At Cornell University, a team of researchers has built a quiet mirror for human judgment — a system that does not decide for us, but watches how we decide and asks whether our choices match our convictions. In an era when artificial intelligence is most often celebrated for replacing human reasoning, Interactive Explainable Ranking takes the opposite path: it holds us accountable to ourselves, surfacing the gap between the values we profess and the preferences we reveal. The tool, tested with film evaluators and teaching assistants, points toward a deeper question that technology rarely dares
Cornell AI System Exposes Gaps Between What People Say They Value and What They Choose
The algorithm doesn't decide. It watches you decide, then asks whether you're being honest with yourself.
Mark
Why does it matter if someone's choices contradict their stated values? Isn't that just how people are?
Mimi
It matters because decisions made this way tend to be unfair. If a teacher says creativity matters most but keeps rewarding technical precision, students don't know what they're actually being judged on. The contradiction isn't just personal—it cascades.
Mark
So the AI is basically a mirror.
Mimi
A very specific mirror. It only reflects the gap between what you said and what you did. It doesn't judge the gap or tell you which one is right.
Mark
In the pilot tests, did people get defensive when the system caught them being inconsistent?
Mimi
The source doesn't say. But the fact that they adjusted their criteria or identified hidden factors suggests they found it useful rather than accusatory. It gave them language for something they were already feeling.
Mark
Could this be used to manipulate people? To force them toward certain decisions?
Mimi
Only if someone weaponizes it. The tool itself is neutral—it just makes contradictions visible. What you do with that visibility is up to you.
Mark
What happens in a high-stakes situation, like hiring or admissions?
Mimi
That's the real test. The pilot was small and controlled. Scaling it to hundreds of applications, where bias is entrenched and the stakes are real, is a different problem entirely.
O Pulso
- Evaluation processes across academia, hiring, and consumer decisions are quietly undermined by a persistent gap between what people say they value and what their actual choices reveal.
- Cornell's Interactive Explainable Ranking system disrupts this comfortable inconsistency by flagging contradictions in real time, forcing evaluators to confront the distance between their stated criteria and their lived decisions.
- Pilot studies with film reviewers and teaching assistants showed the tool reducing emotional bias and producing scores that were both consistent across evaluators and aligned with prior independent assessments.
- Rather than automating judgment away, the system keeps humans at the center while making them answerable to their own declared logic — a rare design choice in a field racing toward automation.
- Presented at CHI 2026, the tool now points toward broader applications in any domain where fairness and rigor depend on evaluators who can honestly account for what they are actually doing.
At Cornell University, a team of researchers has built a quiet mirror for human judgment — a system that does not decide for us, but watches how we decide and asks whether our choices match our convictions. In an era when artificial intelligence is most often celebrated for replacing human reasoning, Interactive Explainable Ranking takes the opposite path: it holds us accountable to ourselves, surfacing the gap between the values we profess and the preferences we reveal. The tool, tested with film evaluators and teaching assistants, points toward a deeper question that technology rarely dares to ask — not what we should choose, but whether we truly know what we value.
Abe Davis, an assistant professor of computer science at Cornell University, noticed the same problem recurring each semester: experienced teaching assistants, clear rubrics, and yet scores that never quite aligned. The inconsistency wasn't carelessness — it was the gap between what evaluators said mattered and what their choices actually revealed.
That observation became the foundation for Interactive Explainable Ranking, developed at Cornell's Ann S. Bowers College of Computing and Information Science. The system works by first asking users to name their priorities and assign them weights — say, price, reliability, and fuel efficiency when choosing a car. It then presents pairs of options and tracks each choice. When a decision contradicts a stated priority, the system flags it, inviting the user to either revise their declared values or acknowledge that something unnamed is shaping their behavior.
Two pilot studies tested the approach. Film evaluators reported that the tool moved them away from gut reactions toward decisions grounded in explicit criteria. Teaching assistants evaluating student projects produced scores that aligned with prior independent grades and showed high consistency with one another — suggesting the method could yield reliable, repeatable assessments.
What distinguishes the tool is its refusal to impose a framework. It enforces internal coherence rather than external standards, keeping the human at the center while making them accountable to themselves. Davis framed the core challenge as making evaluation scalable without sacrificing rigor, and achieving consistency without allowing unconscious bias to enter through the gaps most tools leave unexamined.
Presented at CHI 2026, the system offers something rare in the current landscape of AI development: not a replacement for human judgment, but a demand that human judgment be honest.
Abe Davis, an assistant professor of computer science at Cornell University, spent years watching the same problem repeat itself each semester: hundreds of student projects to evaluate, clear rubrics in hand, experienced teaching assistants working in parallel, and yet somehow the scores never quite aligned. The inconsistency wasn't laziness or incompetence. It was the gap between what evaluators said mattered and what their actual choices revealed they valued.
That friction became the seed for a new tool. Davis and his team at Cornell's Ann S. Bowers College of Computing and Information Science developed what they call Interactive Explainable Ranking—a system that does something counterintuitive in the age of AI: instead of replacing human judgment, it audits it. The algorithm doesn't decide. It watches you decide, then asks whether you're being honest with yourself.
The system works by first asking a user to name what matters. Say you're buying a car. You declare that price, reliability, and fuel efficiency are your priorities, and you assign each a weight. Then the AI presents you with pairs of vehicles and asks you to choose. As you make those choices, the algorithm compares them against the weights you assigned. When it finds a contradiction—you said fuel efficiency was crucial, but you keep picking the thirstier model—it stops and flags the inconsistency. You then have a choice: adjust your stated priorities to match your actual behavior, or recognize that something else is influencing your decisions, something you hadn't named. Maybe it's the color. Maybe it's the brand. Maybe it's something you didn't know about yourself.
Cornell tested the approach in two pilot studies. In the first, four participants used the system to rank short films. They reported that the tool pushed them away from gut reactions and toward decisions grounded in explicit criteria. In the second trial, four teaching assistants evaluated ten student projects from a previous computer graphics course using the system. Their scores aligned with grades those same projects had received before, and the consistency between evaluators was high. The results suggested the method could produce reliable, repeatable diagnoses of work quality.
What makes this different from a simple checklist is that the AI doesn't impose a decision framework. It enforces internal consistency. Davis emphasized the core tension he was trying to solve: how to make evaluation scalable without sacrificing rigor, and how to achieve uniform scoring without letting unconscious bias slip in through the cracks. Most tools try to remove the human from the equation. This one keeps the human in the center but makes them accountable to themselves.
The system was presented at CHI 2026, the Association for Computing Machinery's annual conference on human factors in computing systems. For anyone who has ever sat in a hiring meeting, graded papers, or made a major purchase and felt the nagging sense that their reasoning didn't quite hold together, the implications are clear. The tool doesn't tell you what to want. It tells you whether you know what you want, and whether you're being truthful about it.
Citações Notáveis
The challenge was finding an evaluation process that was not only scalable but capable of combining rigor with fairness, while managing the inherent tension between achieving uniform scoring and avoiding unconscious bias.— Abe Davis, assistant professor of computer science at Cornell University
Quer a matéria completa?
Leia o original em Infobae ↗
Education