From Black Box to Glass Box: Unlocking the True Potential of AI in Construction Disputes

Construction disputes consume billions annually. A delayed handover in Houston, a defective foundation in Denver, cost overruns on a London tunnel—these conflicts drain resources, destroy partnerships, and turn collaborative projects into adversarial battlegrounds. For decades, resolving these disputes has been expensive and slow: dueling experts, mountains of documentation, and outcomes that often feel more political than factual.

Artificial Intelligence promises to change everything. Feed it a million contracts, tens of thousands of delay claims, centuries of case law, case specific evidence and arguments, and it will predict outcomes with mathematical precision. No more waiting six months for a forensic scheduler's report. No more paying high hourly rates for expert testimony. Just clean, fast, objective answers.

Except there's nothing clean or objective about it.

The construction industry stands on the edge of a dangerous misstep—adopting AI systems that promise efficiency while quietly undermining the foundational principles of fair dispute resolution. The risk isn't that AI will make mistakes. The risk is that it will make systematic errors look like scientific truth, turning historical prejudices into algorithmic certainties that nobody can challenge.

The Data Doesn't Show What Really Happens

Every AI model learns from historical data. But in construction disputes, the historical record is fundamentally incomplete—and that incompleteness isn't random.

The Settlement Gap is the core problem. Industry estimates suggest 85-90% of construction disputes settle privately or through confidential arbitration. These resolutions never become public. What does enter the public record? The disasters: disputes so contentious, so broken, that parties chose to endure years of litigation rather than negotiate.

Train an AI on this data and you're not teaching it about construction disputes. You're teaching it about construction warfare.

The model learns that delay equals litigation. That an RFI clarification signals trouble. That any scope ambiguity leads to court. It has no exposure to the thousands of projects where contractors and clients worked through delays collaboratively, where variations were negotiated fairly, where disputes were resolved with a handshake.

This creates a perverse outcome: the AI becomes the pessimist in the room, interpreting normal project friction through the lens of catastrophic failure. A subcontractor's legitimate request for a time extension gets flagged as "high litigation risk" because it pattern-matches to the 10% of cases that exploded into arbitration—ignoring the 90% where similar requests were resolved amicably.

Then there's the context problem. AI models analyzing project communications rely heavily on sentiment analysis and keyword detection. An email from a site manager stating "I need your response by EOD or we're stopping work" might be classified as hostile or threatening. But anyone who's worked a construction site knows that urgency isn't hostility—it's Tuesday.

If the AI correlates "urgent tone" with "legal liability," it creates systematic unfairness against proactive communication. The project manager who clearly flags problems early gets penalized. The one who stays silent until the problem becomes critical gets rewarded.

The Human Cost: When Algorithms Anchor Decisions

But data bias is only half the problem. The other half is what happens when humans interact with AI predictions.

Picture this: A claims consultant receives notification of a delay claim. Before reading the contractor's submission, before reviewing the schedule, before checking weather records, they input the basic facts into an AI risk assessment tool. Seconds later: "Claim Validity: 8% probability. Recommended Action: Reject."

What happens next is predictable and dangerous.

The consultant doesn't approach the claim with an open mind. They approach it looking for reasons to confirm the AI's prediction. This is anchoring bias—the psychological tendency to rely heavily on the first piece of information received. The AI's 8% assessment becomes the truth, and everything that follows is unconsciously filtered through that lens.

An ambiguous contract clause? Interpreted unfavorably. A missing piece of documentation? Proof of weakness, not an administrative oversight. The AI prediction becomes a self-fulfilling prophecy.

This is compounded by automation complacency—the tendency to trust automated systems over personal judgment, especially under time pressure. When a sophisticated AI system delivers a definitive score, the temptation to accept it and move on is overwhelming.

The result? Legitimate claims get rejected not because they lack merit, but because an algorithm—trained on biased data, operating through opaque logic—said they were probably invalid, and nobody had the time or confidence to disagree with the machine.

The Courtroom Crisis: When AI Can't Explain Itself

In traditional dispute resolution, expert witnesses must explain their reasoning. A delay analyst walks through their logic: why they linked specific activities, what evidence supports their conclusions, which methodologies they applied. The opposing counsel can challenge every assumption. The logic is visible and testable.

Replace that expert with an AI model that outputs "Delay liability: Contractor 73%, Owner  27%" and ask "Why?" The AI cannot answer in human-intelligible terms—only by pointing to millions of weighted parameters across neural network layers.

This creates a legal crisis. Standards like Daubert require scientific evidence to be testable with known error rates.

An AI that claims to be 99% accurate but is 0% explainable is worthless in high-stakes disputes. Justice requires transparent reasoning that can be challenged and understood.

The Uncomfortable Truth About Human Experts

Before we crown AI as the villain, we need to acknowledge that traditional expert testimony isn't perfect either. Human experts bring their own biases—a contract manager who spent 20 years working for contractors develops different instincts than one who worked for owners.

The difference is that human bias can be interrogated. You can ask an expert about their career history, their assumptions, their methodology. The bias is visible, which means it's manageable.

AI bias is insidious precisely because it appears objective. The algorithm doesn't have a career history or personal preferences. It just has math. And because it's math, it carries an unearned credibility—the assumption that numbers are neutral, that probabilities are pure.

This is the core danger: AI launders historical unfairness into the appearance of mathematical certainty.

It takes decades of systematically biased dispute outcomes and crystallizes them into an algorithm that simply reproduces those patterns while calling them "predictions." The model isn't discovering truth. It's encoding injustice at scale.

The Path Forward: Glass Boxes, Not Black Boxes

None of this means AI has no place in construction dispute resolution. Speed matters. Efficiency matters. The ability to quickly analyze thousands of contracts or identify relevant precedents could genuinely improve outcomes.

But the industry must demand transparency as a precondition for adoption.

Enter the "Glass Box" approach: AI systems that prioritize explainability over raw predictive power. Instead of just generating scores, these systems provide audit trails:

• "Risk detected based on keyword frequency in contractor emails between March 15-April 2, specifically references to 'critical path' (23 occurrences) and 'your responsibility' (8 occurrences)"
• "Schedule analysis suggests potential concurrent delay based on similarity to Case ID 4482 and Case ID 8231"
• "Contract clause 14.3 language matches high-dispute-risk patterns in 67% of comparable projects"

This transforms AI from an invisible judge into a powerful research assistant. It surfaces relevant information, highlights potential issues, and draws attention to patterns—but leaves the final interpretation to humans who can weigh context, intent, and nuance.

Currently, no major LLM is a true 'Glass Box.' However, the industry is pivoting toward Retrieval-Augmented Generation (RAG) systems that force opaque models to show their work by citing specific source documents and data points—effectively placing a glass pane over the black box. Looking further ahead, emerging architectures like Liquid Neural Networks promise to replace the black box entirely. These adaptive systems use interpretable mathematical equations that change dynamically based on inputs, allowing researchers to trace exactly how the network processes information and reaches conclusions. Unlike traditional neural networks where decision-making happens across millions of inscrutable weighted connections, Liquid Neural Networks offer auditable, mathematical certainty—each computational step can be examined and verified, making them ideal candidates for applications demanding transparent, defensible analytical processes.

The tradeoff is real, but it's a feature, not a flaw. Glass Box systems are less computationally powerful than Black Box counterparts because they constrain their analysis to patterns humans can trace and verify—activating fewer neural connections and employing simpler computational pathways to enable transparency. A deep learning Black Box can identify correlations no human would spot, but those correlations cannot be independently verified, tested, or challenged. When opposing counsel cross-examines your methodology, "the AI said so" isn't defensible. The Glass Box's transparency transforms it from an inscrutable oracle into a defensible analytical tool where every input, assumption, and calculation can be explained and validated. In construction disputes, the critical question isn't whether Glass Box systems match Black Box computational power—it's whether any conclusion, no matter how sophisticated, has value if it cannot be explained, tested, and defended under adversarial scrutiny.

The construction industry must decide: Do we want maximum predictive accuracy in a system nobody can challenge, or slightly lower accuracy in a system that can be interrogated, tested, and refined?

For dispute resolution—where fairness matters as much as efficiency—the answer must be the latter.

Conclusion: What Kind of Industry Do We Want?

For decades, construction has struggled with adversarial relationships. Contractors and clients eye each other with suspicion. Every variation becomes a battle. This culture is expensive, exhausting, and increasingly unsustainable.

Many in the industry push for integrated project delivery, collaborative contracts, partnering approaches that emphasize shared goals over antagonistic positions. They want construction to become less adversarial, more cooperative.

But here's the paradox: if we train AI systems on the adversarial past—on the 10% of disputes that exploded into litigation rather than the 90% that were resolved collaboratively—we're encoding that antagonism into the future. We're building technology that assumes bad faith, that interprets ambiguity as manipulation, that sees every communication through the lens of conflict.

This is the real stakes of the AI debate. It's not just about accuracy or efficiency. It's about whether we use technology to escape the industry's adversarial patterns or to entrench them permanently.

As the industry integrates these powerful tools, one principle must be non-negotiable: If the AI cannot explain its reasoning to a judge, it has no business influencing the dispute.

Justice requires sunlight. Fairness demands transparency. The future of construction technology must be built with glass, not black boxes. The promise of AI isn't speed for its own sake—it's speed in service of better outcomes. But if we sacrifice transparency for efficiency, if we trade explainable reasoning for opaque predictions, we'll achieve speed at the cost of justice.

That's not progress. That's just faster unfairness.

---

Author, Sam Barakat, PE, Esq, FCARB, PSP resolves complex construction disputes as an expert witness, mediator, and arbitrator, bringing the unique perspective of dual credentials as both a Professional Engineer and licensed attorney. For over 27 years, he has provided technical expertise and dispute resolution services on projects ranging from highway infrastructure to industrial facilities with claims valued over $830M. As a Managing Director at GlassRatner Advisory & Capital Group and AAA Construction Panel arbitrator, Sam delivers defensible analysis and equitable outcomes in high-stakes construction matters. Sam may be contacted at sbarakat@glassratner.com.