Why More AI Findings Are Not Better: False Positives in Construction Review

There is a dangerous metric emerging in construction AI:

“Our system found 1,247 issues.”

It sounds impressive.

But without knowing how many findings are valid, unique and actionable, the number tells us almost nothing.

Construction technology has seen this movie before.

BIM clash detection introduced a powerful ability to systematically test model geometry. It also introduced clash reports containing thousands of intersections.

Teams quickly learned that not every detected clash matters.

AI drawing review must learn the same lesson faster.

What is a false positive?

In simple terms, a false positive is a system-generated issue that a competent reviewer decides is not actually a problem.

Examples could include:

·       A penetration that is intentionally sleeved

·       A note that appears inconsistent but applies to a different condition

·       Two overlapping graphics that represent acceptable design intent

·       A duplicate finding already captured elsewhere

·       A standard trade coordination condition

·       A misread dimension

·       An issue resolved by a later revision

False positives are not unique to AI.

Rule-based systems, clash engines and automated QA tools all face them.

BIM research shows why relevance matters

A 2019 study on filtering irrelevant BIM clashes summarized prior research indicating that 50% or more of raw software-detected clashes could be irrelevant.

The researchers described a practical consequence: if every clash must be manually evaluated, the review burden can become too expensive, causing teams to selectively review or ignore reports.

A 2026 review of AI in BIM clash management found the same recurring issue across the literature: substantial volumes of false positives, low-risk clashes and results that do not deserve equal coordination attention.

This is not an argument against clash detection.

It is an argument for prioritization.

Precision and recall in construction language

Two concepts are useful.

Recall

Of all the real issues in the project, how many did the system find?

High recall means fewer misses.

Precision

Of everything the system reported, how many findings were actually valid?

High precision means less noise.

Construction teams usually want both.

But systems often trade one for the other.

A detector configured to report anything remotely suspicious may catch more true issues — but also produce more false positives.

A conservative system may produce cleaner output but miss more real problems.

Why raw issue count can be misleading

Consider a hypothetical drawing package containing 200 real coordination issues.

System A reports 500 findings and correctly identifies 150 of the 200 real issues.

Its precision is 30%.

System B reports 220 findings and correctly identifies 140 real issues.

Its precision is about 64%.

System A found ten more true issues.

But it also created 350 false positives versus 80 for System B.

Which system is better?

The answer depends on the cost of a miss and the cost of review.

That is exactly why raw issue count is a poor comparison metric.

Reviewer time turns false positives into money

Assume each finding takes three minutes to open, understand, inspect the evidence and accept or reject.

For 500 findings:

500 × 3 minutes = 1,500 minutes = 25 reviewer hours.

If only 150 are valid, the team spent 17.5 hours reviewing false positives.

At $100 per loaded reviewer hour, that is $1,750 of review effort for the false alerts on one scan.

Again, this is an illustrative calculation, not an industry benchmark.

But it shows the mechanism.

False positives are not just an accuracy statistic.

They are a workflow cost.

Alert fatigue changes behavior

The bigger risk is not the first false positive.

It is the 200th.

When a reviewer repeatedly opens low-value alerts, confidence drops.

Eventually, users:

·       Skim instead of investigate

·       Bulk-dismiss findings

·       Stop checking low-severity categories

·       Ignore the tool entirely

At that point, even true positives lose value because the system has trained the user not to trust it.

This phenomenon is familiar in other high-alert environments such as cybersecurity and healthcare. Construction coordination has the same human limitation: attention is finite.

Not every true clash is an actionable issue

Construction review has another nuance.

A system can be technically correct and still operationally unhelpful.

Suppose two modeled elements geometrically intersect.

That is a true geometric clash.

But the condition may be a planned penetration requiring a standard sleeve.

So it is not a false positive in the geometric sense.

It is a low-relevance finding in the project-management sense.

This suggests a construction AI system needs more than binary “issue / no issue” logic.

It needs context.

A better quality model for AI findings

A useful construction finding should be evaluated across at least six dimensions.

1. Validity

Is the underlying observation correct?

2. Relevance

Does it matter to the project?

3. Uniqueness

Is it a new issue or a duplicate?

4. Severity

What happens if it is not resolved?

5. Evidence

Can the reviewer quickly verify the finding?

6. Actionability

Is there a clear next step?

A high-quality system should optimize the complete package, not just detection.

Evidence changes the economics of review

Suppose two systems have identical precision.

System A says:

“Possible inconsistency detected.”

System B says:

“Door D104 is shown as 90-minute rated on LS201 but 45-minute rated in the door schedule A601. See highlighted evidence.”

Even if both require human review, System B reduces the cost of verification.

That means AI quality should also be measured as:

Time to validate a finding.

This may be as important as precision itself.

Deduplication matters

Large drawing sets often contain repeated representations of the same underlying issue.

A single coordination problem might appear in:

·       Plan

·       Section

·       Detail

·       Schedule

·       Another discipline

If the system reports five separate findings, the issue count rises but reviewer value does not.

Good systems should group evidence where the findings share the same root condition.

Severity must be calibrated

If everything is “High,” nothing is high.

Severity should reflect a consistent risk framework.

Possible dimensions include:

·       Safety/life safety

·       Structural impact

·       Cost exposure

·       Schedule exposure

·       Constructability

·       Scope ambiguity

·       Ease of correction

·       Proximity to installation

Teams should also be able to override automated severity based on project context.

What about false negatives?

Focusing on false positives should not hide the opposite failure.

A false negative is a real issue the system fails to detect.

This can be more expensive than a false positive.

That is why aggressive filtering can be dangerous.

The solution is not simply “show fewer issues.”

The solution is calibrated detection with transparent confidence and evidence.

A practical scorecard for evaluating AI drawing review

Instead of asking vendors “How many issues do you find?”, ask:

1. What is your measured precision on representative projects? 2. How do you define a valid finding? 3. How do you handle duplicates? 4. Can reviewers see source evidence? 5. Can findings be accepted/rejected? 6. Do you track reviewer verdicts? 7. Can the system explain why it raised the issue? 8. How do you measure misses? 9. Are results benchmarked by discipline and issue type? 10. How much reviewer time is required per accepted finding?

These questions reveal more than a demo showing hundreds of alerts.

The metric that matters: useful findings per reviewer hour

Imagine two tools.

Tool A finds 300 valid issues but requires 100 reviewer hours.

Tool B finds 260 valid issues but requires 25 reviewer hours.

Depending on project risk, Tool B may create more usable value.

This leads to a better operational metric:

Valid, actionable findings confirmed per reviewer hour.

Construction technology should be judged by how much decision-quality it creates, not how much output it generates.

The takeaway

The future of construction AI will not be won by the product with the largest issue count.

It will be won by products that make reviewers faster and more confident.

That requires high precision, strong evidence, sensible severity, duplicate control and human feedback.

BIM clash detection taught the industry that detecting everything can produce overwhelming noise.

AI drawing review should treat that lesson as a design requirement from day one.

Sources and further reading

1. Applied Sciences — Filtering of Irrelevant Clashes Detected by BIM Software. Peer-reviewed discussion of irrelevant clashes and the review burden created by raw clash reports. https://www.mdpi.com/2076-3417/9/24/5324
2. Hu & Castro-Lacouture — Clash Relevance Prediction. Research questioning raw clash precision and using ML to classify relevance. https://pure.psu.edu/en/publications/clash-relevance-prediction-based-on-machine-learning/
3. Journal of Computational Design and Engineering — BIM clash classification. 2026 article noting excessive irrelevant/non-critical clashes and the need for constructability classification. https://academic.oup.com/jcde/article/13/4/227/8537782
4. Buildings — Artificial Intelligence in BIM Clash Management. 2026 review summarizing false-positive/low-priority clash volumes across studies. https://www.mdpi.com/2075-5309/16/13/2623
5. Automation in Construction — Holistic clash detection improvement. Research using dependency networks to reduce irrelevant and duplicate clash burden. https://www.sciencedirect.com/science/article/abs/pii/S0926580519301980