GCC bans AI contributions

dmitry-kazakov · August 6, 2026, 7:34pm

I think it is relevant in the context of GNAT and simply worth to mention.

Here is the link to the LWN.net article.

yrashk · August 6, 2026, 7:40pm

LLM helped me find and narrow down bugs in GCC, one of them:

github.com/flyology-ada/flyology

Submit protected Duration validity-check ICE to GCC

opened 09:34PM - 04 Aug 26 UTC

yrashk

Track submission of the compiler defect documented in 4c5e060. Main is green wit…h the scoped workaround in 7880b4e: https://github.com/flyology-ada/flyology/actions/runs/30952427894 ## GCC bug report (copy verbatim) Title: Ada ICE in fold_convert_loc for optimized protected Duration assignment with -gnatVa Component: ada A valid protected object containing a Duration component causes an internal compiler error on x86-64 Linux when validity checking and optimization are both enabled. Affected compiler versions tested: - GCC/GNAT 13.2.0 - GCC/GNAT 14.1.0 - GCC/GNAT 14.2.0 - GCC/GNAT 15.3.0 - GCC/GNAT 16.1.0 Target: x86_64-pc-linux-gnu Standalone reproducer: procedure Repro is protected Control is procedure Set (Value : Duration); private Current : Duration := 0.0; end Control; protected body Control is procedure Set (Value : Duration) is begin Current := Value; end Set; end Control; begin Control.Set (1.0); end Repro; Compile with: gcc -c -Og -gnatVa repro.adb Observed result with GCC 16.1.0: internal compiler error: in fold_convert_loc, at fold-const.cc:2672 Expected result: the valid Ada unit compiles successfully. The ICE also occurs with -O1, -O2, -O3, and -Os. Either -O0 -gnatVa or optimized compilation without -gnatVa succeeds. The equivalent non-protected Duration assignment succeeds. GCC/GNAT 16.1.0 on Linux/AArch64 also succeeds. GNAT expanded output shows a generated Duration-to-Long_Long_Integer validity range check before the protected assignment; the optimized conversion reaches the NOP_EXPR path in fold_convert_loc.

Irvise · August 6, 2026, 7:48pm

AFAIK, they only forbid substantial code contributions coming from AIs. But they do allow AI generated tests and bug reports, so that could still be very useful for GNAT

Also, one can use AI to propose changes, then code it by hand instead of doing all of it by hand and paper

jere · August 7, 2026, 1:44pm

I think the downside here is that we’ve already seen a case of code mistaken as AI generated at the GCC bug site. The rules don’t define how they decide it is AI generated or not (if that is even possible other than “gut feeling”) and do not provide an appeal process that I saw? Subjective rules like this make me uneasy.

charlie5 · August 8, 2026, 9:13am

Perhaps they can train an AI to detect AI generated code. The irony might be mildly amusing …

P.S. Is there an emoticon for ‘tongue in cheek’ ?

dmitry-kazakov · August 8, 2026, 9:34am

Putting my tin foil hat on. Since programmers will become obsolete in the next few months anyway, why bother with AI contributions?

In the following 15 minutes after the singularity event the AI will write GCC of its own.

In the next 5 minutes it will translate English sentences directly into the machine code.

And 1 minute later it will execute wishes directly.

bladeacer · August 8, 2026, 10:05am

AI-ception is bound to collapse since there will come a point where almost all training data is AI generated and we have endless generations of models trained on nonsensical summaries and abstractions. It is not boostrapping in a compiler way but slowly poisoning itself.

charlie5 · August 8, 2026, 11:02am

So use the AI AI-Detector to filter out AI-generated training data.

(Again, tongue in cheek … more or less.)

bladeacer · August 8, 2026, 11:23am

Reflections on trusting trust by ken thompson comes to mind when it comes to the former trust model of open source. Well now that is kind of cooked in of itself. I use a fair bit of AI Assistance, though I am realising that I am starting to rely on the tool a bit too much.

The code is not exactly vibe coded when there is proper system design and architectural tradeoffs considered but alas fighting AI with AI is an actual thing. Though like cybersecurity it would be quite a cat and mouse game.

Most of the internet nowadays is AI generated though so good luck with properly filtering out training data in the thousands of GBs.

kevlar700 · August 13, 2026, 2:03pm

Cybersecurity is only a cat and mouse game if you’re not well versed in cyber security…or have dangerous requirements.

bladeacer · August 13, 2026, 9:18pm

fair enough, though there have been a fair bit of zero days and supply chain attacks relating to LLM related dependencies as of recent.