Give a frontier model three capacitors drawn in parallel and ask for the total. More than one of the best models available will read them as series, hand back the wrong number, and explain its reasoning in the same calm, fluent voice it uses when it’s right. Not one flag. Not one hedge. Just a confident, specific, wrong answer that looks exactly like a correct one.
That’s the whole problem in miniature. A wrong answer you can spot is a nuisance. A wrong answer wearing the costume of a right one — same tone, same format, no warning — is a board spin waiting to happen. The issue was never the error rate. It’s that the errors are invisible.
What hallucinations look like in electronics
Engineering AI hallucinations almost never look wrong. They look like ordinary answers, which is exactly what makes them dangerous.
- Pin functions. Ask what pin 14 does, and a generic model answers from the average of every datasheet it has ever seen — not this part’s. A supply on your device; a reference on the one it pattern-matched.
- Datasheet values. A maximum current, a threshold voltage, a timing figure — quoted to two decimals, and nowhere in the actual datasheet.
- Absolute-maximum ratings. One misremembered by a volt, so the “safe” margin it reassures you about doesn’t exist.
Each one is plausible, specific, and unverifiable at a glance — the worst possible combination in a review, whose entire job is to catch the plausible-but-wrong.
| Ungrounded LLM | Grounded, bounded assistant | |
|---|---|---|
| Source | The internet, on average | Your board — this netlist, this BOM, these datasheets |
| Answer to “what’s on pin 14?” | A confident, plausible guess | The pin’s actual function, with the reference |
| When it can’t tell | Answers anyway | Refuses, and names what’s missing |
| You verify by | Re-checking everything yourself | Following the citation |
Why a bigger model won’t save you
A general model is trained to produce text that sounds right, from a corpus of the whole internet. It has never seen your board. It doesn’t hold your netlist, your BOM, or the datasheets your parts actually came from. So when you ask about your design, it answers about designs like yours — an average, not a fact. That’s why “paste the schematic into ChatGPT” disappoints: it reasons from your description of the board, not from the board.
The vendors building hardware AI say this out loud. Launching its Fuse EDA AI Agent, Siemens noted that standard tools “struggle with semiconductor and PCB system design because they lack the proprietary domain knowledge needed to interpret dense, physics-based EDA data,” and that the density of modern toolchains “quickly overwhelms generic models, leading to context saturation and hallucinations.” When the company shipping the AI tells you the generic approach hallucinates, believe them.
But notice where that leaves you. The industry’s answer is to make the model know more — more domain data, more specialised agents, more orchestration. Useful. And still, at the end, it hands you an answer you have to take on faith. A better-informed guess is still a guess. The gap between a plausible answer and a verified one doesn’t close by making the model smarter. It closes by making the model prove it.
The fix isn’t a smarter model. It’s a boundary.
Grounding is the first half: tie every answer to the actual design data, so the assistant reports what’s there instead of what’s usual. It reads this netlist for connectivity, this part’s datasheet for values, and points at this layout location — so a human can confirm any claim in seconds. “The AI said so” becomes “the netlist says so — here’s the line.” You stop trusting the model and start trusting the citation, exactly as you’d treat a human reviewer’s finding: don’t trust it, verify it.
The second half is the part almost nobody builds: a claim that can’t be traced to the data doesn’t get to exist. Not softened, not hedged — refused. An answer with nothing behind it is flagged as exactly that, instead of dressed up as fact. Grounding lets the AI cite its sources. The boundary stops it from writing anything it can’t.
How Probe does it
Probe holds AI to the rule every reviewer already lives by: show your working. Over MCP — the open Model Context Protocol — your assistant (Claude, Copilot, whichever you run) queries the actual board instead of answering from memory. It asks the design, gets back facts, and every finding carries the same references a human reviewer’s would: the netlist line, the layout location, the datasheet value.
And this isn’t a prompt convention you hope the model honours. A finding driven from anything but the real connectivity is rejected at the write boundary — the model is not permitted to record a conclusion the board can’t support. Citations are checked against their source, and an answer with nothing to trace to is flagged rather than hidden. The point isn’t an AI that’s always right. It’s an AI whose answers you can always check — and one that isn’t allowed to hand you a guess with a straight face.
That’s the difference between the field’s direction and ours. Everyone else is teaching the model to know more. We’re refusing to let it claim more than it can show.
AI for PCB design review, once you can check it
AI for PCB design review is genuinely useful — the instant you can tell a fact from a fluent guess. The problem was never that the model is sometimes wrong; it’s that you couldn’t see which times. Ground the answers in the real design, attach the evidence, refuse what has none, and that problem simply goes away. The assistant stops being a confident stranger and becomes a reviewer that cites its sources.
Probe is in development, launching later in 2026. Every answer about your board comes with the netlist line, layout location, or datasheet value it came from — checked, not asserted. And the ones that can’t? It won’t pretend.
