AI for design review: make it prove it
The danger of AI in board review isn't that it's wrong, but that you can't see when. Ground every answer in the board, and refuse the ones it can't prove.
Hardware engineers are right to distrust AI for board review — but not for the reason they usually give. The problem isn’t the error rate. Errors are invisible. A mistake you can spot is an annoyance; a mistake that looks correct is a potential respin. The danger isn’t that it’s wrong. It’s that you can’t see when.
Give a frontier model three capacitors in parallel and ask for the sum. Several strong models treat them as in series, answer wrong, and explain it confidently without a hedge.
What hallucinations look like in electronics
All three are plausible and hard to check in passing. Pin functions: asked about pin 14, a generic model answers from the average of many datasheets, not this device’s. Datasheet values: a max current, threshold voltage or timing given to two decimals that isn’t in the actual datasheet. Absolute maximum ratings: a value misremembered by a volt, so the claimed margin doesn’t exist.
Why a bigger model doesn’t solve it
General models are trained to produce text that sounds right. They haven’t seen your board, so they answer about similar designs, not the actual one. Pasting a schematic into a chat gives reasoning from your description, not from the board. The industry’s answer — more domain data, specialised agents, orchestration — helps, but the output is still an unverifiable guess. The gap closes only when the model can prove its answers.
The solution is a boundary, not a smarter model
It has two halves.
Grounding. Every answer is tied to the actual design data: the netlist for connectivity, the device’s datasheet for values, and a layout position. The difference between “the AI said so” and “the netlist says so, here’s the row”.
Refusal. A claim that can’t be traced to the data isn’t produced. It’s flagged — not softened, not hedged.
How Probe does it
Probe applies the reviewer’s rule of showing your work. Over MCP (Model Context Protocol), your assistant — Claude or Copilot — asks the actual board instead of answering from memory. Every finding carries the same references a human would give: netlisted, layout position, datasheet value. Findings that don’t rest on real connectivity are rejected at the write boundary, citations are checked against their source, and untraceable answers are flagged.
The field teaches models to know more. We stop them from claiming more than they can show. AI is useful for design review only when facts can be told apart from fluent guesses — when every answer comes with its netlisted, layout position or datasheet value.
See an AI answer with evidence attached — or read why ERC and DRC aren’t enough.