Can an AI system know when it does not know? That first question led our team a decade ago to explore uncertainty quantification, explainability, causality, law and ethics towards a larger ambition: AI that strengthens the people who must understand, govern and take responsibility for it.
How the research question expanded
Chapter 1: Mathematics + computer science
Teaching AI to doubt
Can an AI system know when it does not know?
The foundational problem was uncertainty: a model can be wrong because the case is outside its operational domain, because the model has not learned enough from representative evidence, or because the world remains genuinely variable even with good evidence.
Our scientific contribution was to leverage those sources of uncertainty to inform decision-making. Understanding uncertainty became useful to determine what happened next: collecting better evidence, redesigning the operating domain, or escalating to accountable human judgment.
A well-developed uncertainty framework allows for joint human-AI decision processes, in which refusals, reviews, and responsibility are embedded into the system.
The case may sit outside the distribution or operating conditions where the model should be trusted.
Epistemic uncertainty
Learn & compare models
The model lacks enough representative knowledge; better evidence, data or models may reduce it.
Aleatoric uncertainty
Show scenarios to human
Several outcomes remain plausible because the observation is partial, noisy or intrinsically variable.
Chapter 2: Engineering + public projects + critical systems
A shared underlying problem
Where can we integrate uncertainty quantification into decision-making?
Our research on decision-making informed by uncertainty quantification moved across finance, safety-critical systems, autonomous mobility, pharmaceutical research, information integrity and other applied contexts.
The wide diversity of these domains reveals a common challenge of determining how incomplete evidence, changing operating conditions, or asymmetric consequences shape decisions. Uncertainty quantification proves essential for navigating these complexities and ensuring responsible decision-making.
Authorship, attribution, copyright, cultural work.
Chapter 3: Causality + explainability
Looking for explanations
What caused the prediction? What would change under a different intervention?
Uncertainty tells us when confidence is limited. but it does not explain alone what produced an outcome or what would change under a different intervention. To make system more trustworthy, it becomes necessary to examine limits, reasons and alternatives together.
Causal reasoning, counterfactuals and explainability provide this layer of reasons and alternatives that can be inspected, challenged, and used in decision-making. At this point, the scope of our research broadened from Safe AI to Trustworthy AI.
Does the explanation reflect the model and the decision?
Chapter 4: Law + regulations + standards
When technical trust is not enough
Combining technical and legal expertise for AI governance
An AI system may perform reliably and still be unsuitable for a particular deployment when it needs to address legal duties like risk management, data governance, documentation & standards, fundamental rights assessments, oversight, or accountability.
TRAILAB brings these requirements into the research process, translating them where possible into testable requirements and evidence. The result is a governance framework that supports and complements system development.
The Commission frames the AI Act as a risk-based legal framework with staged application: prohibited practices and AI literacy from February 2025, GPAI governance from August 2025, and high-risk obligations now sequenced toward December 2027 and August 2028 under the simplification package.
The Commission describes how harmonised standards translate AI Act requirements into technical language for risk management, data governance, logging, transparency, human oversight, accuracy, robustness, cybersecurity, quality management and conformity assessment.
CEN-CENELEC JTC 21 develops European AI standards, including harmonised standards in support of the EU AI Act and related work on trustworthiness, risk management, quality management and conformity assessment.
ISO/IEC 42001 specifies requirements for establishing, implementing, maintaining and continually improving an AI management system for organisations providing or using AI systems.
Risk-based duties, AI literacy, GPAI governance and high-risk system obligations.
02
CEN-CENELEC
Harmonised standards that translate legal requirements into technical evidence.
03
ISO/IEC
AI management systems and AI risk management: ISO/IEC 42001 and ISO/IEC 23894.
04
Sector standards
Assurance in the automotive, aviation, automation, medical, and critical-system industries.
Chapter 5: Ethics + philosophy + social legitimacy
Reliability needs a purpose
How can technical capabilities strengthen responsible judgment?
A system that meets technical and legal requirements still needs scrutiny in the ways it shapes autonomy, responsibility, authorship or institutional purpose. Ethics and philosophy enter to determine what should remain human, where responsibility sits, which trade-offs matter, and what forms of progress are worth pursuing.
Capability
Can it work?
Compliance
Can it be justified?
Legitimacy
Is it worth pursuing?
01
Human agency
Autonomy, authorship and control.
02
Accountability
Responsibility, liability and fairness.
03
Social grounds
Dignity, care, vulnerability and epistemic justice.
04
Responsible judgment
Purpose and public interest.
Chapter 6: Science + culture + society
AI reshapes the conditions for thinking
What happens when AI becomes judgment infrastructure?
AI is increasingly embedded in the workflows of reasoning, creating and investigating. This interaction is shaping how people frame questions, inspect alternatives and make commitments. While AI can provide readymade solutions in diverse domains, the risk is that understanding, critical thinking, authorship or responsibility are bypassed in the process.
Reflective AI asks how AI changes the human capabilities on which good judgment depends. TRAILAB proposes RAIO, the Reflective AI Orchestrator, as a research architecture that coordinates clarifications, assumptions, uncertainty, alternatives and trade-offs during a human-AI interaction. Reflective AI becomes a research objective for strengthening judgment, critical thinking, creativity, responsibility and explicit human commitment.
Narayanan's ICML 2026 keynote frames AI as changing what humans work on, motivating tools and roles that turn AI-generated solutions into human understanding and accountable action.
Narayanan and Kapoor argue for treating AI as a powerful general-purpose technology whose adoption, deployment and human control matter as much as model capability.
Messeri and Crockett warn that AI tools can accelerate scientific production while creating illusions of understanding and scientific monocultures if explanation, validation and critique are bypassed.
01 / understandingClarify assumptions before acting
02 / agencySurface uncertainty without surrendering control
03 / creativityExplore alternatives before closure
04 / responsibilityReview values and trade-offs before commitment
Human judgment
Chapter 7: An independent multidisciplinary institution
This scientific journey needed an institution
Why create TRAILAB?
The problems we address have outgrown any single discipline, but the relevant expertise remained distributed across universities, research centres, industry, public institutions and civil society.
We founded TRAILAB in 2026 as an independent non-profit where all those capabilities can be combined around shared research questions without replacing the primary affiliations of the people and institutions involved. It consolidates, extends and transfers the research trajectory behind it, building methods, evidence, pilots and institutional capacity for Trustworthy and Reflective AI.
