What a Cursory Review Cannot Catch
In July 2024 the NSPE Board of Ethical Review published Case 24-2, on the use of artificial intelligence in engineering practice. What I want to pull out of it is not the AI holding, which has been summarized plenty. It is two of the defects that got through, because both of them are the kind a faster look was never going to find.
The situation first. Engineer A is a licensed environmental engineer, several years of experience, technically strong, less confident writing. Previously they had relied on their mentor and supervisor, Engineer B, for guidance and quality assurance review of their drafts. Engineer B retired.
Facing a report and a set of design documents with nobody left to check the work, Engineer A used open-source AI to draft the report and AI-assisted drafting tools to generate preliminary plans and specifications. On the report they were careful: a thorough review, key facts cross-checked against professional journal articles, search-engine queries run to confirm the phrasing was original, wording adjusted. On the drawings they completed what the case calls "a cursory review" and adjusted certain elements to align with site-specific conditions.
The client's reaction to the report was that it "read as if written by two different authors but was otherwise satisfactory." On the drawings the case says the client "noticed several issues with the AI-generated design documents, including misaligned dimensions and an omission of key safety features required by local regulations."
Several issues, including those two. The case does not enumerate the rest, so I have no idea what else was in there. But those two are named, and they are worth sitting with, because of what they have in common.
Neither one rewards a faster look
A cursory review is a looking review. You open the sheet, you scan it, things appear to be where they should be, you move on. That method finds a great deal. It is poorly suited to both of these.
Misaligned dimensions. The case does not define the phrase, so I am going to tell you the reading I have in mind and you can discount it accordingly. It can mean dimensions that do not physically line up with the geometry they annotate, and that version is visible: a witness line landing off the face it claims is something you can see. The version I find more interesting is the string that does not close. Every individual dimension on it is a plausible number, correctly formatted, in the right place. The defect is in none of them. It is that adding the individual dimensions along the wall does not give you the overall dimension carried at the outside.
That one has no visual signature. Nothing is crooked, nothing is missing, no number looks odd on its own. You find it by adding, and adding is not looking. It is also the sort of thing worth checking on any generated drawing set, because whether the numbers are constrained to sum depends entirely on how the tool produced them, and a set that was assembled rather than derived from geometry has nothing enforcing it.
An omitted safety feature is harder, for a reason that has nothing to do with AI. You cannot see an absence.
Be precise about what is and is not visible here, because the obvious version of that claim is too strong. The triggering condition is usually drawn. A level change is on the plan and on the section, and a reviewer scanning for guard conditions will find it. What carries no mark is the response. A guardrail that should be there and is not leaves the drawing looking like an ordinary level change, and a required clearance that was never held looks like a layout.
So the reviewer has to arrive already carrying the requirement, then go looking for whether it was answered. That is a checklist run against a body of requirements, and it is a different activity from reading a drawing. The case's phrasing points at why it is hard: the omitted features are ones "required by local regulations." When the Board restates the facts later it writes "key safety features (including those necessary for compliance with local regulations)," which makes local compliance a subset rather than the whole, so I do not want to lean on that qualifier too heavily. But to whatever extent a requirement is jurisdictional, the reviewer either has that jurisdiction loaded or does not, and no amount of care with the drawing substitutes for it.
One defect gets through because the numbers look fine. The other gets through because there is nothing to look at. Whatever else was in that set, those two are not going to yield to reading faster.
What the board held
Briefly, because it is well covered elsewhere and it is not the point of this piece.
Using AI was not the violation. The Board has a long record of welcoming tools, and it walks its own precedent to say so. Case 90-6, about thirty five years ago, held that an engineer could sign and seal documents prepared with CADD. Case 98-3 dealt with a mailed solicitation for a CD-ROM promising that "specifying, designing and costing out any construction project is as easy as pointing and clicking your mouse," with the line "never designed a highway before? No problem." The Board's conclusion there was that technology "must never be a replacement of a substitute for engineering judgment." The typo is in the original.
Disclosure was not required either: "Similar to other software used in the design or detailing process, Engineer A has no professional or ethical obligation to disclose AI use to Client W (unless such disclosure is required under Engineer A's contract with Client W)." The Board adds that transparency is nonetheless favored where AI plays a substantial role.
What was unethical was proceeding to seal: "Engineer A's misuse of the tool, by failing to maintain Responsible Charge over the AI tool and its output before sealing the document and providing it to Client W, was unethical."
On the report the finding was mixed, on two grounds. Engineer A did not obtain client permission before putting the client's information into an open-source tool, which the Board describes as "tantamount to placing the Client's private information in the public domain." And they did not document required technical citations, which the Board ties to Code III.9 on giving credit. The careful review did not cure either one, because neither was a review problem.
The sentence that does the work
Responsible Charge has a written definition. NSPE anchors it to Position Statement No. 10-1778: "being actively engaged in the engineering process, from conception to completion," with engineering decisions personally made by the professional engineer or by others under their supervisory direction and control.
Then this:
"Reviewing drawings or documents after preparation without involvement in the design and development process does not satisfy the definition of Responsible Charge."
That is stricter than most people's working assumption, and it is worth reading by anyone who thinks about review for a living, which includes me. It does not say after-the-fact review is worthless. It says it is not the act the seal certifies. Those are two different jobs and the profession wrote down which one carries the stamp, well before the current argument started.
The Board goes further in a sentence I nearly left out because it is the least convenient one available. The engineer in Responsible Charge, it says, is required to provide "an experienced-based quality assurance review, engaging in critical discussions, mentorship, and professional development," and it closes that sentence by calling those elements that AI cannot replicate.
Mentorship is in that list. Which brings it back to Engineer B.
Texas got to the same place and declined to write a rule
On November 14, 2024, the Texas Board of Professional Engineers and Land Surveyors took up the same question, having received "numerous inquiries from licensees regarding the use of AI software and how it fits into the Board's regulatory framework."
Three points of guidance. AI "should not be relied upon without oversight and review by the licensee," and "licensees are ultimately responsible for any work product they sign and seal," citing Board Rules 137.33(b) and 137.55(b). Licensees "should not use AI software to work on a project in which they do not already have established competence through education or experience," citing Rule 137.59. And licensees "should also be aware of how any AI software utilizes and stores data that is provided to it," citing Rule 137.61, which lands on the same confidentiality point NSPE reached from a different direction.
Then the conclusion: "No new Policy Advisory Opinion will be developed for this request as the Act and Board rules adequately address the use of artificial intelligence software at this time." The Board adds that it "will continue to monitor artificial intelligence software and its impact on the Board's licensees and update and amend this policy statement as warranted," so this is a current position rather than a closed file.
A state board looked at AI and found its existing rules sufficient. That is a claim that the obligation attached to a seal already covered the question.
Source: https://www.pels.texas.gov/nm/2024/pao-71-response.pdf
Why any of this is live in 2026
Both documents are from 2024. Neither is new. What has changed is how common Engineer A's situation is becoming, and the reason is not what you would guess.
The ACEC Research Institute published its Q2 2026 Engineering Business Sentiment survey on May 21, from 583 firm executives. Two figures matter here.
The staffing squeeze is easing: "only 33% of firms turned down work due to staffing shortages in the past six months, down from 51% in Q4 2024." And a cost is rising: "One emerging pressure point is healthcare costs. Seventy-eight percent of firms saw insurance premium increases at their last renewal, with an average increase of 14%, and 92% of respondents expressed at least some concern."
Board chair Steve Lefton: "The sector is shifting from a growth posture driven by demand and labor constraints to one defined by cautious expansion, cost management, and strategic selectivity. Firms are healthy and they're planning to grow with greater discipline and a closer eye on the horizon."
Cost management in a professional services firm rarely arrives as layoffs. Hiring intentions in this survey are strong. It arrives as pressure on hours per project, and inside a project the internal check is the only task with no external deadline attached to it. The client has a submittal date. The permit has a filing date. The coordination meeting is on other people's calendars and they show up to it. The internal review is a date the firm set for itself, and when the week goes sideways it is the one commitment that moves without a phone call from outside the building.
It also has no artifact when it works. Nobody writes a thank you note for the error that did not occur, so the honest internal accounting of a shortened review is that you did it four times and nothing went wrong.
Engineer A did not lose their reviewer to cost management. They lost him to retirement. The position is the same either way: work due, no second set of eyes, and a tool that will produce something that looks finished. And the two defects that got through are the two that a faster look was never going to catch.
What is wrong with all of this
A BER case is not necessarily hypothetical. It is tempting to assume these are all constructed teaching examples and that Engineer A never existed. NSPE's own note says otherwise: the Board "considers ethical cases involving either real or hypothetical matters submitted to it from NSPE members, other engineers, public officials, and members of the public." The facts are submitted to the Board rather than written by it, and the note says only that they "do not necessarily represent all of the pertinent facts submitted to or reviewed by the BER and shall not be considered a direct evaluation of any real matter that may have been submitted." So I cannot tell you whether this happened. Neither can anyone outside NSPE. The opinion is also "for educational purposes only and is not legal advice," in its own words.
The ACEC survey cuts against my framing in places, and I should say where. Sixty four percent of those firms expect hiring to increase over the next twelve months. Forty nine percent hold a year or more of backlog, at a median of eleven months. ACEC's own framing of the healthcare figure adds that most firms, 68 percent, report no change to hiring plans as a result. That is not a picture of firms in retreat, and any reading of mine that implies engineering is cutting back is going further than the data. The narrow claim I would defend is that cost management and hours-per-project pressure can coexist with strong hiring, not that one is causing the other.
The two ACEC comparisons run over different windows. The 33 percent against 51 percent spans Q4 2024 to Q2 2026. Other figures in the survey are quarter over quarter. They are not simultaneous measurements.
ACEC figures are sentiment, not dollars, and I have the summary only. The full report sits behind a member portal, so these are the numbers the association chose to publish, without the question wording behind them.
The connection between the survey and the case is mine. ACEC does not measure review practice. The ethics opinion says nothing about market conditions. I am putting two documents next to each other because the situations rhyme, and that is an argument rather than a finding.
The case does not define "misaligned dimensions," and my reading of it is the load-bearing part of this article. The phrase appears three times in the case and is never elaborated. It can mean dimensions that do not line up with the geometry they annotate, which would be a visible defect. The string-that-does-not-close reading is mine, flagged as mine where I use it, and if the submitted facts meant the other thing then that section describes a real failure mode that is not the one the case is about.
The Board thinks these were catchable, and says so. Its third use of the phrase is the one that cuts hardest against this piece: "failure to detect misaligned dimensions and omitted safety features further indicates that Engineer A did not exercise sufficient diligence." So the Board's position is that adequate diligence would have found both. I do not think that contradicts the argument here, because diligence is not the same thing as reading faster, and the whole point is that these two require a different method rather than more speed. But it is the Board's view, it is in the same document, and a reader deserves it from me rather than from the comments.
The case names two defects among "several issues" it does not enumerate. So nothing here is a claim about what the complete defect list was, or about the two named ones being representative.
The dimension and guardrail explanations are how the work goes, not citations. The case names the defects and does not explain why they are hard to catch. That part is mine. And it is worth being exact about one thing in it, because the stronger version is wrong: a missing guardrail does not leave no trace. The level change that triggers the requirement is drawn, and it is exactly what a reviewer scans for. What carries no mark is the response, not the condition.