AI speak-up culture should be measured using several forms of evidence rather than a single employee survey or incident count. A useful assessment combines what people believe about raising concerns, what they actually do, how the organization responds, and the context in which AI is being used. The most informative findings often appear where those evidence sources disagree.
This matters because the most visible metric is not always the most meaningful one. An organization may report very few AI concerns because its systems are well controlled, because employees do not recognize weak signals, because reporting routes are unclear, or because people have learned that raising an issue achieves very little. The number is real. Its meaning still has to be established.
Measurement therefore has to answer a harder question than “How many problems were reported?” Senior leaders need to understand whether their organization can reliably detect, move, interpret, and act on human observations about AI risk.
A mature measurement approach needs evidence from several parts of the system.
At a high level, organizations should examine four types of evidence:
| Evidence type | What it helps reveal | Example questions |
|---|---|---|
| Perception evidence | What employees believe about speaking, challenge, authority, and organizational response | Do people believe they can raise uncertainty safely? |
| Behavioral evidence | What people actually do around AI | Are employees challenging, escalating, overriding, or intervening? |
| Process evidence | What happens after a signal enters the organization | Are concerns routed, reviewed, resolved, and closed effectively? |
| Context evidence | What level and type of signal would reasonably be expected | How much AI exposure, autonomy, consequence, and organizational change exists? |
None of these provides a complete diagnosis on its own.
The purpose of combining them is to determine whether the evidence tells a coherent story about how risk information moves through the organization.
Surveys are useful because culture includes beliefs that cannot be observed directly. Employees can tell you whether they feel comfortable raising concerns, whether managers respond constructively, whether they believe intervention will be supported, and whether using an escalation route feels professionally safe.
Amy Edmondson's foundational research on psychological safety demonstrates why those perceptions matter. Psychological safety influences whether people engage in learning behaviors such as asking questions, admitting mistakes, seeking help, and challenging assumptions.
For AI risk management, however, self-reported confidence should be treated as evidence rather than proof of control effectiveness.
Employees may report that they feel comfortable speaking up while formal AI concerns remain almost nonexistent. They may believe their manager welcomes challenge while still assuming that only confirmed failures belong in the reporting system. A team can be psychologically safe and still lack clear escalation routes, sufficient AI literacy, or meaningful intervention authority.
That is why psychological safety in AI risk management should be measured alongside behavior and organizational conditions rather than used as a substitute for them.
The survey tells you something important about the environment. It does not tell you the whole system works.
Perception evidence should help explain whether employees believe they can participate effectively in AI risk management.
Relevant areas include whether people feel able to question AI-generated recommendations, acknowledge uncertainty, report mistakes, challenge people with greater authority, and raise concerns before they have complete evidence. Organizations should also understand whether employees believe reported concerns receive serious attention and whether exercising legitimate intervention authority will be supported.
The wording matters because generic culture questions can miss AI-specific conditions. Someone may strongly agree that their team is open and supportive while feeling considerably less comfortable challenging a high-profile AI deployment backed by senior leadership.
Context-specific questions can reveal that difference.
This is also where organizations need to distinguish employee sentiment from organizational capability. “I feel safe speaking” and “I know how to escalate an AI concern beyond my manager” measure different things. Both can influence the same outcome.
Behavior provides another view of the system because it shows what happens when people encounter AI in real work.
Depending on the environment, relevant evidence might include:
The goal is not to maximize any of these behaviors. High override rates, for example, could indicate healthy skepticism, poor model performance, badly designed workflows, or inappropriate use of automation.
The behavior becomes useful when interpreted in context.
A consequential agentic workflow with substantial autonomy might reasonably generate more intervention and escalation activity than an internal tool used to help employees draft routine text. Comparing the raw counts without understanding that difference would produce a very poor risk assessment.
Behavioral evidence tells you what people are doing. Interpretation tells you why it matters.
A strong AI speak-up system needs visibility into concerns before they become obvious incidents.
Weak signals can include unusual outputs, unexplained changes in behavior, unexpected system access, repeated workarounds, patterns of human override, or a person simply noticing that an automated process no longer behaves as expected.
These observations often arrive with uncertainty attached. That is part of their value.
The FAA's voluntary reporting programs illustrate why safety disciplines invest in mechanisms for surfacing unusual events, errors, and near misses before they become larger failures. The information creates an opportunity to identify patterns and improve the system.
The same principle applies to AI. If an organization only records confirmed incidents, it loses visibility into the early stages of risk formation.
Our article on the warning signs of a weak AI speak-up culture explores this problem in more detail. One of the most important signals is a reporting environment where employees wait until the evidence is overwhelming before concerns become visible.
That produces clean incident data and poor early warning.
Once someone raises a concern, the organization itself becomes part of the measurement problem.
A strong reporting culture can still produce poor risk outcomes when concerns enter a slow, confusing, or ineffective process. Organizations therefore need to examine how signals move after they are raised.
Useful evidence may include the time between reporting and initial response, the functions receiving concerns, whether issues reach people with appropriate expertise, the use of independent escalation routes, the proportion of issues that receive closure, repeated concern categories, and whether validated problems produce corrective action.
Response time deserves particular attention because AI systems operate at very different speeds.
A two-day review may be perfectly reasonable for one use case and meaningless for an agent capable of taking consequential actions in seconds. Measurement has to reflect the time available for human intervention before the outcome becomes difficult or impossible to reverse.
This is why AI escalation paths need to be assessed as operating mechanisms rather than simply documented processes. The existence of a route tells you very little about whether information reaches the right person early enough to matter.
Intervention authority provides a particularly useful example of why multiple evidence sources matter.
The organization can begin with formal design. Which roles are allowed to pause an AI-enabled process? Who can require human review? Who can restrict an agent's permissions, reject a recommendation, or stop execution?
That establishes authorized intervention.
The organization can then look at behavior. When relevant conditions actually occurred, did people use that authority?
That establishes evidence about exercised intervention.
The gap between the two is potentially important.
As discussed in What Is Intervention Authority in AI Governance?, low use of intervention authority can have several explanations. The system may simply perform well. Employees may fail to recognize relevant conditions, misunderstand their authority, lack practical ability to stop the workflow, or believe that intervention carries too much organizational cost.
A policy review cannot distinguish among those possibilities.
Measurement can reveal the gap. Diagnosis still requires interpretation.
Risk signals do not have universal meanings.
Ten AI concerns in a year could represent a lot of activity or almost none. The answer depends on the size of the workforce, the number of AI users, the type of systems being used, the autonomy of those systems, the volume of decisions, and the consequences attached to the work.
Context also includes organizational conditions.
A team operating under severe productivity pressure may behave differently from one with greater time for review. A newly deployed system may generate more uncertainty than a familiar one. A function where employees routinely challenge technical decisions may have a different baseline from one with strong hierarchy.
NIST's AI Risk Management Framework reflects this broader view of measurement. The AI RMF Core combines governance, mapping, measurement, and management while explicitly connecting AI risk to organizational culture, roles, human oversight, and the context in which systems operate. NIST has also described the framework as intended to support an AI risk-aware culture rather than function as a compliance checklist.
For human risk, this means measurement needs enough organizational context to explain the evidence rather than merely count it.
Some of the most useful insights appear when different evidence sources agree or disagree.
Cybermaniacs uses concordance as a useful way of thinking about whether different forms of human-risk evidence tell a compatible story. The public concept is straightforward: if perceptions, behavior, processes, and outcomes point in different directions, the disagreement is worth investigating.
Consider a few examples:
| Evidence pattern | What it may suggest |
|---|---|
| High psychological safety + low reporting | Recognition, reporting routes, or perceived usefulness may be weak |
| High formal intervention authority + very low intervention | Authority may be poorly understood, difficult to exercise, or culturally inhibited |
| High reporting + poor closure | Employees are speaking, but organizational response may be weak |
| Low reporting + high AI exposure | Strong controls are possible, but weak visibility should also be considered |
| Leaders report openness + employees report retaliation concern | Experienced culture may differ materially from leadership intent |
| Frequent override + recurring system pattern | Local human correction may be masking a systemic problem |
These patterns are hypotheses to investigate rather than automatic diagnoses.
That distinction is important. A mature measurement system should create better questions before it creates stronger claims.
Not necessarily.
Higher reporting can reflect more underlying problems, greater exposure, better detection, stronger psychological safety, easier reporting mechanisms, or several of those things at once.
The same ambiguity exists in other safety disciplines. More visible incidents can sometimes accompany stronger learning because conditions that were previously hidden have entered the reporting system.
This makes simplistic benchmarking difficult.
An executive dashboard that ranks business units according to who reported the fewest AI problems could create exactly the wrong incentive. Teams may learn that visibility makes them look worse, while quiet teams appear safer without demonstrating anything about the quality of their risk environment.
Measurement should reward understanding rather than silence.
The question is whether the volume and type of signals make sense given the work being performed and whether those signals lead to appropriate action.
A summary score can be useful for communication and trend tracking, but it should not be mistaken for the underlying risk.
Human Risk Management often suffers when organizations compress complex evidence into one number and then begin managing the number rather than the condition it represents. AI speak-up culture is especially vulnerable to that problem because several different failure mechanisms can produce the same headline result.
Two organizations could receive similar culture scores while having completely different issues. One may have low psychological safety. Another may have confident employees and ineffective escalation routes. A third may have good reporting and poor organizational response.
The intervention should differ accordingly.
Cybermaniacs' measurement work therefore distinguishes the analytical output from the underlying condition. Scores, metrics, rates, surveys, signals, patterns, and risk indicators can all contribute useful information, but they perform different jobs.
The goal is to understand enough of the system to explain the risk and choose an appropriate response.
The dashboard should help a CISO, CIO, CAIO, risk leader, or board understand whether the human sensing and response system appears healthy.
A useful executive view might include trends around employee confidence in raising concerns, weak-signal and near-miss visibility, escalation activity, intervention patterns, response and closure, recurring issues, and meaningful variation between populations.
The temptation will be to reduce all of this to green, amber, and red.
Some simplification is unavoidable at executive level, but the dashboard should preserve enough context to prevent easy misinterpretation. A sudden increase in reporting after a new speak-up initiative, for example, may represent improved visibility rather than deteriorating culture.
Similarly, low escalation can be healthy when local decision rights work well. It can also indicate that concerns never travel.
The dashboard should therefore support investigation, not replace it.
Improvement should appear across more than employee sentiment.
An organization might reasonably expect stronger confidence in raising concerns, better recognition of what should be reported, more appropriate weak-signal visibility, clearer use of escalation routes, more consistent intervention behavior, improved response times, stronger feedback closure, and fewer repeated problems after corrective action.
The pattern matters more than any single movement.
Improvement might initially produce more reports. Later, better controls could reduce recurrence while reporting confidence remains high. That would tell a much more useful story than simply showing a falling incident count.
Time also matters because culture changes through organizational experience. People update their beliefs based on what happens when they or their colleagues raise concerns. Measurement should therefore look for persistent change rather than rely entirely on one-time surveys.
The organization needs to understand whether the conditions influencing behavior are actually changing.
AI speak-up culture illustrates why Human Risk Management has to extend beyond training activity and isolated employee risk scores.
The relevant evidence crosses competency, psychology, behavior, culture, organizational conditions, technology use, and the context of the work itself. The task is to understand how those factors interact well enough to identify meaningful risk conditions and determine what kind of intervention could change them.
That is also why the wider AI speak-up culture cluster distinguishes psychological safety, escalation, intervention authority, and organizational response rather than treating “culture” as one large variable.
Cybermaniacs has been developing the models, evidence structures, diagnostics, and measurement approaches required to examine human risk across these layers. Our AI Enablement & Culture Management (AIECM) work applies that approach to workforce AI adoption, culture, capability, and behavior, while Agent Risk & Control (ARC) extends it into environments where humans supervise and collaborate with increasingly autonomous systems.
The public framework helps organizations understand what needs to be measured.
Turning those observations into reliable evidence about risk, determining how confident to be in the interpretation, and selecting an appropriate intervention requires a deeper measurement architecture.
Treat every metric as evidence about the system rather than as the system itself.
A survey result is evidence. An escalation is evidence. An override is evidence. A near miss is evidence. A lack of reports is evidence. None of them carries a single fixed meaning outside the context in which it occurred.
Strong measurement brings those signals together and asks whether they form a credible picture of how people recognize, communicate, and act on AI risk.
The central measurement problem is visibility. Organizations need enough evidence to understand whether weak signals can become visible, whether they can travel to someone with appropriate authority, and whether the response strengthens or degrades future reporting.
That is a more demanding standard than counting incidents.
It is also much closer to understanding whether the human part of AI governance is actually working.