generative AI training data.
Bertrand Meyer published a July 21 follow-up to his May 7 ACM Tech Talk, "Software Verification in the Age of Artificial Intelligence," after the session ended with more attendee questions than he could answer. Meyer posted the unanswered questions and said he plans to address those he can over the next two or three weeks.
From a webinar to an open Q&A
Meyer wrote that ACM organizer Yan Timanovsky supplied the remaining chat questions after removing participants' names and email addresses. The post turns the one-hour webinar into an ongoing public discussion rather than presenting a finished paper, product, or benchmark.
The questions span several practical fault lines in AI-assisted development:
- •Whether using generative AI for coding or testing is ethical when authorization for training data is disputed.
- •How safety-critical software written with AI can remain safe during operation after pre-deployment verification.
- •What role testing retains if formal methods and design by contract become more widely used.
Meyer also invites additional questions through the page's comments, so the published list may expand as he adds answers.
The verification model behind the discussion
ACM's official description of the May 7 talk asks whether "vibe coding" can scale to production-quality systems. It frames modern verification as a combination of dynamic techniques such as tests, static techniques such as proofs, and advanced tools and AI agents that support both.
That framing distinguishes code generation from evidence that code meets a requirement. Tests can exercise selected behavior, while static analysis, contracts, and proofs can establish bounded properties under stated assumptions. None of those methods makes generated code automatically trustworthy, and the retrieved material does not offer new empirical results showing that one combination works best.
For engineering teams, the immediate value is the agenda rather than a new prescription. AI coding workflows still need independent controls appropriate to the risk: review, repeatable tests, dependency and security checks, static analysis, production monitoring, and stronger specifications where failure costs justify them. The follow-up is useful because it surfaces unresolved questions about how those controls should evolve, but its answers are still in progress.
Key Points
- 1Meyer published unanswered questions from his May 7 ACM Tech Talk and plans to address those he can over the next two or three weeks.
- 2The questions cover AI training-data ethics, safety-critical software in operation, testing, formal methods, and design by contract.
- 3ACM framed modern verification as a combination of dynamic tests, static proofs, advanced tools, and AI-agent support.
Scoring Rationale
The follow-up surfaces concrete professional questions about verification for AI-assisted development but does not introduce a new tool, study, benchmark, or completed set of recommendations. Its value is the engineering agenda and official ACM context.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
