Can a licensed engineer seal a drawing that an AI tool generated?

The question a board will ask is not whether a tool was involved but whether the work was prepared under the sealer's responsible charge — that is, whether the licensee exercised direct control and personal supervision over the engineering decisions embodied in the document. A licensee who defined the design criteria, made or verified the governing decisions, checked the outputs against independent expectations and can explain every value on the sheet is in a familiar position: engineers have sealed work produced by drafters and by analysis software for decades. A licensee who received a complete generated package from a tool they cannot interrogate and applied a seal to it is doing something the rules describe unfavourably, because the seal is a personal representation and there is no version of it that means 'the software said so.' The distinguishing test is whether the engineering judgment happened before the output or after it.

Is using AI a departure from the standard of care?

Using it is not; relying on it without verification is likely to be. The professional standard of care asks what a reasonably prudent practitioner would do under similar circumstances in the same locality at the same time, and that standard moves with practice. Tools have never been the subject of the inquiry — outcomes and process are. The realistic exposure runs in two directions at once. In the near term the risk is unverified reliance: an output adopted without an independent check that a prudent practitioner would have performed. Over a longer horizon the inverse becomes plausible, where failing to run an available check that has become routine looks like the omission. Both point to the same defensible practice: document what the tool produced, what you independently verified, and what you rejected.

Does an AI-assisted deliverable affect our errors and omissions coverage?

It can, and this is moving faster than most firms have noticed. Professional liability policies for design firms are being revised with AI-related conditions, exclusions and disclosure questions on the renewal application. Three specific things to check: whether the policy contains any exclusion or sublimit for claims arising from automated or generative tools; whether the application asks about AI use, because an inaccurate answer creates a rescission argument entirely separate from the underlying claim; and whether your subconsultant agreements require the same disclosure downstream, since a subconsultant's undisclosed use lands in your deliverable. Read the renewal application questions carefully before answering them firm-wide, since in most firms the honest answer is 'yes, in ways we have not inventoried.'

Do we have to tell the client that AI was used in producing their design?

Check the contract before you check your instincts, because many owner agreements already answer it without mentioning AI. Standard-form and owner-drafted agreements commonly contain provisions on personnel, subcontracting, confidentiality of owner information, ownership of instruments of service, and warranties about how the work will be performed. Feeding owner-confidential program documents into an external tool can breach the confidentiality clause on its own terms, regardless of the design question. Public-sector and institutional owners increasingly add explicit AI clauses requiring disclosure or consent. The practical approach is to inventory what your active contracts already require, adopt a firm position on disclosure that you can apply consistently, and negotiate the clause openly on new work rather than discovering the answer during a claim.

What is the exposure for the firm as opposed to the individual licensee?

They are separate tracks that run at the same time, and firms consistently underestimate the first one. The individual licensee holds the seal and faces board discipline personally — reprimand, conditions, suspension or revocation — and that proceeding turns on responsible charge, not on the firm's policies, although a firm policy that pressured the licensee will be visible in the record. The firm faces the civil claim, the contract claim from the owner, and in many jurisdictions its own certificate of authorisation to offer professional services, which can itself be subject to board action. There is also an internal fracture worth planning for: a firm production mandate that encourages generated deliverables while the sealing engineer bears the personal licensure consequence creates a conflict that surfaces at the worst possible moment. Resolve it in writing, by giving the sealing licensee unambiguous authority to refuse.

AI and the PE Seal 2026: Responsible Charge When the Model Drew It

Why This Is Not the Same as CAD or FEA

Design professionals have used software to produce work product for forty years, and the profession settled that question long ago: the tool does arithmetic and drafting, the licensee makes the engineering decisions, and the seal covers the decisions. Analysis packages are deterministic, their assumptions are declared, and a competent reviewer can reproduce the result by hand for a governing case. That reproducibility is what made the supervision doctrine workable.

Generative tools break the arrangement in three specific ways. They produce plausible output for inputs outside their competence rather than failing visibly. They do not expose the assumption set behind a result, so there is nothing to check the way one checks a load path or a boundary condition. And they produce complete-looking deliverables — a full sheet, a specification section, a calculation package — which invites review-as-proofreading rather than review-as-engineering.

The doctrinal point is unchanged and worth stating plainly: no licensure framework has ever permitted sealing work you did not supervise, and none of them contain an exception for tools. What changed is only how easy it has become to arrive at a deliverable nobody supervised.

The Deliverable Chain, Stage by Stage

Responsible charge is judged across the whole production chain, not at the signature. Each stage below has a defensible use and a use that undermines the seal.

1. Criteria and code basis

Consistent with responsible charge: Using a model to assemble a candidate list of applicable code sections, jurisdictional amendments and reference standards for the licensee to verify against the adopting ordinance.

Undermines the seal: Accepting the generated code basis as the design basis. Locally adopted amendments and effective dates are exactly what a general-purpose model gets wrong, and the error propagates into every downstream decision.

2. Concept and layout generation

Consistent with responsible charge: Generating alternative layouts, framing schemes or routing options as candidates, with the licensee selecting and then developing the chosen scheme on the merits.

Undermines the seal: Advancing a generated layout to documents because it looks resolved. Generated geometry frequently satisfies the visible constraint and quietly violates the unstated ones — clearances, access, constructability, coordination with other disciplines.

3. Calculations and sizing

Consistent with responsible charge: Using the tool to set up or check arithmetic that the licensee has framed, with independent verification of governing cases by hand or by a validated package.

Undermines the seal: Accepting generated numerical results as the calculation of record. This is the stage where the difference between plausible and correct is invisible on the page, and it is the stage a board will focus on.

4. Drawing and detail production

Consistent with responsible charge: Generating boilerplate details, sheet setup, annotation and schedules from licensee-specified content, checked in the same way drafted work has always been checked.

Undermines the seal: Generated details whose provenance nobody can state. A detail that no one on the team can attribute to a source or a decision is unverifiable by definition, and it is the classic origin of a construction-phase failure.

5. Specifications

Consistent with responsible charge: Drafting and editing sections from the firm's master, with the licensee reconciling every product, standard and test reference against the drawings.

Undermines the seal: Generated specification text citing standards that were withdrawn, renumbered or never existed, and products that do not meet the performance the drawings assume. Specification fabrication is the most reliably reproducible AI failure in design work.

6. Seal and issue

Consistent with responsible charge: Sealing after a documented review in which the licensee can state the basis for every governing decision on the sheet and identify what they independently verified.

Undermines the seal: Sealing to a schedule. If the answer to 'why is this member this size' is a reference to the tool, the representation the seal makes is not accurate — and that is the finding, independent of whether the design turned out to be adequate.

The Plan-Stamping Analogy Boards Will Reach For

Licensure boards already have a well-developed offence that fits this fact pattern closely, and it is the one to reason from: plan stamping — applying a seal to work prepared by someone the licensee did not supervise. It is among the most consistently disciplined violations in the profession, and the reason is structural rather than punitive. The seal is the only mechanism by which a jurisdiction that cannot review every drawing obtains an accountable person, so a seal applied without supervision defeats the entire scheme.

A board evaluating an AI-produced deliverable will ask the same questions it asks in a stamping case: what did you do, when did you do it, what would you have caught, and can you explain the decisions. "A model produced it and I reviewed it" is structurally the same answer as "another firm produced it and I reviewed it" — which is the answer that loses those cases unless the review was substantive and documented.

What a defensible file contains

☐A stated design basis authored by the licensee, dated before the generated content exists
☐Independent verification of governing cases, with the check preserved rather than discarded
☐A record of what was generated, what was modified and what was rejected
☐Named reviewers with dates, not a single approval at issue
☐Confirmation that every cited standard and product was verified as current and applicable
☐The licensee's own explanation, in the file, of the decisions the seal covers

Firm-Level Policy That Actually Holds

Most design-firm AI policies are a paragraph about confidentiality attached to the employee handbook. That paragraph does not address the licensure exposure at all, because the exposure is personal to the sealer and arises from workflow rather than from data handling. A policy that holds up has four elements:

An explicit refusal right. The sealing licensee has unreviewable authority to decline to seal, to require additional verification, or to discard generated content, without schedule or utilisation consequences. Without this, the firm has created pressure that ends up in a board record.
Approved tools and prohibited uses, by stage. A blanket ban gets ignored, and a blanket permission is worse. Name the stages where generation is acceptable and the stages where output may not enter the deliverable at all — calculations of record and code-basis determinations being the usual prohibitions.
Confidentiality routing. Owner program documents, geotechnical reports, tenant information and security-sensitive drawings should not leave the firm's controlled environment. This is a contract question that exists whether or not the design question is resolved.
Flow-down to subconsultants. Your seal covers deliverables incorporating their work. If your agreements are silent, you are relying on their internal practice and inheriting the result.

Frequently Asked Questions

Our design software vendor added AI features we did not choose to adopt. Where does that leave us?

In the position of having to find out what they do, which is a real burden the vendor has transferred to you. Generative features are arriving inside established platforms as assistants, auto-completion, generated schedules and suggested details, and they are frequently enabled by default and framed as productivity rather than as analysis. Three practical steps: inventory which features in your production platforms are generative and whether they can be disabled at the deployment level; determine whether their output is visually distinguishable from user-authored content in the file, since an assistant-produced schedule that looks identical to a drafted one is invisible to your reviewers; and read the release notes on every update rather than only the ones you requested. The exposure is not that the feature exists — it is that the seal covers content whose origin the licensee did not know about.

Is it enough to add a disclaimer to the drawings saying AI tools were used?

No, and it may make things worse in two ways. A seal is an affirmative representation about supervision, and a note disclaiming responsibility for content within the sealed set is in tension with the representation itself — boards read that tension unfavourably, because the licensee is simultaneously claiming and disclaiming the same work. It is also unhelpful to the people who receive the drawings: a contractor cannot act on a general disclaimer, and building officials do not accept caveated sealed documents in any other context. If a portion of the work genuinely falls outside your responsible charge, the correct mechanism is a scope limitation defining what is and is not covered by the seal, with the excluded work sealed by whoever did supervise it — not a general note about tooling.

How is this different for architects and other licensed design disciplines?

The vocabulary differs, the structure does not. Architecture, surveying, landscape architecture and the engineering branches all operate under statutes with an equivalent supervision requirement — 'responsible control,' 'direct supervision,' 'responsible charge' — and all of them treat the seal as a personal representation about that supervision. Where meaningful differences appear is in what the deliverable does downstream: a surveyor's sealed plat has legal consequences for boundaries and title that a generated geometry error carries directly into the land records, and life-safety systems have inspection regimes that catch some errors and not others. The analysis in this article transfers across disciplines; the specific wording of your obligation is in your own board's statute and rules, which is where to read it rather than in a general article.

Can we defend a claim by pointing to the tool's own accuracy documentation?

It will not do the work you want it to do. Vendor documentation describes the tool's performance under the vendor's evaluation conditions, and vendor terms for these products almost invariably disclaim fitness for professional use and require the licensed professional to verify all output — which means the vendor's own contract is evidence against the reliance you are asserting. The standard-of-care question is about what a reasonably prudent practitioner would have done, and a prudent practitioner does not delegate a sealed judgment to a general-purpose tool on the strength of a benchmark. Where the documentation does help is narrower and worth preserving: it can support the reasonableness of using the tool at a particular stage, and it can support a claim against the vendor if the tool was represented as validated for something it was not.

The client is demanding AI-accelerated delivery at a reduced fee. How do we respond?

By separating the two things they have combined, because the schedule is negotiable and the seal is not. Generation genuinely compresses parts of the production chain — drafting, boilerplate, coordination checks, first-pass schedules — and it is fair to price that. It does not compress the verification, the code-basis determination or the engineering judgment, and those are the parts the seal actually covers. A fee that assumes the verification also disappeared is a fee that assumes an unsealed deliverable. The productive conversation names which stages are compressed and which are not, and it is worth having explicitly, because a firm that quietly absorbs the reduction ends up making the compression somewhere — and it will be in the review stage, which is the one that produces board proceedings.

Ask the Question That Decides It

Take the last sealed set your firm issued and pick three governing decisions from it — a member size, a system capacity, a code-basis call. Ask the sealing licensee to state the basis for each without opening the model.

If they can, responsible charge is intact and the tooling is a productivity question. If they cannot, the exposure is not hypothetical and it is not the firm's — it is on the person whose seal is on the sheet.