EngineeringID Staff
An AI Can Draft Your Drawings. It Cannot Hold Your License.
Generative tools are producing plan sets, and the profession is asking whether that is allowed. Wrong question. The seal was never a claim about who drew the lines — it is a claim about who is answerable for them.
An engineer feeds a program a site plan and a design brief and forty minutes later has a plan set that looks like it took a week. The profession's reaction has largely been to ask whether that is allowed. That is the wrong question, and asking it reveals a misunderstanding of what a seal has ever meant.
Your seal is not a claim of authorship. It never was. It is a claim of answerability.
What the seal actually asserts
Read the rules that govern sealing and notice what is absent. Boards specify geometry, required text, the licensee's number, whether an electronic seal is permitted. Colorado requires a circular seal between 1.5 and 2 inches under Colo. Rev. Stat. § 12-25-217. California specifies 1.5 inches under Cal. Code Regs. tit. 16, § 411. New York wants 1.75 inches under 8 NYCRR § 68.13. Ohio and Texas both allow 1.5 to 2 inches, under Ohio Rev. Code § 4733.14 and 22 TAC § 137.31. Ten distinct diameter specifications across the 51 US jurisdictions, and not one of them says anything about whose hands moved the mouse.
What they say instead — in different words, jurisdiction to jurisdiction — is that the sealing professional must have been in responsible charge. That is the load-bearing concept, and it is a statement about control and review, not production. If you have wondered what a PE stamp can and cannot cover, the boundary is drawn by what you personally directed and verified, not by what you personally drew.
This is not a novel reading invented to accommodate AI. Sealed work has essentially never been produced entirely by the sealing engineer. A junior drafts. A detailer redlines. A geotechnical subconsultant hands you bearing capacities. A fabricator's software generates connection designs. You review, you direct, you accept, you seal. The reason boards require both a seal and a signature is that one identifies the credential and the other binds a specific human being to a specific document on a specific date — Washington's board rules on professional stamps and seals still describe that binding in its most literal form, a signature in permanent ink across the face of the seal.
The AI question is a delegation question
Once you see the seal as an answerability claim, generative tooling stops being a category problem and becomes a familiar one: how much do you trust this delegate, and how do you verify their output?
Every engineer already has a calibrated answer for human delegates. You know which drafter puts dimension strings on the wrong side. You know the subconsultant whose appendices need checking. You review differently for a first-year EIT than for a colleague with fifteen years in the same building type. That calibration is professional judgment, and it is the thing your license actually certifies you to exercise.
What generative tools break is the calibration, not the framework.
Why AI output is harder to review than a junior's
A junior engineer's errors are legible. They cluster. Someone who does not yet understand lateral load paths produces a plan set that is wrong in a patterned, diagnosable way — you spot one symptom and you know to check five related things. Human error correlates with human understanding, which means finding one mistake tells you where the others live.
Generative output does not behave that way. It is optimized to be plausible, which is not the same as correct, and its failures do not track any underlying misunderstanding you can model. A schedule can be internally consistent and reference a member size that does not exist. Notes can be perfectly conventional in tone while contradicting the detail they annotate. A calculation can carry the right units, the right order of magnitude, and the wrong governing load case. The output has the surface texture of competent work everywhere, including where it is wrong.
Human review is heavily surface-driven. We scan for things that look off. When nothing looks off, review speeds up — not laziness, but the adaptive heuristic that makes reviewing a hundred sheets possible at all. Feed that heuristic output with no tells and it stops working while continuing to feel like it is working.
So the honest conclusion runs opposite to the convenient one. AI-assisted production does not reduce your review burden. It increases it, because it strips out the error signals your review process was implicitly tuned to detect.
The objection worth taking seriously
The strongest counter-argument is not that AI is unreliable. It is that responsible charge has always contained a fiction, and AI exposes the fiction rather than creating it.
The objection runs like this: in a great many practices, production volume already exceeds what genuine line-by-line review would permit, and principals seal sheets they have spent minutes with. Anyone who has worked a deadline knows the pattern. If we insist that AI-assisted work demands rigorous personal review, we are holding a new tool to a standard the profession does not consistently meet with its existing ones. That is a fair hit, and any argument that dodges it is not serious.
Two responses.
First, the existing gap is bounded by the throughput of human production. A firm can only generate so many sheets before the review deficit shows up in its own economics. Generative tooling removes that governor. What a small office can produce is no longer limited by drafting hours, so the gap between what gets sealed and what gets reviewed can widen without anything internal signaling that it has.
Second — and this is the part that should concentrate the mind — an existing weakness in professional review is not a defense. It is a description of exposure. When something fails, the question a board asks — and the 69 member licensing boards NCEES represents across the states and territories all exist to ask it — is what you did before applying the seal, and "this is how the industry operates" has never been a satisfying answer. Knowing when a seal is legally required in the first place is table stakes; knowing what you are personally asserting each time you apply one is the professional obligation.
The real risk is not AI drafting. It is unreviewed sealing.
Strip the technology out and the failure mode is ancient: a seal applied to work nobody meaningfully checked. That has been the shape of the problem for as long as there have been seals — schedule pressure, an over-extended principal, a plan-stamping arrangement, or outright misuse of a professional seal by someone not entitled to apply it. Generative tools are the newest route to that failure. They are not a new kind of failure.
The corollary matters for where you spend attention. Boards converged on permitting electronic seals — all 51 US jurisdictions now allow digital or electronic seals in some form — and that transition mostly went fine, because the profession worried about the right thing. It asked how to bind identity to a document and detect tampering, questions with concrete technical answers. The difference between an electronic signature and a cryptographic digital signature turned out to be the load-bearing detail, and once firms understood how the underlying cryptography establishes integrity, the anxiety about whether an electronic seal was as good as a wet one resolved on its own. The mechanics of a compliant electronic seal are settled; the mechanics of adequate review are not.
The AI conversation deserves the same discipline. Not "is it allowed" — the framework already answers that — but "what does adequate review look like when the output has no tells."
What responsible charge has to mean now
Review by independent derivation, not by inspection. When output is uniformly plausible, reading it is not review. For anything governing — load paths, member sizing, capacity checks, controlling dimensions — the check has to be an independent path to the same number. An order-of-magnitude sanity check, a hand calc on the critical case, a second method that should converge. If your process is fundamentally "read it and see if anything looks wrong," it is not calibrated for this input.
Know which decisions are load-bearing and own those personally. Not every line carries equal consequence. The professional act is identifying the small set of decisions where being wrong is unrecoverable and confirming those yourself, in detail, every time. Delegate the rest with appropriate spot-checking, as you always have.
Document the review, not just the result. If your file shows only a sealed output, you have no record of the judgment you exercised. Keep the check calcs, the markups, the notes on what you rejected and why. It becomes essential when the production tool cannot itself testify to what it did or why.
Treat seal application as a deliberate act with friction. A seal should require a decision, not a keystroke. Whatever your workflow — and there are several ways to apply a seal to PDF drawings — the moment of sealing should be when you consciously affirm responsible charge, not a batch operation at the end of a long day.
Keep the credential side clean. None of the above matters if the mark itself is non-compliant or the license behind it has lapsed. Requirements diverge more than people expect, which is why it is worth checking how seal requirements differ state to state rather than assuming your home jurisdiction's rules travel, then confirming against the board itself for each state you seal in, such as the Florida Board of Professional Engineers. A Texas engineer seal and a California engineer seal are not interchangeable objects. Ohio requires 30 PDH per two years; New York requires 36 PDH per three years; Texas requires 15 PDH per year; Colorado has no continuing-education requirement for engineers at all. If you seal across jurisdictions, that variance is yours to track. The same discipline applies to architect seals and land surveyor seals, which follow their own board rules.
Do this today
Pull the last set you sealed. Not one you are proud of — the last one. For each governing decision on it, write down how you personally verified that number. Not who produced it. How you confirmed it.
If you can do that for every load-bearing decision, your review process will survive AI-assisted production, because it never depended on error signals in the first place. If you cannot — and most engineers cannot, for at least a few items — you have just located the exact place where new tooling will hurt you, and located it before a board does.
Then do the boring housekeeping. Confirm the mark you apply matches your current jurisdiction's specification against the PE stamp requirements guide, or see how the geometry renders in the free stamp preview tool, no signup required. Verify your status and that of anyone sealing under your firm's name through a state license lookup — a lapse you did not notice is a much duller way to lose a career than anything an algorithm will do to you.
The tools will keep getting better at drawing. They will never answer for the drawing. That gap is not a gap in the technology. It is the definition of your profession, and it is the only part of the job that was ever actually yours.