Covert AI Use Is an Organizational Governance Problem
I have been thinking about the rise of “covert AI use” in workplaces, especially in smaller settings. Nonprofit social work settings that focus on mental health, education, and social services, especially. It’s just as relevant to other settings, but small agencies tend to get overlooked when considering the broad-scale implications of problems. They also can’t always afford consultants, access to research, and expert technical advice. This has led many program staff members, including leaders, to turn to AI tools to augment their capacity and skill sets.
Many people work to conceal the use of AI software, largely because others perceive AI content negatively. This is especially true in the generation of written work, where our writing signals far more than its content. Our written work signals confidence, skill, and commitment. Conversely, when we see AI telltale signs, such as formulaic bullet points, triplicate phrasing, or the dreaded antithesis, we tend to assume negative things about the writer and the work itself. These forms of social perception push much AI use underground, but at work, it becomes more complicated.
“Covert AI use,” for the purposes of this discussion, is when workers use AI tools to complete job tasks without disclosing that use to supervisors, coworkers, clients, students, families, or the organization. Sometimes this is explicitly against policy, and sometimes there is no policy to attend to. It is valuable to distinguish this work-based pattern from other forms of AI use, where users hide their overall intent, primarily because work-based use carries different implications. Even when there is no clear policy or structure, perceived AI use without sanction or citation carries the same negative overtones as it does in our newsfeeds and daily emails. Moreover, at work, when discovered, it is commonly framed as an employee conduct problem: someone used a tool they were not supposed to use, hid it, and could have created a liability.
In many human service environments, covert AI use is a response to high workload expectations, unclear organizational rules, and negative human impact consequences of failure. The high demand for personalized, human responses to clients, along with a variety of documentation expectations, ratchets up the pressure on staff members. Often, grant funders and oversight agencies require organizational leadership to track large volumes of technical data and to respond to formal and informal data requests.
Some years ago, while working in a K-12 educational environment, a colleague joked that when the 3:00 school dismissal bell rang, it was the “principal’s lunch hour,” alluding to the constant demands placed on school administrators to be fully available when students are in the building. This is not an aberration of the public school environment. Healthcare clinic managers often cover for their front-desk staff when they are sick, leading to a backlog of their own administrative work. Social service agency staff often have full caseloads and share coverage for staffing phone lines or drop-in spaces. Mental health staff commonly cover their regularly scheduled clients while also maintaining a “crisis counselor” duty. The limited resources increasingly placed on our modern human service agencies push staff to the limit to look for efficiencies.
The informal incentive systems quickly become perverse in these resource-limited spaces. While a staff member may understand that there are risks to how AI generates its outputs, it also helps them meet the expectations placed on them. They also probably understand that admitting to using AI could make them look careless, lazy, unethical, or noncompliant. A leader in a university system I work with recently generated a college-wide summary of an upcoming meeting. When I asked about it and how they had not disclosed that as AI use, they expressed puzzlement about why they would do that, even though the system has a disclosure expectation. The knowledge that disclosure signals something “gray” or inconsistently permitted is deeply ingrained in our social perception. Similarly, the pressures to produce are intense and carry their own consequences for failure. Organizations, like their staff, benefit from faster writing, cleaner documentation, better-formatted training materials, and quicker communication, resulting in mixed policing of AI use. However, the organization also has a plan in case an issue arises, and it can push the responsibility onto individuals.
This is the perverse incentive: use the tool, get the work done, and stay quiet.
This is not only a technology problem; it is an organizational governance problem. In asymmetric workplace power dynamics between staff and systems, the person with less power often bears the greatest risk. The organization may benefit from the increased output, the supervisor may reward the polished product, and the worker may still be the one exposed if the use of AI is later questioned. The smaller the human service organization, the more it often runs on informal trust and relational accountability. Truthfully, there is also a lot of “just figure it out” mentality. While those agile cultures can be a strength, they also leave workers guessing about where the ethical lines are.
In mental health, education, social work, and social services, the use of covert AI adds complexity because the work often involves sensitive information, vulnerable people, and professional judgment. There is a meaningful difference between using AI to clean up the wording of a general training handout and entering identifiable client or student information into a public AI tool. There is also a meaningful difference between using AI to organize a rough draft and asking AI to generate clinical impressions, risk language, eligibility decisions, placement recommendations, or documentation for an official record.
We cannot allow these lines to be blurred, as there are significant implications to this outcome.
This is where many organizations struggle. The easiest response is vague fear or blanket prohibition. Both may sound appealing and protective to the organization, but in practice, they tend to push AI use further underground. When staff believe that any AI disclosure will be treated as suspicious, they are less likely to speak openly about the tools they use, the information they enter, and where they need guidance. Supervisors then lose the opportunity to coach the work, organizations lose the opportunity to manage risk, and teams lose the chance to build shared standards. Furthermore, when we do not openly disclose the use and components of AI, we reinforce secrecy. We can inadvertently let people think that they have to pretend that it is also human work, even when it isn’t
A better approach would begin openness: make safe disclosure easier than secrecy.
Staff should know which kinds of AI use are allowed, which require review or approval, and which are clearly off-limits. A worker should be able to say, “I used AI to help organize this draft, but I did not enter identifying information, and I reviewed the final version myself,” without fearing that the disclosure alone will be treated as misconduct. That kind of statement should open a supervisory conversation, not automatically create a disciplinary one. This transparency is beneficial in the workplace culture and the community at large. Clients deserve to know what has been used and what hasn’t; where the data is safe.
This does not mean treating AI use casually. In fact, it means the opposite. Organizations need clearer distinctions between low-risk administrative support, higher-risk professional drafting, and prohibited uses involving private data, clinical judgment, eligibility decisions, or official records. Without those distinctions, workers are left to interpret ethical risk on their own, often under time pressure, while the organization still benefits from the speed and polish the tool provides. However, these kinds of conversations and discussions require the psychological safety fostered by open disclosure policies.
Covert AI use should also be used as a quality improvement measure. It indicates that workload, documentation burden, supervision, policy, and technology adoption are misaligned. It can also communicate the early warning signs of worker burnout.
The goal should not be to catch people using AI.
The goal should be to bring AI use into the open, where it can be guided by ethics, supervision, professional judgment, and shared responsibility.
To that end, I will disclose that the text in this document is human-drafted, with spelling and grammar revisions provided by the Grammarly AI tool. Brainstorming and revisions: suggestions were developed through both human and AI reviews of the text fo the purpose of improving the overall clarity and readability of the final human-written draft.
