Trustworthy AI

01 Problem

Why capability is no longer sufficient

Artificial intelligence was defined at its inception as the science and engineering of intelligent machines, and for several decades capability served as the field's principal measure of progress. As these systems have been placed in consequential settings — hiring, clinical diagnosis, public administration, examination — capability alone has ceased to be an adequate criterion.

A model acquires its behaviour from a particular corpus of training data. Bias or error present in that corpus propagates into the system's outputs, and the system possesses neither an independent standard of fairness against which to assess a result nor any means of verifying its own work. Scrutiny, if it is to occur at all, must be applied from outside.

Failure of this kind is seldom visible in aggregate performance. It concentrates in edge cases: rare operating conditions, populations thinly represented in the training data, circumstances no designer anticipated. In aerospace, medicine and finance a single such failure may be severe, and the absence of demonstrated trustworthiness is frequently what prevents an otherwise functional system from entering service.

The institute takes its name from the Latin judex, a judge. Its object is not the advancement of model capability but the verdict: an assessment of whether a given system, in a given use, can be relied upon. Because what a model can and cannot do is in principle measurable, such an assessment can rest on evidence rather than on expectation or apprehension. The operative standard is not infallibility; it is that errors be discoverable, explicable, and attributable to a responsible party.

02 Research

Criteria of assessment

Four criteria are applied to the system under examination. They are not independent of one another, and a system may perform accurately while failing against any one of them.

• 1

  AI Ethics

  Whether the system performs equitably across populations A model trained predominantly on one population tends to degrade on others. Establishing equity requires knowing which groups the training data represents, in what proportion, and in which direction any imbalance displaces the output. Fairness, Representation, Bias
• 2

  Explainable & Reliable AI

  Whether its reasoning can be reconstructed and its performance sustained A system that yields only a conclusion cannot be audited. Our work concerns the retention of inference in a form open to inspection, and the characterisation of behaviour under rare conditions rather than average ones. Explainability, Reliability, Edge Cases
• 3

  AI Governance

  Where oversight is situated and accountability assigned A model bears no accountability of its own. Oversight must be designed into the surrounding process: the point at which a person intervenes, the information available to them, and the grounds on which a result may be overturned. Human Oversight, Human-in-the-Loop, Accountability
• 4

  AI Evaluation

  By what measures trustworthiness is to be established Accuracy figures alone do not establish trustworthiness. An adequate standard takes account of the setting in which a system operates and the population on which it operates. Assessment, Benchmarking, Impact

03 Assessment

From principle to inspection

That artificial intelligence ought to be trustworthy is widely agreed, and guidelines to that effect are plentiful. The difficulty lies in determining whether a particular system satisfies them. Such guidelines are formulated at a level of abstraction to which a running system does not directly answer, and a checklist reaches only those failure modes that someone thought to anticipate.

The development of that procedure is the institute's principal work. An assessment undertaken once a system is complete arrives too late; the relevant questions belong at every stage of the lifecycle — intended use, design, development, deployment and operation.

• 1

  Context of use

  The population on which the system operates, and the setting Identical models carry unequal risk according to where they are placed. Assessment proceeds from the deployment setting rather than from the technical specification. Deployment setting, Risk profile
• 2

  Scenarios

  Tracing a concrete use through to its conclusion Following a single realistic situation to its end surfaces failure points that no enumerated checklist contains. Socio-technical scenarios, Walkthroughs
• 3

  Multidisciplinary review

  Technical evaluation alone is insufficient Law, ethics and the practice of the domain must examine the same system. A design that satisfies one of these frequently fails against another. Law, Ethics, Domain practice
• 4

  Tensions

  What has been conceded, and in return for what Accuracy and explainability, privacy and performance, cannot be maximised simultaneously. Such trade-offs warrant documentation rather than silent absorption into the build. Accuracy vs. explainability, Privacy vs. performance

04 Field Study

Grading essays at national scale

In August 2026 Korea's National Education Commission announced its plan to introduce constructed-response and essay items into the College Scholastic Ability Test, the national university entrance examination taken by approximately half a million candidates on a single day. The final plan is due in March 2027.

An automated grader in this setting presents the trustworthiness problem in concentrated form. The score governs university admission. The affected population is an entire cohort rather than a sample. The output is contested by design, in that every candidate has standing to challenge it and a challenge must be answerable. Accuracy alone would not render such a system fit for service.

The governing question is accordingly not whether a model can grade competently, but what would have to be demonstrated, and to whom, before a system of this kind could be placed in service at all.

The case carries a further implication. The system under examination is itself an instrument of judgment, applied to half a million people. The institute is developing one, and holds it to the criteria it would apply to any other — the most demanding test to which its own procedure can be put.

1. ResearchGrading criteriaWhat constitutes a strong argument
3. FrameworkArgument structureClaim, reason, evidence, counterargument, rebuttal, conclusion
5. SystemAssessment engineReturns the grounds alongside the score

05 Publications

Publications

• The Architecture of Reason A Neuro-Symbolic Approach to Solving the Black Box Problem

  An account of systems that deliver conclusions without reasons. The book argues for situating the computational capacity of neural networks within a logical structure open to human inspection, and sets out the OREO framework — Opinion, Reason, Example, Opinion — by which an output may be rendered as a traceable argument.

  In English · Kindle edition

• The Last Enter Key An architecture for the accountability of judgment

  Proceeds from the principle that an artificial system proposes while the decision, and the responsibility attaching to it, remain with a person. Across four parts — metaphysics, epistemology, logic and ethics — it examines the statistical limits of contemporary AI and advances a glass-box design in which the reasoning is exposed.

  In Korean, as «마지막 엔터키» · February 2026 · 214 pages