Last updated August 3, 2026
The video above is from March 2023. We pointed Bing’s brand-new “Artificial Intelligence” at AutoLISP to see whether it could write anything usable for AutoCAD drafting, and published the result as a curiosity — the sort of thing you show someone at smoko.
Three years is a long time in this business. The question has stopped being a novelty and started turning up in job interviews, so it is worth revisiting properly: what actually changed between then and now?
The answer is not the one either camp expected. The tools did get better. But the failure mode moved, and it moved somewhere much harder to see.
What follows is drawn from the public record — the Autodesk forums and CADTutor, December 2022 to July 2026 — because that is where working drafters wrote down what happened when they tried it, at the time, without a product to sell.
2022–2023: the wonky era
The first wave of results was bad in a way that was at least honest about itself.
On the Autodesk Visual LISP board in December 2022, asimjian posted some ChatGPT-generated code and, after a few rounds with the regulars, concluded:
Look like the AI algorithm writes the code all wonky and unusable.
martti.halminen, a few days later, put his finger on exactly what was wrong, and the diagnosis has aged well:
Seems the fundamental problem here is that the system doesn’t know the syntax and semantics of IF, so it doesn’t know when to close the parentheses, and when to use PROGN to group commands for the correct semantics.
That is a structural failure. The code did not load. You knew within seconds.
Over on Autodesk’s own community blog in January 2023, a poster going by TheCADnoob ran a similar experiment and reached the conclusion everybody reached that year:
At first glance, it was clear that my job was safe for now. This AI is not going to take my place in the drafting world soon.
He immediately walked it back, though, and the fuller version is the fairer one to quote:
Overall I’m pretty impressed. I’m not sure this technology is going to take my drafting job yet, but I can definitely see it being an absolute asset for my productivity.
By May 2023, Henry C. Francis — an Authorized Autodesk Developer, posting on the Civil 3D board — had identified the specific shape of the confusion, and this is the thread that runs all the way through to today:
It will even suggest code although it doesn’t yet have a full knowledge of the limitations/differences in Autolisp vs LISP and VLX vs ActiveX.
AutoLISP looks like Lisp. The world contains a great deal of Lisp and comparatively little AutoLISP. That gap never fully closed.
October 2024: it starts working, and the problem gets quieter
Eighteen months on, the tone changes. A thread opened on the Autodesk forum in October 2024 titled “Microsoft CoPilot AI can write AutoLisp coding” — and the opening post is a success story. The poster had asked for a routine to convert survey figures to polylines at zero elevation:
It’s not perfect. Sometimes theses a back and forth to correct and error check. But the AI will try to fix the code until it does work. … Took a few fixes, but eventually worked. Now it’s a Button Macro.
That is a working drafter with a working tool. But read the replies, because this is where the story turns.
Kent1Cooper — a name anyone who has searched an AutoLISP problem in the last twenty years will recognise — replied:
There have been numerous Topics raised here based on problems with AI-generated AutoLisp code [more commonly from ChatGPT] that was really way off — using assumed but non-existent function names, supplying incorrect arguments, etc. I guess we should assume that AI will get better at it over time, but so far….
There is the shift, in one sentence. Not “it does not compile”. Non-existent function names.
That is a categorically worse failure and it is worth being clear about why. A syntax error stops the routine at load. A function that does not exist loads perfectly, sits in the file looking like every other line, and fails only when the branch containing it actually runs — which might be next Thursday, on someone else’s drawing, after your routine has already changed the current layer and switched the object snaps off.
AutoLISP has no compile step, no import resolution and no linter. Nothing checks a function name until it is called.
DGCSCAD, on the same thread, had built a passable find-and-replace routine and then hit the wall:
I had it add MLeaders, but that was full of made-up function definitions. I’ll be trying a different approach.
Two other things from that thread are worth keeping. CodeDing noted something in October 2024 that turned out to be a leading indicator: “I have personally found recently that Claude.ai is better at writing code than ChatGPT. Might be worth looking into.” And Sea-Haven offered the most practical tip on the whole page: “For anyone with Copilot make sure last 2 words are “Autocad lisp”, but it can struggle on some tasks.”
There was dissent, too, and it deserves printing. In March 2024, Gilles Chanteau — who has been answering AutoLISP questions for a very long time — responded to a drive-by AI-sounding post with:
Please don’t offend the natural intelligence of the people who provide help here by posting AI-generated code (the Stack Overflow site has banned generative AI).
Aimed at a specific behaviour rather than the tools, but a fair shot.
December 2025: somebody tries the obvious fix
The obvious answer to invented function names is to give the model the real list. In December 2025, on that same thread, SamBensonL4GSH asked about knowledge sources for Copilot, and Kent1Cooper followed up with the question — in his own square brackets — that anyone who has hit this problem eventually arrives at:
[Is it possible to embed the entire AutoLisp Reference in there as a “knowledge source,” so it won’t make up functions, use incorrect arguments, etc.?]
Nine minutes later, CodeDing answered — and this reply is the single most useful thing in three years of these threads, because it is somebody reporting an actual result rather than a theory:
I’ve tried. It makes it better, but still often confuses functions from Common Lisp as useable for AutoLisp. But I have noticed that as these newer AI models come out they do get much better at writing AutoLISP.
Both halves matter. Constraining the vocabulary helps. It does not cure. And the models are still improving underneath the whole argument.
Nobody, in the seven months since, published the reference in a form you could actually paste in. So we did — 485 function names from Autodesk’s own AutoLISP Reference, grouped by family, with the Express Tools and AutoCAD LT traps labelled. It is a free text file. CodeDing’s caveat still applies, and we say so on the page.
June–July 2026: where it stands now
The most recent evidence is a CADTutor thread called “AI taking over”, started by BIGAL on 19 June 2026 and still running on 25 July. It is the best snapshot available of what working drafters think today, and what is striking is how little of it is about whether the code runs.
BIGAL, opening: “Even I try AI now and again when stuck, it is successful some times.”
The clearest account of the current state came from a poster called Least, and it is worth reading whole because it draws the line in exactly the right place:
AI is a mixed bag for me. Copilot can be infuriating at times — it keeps using LISP functions that aren’t supported in AutoLISP, or ACAD-only functions that don’t work in BCAD. It can wander off into rabbit holes even after you’ve told it not to, and sometimes it changes working code for no reason and breaks it.
But with some persistence and back-and-forth, it can produce really good functions.
I’ve also tried Claude, which is much better in lots of ways — the code tends to run first time. If i run out of credits then I continue on with co-pilot. I used Fable just before it was banned, and that was pretty amazing too.
AI still needs a lot of guidance to produce anything worthwhile. It can be overconfident, sometimes it doesn’t listen and it doesn’t always think clearly about the end goal.
Note “it keeps using LISP functions that aren’t supported in AutoLISP” — present tense, June 2026, twenty months after Kent1Cooper described the same thing. That is not a solved problem. It is a managed one.
GLAVCVS was blunter about ChatGPT specifically: “However, the code suggested by Chatgpt is almost always lengthy and often fails.” rlx, describing Google’s AI — which he nicknames Clippy — found it fine until it wasn’t: “But as things got complexer it really started to mess things up.”
Danielm103 gave the most useful framing on the thread:
I use the heck out of AI though; I even run local models. The issue is, people expect LLMs to “one shot” solutions, it’s unlikely to happen. … If you spend time and find a system to work with the models, they are amazing.
And then there is the post that shows the conversation has genuinely moved on. ryanatkins49056, describing what his office actually does about it:
I’ve been pushing for not an outright ban on AI coded AutoLISP but trying to get all code passed through myself or the other guy as a few I’ve seen so far when help was asked for were deleting files out of shared directories.
He is careful about the detail, and so should we be: the deleted files were the routine’s own temporary files, and he says so in the next sentence. Nothing was lost. His concern was that untested code was running with that level of filesystem access at all — enough to worry him, his colleague and two CAD managers.
That is not a debate about whether AI can write AutoLISP. That is a drafting office writing a code review policy.
What actually changed
| 2023 | 2026 | |
|---|---|---|
| Typical failure | Unbalanced parentheses, misused IF/PROGN | Invented function names, missing production hygiene |
| When you find out | At load, immediately | When that branch runs — possibly weeks later |
| Cost of the failure | Nothing. It never ran. | Aborts mid-job, often after the drawing has been changed |
| What drafters were doing | Posting it to a forum as a curiosity | Writing office policy about reviewing it |
The honest three-year verdict: the answer moved from “no” to “yes, with a competent human in the loop”. Nobody in the current threads argues the code cannot work. They argue about how much guidance it needs and how carefully it has to be checked.
And the checking is the part that is genuinely learnable. The failures cluster somewhere specific — error handlers, undo groups, system variables left unrestored, variables never declared local, vl-load-com missing before ActiveX calls. Every one of those is a finishing step a human adds after the routine already works, which is exactly the phase a model that hands you finished-looking code in one shot has skipped.
Which is good news, because it means the review is a reading exercise, not a testing exercise. It takes about ninety seconds and you can do it at your desk.
We have written it up properly: the full failure list, a line-by-line red-pen review of a generated routine with a twelve-point checklist, the function allowlist for the one failure reading cannot reliably catch, and — if you are on the budget seat — what runs and what does not on AutoCAD LT. If you are wondering whether Autodesk is about to solve all this for you, we checked: what AutoCAD 2027 actually ships.
Three years ago the interesting question was whether it could write AutoLISP at all. It can. The interesting question now is whether you can spot what it left out — and that has always been the job.
References
- Help with OpenAI ChatGPT generated Autolisp code — Autodesk Visual LISP, AutoLISP and General Customization Forum. asimjian, 11 December 2022; martti.halminen, 14 December 2022; Gilles Chanteau (_gile), 20 March 2024.
- Talking to AI about a Lisp — Autodesk community blog, TheCADnoob, 6 January 2023.
- How ChatGPT explained my Autolisp problem — Autodesk Civil 3D Customization Forum. hencoop (Henry C. Francis), 2 May 2023.
- Microsoft CoPilot AI can write AutoLisp coding — Autodesk forum, October 2024 to December 2025. chlaksavery, Kent1Cooper, CodeDing, DGCSCAD, Sea-Haven, SamBensonL4GSH.
- AI taking over — CADTutor, AutoLISP, Visual LISP & DCL. Started by BIGAL, 19 June 2026, running to 25 July 2026. Posts by Least, GLAVCVS, rlx, ryanatkins49056, Danielm103.
- AutoLISP Reference — Autodesk, AutoCAD 2026. Source of the function allowlist.
- Claude + AutoCAD via MCP: AI Drafting on AutoLISP (2026 Guide) — blog.draftsperson.net.
- AutoLISP for AutoCAD: Tutorials and 139 Free Routines — blog.draftsperson.net.