wro.cpp
short

How the C++ community is actually using AI in 2026

· english· audience: working-cpp· AI-generated, reviewed by Filip Sajdak

Something changed in the C++ community’s relationship with AI over the last few months. The conversation stopped being “should I use an LLM for C++” and became “how do I get it to produce code I would actually merge.” Here is where that discussion stands, gathered from the people the community listens to.

Jason Turner’s three-part arc on C++ Weekly

Turner has been working the problem in public across several episodes, and the titles tell the story of where he landed.

  • Ep 531, “You Are Using AI Wrong” is the provocation: most of the frustration developers have with AI-generated C++ comes from how they are driving it, not from the model. The workflow is the variable you control.
  • Ep 535, “Getting The Best AI Generated Code” is the practical follow-up: how to set up the context, constraints, and feedback loop so the output is closer to what you want on the first pass.
  • Ep 539, “Modernizing C++ with AI” is the payoff: driving an LLM through iterative refactoring of real C++, modernizing older code in small, reviewable steps rather than one giant unreviewable rewrite.

The through-line is consistency with how good engineers already work: small diffs, tight feedback, you stay the reviewer. The AI is a fast junior pair, not an oracle.

Teaching C++ when the students have LLMs

CppCast Ep 408 took the same question into education, with guest Tom Tesch on teaching C++ to game-development students in the age of LLMs. The hard part is no longer “how do you explain pointers”; it is how you assess understanding when every student has a tool that will produce a plausible answer. The discussion ranges from assessment design to why constrained, retro hardware (down to the NES) is still a useful teaching environment precisely because the model cannot lean on infinite resources.

The question nobody wants to skip

Not all of it is upside. Back in April, Herb Sutter ran a poll asking whether readers had observed AI agents doing harm in practice. It is a deliberately uncomfortable question, and it belongs in this roundup: the same agents that refactor your legacy code can also confidently introduce a subtle bug, a security regression, or a license problem. Turner’s “you stay the reviewer” and Sutter’s “watch for harm” are the same lesson from two angles.

The takeaway

The honest state of the art in mid-2026: LLMs are genuinely useful for C++ when you treat them as a workflow to be engineered rather than a button to be pressed. Drive in small steps, keep the feedback loop tight, and never stop being the reviewer. That is not a hype take or a doom take; it is where the practitioners have actually converged.

For a concrete worked example on this site, see the fastest JVM is the C++26 compiler, on what compile-time C++ can do that an LLM cannot guess at.

Update, 7 September 2026: a worked example of the review going wrong

Giovanni Dicanio published a case study on 4 September that is worth more than the general advice above, because it is specific.

He asked an AI to review WinReg, his Windows registry wrapper. It came back with a crash report: zero-length REG_SZ and REG_EXPAND_SZ values would break GetStringValue, GetExpandStringValue, and their two Try variants. It even supplied the corroborating detail, that the binary getter GetBinaryValue guards against dataSize == 0 while the string getters do not.

The bug was not there. Dicanio checked and the code handled the case.

What makes this worth reading is not that a tool was wrong. It is the shape of the wrongness. The report named real functions in a real file, drew a real distinction between two families of getters, and offered exactly the kind of asymmetry that a careful human reviewer would also find suspicious. Nothing about it looked like a hallucination. A reviewer who trusted it would have added guards that do nothing, and gone looking for the same omission everywhere else in the codebase.

That is the cost model the section above understates. The expensive failure is not the obviously wrong answer, which you discard in seconds. It is the plausible one that survives a skim, because a plausible false positive spends your attention at exactly the rate a true positive would.

Since this site is drafted with AI assistance and says so on every post, the honest version of the lesson is the one we apply to ourselves: every code claim here gets compiled before it ships, and the ones that cannot be compiled get labelled as unverified. Dicanio’s case is what the alternative looks like.

Sources: C++ Weekly with Jason Turner, Ep 531 / Ep 535 / Ep 539. CppCast Ep 408 with Tom Tesch. Herb Sutter’s AI-harm poll.