White Papers

  • The papers make the case and the frameworks do the measuring. Each paper starts from a question that technical approaches leave open: why governance fails when people don’t follow it, why oversight of autonomous systems is hard to sustain, and what it takes to turn a behavioral measurement into a claim someone can rely on. Each is written for practitioners and decision-makers, and each cites the research it draws on.

    Each paper by Behavieural takes a position on a problem that organizations, regulators and assurance providers are working through now. They draw on decades of behavioral science and show the reasoning behind the frameworks and engagements.

  • Seeded-Error Testing: A Protocol for Assuring Human Oversight of AI

    12 October 2026

    Oversight isn’t proven by a person in the loop, only by what that person catches.

    Human oversight of AI is widely required and rarely tested. Assurance reports typically describe policies, roles, workflows and review logs, but seldom show whether the person in the review step notices an error when one occurs. This paper describes a method for finding out. In seeded-error testing, a set of AI outputs containing known, deliberately placed errors is put into the stream of cases that reviewers handle, and the assessor records what the reviewers do. The approach borrows from penetration testing and red-team exercises in security, where a control is challenged to see whether it holds, and from signal detection research, which provides standard measures of how well people distinguish problems from non-problems. The results produce the highest level of the evidence hierarchy proposed in the companion paper From Attestation to Evidence: direct measurement of whether reviewers detect the errors that matter.

    The protocol has six steps: define the decision point and what counts as an error, design the seeded errors and the case library, run the test using an insertion mode that fits the risk, measure detection, false alarms, sensitivity and response criterion, interpret the results against thresholds set in advance, and report with a stated evidence level and re-test triggers. A central point is that a catch rate alone is ambiguous. A team that rarely flags anything may be unable to recognize errors, or it may be able to recognize them and simply set a high bar for acting, perhaps because of workload, incentives or trust in the tool. Separating the two points an organization toward the right fix, whether that is training and interface design or the conditions in which reviewers work.

    The paper also addresses what makes such a test credible and safe. It sets out threats to validity, including reviewers who behave differently because they know they are being tested, unrealistic error rates, and organizations training staff to the test. It covers ethical safeguards, which include protecting the people affected by decisions, reporting results at the group level and not using them to discipline individuals, and it maps the method to the EU AI Act, the NIST AI Risk Management Framework and ISO/IEC 42001. It closes with candidate requirements that standards drafters could adapt. The method is presented as a proposal to be piloted, criticized and improved, and the paper is explicit that it tests only the error types and conditions it includes and is not sufficient on its own.

  • From Attestation to Evidence: Assuring That Human Oversight of AI Works

    12 October 2026

    Attestation shows that oversight of AI is intended; only measured performance shows that it works.

    Human oversight is now a requirement in the leading frameworks for governing high-risk AI, including the EU AI Act, the NIST AI Risk Management Framework, and ISO/IEC 42001. Yet the evidence organizations typically offer that oversight works is a policy, a documented design, or a statement that a person reviews the system’s outputs. These show that oversight was planned, not that it functions. Good decision outcomes are no better a guide, because an accurate system can produce good outcomes even when reviewers are not catching its errors. Assessors, auditors, and standards bodies are therefore often asked to certify something that has never been tested.

    From Attestation to Evidence proposes a five-level evidence hierarchy for assuring human oversight: policy and attestation, documented design, process records, condition indicators, and measured performance. Each level answers a different question, and only the highest asks whether reviewers detect and correct errors in practice. The hierarchy does not depend on any single account of what effective review requires, so assessors can apply it with the model of oversight they already use. The paper links the level of evidence expected to the risk tier of the system, so that scrutiny is proportionate, and maps each level to the oversight provisions of the major regulatory and standards frameworks.

    A worked example of two insurers with identical attestations but very different measured performance shows why the upper levels matter. The paper closes with guidance on what assessors and standards bodies should request, what they can reasonably accept at each tier, and where current assurance practice falls short.

  • The Behavioral Premise: Why AI Adoption Failures are Misdiagnosed

    11 October 2026

    Why AI adoption failures are diagnosed wrongly.

    When AI adoption stalls, organizations tend to reach for the same set of explanations in the same order: the technology, the organization, the governance. This paper calls that habit the diagnosis reflex and tests each explanation against the evidence. It argues that a fourth explanation, a strained behavioral condition, is often what accounts for the behavior actually observed, and that fixing the other three does not resolve a failure rooted there.

    The paper states the premise formally, grounds it in behavioral theory, and makes it falsifiable by setting out two predictions and the conditions under which they would fail. It then works through worked examples and addresses the objections a skeptic would raise.

  • From Indicator to Claim: Validating the Behavioral Constructs in AI Assurance Frameworks

    7 October 2026

    A method for reviewing standards and the evidence when the claim concerns human behavior.

    Standards, frameworks and guardrails for AI increasingly ask organizations to show that a claim holds. Many of those claims concern people: that oversight is meaningful, that disclosures are understood, that users rely on a system to the right degree. Each is only as strong as the measure behind it. Leading frameworks such as the NIST AI Risk Management Framework already ask teams to show that an indicator measures what it claims to. The open question is what a defensible argument looks like when the thing measured is behavior, who judges it, and how much evidence a given level of risk warrants.

    The paper offers a method for that gap. It adapts a measurement approach proposed for evaluating generative AI and applies it in two directions: to reading a requirement in a standard, and to reading the evidence offered under it. Behavioral science contributes the known threats to validity in behavioral evidence, such as reactivity, prevalence effects and the gap between stated and revealed behavior, along with a way to scale the evidence to the consequence of being wrong.

    The paper is explicit about its limits. It does not claim that current assurance practice is deficient, and it describes the method as unvalidated and in need of piloting before it is relied on.

  • Governing Agentic AI: The Human–Agent Relationship as the Object of Governance

    9 September 2026

    Why the object of governance must expand from the model to the human–agent relationship.

    Agentic systems execute extended sequences of actions with only intermittent human oversight. Most governance frameworks were built for systems in which a person sat at the point of every decision, and they emphasize model performance, output accuracy and regulatory compliance. This paper argues that those remain essential but do not address the behavioral risks that arise when organizations delegate meaningful authority to agents.

    It identifies five behavioral mechanisms that shape the human–agent relationship: automation bias, responsibility diffusion, anthropomorphic trust, the autonomy paradox and failure recovery. A central idea is the repair window, the interval between an agent’s first erroneous action and the moment a human notices it. Errors that compound across sequential actions become more costly the longer that window stays open, which is why the speed and quality of detection, not raw model accuracy, often separates a contained error from an institutional incident.

    The paper complements the OWASP Top 10 for Agentic AI by supplying the behavioral layer that its technical categories assume without modeling. It maps each behavioral dimension to the threat categories it can enable or suppress, and it ends with a Behavioral Agentic Threat Model as an appendix.

  • Behavioral Governance: Designing for the Behavior Policy Assumes

    1 June 2026

    Designing institutions so that intended behavior is the default.

    AI governance usually fails because people do not behave the way policy assumes, not because the policy is badly written. This paper calls that distance the assumption gap. Traditional governance models rest on rational-actor assumptions, but people act according to cognitive load, heuristics, trust dynamics, incentives, friction and social norms. A policy that ignores those forces can look complete and still be bypassed.

    Behavioral governance takes the opposite starting point. It uses established research on how people process information, respond to uncertainty and decide under pressure to design an institutional architecture in which intended behavior is the default, unsafe behavior is structurally constrained, oversight is reliable, trust is calibrated and drift is detectable. It does not try to perfect human behavior, only to make it predictable and aligned with governance intent.

    The paper sets out the theoretical foundations, introduces the AI Trust Axis as a way to assess the human side of governance, works through case scenarios, and offers a maturity model that organizations can use to locate themselves. It is the starting point for the B-GRIT engagement.