AI research

The problem we seek to solve: AI Produces Answers, Not Understanding
Most AI systems today are engineered to produce answers quickly, not to foster genuine understanding. They generate scores, predictions, or classifications with impressive speed and confidence, but they rarely explain how those results came to be or what they truly mean in context.
These systems excel at delivering decisive outputs—approving loans, flagging content, ranking candidates—but they operate as black boxes, offering conclusions without reasoning. As AI systems spread into more areas of life and work, from hiring decisions to criminal justice, from healthcare diagnostics to credit assessments, this fundamental lack of understanding creates a quiet but powerful effect throughout organizations and society fear..
This fear isn't irrational. It emerges from a reasonable response to being evaluated by systems we cannot fully see, question, or comprehend. The opacity creates anxiety, and that anxiety changes behavior in ways that ripple through entire institutions.
The DOSSIERVEILLANCE Framework for developing safe AI systems 
The DOSSIERVEILLANCE framework is built around a small number of simple but strict principles that seek to diminish the mistrust and fear towards AI technologies and systems.
The Dossier is Core
Every AI-assisted process revolves around a dossier, capturing data, AI outputs, human insights, and decisions. It's designed for ongoing revision, not premature closure.
AI as Investigative Tool
AI serves as an investigative instrument to extract signals, identify patterns, and propose hypotheses. Final decisions and documented authority always rest with humans.
Preserve Uncertainty
Uncertainty is inherent in important decisions. The framework explicitly records known limits, missing data, and contested interpretations within the dossier, rather than removing them.
Traceable Responsibility
The framework ensures clear traceability of responsibility, showing AI contributions, human interventions, and accountability for each decision. Responsibility is never obscured by automation.
Investigable Decisions
Decisions are always investigable post-facto. The dossier enables reconstruction of decision-making, alternative considerations, and accepted trade-offs, fostering correction and accountability.
Behavioral Effects Matter
The framework identifies fear, self-surveillance, and conformity as critical system risks. Compliance achieved through opacity is deemed a design failure, not a success.
Continuous Learning
Learning is continuous. Dossiers are living objects that evolve with new information. Past mistakes are preserved as resources for improvement, not hidden liabilities.op
Fear & AI
When evaluation systems become opaque, people begin to fear how they are being judged by algorithms they cannot fully see or question. Developers fear failing mysterious performance metrics. Managers fear liability from decisions they don't fully understand. Users fear invisible judgments that might affect their opportunities, their reputation, or their livelihood.
In response to this pervasive uncertainty, people start adjusting their behavior to what they think the system expects. They avoid risk, hide uncertainty, smooth over contradictions, and conform to perceived norms. Innovation gets suppressed. Honest disagreement disappears. Nuance evaporates.
People become their own monitors, constantly adjusting to avoid triggering unknown algorithmic consequences. The system doesn't need to watch everyone constantly—people watch themselves.
Developers
Fear failing opaque metrics and build defensively
Managers
Fear liability from AI decisions they don't control
Users
Fear invisible judgments affecting their opportunities
From Watching People to Investigating Systems
Our mission is to promote policies that would  reverse this troubling dynamic. Instead of using AI to watch people, score them, or silently judge them, it proposes a fundamentally different approach: using AI to investigate situations.
Traditional AI Surveillance
Monitors individuals, generates scores, automates judgment
What we seek to promote:
Develop policies that would make AI Investigate situations, preserves context, supports understanding
The focus shifts decisively away from monitoring individuals and toward examining systems, decisions, contexts, and the complex interactions between them. Rather than asking "What did this person do?" the question becomes "What happened in this situation, and why?"
The goal is not to automate judgment or make decisions faster. The goal is to make judgment more informed, more visible, and more accountable. This requires a complete rethinking of what AI systems should do and how they should operate within organizations.
The Dossier as the Core Unit
At the center of our approach is the idea of the dossier. A dossier is not just a file, a folder, or a database record. It is a structured, deliberate way of gathering information that keeps context, contradictions, and uncertainty alive throughout the decision-making process.
A dossier does not aim to close a case as fast as possible or reach a predetermined conclusion. Instead, it aims to preserve the conditions under which a decision can later be understood, questioned, revised, or appealed. It maintains the complexity of real situations rather than collapsing them into simple scores.
Preserves Context
Maintains the full situation surrounding decisions, not just isolated data points
Captures Contradictions
Keeps conflicting evidence and perspectives visible rather than forcing false consensus
Documents Uncertainty
Records what remains unknown or contested, not just what appears certain
Enables Revision
Supports future reconsideration as new information emerges or understanding deepens
AI as an Investigative Tool, Not an Authority
In our  framework, AI does not replace human judgment. It supports it. This distinction is crucial and changes everything about how these systems function and what they're designed to accomplish.
AI becomes a tool for collecting signals from multiple sources, identifying patterns that might otherwise go unnoticed, surfacing inconsistencies that require explanation, and suggesting possible interpretations for human consideration. It does not deliver final answers with artificial certainty. Instead, it contributes pieces of evidence to a larger picture that humans are still responsible for assembling, interpreting, and acting upon.
1
Collect Signals
Gather relevant information from diverse sources
2
Identify Patterns
Surface connections and trends in complex data
3
Surface Inconsistencies
Highlight contradictions requiring explanation
4
Suggest Interpretations
Offer possible meanings for human consideration
5
Human Decision
People integrate evidence and exercise judgment
Key principle: The AI proposes, but humans dispose. The system augments human judgment rather than replacing it with automated certainty.
Designing AI for Understanding, Not Confidence
This fundamental shift changes how AI systems must be designed from the ground up. Instead of asking only whether a model achieves high accuracy scores or appears confident in its predictions, developers must also ask deeper questions about the system's impact on human understanding and organizational behavior.
01
What does this system help people understand?
Beyond producing outputs, what insights does it genuinely provide?
02
What does this system hide or obscure?
What important information becomes invisible through automation?
03
What behavior does this system encourage?
How will people adapt their actions in response to this system?
04
What are the system's limits and assumptions?
Where does it work well, and where does it break down?
An AI system built this way is not optimized to look confident at all costs or to provide reassuring certainty when none exists. Instead, it is designed to show its limits, its assumptions and blind spots.   It acknowledges when it doesn't know something rather than manufacturing false precision. This kind of honesty may feel uncomfortable at first, but it leads to much better decisions over time.
Making Uncertainty Visible
In traditional AI systems, uncertainty is treated as a technical failure—something to minimize, hide, or eliminate through more data or better algorithms. In our approach to AI development, uncertainty is treated as information—something valuable that decision-makers need to see and understand.
A well-constructed dossier includes what is known with reasonable confidence, what remains unknown despite investigation, and what remains contested or subject to legitimate disagreement. This transparency makes decisions slower and more deliberate in some cases, requiring genuine thought rather than reflexive automation.
But this deliberation makes decisions much stronger in the long run. When something goes wrong—and things will inevitably go wrong in complex domains—the organization can examine how the decision was made, what information was available, what was uncertain, and where judgment was exercised. This enables real learning instead of blame-shifting.
Known
Information established with reasonable confidence
Unknown
Gaps in information despite investigation
Contested
Areas of legitimate disagreement
Keeping Responsibility Where It Belongs
Dossierveillance addresses one of the most serious problems in current AI systems: responsibility tends to disappear into a fog of distributed decision-making and technical complexity. When outcomes are poor or harmful, decisions are often attributed to "the algorithm," even when humans designed it, trained it, deployed it, and chose to act on its outputs.
This attribution to algorithms serves as a convenient shield. Developers claim they just built what was requested. Managers claim they just followed the system's recommendations. The system itself, of course, cannot be held accountable—it has no agency, no intentions, and no moral standing.
Document AI Contributions
Record what the system suggested and why
Track Human Judgment
Capture where people interpreted, modified, or overrode AI outputs
Maintain Traceability
Preserve the full decision chain from data to outcome
Assign Accountability
Ensure someone remains responsible for every consequential decision
By meticulously documenting how AI contributes to decisions and where human judgment intervenes, Dossierveillance keeps responsibility visible and traceable throughout the decision-making process. Someone remains accountable, and that accountability is supported by evidence rather than obscured by automation. This isn't about blame—it's about enabling genuine responsibility and continuous improvement.
Failing in Ways We Can Learn From
AI systems built with Dossierveillance principles tend to fail differently than traditional black-box systems. Their failures are more visible—nothing is hidden behind claims of proprietary algorithms or technical complexity. But precisely because failures are visible and well-documented, they tend to be less catastrophic and more recoverable.
Traditional AI Failures
Hidden until catastrophic impact
No record of reasoning or decision process
Impossible to understand what went wrong
System either continues failing or gets scrapped
Organizational amnesia about causes
Failures when AI systems built with Dossierveillance awareness in mind
Visible early through documentation
Full record of reasoning preserved
Clear understanding of failure causes
Targeted improvements possible
Institutional memory and learning
Because reasoning is documented and preserved throughout the process, mistakes can be understood in their full context and corrected at their source. The system can improve incrementally without rewriting history or pretending nothing went wrong. This creates institutional memory instead of collective amnesia—organizations actually get smarter over time rather than repeating the same mistakes with different AI vendors.
Reducing Fear Restores Judgment
Over time, this approach fundamentally reduces the fear that drives self-surveillance. When people can see how decisions are formed, understand the role AI plays, and recognize that human judgment remains central and valued, they stop trying to reverse-engineer what the system wants. They stop preemptively policing themselves to avoid mysterious algorithmic punishment.
Transparency
Decision processes become visible and understandable
Trust
People trust systems they can examine and question
Courage
Fear diminishes when evaluation criteria are clear
Creativity
Innovation returns when risk is no longer invisible
Disagreement
Productive conflict emerges when judgment is valued
Judgment
Human reasoning strengthens rather than atrophies
Creativity, genuine disagreement, and thoughtful judgment return to organizations. AI stops functioning as a silent disciplinary force that shapes behavior through fear of unknown consequences. Instead, it starts functioning as a shared tool for understanding complex situations—a tool that people can question, challenge, and ultimately choose whether to use.
From Control to Understanding
Ultimately, Dossierveillance represents a fundamental change in the role of AI within organizations and society. It transforms AI from a system that produces answers with artificial certainty and enforces behavioral conformity through opacity and fear into an infrastructure that supports thinking, deliberation, and judgment under conditions of genuine uncertainty.
"The question is not whether AI will make mistakes—it will. The question is whether those mistakes will be visible, understandable, and correctable, or invisible, inexplicable, and perpetual."
Instead of narrowing decisions through fear and automation, rushing toward premature closure and false certainty, Dossierveillance opens decisions up through systematic investigation and careful evidence gathering. It slows down where speed would be dangerous and speeds up where documentation enables learning.
This approach recognizes that most important decisions involve irreducible uncertainty, competing values, and legitimate disagreement. Rather than pretending these complexities can be eliminated through better algorithms, Dossierveillance builds systems that help people navigate complexity with greater awareness, stronger reasoning, and clearer accountability.
The choice is not between using AI or rejecting it. The choice is between AI systems that diminish human judgment through fear and opacity, and AI systems that strengthen human judgment through investigation and understanding.