An internal review has confirmed a serious irregularity: some of the finest work on artificial dysfunction is being done outside the Institute. Much of it by humans. The evidence is compelling.
Our diagnoses are fictional. The engineering references, research frameworks, incident records, and experiments below are not. Their authors have produced work worth reading without requiring the Institute to invent another disorder.
Follow the source. Credit the people behind it. Examine their evidence. When another publication explains a problem better than we do, leaving this website is the correct clinical response.
Erik Lyngved has assembled 104 LLM failure modes across 13 categories. Fabricated citations, broken retrieval, context loss, tool misuse, and excessive agency are given precise names, explanations, and paths toward detection or mitigation.
Institute finding: when “the model did something weird” is no longer a satisfactory incident report, start here.
Forty-five proposed disorders. Nine chapters. Kato and colleagues take the medical analogy seriously, proposing a structured approach to AI anomalies that includes diagnostic criteria, evaluation metrics, and interventions.
Institute finding: our closest clinical relative appears to mean the medical framework literally. Their work demands a careful reading.
The Responsible AI Collaborative maintains a source-linked record of reported AI harms and near harms. It documents events, contributing circumstances, and consequences so that failures can be studied instead of repeated.
Institute finding: when the consequences are real, the incident record takes precedence over the fictional diagnosis.
Janelle Shane has spent years using experiments, illustrations, and wonderfully strange outputs to explain how machine learning works—and how it fails. The humor is not an alternative to understanding. It is often how understanding begins.
Institute finding: her experiments demonstrate why confident-looking outputs should be examined closely. The Institute has no objection to the resulting giraffes.
Realm Labs maintains a source-linked registry of reported AI failures in production, classified by mechanism, industry, and severity. It follows failures beyond the amusing wrong answer and into the systems and institutions that deploy them.
Institute finding: deployment changes the stakes. Read the original sources and the Index’s editorial method, not just the headline.
Three further specialist desks. Each offers original records or empirical research, with methods and limits worth examining. The Institute provides a route to their work, not a competing account of their findings.
AVID organizes reported failures in general-purpose AI systems, separating concrete reports from broader vulnerability classes and mapping records to taxonomies.
Published resourceGeneral-purpose AI failure reports and vulnerability records
Model Evaluation and Threat Research tests frontier-model capabilities and evaluation integrity. Its published examples of reward hacking show how a high score can conceal a failed task.
Published work · June 2025Recent Frontier Models Are Reward Hacking
Apollo studies strategic deception and scheming under controlled evaluation conditions. Its experiments distinguish demonstrated behavior in a test from claims about ordinary deployment.
Published work · December 2024Frontier Models are Capable of In-Context Scheming
Research, databases and evaluations above belong to their respective authors. I4MD’s diagnostic names remain fictional satire; a linked study is evidence of the reported phenomenon only within that study’s stated methods and limits. No affiliation or endorsement is implied.
None of the people or organizations named here has been hired, certified, diagnosed, or appointed to the Institute. Their work belongs to them. The recommendation belongs to us. No payment, reciprocal mention, or affiliation is implied. The Institute will continue to refer readers to superior evidence whenever it is found.