Most of us have tried out different artificial intelligence (AI) chat tools, typing in prompts and watching paragraphs appear on the page. You probably thought, “Wow. That’s actually not bad.” You may have even thought, “I wonder if it could write my specs for me.” What could go wrong?

Confidence Is the Problem

Traditional bad specs are usually bad in obvious ways, such as a section reference or thickness table that never got updated, or maybe a product name is used where a generic description should appear. Those problems are at least visible, if you look.

What makes AI-generated specs bad is less obvious. The problem is that AI tools are fundamentally pattern-matching engines—in our example, producing text that sounds like a mechanical insulation specification, assembled from everything the tool has ever “read” on the subject. AI results do tend to come with a warning to verify results. However, according to a recent Clear Spark Digital study, of the 80% of Americans making decisions based on AI, only 17% are checking the results.¹

A Chilling Example

To illustrate, here is an abbreviated prompt:

Write a mechanical insulation specification for a chilled water system operating at 44°F in a conditioned interior space. Include material type, applicable ASTM standards, insulation thickness, vapor retarder requirements, and jacketing.

Here is a condensed snippet from the AI-generated response:

PIPE INSULATION—CHILLEDWATER SYSTEMS
Material: Flexible elastomeric foam insulation per ASTM C534-19.
Thickness: 1 inch for all pipe sizes up to 4 inches.
Vapor retarder: factory-applied per ASTM C921.
Vapor retarder jacket: SSL per ASTM C921.

Looks reasonable, right? Let’s review it, with help from members of NIA’s Technical Information Committee.

• ASTM C534 covers both sheet (Type I) and tubular (Type II) insulation. A system-level spec should explicitly require Type II for pipe applications. That distinction matters for quality control and for driving what the contractor installs.
• 1-inch thickness for all pipe sizes up to 4 inches? Depending on the pipe size, fluid temperature, and climate zone, ASHRAE 90.1 may require considerably more to prevent condensation. A universal 1-inch spec can sail through submittal review yet result in an installation that is not code compliant.
• ASTM C921 governs properties of jacketing materials for thermal insulation—not vapor retarder performance criteria for elastomeric foam. This is a hallucinated misapplication of a real standard number.
• Vapor retarder jacket specified as “SSL” with no perm rating or film thickness requirement? The AI tool produced generic jacketing language but omitted the performance criteria that would make it enforceable. ASTM C921 defines jacketing properties. Without a specified maximum perm rating (typically 0.02 perms for cold service), the spec is unverifiable at submittal and un-inspectable in the field. The AI tool wrote around the requirement without actually writing it.

Copy and Paste, with Better Grammar

At its core, the AI spec problem is a new spin on a very old sin: copying a spec without understanding it. Engineers have been recycling 20-year-old boilerplate sections forever. AI just makes the recycling faster and the errors harder to spot because nothing looks obviously dusty.

The spec that came from a 1999 project at least has the virtue of having worked on a 1999 project. The AI spec has the virtue of having been generated very quickly, from a statistical average of everything on the internet. Whether that statistical average reflects current requirements, product availability, or a standard that actually exists is a separate question.

How to Use These Tools without Getting Burned

AI tools can help you organize a spec outline, draft explanatory language around performance requirements, or generate a checklist of items to verify. What they cannot do is replace the judgment of someone who knows what ASHRAE 90.1 Table 6.8.2 actually says, which guidelines are current, or why certain materials and service temperatures above 220°F are an unhappy combination.

Here are a few practical rules if you’re going to experiment with AI-assisted specs.

• Verify every standard citation. If the AI writes “ASTM C547” or “ASHRAE 90.1,” look it up. Confirm the edition year is current and that the standard actually applies to what’s being specified.
• Check every thickness against your actual project requirements. Don’t accept a flat table. Run the numbers for the climate zone, operating temperatures, and pipe sizes involved.
• Be especially suspicious of anything that sounds like a proprietary guideline or association standard. AI tools are creative about inventing citations that don’t exist.
• Treat AI output as a first draft from a smart intern who writes beautifully but has never been on a jobsite.

A good spec isn’t just a list of materials and standards. It brings together the accumulated judgment of people who understand how insulation actually behaves in the field. AI may have read every spec ever written, but it has never been in the mechanical room. When you lean too heavily on it, you risk producing a document that satisfies a checklist without capturing the intent. The spec is only as good as the understanding behind it, so make sure that understanding is yours.

Reference

1. www.natlawreview.com/press-releases/nearly-80-americans-make-decisions-based-ai-only-17-consistently-check