Safety and Alignment Leadership
AI Safety for the Real World
Stephen Lieberman builds the governance architectures that determine whether AI safety survives contact with operational reality.
Most organizations deploying AI at scale already have safety policies. What they lack is leadership that can hold the line between those policies and what actually happens when capable systems meet real institutional pressure. That is the work I do: identifying where alignment breaks down before it breaks down, and building the technical and organizational conditions that make safety durable.
Available for fractional, remote, and in-house executive roles through 1023AI.
Selected organizations
U.S. Department of Veterans Affairs
U.S. Department of Defense
U.S. Department of State
Federal Emergency Management Agency
Naval Postgraduate School
University of Southern California
United States Army
Defense Manpower Data Center
The Challenge
Deploying AI at scale creates a leadership challenge that software evaluation cannot solve
As capable systems scale, they produce emergent capabilities that were not designed, emergent risks that were not anticipated, and interaction effects with human systems that no evaluation framework fully captures. The gap between what the model was tested on and what it actually does, in messy organizational contexts, across human-AI teams, under real-world deployment pressure, is where the most serious safety failures live.
This is why safety at scale is not just a research problem. It is a leadership problem.
Static controls, compliance checklists, and point-in-time evaluations are designed for systems with stable behavior and enumerable failure modes. They were not designed for systems whose capabilities and risks emerge at scale, shift with deployment context, or interact with human behavior in ways that are invisible during testing.
The robustness gap
A system can appear safe in theory and still fail under real deployment pressure. I use the term robustness gap to describe the distance between nominal safety and real-world resilience. In high-stakes AI deployment environments, safety claims are exposed to shifting incentives, changing contexts, adversarial pressure, organizational fragmentation, and downstream effects that do not appear in controlled settings.
In AI systems deployed at scale, emergent misalignment is the most significant expression of the robustness gap. Alignment that appeared solid at one capability level quietly breaks down as the system becomes more capable, through a process that standard evaluation pipelines were not designed to detect. Closing this gap requires leadership that understands how emergent properties, nonlinearity, and sociotechnical systems behave in the real world.
The Gap
This is also where AI iatrogenics becomes dangerous. In medicine, iatrogenics refers to harm caused by the treatment itself. A narrow intervention can reduce one visible risk while creating new harms elsewhere, distorting incentives, increasing brittleness, or destabilizing the broader system.
AI Iatrogenics
When a safety intervention introduces new failure modes elsewhere in the system, reducing one visible risk while creating harms that were not present before, distorting incentives, or destabilizing the broader sociotechnical environment. In high-stakes AI deployment, iatrogenic failure is often invisible until scale amplifies it.
The Orchestration Requirement
No two organizations face the same version of this problem. The specific failure modes in your deployment environment are a function of your technical architecture, your institutional incentives, your capability trajectory, and the human systems operating around your models. Arriving with a predefined playbook is precisely the wrong response.
My role is to move quickly from diagnosis to structure, mapping where your safety posture is genuinely robust and where it is nominally compliant but operationally brittle. The interventions that follow are built for your organization's specific sociotechnical conditions, not for a generic deployment scenario. That is what it means to govern AI as an adaptive system rather than a static product.
This is not theoretical. The same cascade dynamics that make AI systems dangerous at scale, emergent failure modes in human-automated systems, nonlinear amplification under stress, and governance gaps invisible until consequence arrives, were the operating environment of an eight-year applied complexity research program in US equity index derivatives markets. That work produced something most AI safety frameworks lack: empirical validation of the core thesis under conditions where the cost of being wrong was immediate and measurable.
When to bring me in
AI safety and alignment become a leadership function when:
If more than one of these describes your situation, a conversation is worth having.
Start a Confidential Conversation →Approach
How I approach AI safety and alignment
My work produces two things simultaneously: technical systems, and the organizational architectures that make those systems safe under real-world pressure. They are developed together, because the technical system shapes what the human system can do, and the human system shapes what the technical system needs to be.
I approach AI deployed at scale as a systems problem, a leadership problem, and a human problem. In practice, that means working from principles most safety reviews treat separately.
Complex Adaptive Systems
AI deployed at scale does not operate in isolation. It interacts with organizations, incentives, feedback loops, and people in ways that produce behavior no single component was designed to generate. Safety is a property of the whole sociotechnical environment, not just the model.
Epistemic Uncertainty
Leaders deploying AI at scale make consequential decisions under genuine uncertainty. At the frontier, that uncertainty is not a gap to be closed by better evaluation. It is a structural feature of the domain. My work is grounded in the discipline of reasoning clearly about what we do not yet know, and building governance structures that remain sound as the picture continues to change.
Emergent Foresight
Emergent foresight is the capacity to govern for what a system is becoming, not just what it currently is. Capability levels shift. Risk surfaces expand. The alignment properties that held at one threshold can degrade quietly at the next. Closing that gap requires anticipatory leadership, not retrospective review.
Safety at Scale
The real test is whether safety survives growth, speed, strategic pressure, and social consequence. That standard cannot be met by evaluation frameworks alone. It requires leadership that can govern the whole sociotechnical system as it scales, holding technical architecture, organizational design, human-AI teaming, and institutional governance in view simultaneously. That capacity was built across three environments, defense, financial markets, and institutional governance, where the standard was not nominal compliance but operational survival.
Human Systems as Causal Variables
Most AI safety frameworks treat human systems as context. Organizational dynamics, institutional incentives, and social structures are not background conditions for AI safety failures. They are causes. Interventions that ignore this dimension do not simply miss a variable. They introduce failure modes that were not present in the original system and cannot be fully anticipated in advance. Governing AI at scale requires the same analytic discipline applied to the technical system applied equally to the human system around it.
Leadership as a Safety Variable
Safety frameworks do not implement themselves. The decisions that determine whether governance holds under pressure, the interpretations, tradeoffs, and escalations in conditions no policy document fully anticipated, are leadership decisions. My approach treats leadership architecture as a core safety variable: who holds authority at which decision points, what information reaches them, and whether the institutional structure allows them to act on it. Technical safety work that does not account for the leadership environment it operates within is incomplete.
About
Stephen Lieberman
I lead AI safety and alignment work for organizations where capable AI is moving beyond what standard governance was built to handle. With more than 20 years leading technical and operational teams in mission-critical environments, I bring the institutional discipline and operational judgment that high-stakes AI governance demands.
The core argument: capable AI cannot be governed as if it were ordinary software. As systems scale, they move beyond what traditional analytic methods can model, producing emergent capabilities, emergent risks, and interaction effects with human systems that no evaluation framework fully captures. Closing that gap is a leadership problem before it is a technical one.
Through 1023AI, I work with organizations navigating the transition from controlled research conditions to real-world deployment at scale.
Mission-Critical Technical and Operational Leadership
More than 20 years leading technical and operational teams across government, defense, academia, nonprofit, and industry. Senior leadership on Department of Defense and Veterans Affairs programs within funding environments exceeding $100 billion ⓘ, spanning enterprise architecture, decision-support systems, security and compliance, electronic health records, cloud systems, and data strategy. Led programs with multimillion-dollar budgets and worked directly with senior leaders across defense, government, and institutional settings.
Defense, Security, and International Systems
At the Naval Postgraduate School, served as a DoD civilian program leader and Principal Investigator responsible for programs in defense technology, modeling and simulation, collaboration platforms, and decision-support systems, with more than $9.8 million ⓘ in fully funded federal proposals across six programs. Every proposal was funded at the amount requested. Work included counterterrorism, counterinsurgency, peacekeeping operations, and international collaboration across more than 100 countries. Recognized directly by the Undersecretary of Defense for Intelligence:
“Steve, you and your team have performed superbly. Your collective skills, resourcefulness, and creativity have created a ground-breaking tool that will benefit the U.S. government and our allies.”
Michael G. Vickers
Undersecretary of Defense for Intelligence, U.S. Department of Defense
Research in Complex Adaptive Systems
The research foundation of this work is CASS, the Complex Adaptive System Simulation framework, a federally validated, VV&A-certified agent-based simulation platform built and published over nearly two decades. Originally developed at the Naval Postgraduate School for counterterrorism, peacekeeping, and international conflict environments, CASS has since extended to organizational behavior, social policy, and human-machine interaction. The framework models the emergence of beliefs, values, and interests across whole societies at scale, applying network science, discrete event simulation, and behavioral survey data to forecast group-level dynamics in high-consequence environments.
Published work spans agent-based modeling methodology, social simulation validation, cognitive modeling from behavioral data, and violent extremist network dynamics. More than 20 peer-reviewed publications, book chapters, and conference papers, with an H-index of 7, more than 100 citations, and 8 highly influential papers indexed on Semantic Scholar.
Selected publications
Lieberman, S. (2012). Extensible Software for Modeling Whole Societies: Framework and Preliminary Results. Simulation, 88(5).
Lieberman, S. and Alt, J. (2010). Developing Social Networks for Artificial Societies from Survey Data. In Advances in Social Computing, Lecture Notes in Computer Science, Vol. 6007. Springer.
Alt, J. and Lieberman, S. (2010). Developing Cognitive Models for Social Simulations from Survey Data. In Advances in Social Computing, Lecture Notes in Computer Science, Vol. 6007. Springer.
Alt, J., Lieberman, S., and AlRowaei, A. (2010). Exploring the Implications of Time in Discrete Event Social Simulations. AAAI Spring Symposium Proceedings.
Applied Complexity in Live Equity Derivatives Markets
From 2015 to 2023, the CASS framework was extended into live US equity index derivatives markets as a rigorous, high-stakes validation of the methodology. Over eight years, agent-based simulation of human-automated market dynamics was applied to identify and operationalize the emergent cascading failure patterns produced when human traders and automated execution systems interact under volatility stress. Three years of simulation-only operation preceded any live market participation, applying the same pre-deployment validation standard that now grounds the AI safety governance work.
The operation was formally structured, with institutional-grade exchange access, multi-platform infrastructure across futures and options on ES and NQ contracts, and legal and tax architecture appropriate to the complexity of the instruments. The results validated the core thesis: that complex adaptive systems producing emergent behavior under stress are modelable in advance, governable through disciplined simulation, and consequentially different from what static evaluation frameworks can detect. During this period, formal participation in sessions convened by the CFTC, SEC, CME, and CBOT on the development of regulatory frameworks for cryptocurrency futures and derivatives contributed practitioner perspective at the formative stage of that policy development.
The intellectual through-line to AI safety governance is direct. Emergent behavior in live financial markets and emergent misalignment in AI systems at scale share the same underlying structure: nonlinear dynamics in human-automated systems that static testing cannot anticipate, that manifest only at scale and under stress, and that carry asymmetric consequences when governance fails. This research program is ongoing. Partnership with organizations seeking to extend this work into AI-relevant domains is actively welcomed.
Go Deeper: Live Market Operations →AI Governance at the Intersection of Technology and Human Systems
The most significant gaps in real-world AI safety governance are not purely technical. They are organizational, institutional, and deeply human. Doctoral research at USC's Suzanne Dworak-Peck School of Social Work centers on governance architecture for AI systems deployed in professional and high-stakes institutional contexts. This builds on a research lineage running from graduate work on network-theoretic organizational resilience at the University of Connecticut, through doctoral training in computational modeling and simulation at the Naval Postgraduate School, to current work on the governance conditions that determine whether AI safety frameworks hold under real institutional pressure.
The capstone examines LLM governance in professional practice environments, arguing that the decisive barrier to safe AI integration is a governance architecture problem rather than a technical one, grounded in antifragile systems theory, community-based participatory design, and the organizational conditions that determine whether safety holds under real institutional pressure.
Approach draws on sociotechnical systems theory, organizational behavior, industrial psychology, human-centered design, and macro social work: disciplines that illuminate how people actually act inside institutions and how to intervene at the level of structural conditions rather than surface behaviors. That is precisely the level at which real-world AI governance must operate.
Go Deeper: Research, Publications, and Academic Work →Strategic Integration
The First 90 Days
When I step into an organization as a fractional or in-house executive, the first priority is clarity, not process. Before any governance framework can hold, I need to understand the gap between your stated safety position and your operational reality.
Days 1–30
Contextual Discovery
I conduct structured conversations across the board, research, product, and operations, not to audit, but to map the lived experience of safety inside the institution. Where does the policy stop and the workaround begin? Where are the accountability gaps that no one has named? This surfaces the friction points where governance is weakest under real pressure. Output: a friction map of where governance is weakest under real operational pressure, delivered to leadership at the 30-day mark.
Days 31–60
Sociotechnical Alignment
Working from the discovery findings, I identify the specific robustness gaps in your deployment pipeline and design interventions that address both technical safeguards and the organizational conditions that determine whether those safeguards hold. Complex adaptive systems fail at the seams between components. That is where I focus. Output: a prioritized intervention architecture addressing both technical safeguards and the organizational conditions that determine whether those safeguards hold.
Days 61–90
Operational Safety Case
I deliver a prioritized safety case, a living document grounded in your actual capability thresholds, your specific risk surface, and the institutional conditions I have observed directly. This is not a compliance checklist. It is a durable foundation for governing AI as your systems continue to scale.
Why 1023AI
The name references Avogadro's number (6.022 x 1023), the precise mathematical boundary where immense collections of microscopic interactions forge emergent macroscopic behavior. That is not a metaphor for AI. It is a description of what actually happens. Scaling does not simply improve performance. It changes what the system is, what it can do, and what it can get wrong. Beyond a certain scale, aggregate behavior changes qualitatively, demanding a different approach.
The European Commission's official Guidelines under the EU AI Act arrive at the same number, establishing 1023 floating-point operations of training compute as the precise threshold at which AI models qualify as General Purpose AI triggering mandatory regulatory oversight. That convergence is not coincidental. It marks the boundary where AI generality becomes real, emergent capabilities and emergent risks become the dominant safety challenge, and governance must cross the same threshold the model does. Safety at that scale requires leadership that understands emergence, not just evaluation. That is what my work is about.
Start a Confidential Conversation
If your organization is navigating the gap between evaluated safety and real-world resilience, or the human and institutional conditions that determine whether safety holds at scale, reach out. Every engagement is shaped by the specific organization, its specific challenges, and the specific sociotechnical system it is operating within.
Safety at scale. That is what I do.