Search "ChatGPT PCB design" and you find two camps. One says it can do everything. The other says it is useless. Both are wrong, and the useful answer sits in a specific place: the netlist.
The short answer
ChatGPT cannot design a manufacturable PCB on its own. It can help you choose parts, explain a topology and draft a netlist in text. It cannot place components, route copper, run a design rule check or produce Gerber files, because it has no geometry engine and no rule checker.
ChatGPT is good at the half of PCB design that is judgement, and structurally unable to do the half that is geometry. Knowing which is which saves a lot of wasted effort.
What ChatGPT is genuinely good at
- Picking parts. Describe what the circuit must do and it will suggest a regulator, a microcontroller or a sensor family worth considering.
- Explaining. It can explain a topology, sanity-check a power budget, or talk you through why a decoupling capacitor belongs close to a supply pin.
- Drafting. It can draft a rough netlist in text form, which is a reasonable starting point for a design you intend to verify yourself.
If you are learning electronics, that is a patient engineer who never gets tired of "why". Use it.
The four things it cannot do
- ChatGPT has no geometry engine. It cannot place components on a board with real dimensions, or produce copper that respects clearance.
- It cannot route a multi-layer board. Routing is a search problem over physical space with hard constraints, not a text-generation problem.
- It cannot run a design rule check. Without one there is no way to know whether the result is manufacturable.
- It cannot output fabrication files. Gerber, Excellon drill and pick-and-place data are exact formats that a fab will reject if they are approximated.
None of that is controversial once you have tried it. The part people underestimate is the one ChatGPT can do.
The netlist looks finished. It usually is not.
A netlist is a list of which pin connects to which. It is text, so a language model produces it fluently, and it reads like the hard part is done.
I build an AI PCB tool, and language models draft netlists in our pipeline every day. The numbers below come from the models we run, not from ChatGPT itself. But none of these failures depend on which model you use. They come from asking any text generator to recall pin-level facts.
1. The same pin on two nets. In 4 of 6 boards we sampled, the model's netlist listed at least one pin on two different nets. That is a short circuit. The pattern was always the same: a power pin listed on its rail, then listed again on a signal net.
+3V3 -> ..., U1.1, ...
QSPI_CS -> U1.1, U2.1
On an ESP32 board it was the pull-up resistor: R6.2 on both +3V3 and EN. Read as prose, each line looks reasonable. Only a check that asks "does any pin appear twice?" catches it.
2. Pinouts from memory. A battery charger IC, the MCP73831, came back as an 8-pin part with a "VPROG" pin. The real part is a 5-pin SOT-23-5. The model described a plausible chip that does not exist, with total confidence. A wrong pin map produces a board that cannot work, and nothing about the text tells you so.
3. The value field names the wrong thing. Asked for an Arduino Nano soldered to a board, the model wrote the Nano's value as "ATmega328P", the chip the Nano carries. Anything downstream that trusts that field builds the board around a bare 32-pin chip instead of the module. It is a reasonable sentence and the wrong part.
4. Same prompt, different circuit. Temperature 0 does not make a hosted model repeatable. Two identical runs of the same prompt gave us 48 nets and then 50 on one board, and 13 and then 11 on another. If you re-ask ChatGPT for the netlist, expect a slightly different circuit, and check the one you actually build.
Every one of these is a pin-level fact that has to be looked up, in a symbol library or a datasheet. Generating it is the wrong operation, and a better model does not change the operation.
How to use ChatGPT for PCB design anyway
This is the workflow that holds up:
- Let it choose. Describe the board, ask for parts and the reason for each. This is the judgement half, and it is good at it.
- Ask for the netlist, then distrust it. Treat it as a draft.
- Check every pin against the datasheet, not against the model's memory. Especially power, enable, reset and anything on a module.
- Check that each pin is on exactly one net. A ten-line script does this. It catches the most common short.
- Hand the result to a tool that has geometry and rules: placement, routing, DRC and Gerber export are computed, not written.
What we built around that split
The split we landed on in PCBEditor is simple to say: AI decides intent, algorithms verify physics.
- A language model chooses the parts and what the board is for.
- Pins come from library data (KiCad's symbol library and the part's land pattern), never from the model's memory.
- Nets are formed by rules over those pins, so the same parts give the same circuit every time.
- Placement is constrained, routing is a deterministic router with no AI in it, and DRC is a rule engine.
The test we hold it to: if we swap the model tomorrow, the boards must still be correct. That is only true if correctness lives in code.
It is not magic, and this week proved it. A board passed routing and DRC with zero errors and was still electrically wrong: an amplifier's inputs were tied together. Routing and DRC check geometry, not whether the circuit works. So we added checks for exactly that: two outputs driving one net, a chip's own inputs shorted, a regulator bypassed. The lesson is the same one as above. Anything that must be true needs a check that measures it.
Quick answers
Can ChatGPT generate a schematic?
It can describe a schematic and produce a netlist in text form. It cannot produce a verified schematic with real footprints and pin mappings, because it has no component library to check against.
What is ChatGPT actually good for in hardware design?
Part selection, explaining trade-offs, reviewing a design decision, and drafting a first-pass netlist you intend to verify. Those are judgement tasks, which is what language models are good at.
Will a newer model fix this?
It will get better at the judgement half. Placement, routing, DRC and fabrication output need a geometry engine and a rule checker, and pin maps need a lookup. A newer model doesn't give you any of those, however good it is. GPT-6 Astra is a different question: it can drive KiCad on your desktop, which is new, but when it presses Auto-Route it is still KiCad's router doing the work. More on that in GPT-6 Astra for PCB design.
What should I use for the rest?
A tool that keeps AI for judgement and uses deterministic algorithms for placement, routing and DRC. That is the split this whole post argues for, and it is the one PCBEditor is built on.
A longer version of the can/cannot split, kept up to date, lives at pcbeditor.com/ai-pcb-design/chatgpt-for-pcb-design.
Have you built a board from a ChatGPT netlist? I would like to know what it got wrong for you, and whether you caught it before or after the boards arrived.
Top comments (0)