When AI Tests the Values We Claim to Hold
I keep coming back to a question that sits uncomfortably beneath many of our conversations about artificial intelligence: what, exactly, do we expect an ethical commitment to change when the most profitable, convenient, or competitively attractive decision takes us in another direction?
It is relatively easy to speak about responsible innovation while responsibility appears compatible with everything else we want to achieve. We can support transparency when disclosure does not weaken our position, defend human oversight when it does not affect the savings we have promised, and express concern for privacy while the information we would like to retain remains available to us. The more revealing moment comes when those commitments begin to constrain a decision, because this is where we discover whether our values have any influence over the way we intend to operate.
Imagine an organization preparing to introduce an AI service that will handle a substantial part of its customer interactions. In this hypothetical situation, the early results are encouraging: routine requests are processed quickly, the responses appear coherent, and the business case suggests that the organization could significantly reduce the work assigned to its support teams. Yet testing also reveals that some complex situations are handled poorly, particularly when a customer’s circumstances do not fit the assumptions around which the service was designed. Additional human support could address part of that weakness, although retaining that support would reduce the expected savings.
The leadership decision now concerns much more than the performance of a model. It concerns whose experience will count when performance is assessed, who will absorb the consequences of the remaining limitations, and whether the organization is prepared to revise an attractive business case because the evidence no longer supports it in its original form. A reduction in internal workload may still represent a useful improvement, but if customers must spend more time correcting misunderstandings, repeating their circumstances, or finding someone with the authority to help, some of the cost may simply have moved outside the organization’s preferred field of measurement.
This is where Congruence becomes more demanding than the reassuring idea of being consistent. As developed in the Ethical Leadership Practice Guide, congruence concerns the relationship between what we believe, what we say, and what we do; its significance lies in whether those three dimensions remain coherent when circumstances make alignment inconvenient (Ethics & Leadership Association [ELA], 2026a, p. 7). A leader who revises a launch commitment because testing has exposed a serious weakness may therefore be acting with greater congruence than one who protects the original date, particularly if the commitment that matters most is the promise to serve people responsibly.
That distinction matters because we can become remarkably attached to the visible expression of a decision while losing sight of the purpose that justified it. The date becomes a promise, the promise becomes a measure of credibility, and eventually the ability to preserve the promise begins to matter more than the conditions under which it can responsibly be fulfilled. By the time someone raises an inconvenient concern, the organization may already have invested enough reputation in the preferred outcome that reconsideration feels like a failure of leadership.
The Integrity Model offers a useful way of examining this tension because it separates four dimensions that are too easily collapsed into a single judgment: Ability, Legal, Ethics, and Values. Ability concerns whether we have the practical means to do something, including the competence, resources, information, and operational capacity required to do it responsibly. Applied to our hypothetical AI service, that would include the ability to support the situations the technology cannot handle well, rather than stopping the assessment at the point where the software produces an answer.
The Legal dimension concerns the obligations and permissions governing the activity, while Ethics asks how the decision relates to the standards and codes through which a profession or organization defines responsible conduct. Values brings the examination back to the convictions through which people understand what they are prepared to stand for, including the relationship between those convictions and the commitments expressed by the organization. These dimensions influence one another, but they answer different questions; the fact that something is technically possible, for example, does not establish that it is permitted, appropriate, or consistent with the principles we claim to hold (ELA, 2026a, pp. 54–55, 79–80).
What makes the model useful, in my view, is precisely its resistance to the temptation to let one attractive answer settle the whole discussion. A capable system may still be used for a poorly justified purpose, a legally defensible arrangement may still treat people in ways that conflict with our professional commitments, and a sincere conviction may still require examination when its consequences become clearer. The model helps us expose those tensions; the difficult work of judgment, explanation, and ownership remains with the people making the decision.
This brings me back to a point I explored in How We Show Up: we often speak about organizations as though they were independent beings capable of deciding, permitting, refusing, or taking responsibility, when the decisions through which an organization acts are made and sustained by people (Lazar, 2026a). AI can make that distinction harder to see because an outcome may pass through a provider, a development team, a business sponsor, a configuration process, and an automated workflow before it reaches the person affected. Each participant may have fulfilled a particular responsibility, while the obligation to explain and answer for the resulting decision remains surprisingly difficult to locate.
Responsibility and accountability therefore deserve more careful treatment than simply assigning a name to the last approval step. Different participants need to answer for different choices: what was designed, what limitations were disclosed, what use was authorized, what operating conditions were provided, and what happened when evidence suggested that the arrangement was causing a problem. A provider’s technical competence does not establish that a particular deployment is appropriate, just as a deploying organization’s approval does not relieve those developing the system of responsibility for the choices and representations within their control.
Courageous Responsibility begins to acquire substance here, because owning a decision involves being willing to make its unresolved consequences visible before circumstances make them impossible to avoid. A fictional teaching case, The Warning No One Wanted, illustrates this through Jane, a team leader who identifies readiness weaknesses before a service platform is launched. Her concerns are recorded and acknowledged, yet their meaning is progressively softened as they move through the organization, until she asks the leadership team to make an explicit decision about the remaining risk and identify who will answer for accepting it. The case concerns a service platform rather than an AI deployment, but the distinction it exposes is directly relevant: a warning can remain present in the documentation while having very little influence over what happens next (ELA, n.d.).
That possibility becomes especially troubling when people hold sincere ethical commitments while working within arrangements that penalize the conduct those commitments require, a tension also raised in ELA’s recent reflection on hidden dependency (ELA, 2026b). If a sponsor’s standing depends on delivering a particular saving, if managers are rewarded primarily for reducing handling time, and if reviewers must justify every intervention that slows the process, the organization has already shaped the conditions under which its ethical commitments will be interpreted.
This is also why Service Before Status has a practical meaning that extends well beyond the expectation that leaders should be considerate. It asks what authority is being used to protect, particularly when protecting the interests of those being served requires someone in a senior position to accept embarrassment, revise a public commitment, or acknowledge that an earlier judgment was incomplete. Expecting an employee to demonstrate moral courage while leaving that employee to carry the professional cost of challenging a powerful sponsor would be a rather selective interpretation of ethical leadership.
Amy Edmondson’s work on psychological safety helps explain why the conditions surrounding such a challenge matter. Psychological safety concerns whether people believe they can take interpersonal risks within a team, including admitting uncertainty, asking for help, or discussing an error; her 1999 study found an association between that shared belief and team learning behavior. Applied to AI, the implication I would draw is that an organization needs people who can question both a system’s output and the assumptions of those sponsoring it, with a reasonable expectation that the concern will be examined on its merits. If raising a limitation is interpreted as resistance to innovation, the organization may weaken its own ability to discover what responsible use requires. (Edmondson, 1999)
The familiar promise of human oversight needs to be examined with the same attention to operating conditions. Returning to our hypothetical service, suppose the organization retains a person to review difficult cases, but also reduces staffing, increases the number of cases each reviewer must handle, and measures performance through the speed at which those cases are closed. The presence of a human reviewer may satisfy the description of the process while leaving that person with very little practical capacity to exercise the judgment upon which the process supposedly depends.
There is a useful historical perspective here. In Ironies of Automation, Lisanne Bainbridge examined how the automation of industrial processes could expand the problems faced by human operators, particularly when they remained responsible for handling abnormal conditions. Although her subject was industrial automation, the question carries into the discussion of AI: when we rely on people to manage the situations that automation cannot adequately resolve, what competence, information, time, and authority are we preserving for them? Treating those conditions as optional overhead would weaken the very mechanism we are relying upon to make the arrangement responsible. (Bainbridge, 1983)
At the same time, we should be careful about assuming that adding a person necessarily improves every outcome. The NIST AI Risk Management Framework’s discussion of human–AI interaction recognizes that the results depend on how the relationship is configured, and that human and system biases may sometimes reinforce one another. Meaningful oversight therefore requires an examination of the work being performed, the consequences involved, and the particular contribution a person is expected to make; an approval click, by itself, tells us very little about any of those things. (Tabassi, 2023, Appendix C)
Transformational Humility provides an ethical foundation for maintaining that examination. It involves recognizing that confidence in a decision does not remove the possibility that our understanding is incomplete, and that new evidence may require us to change the way we proceed. For AI deployment, this could mean treating the initial authorization as a judgment that remains open to review, with explicit conditions for expanding the use, narrowing it, correcting it, or bringing it to an end. The commitment to learning has greater credibility when those possibilities have been considered before the organization becomes dependent on the preferred result.
Privacy introduces another version of the same tension. Consider a second hypothetical situation, an AI service intended to help people think through difficult professional circumstances. Detailed conversations could potentially help improve the service, while also containing information about colleagues, clients, or other people who never chose to interact with it. The question of whether to retain those conversations would therefore involve more than the value of the resulting dataset, because the people creating value for the service and the people exposed by that information may not be the same.
This is where Shared Stewardship broadens the scope of leadership responsibility. It asks us to consider the trust, relationships, resources, and consequences placed in our care, including those that extend beyond our immediate role or direct transaction. A decision about information retention would then require a clear account of what is needed, for which purpose, who can access it, how long it remains necessary, and what commitments can actually be supported about its handling. Accountability may require records of decisions, tests, and corrective action, but that requirement does not automatically settle whether every underlying conversation should be kept (ELA, 2026a, pp. 7, 79–80).
Across these examples, the ethical cost becomes easier to understand once we stop assuming that the organization’s own expenditure captures the whole consequence of its decision. A saving in support capacity may become additional effort for a customer, a faster release may become uncertainty carried by employees, and a more useful dataset may create exposure for people whose circumstances were never part of the business case. Shared Stewardship asks leaders to bring those consequences into the discussion, even when the people carrying them have little influence over the decision.
There is, however, another risk that deserves equal attention, because an organization can respond to these concerns by accumulating reviews, committees, declarations, and reporting requirements without becoming much better at addressing them. In When Governance Forgets Its Purpose, I explored how the arrangements intended to support responsibility can gradually become more invested in their own preservation than in the mission they were meant to serve (Lazar, 2026b). AI governance is vulnerable to that same drift when producing evidence of oversight becomes easier to recognize and reward than correcting the conditions that make oversight ineffective.
A useful governance process should therefore help people make and revisit decisions with a clearer understanding of their consequences. Its contribution should be visible in what changes: an unsuitable use is restricted, a concern reaches someone with the authority to act, a person receives meaningful reconsideration, or a misleading performance measure is revised. Where another layer of approval adds little to that capacity, its existence deserves examination alongside every other claim made in the name of responsibility.
Nor should this argument be understood as a presumption that delay is always the more ethical choice. Delaying a useful service can carry consequences of its own, particularly when people continue to depend on an existing arrangement that is slow, inconsistent, or inaccessible. The task is to examine the available alternatives honestly, including the cost of proceeding, the cost of waiting, and the distribution of both; uncertainty cannot always be removed, but it can be made visible enough for a decision to be explained, challenged, and revised.
What concerns me most is the possibility that we may ask AI systems to express values that we have not made operational within the organizations deploying them. We can ask for transparency while protecting opaque decisions, insist on human agency while rewarding compliance, or declare that people matter while making their needs disappear inside an aggregate measure of efficiency. In such circumstances, improving the technology may be necessary, but the questions of leadership would remain.
This is why Congruence, Courageous Responsibility, Transformational Humility, Service Before Status, and Shared Stewardship deserve to be understood as connected disciplines. They help us examine whether our commitments remain coherent, whether consequences are owned, whether judgment can evolve, what authority is protecting, and whose interests enter the decision. Their relevance to AI becomes clearer when they affect the conditions under which a system is developed and used, including the occasions when respecting them requires us to accept a less convenient outcome.
Perhaps, then, the question to bring into the next discussion about responsible AI is whether we can identify a decision that our ethical commitments have actually changed. And if the people affected could see how that decision was made, what we chose to preserve, and which costs we were prepared to carry ourselves, would they recognize the values we say are guiding us?
References
Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. DOI: 10.1016/0005-1098(83)90046-8.
Edmondson, A. (1999). Psychological safety and learning behavior in work teams. Administrative Science Quarterly, 44(2), 350–383. DOI: 10.2307/2666999.
Ethics & Leadership Association. (n.d.). The warning no one wanted [Written teaching case]. In Courageous Responsibility: Case Study Analysis. ELA online learning.
Ethics & Leadership Association. (2026a). The Ethical Leadership Practice Guide (1st ed.). Particularly pp. 7, 54–55, and 79–81.
Ethics & Leadership Association. (2026b, September 9). What does hidden dependency actually cost an organization? [Post]. LinkedIn.
Lazar, O. (2026a, March 3). How we show up: The beginning and the end of leadership. Ethics & Leadership Association.
Lazar, O. (2026b, March 22). When governance forgets its purpose. Ethics & Leadership Association.
Tabassi, E. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0) (NIST AI 100-1). National Institute of Standards and Technology. Particularly Appendix C, “AI Risk Management and Human-AI Interaction.”
Comments