AI in AEC - What Crossed the Line and What Didn’t (Part 3 of 3)

📬 Intro: The Line Moved in May

I sat through an Autodesk Assistant demo in Revit two weeks ago. You type “how many doors are narrower than 90 cm” into a chat panel docked inside Revit 2027, and it builds the sheet for you. The Tech Preview started in April. It runs on the same Model Context Protocol the new Claude agents use.

Same morning, I read that an Oregon judge sanctioned two lawyers $110,000 for filing briefs with 23 AI-fabricated citations. Largest AI hallucination penalty in American legal history…so far. A federal court database tracking these AI error cases is past 1,300.

Both happened in the last 30 days. Part 1 was the moves. Part 2 was the agents. Part 3 is the map.

🌍 Field Notes: What Crossed the Line in May (and What Didn’t)

🛠️ Revit Just Got a Chat Panel

Autodesk Assistant in Revit dropped as Tech Preview on April 22. It is the first Revit version with a built-in MCP server, on the open protocol Anthropic released last year. The panel docks inside Revit 2027 and takes natural-language commands across six buckets: model queries, sheets, rooms, schedules, exports, and element operations.

A designer types “create a door schedule sorted by level” and Revit builds it. Not a documentation lookup.

Revit just got a write API in plain English. The chat panel has evolved from ‘oracle’ to ‘operator’.

⚖️ The Case That Doesn’t Exist

The Oregon sanction is from a winery dispute called Couvrette v. Wisnovsky. Two lawyers filed three briefs with 23 fabricated citations and eight invented quotations. The judge handed down the largest AI hallucination penalty in American legal history (so far).

The Suprmind May benchmark explains the pattern. Gemini 3 Pro hallucinates 88% of the time on questions it does not know. GPT-5.5 hits 86%. Only Claude 4.1 Opus comes close to 0%, and only because it refuses to answer when uncertain. For the time being default model behavior is to invent a confident answer.

A chatbot that will make up a court case to a federal judge will make up an ASHRAE chapter to your code reviewer.

🤖 Agents That Got Smarter This Month

Anthropic shipped three new Managed Agents features three weeks ago on May 7. Dreaming reviews session memory between runs. Outcomes grades a run and pings a webhook. Multiagent orchestration delegates from a lead agent to parallel specialists. (Netflix is running multiagent in production).

Then on May 25, Anthropic previewed Mythos. The model already found 6,202 high or critical-severity vulnerabilities across 1,000+ open-source projects. Anthropic kept it locked to Project Glasswing partners. The company says no firm, including itself, has built safeguards strong enough. All Skynet jokes aside, now is a great time to ensure your firm has air-gapped physical backups of critical data and history.

My takeaway: agents got smarter this month, guardrails did not. Plan accordingly.

A MESSAGE FROM BUILDINGPLAQUES.COM

The Better Way To Present Your LEED Certified Project to The World

Using buildingplaques.com online Plaque Designer, you can quickly and easily design and order customized LEED recognition plaques that features the details of your specific LEED certified project.

Our architectural grade anodized aluminum signs are available in multiple sizes, and identify the elements of your design that earn LEED certification and acknowledge key project members, all while being less expensive than comparable LEED recognition plaque options.

Ready to create your own LEED recognition plaque?

Check Out BUILDINGPLAQUES.COM

🏛 Built to Last

The Revit panel changes the model on command. Managed Agents grade themselves. Mythos finds vulnerabilities no human team would catch. And two lawyers still filed 23 fake citations to a federal judge, because the same tool that drafts the brief makes up the case.

Add “cite the exact sentence on the exact page” to every prompt. Every code lookup gets a second model checking it. Be sure drawing review catches the conflict before the change order. Build the habit before the next preview ships.

📋 LEED consultants: Add “cite the exact section” to every code question. Verify before you submit.

🏗️ Project architects: Autodesk Assistant in Revit Tech Preview is live. Get the chat panel into your workflow before the production release.

📋 Project managers: Cite the source on every AI output before it leaves your inbox. The Oregon sanction was three briefs.

🏢 Principals: Mythos shows what is coming. Build the verification standard now, not after the first malpractice notice.

Limited Sponsorship Opportunities Available

Green Building Matters Podcast is now exploring a select number of brand partnerships with organizations aligned with the future of sustainable building.
Reach the professionals shaping the industry—and do it with intention.

[email protected]

Now Open: Limited Brand Partnerships

✍️ Brian Bollinger, our Head Writer, helps sustainability professionals connect performance, workflow, and real-world results.

🌍️ Charlie Cichetti, our Fearless Leader and CEO of Skema and LEED Fellow, has guided more than 150,000 professionals as building practice continues to evolve.

Let’s Green Up together.

👉 [Explore Sustainability Credentials at GBES.com]
👉 [See what Skema’s building for architects]
👉 [Book a Waste2Zero audit before your next ESG Renewal]