Speaker Bios

 

​

Robert Gibbons

The University of Chicago, USA

The Future of Mental Health Measurement

​​

Mental health measurement has been based primarily on subjective judgment and classical test theory. Impairment is determined by a total score, requiring that all respondents be administered the same items. An alternative is adaptive testing in which different individuals may receive different scale items that are targeted to their specific impairment level. We have developed a multitude of adaptive tests for psychopathology, substance use disorders, suicide risk, autism, cognitive impairment for youth and adults based on multidimensional item response theory. The shift in paradigm is from small fixed length tests with questionable psychometric properties to large item banks from which an optimal small subset of items is adaptively drawn for each individual, targeted to their level of impairment. Results to date reveal remarkable increases in precision of measurement and dramatic decreases in patient burden. For example, depressive severity can be measured using an average of only 10 items in 2 minutes from a bank of 400 items, yet maintains a correlation of r = 0.95 with the 400 item scores. Applications in emergency departments and primary care centers, community settings, child welfare systems, college mental health screening, judicial systems, and epidemiologic studies are widespread.  Results of these tests benefit from an AI interface that communicates the results of the testing session to clinicians and patients.  Finally, I discuss the use of this technology to AI model benchmarking experiments.

​

Eren Kurshan
Princeton University, USA

​

​

Chung-Sheng Li

Genpact, USA

The Cognitive Enterprise: Why Agentic AI Belongs Only at the Exception Layer for Shared Services — and What Compounds Above the Model

 

After roughly half a century of automation through ERP, RPA, BPM, and case management, the deterministic majority of enterprise work in finance, accounting, procurement, and supply chain is already handled. What remains — and what agentic AI is now being pointed at wholesale — is largely exception handling: the residue those systems were never designed to resolve. This keynote argues that the prevailing instinct to "agentify the enterprise" mistakes the level at which cognition should be organized, and proposes a more disciplined architecture for the cognitive enterprise.

​

The argument proceeds from a single formalism-first principle: only genuinely partially observable, probabilistic decisions — POMDP problems — warrant agentic treatment, while fully observable, deterministic work should remain in the finite-state systems that already own it. From this principle follows a decomposition of enterprise responsibility into three orthogonal planes — execution (deontic, non-agentic), reasoning (epistemic, agentic), and governance (normative) — whose independence explains why a coherent platform cannot emerge from accumulated use cases but must be deliberately co-designed. Building on this, the talk distinguishes the durable competitive asset in enterprise AI from the commodity one: not the transaction log, which records what happened, but the accumulated causal map of why exceptions occur — an asset that compounds above the model layer and survives model-generation turnover. We show that this compounding follows a dual-strand "generative helix" of causal ascent and structural descent, structurally an analysis–synthesis loop, and that its terminus in counterfactual reasoning is reached not through model scale but through construction of a sufficiently rich world model — the cognitive digital twin — a route classical AI did not anticipate. We close with the corresponding risk: the same mechanism that compounds institutional intelligence also encodes the conditions for organizational blindness, making deliberate governance permeability an architectural requirement rather than a compliance afterthought.

​

Drawing on cross-domain analogies from semiconductor manufacturing, evidence law, and molecular biology, the talk offers enterprise architects and researchers a formal vocabulary for deciding what to automate, what to leave alone, and where the compounding value of enterprise AI actually accumulates.

​

Shawn Powers
Southern New Hampshire University, USA

Olé, to You, Nonetheless 

​

​“Olé, to You, Nonetheless” challenges the idea that the future of AI technology can be understood simply as a “country of geniuses” in a datacenter, arguing through historical creative clusters that transformative work emerges from intelligence that is fostered by constraint, existential stakes, and productive difference. The keynote proposes that our most important role in an AI-infused world is to deliberately build and protect those spaces of seminars, studios, classrooms, and institutions to allow consequential ideas and creative work to emerge alongside, and in collaboration with, artificial intelligence.