The conversation focuses on how to build a data strategy addressing the needs of a business first, not what is available on the market in terms of an AI model.
Anthony Scriffignano, PhD is one of the world's best data scientists — a thinker who has spent more than 40 years at the intersection of linguistics, advanced algorithms, business intelligence, and the very human problem of making good decisions from imperfect information. Anthony is currently a Distinguished Fellow and member of the Alfred Lee Loomis Innovation Council at The Stimson Center. He also serves on advisory committees in government, the private sector, and academia, and has briefed the National Security Telecommunications Advisory Committee and contributed to reports to the President on Big Data Analytics and Emerging Technologies. He served on the OECD Network of Experts group on implementing Trustworthy AI.
His most recent publication is a co-authored chapter — "Illuminating the Invisible: AI's Role in Understanding Complex, Opaque, and Dynamic Ecosystems" — in the Springer volume Artificial Intelligence Revolution (2026, edited by Max Rangeley and Nicholas Fairfax), co-written with Dr. David Bray.
Data before models. Most AI failures begin in the data, not in the model. Ask the three Ps before you touch a model: Provenance, Permissible use, Perturbation.
Decision elasticity. The right question is not "is this data perfect?" but "how wrong could it be and would I still make the same call?" If you cannot answer that, you are not ready to decide.
The walled garden method. You can detect disruption inside a system you cannot observe directly by modelling the ecosystem immediately outside it — the haversine annulus, Z-scores, Jensen-Shannon divergence.
Start with needs, not tools. Four-box needs analysis: known met, known unmet, unknown met (the danger box), unknown unmet (the blue sky box
Synthetic data compounds error. Pre-LLM synthetic data let you estimate error. Post-LLM, models are generating training data that other models consume, and error compounds multiplicatively rather than averaging out.
https://www.stimson.org/ppl/anthony-scriffignano/
Anastassia Lauterbach - LinkedIn
First Public Reading, Romy, Roby and the Secrets of Sleep (1/3)
First Public Reading, Romy, Roby and the Secrets of Sleep (2/3)
First Public Reading, Romy, Roby and the Secrets of Sleep (3/3)
The haversine annulus — a geodesically corrected ring around a walled garden used to model its immediately dependent ecosystem.
Jensen-Shannon divergence — an information-theoretic measure used to compress multiple differently-scaled signals into a single change detector.
Bayesian inference / updating: start with conditional probability, update as evidence accumulates.
Canonical models of malfeasance — Anthony's term for the well-documented patterns of fraud that recur around disasters, insurance events, and public contracts.
Dispositive threshold — the point at which you have enough data to make a decision — but just because you can does not mean you should.
Shadow AI — unsanctioned AI tools used by employees without employer knowledge or governance; 64% prevalence in Anastassia's SME survey.
