Shadow AI

Shadow AI — the use of AI tools that an institution has not approved, vetted or even catalogued, by the people inside it. The term descends from shadow IT, but the object is different: shadow IT substituted unauthorized systems for approved ones, mostly in information processing, while shadow AI covers tools that generate content, support decisions and perform knowledge work. Control therefore shifts from managing access to regulating use, and the person using the tool is usually the one making the judgment call.

Questions to Consider

Can you name the AI tools your staff and faculty used last month, or only the ones you bought?
What does your institution require someone to do to get a tool approved, and how long does that take compared with the week in which a new model ships?
If a teacher uploads student work to an unapproved tool today, who finds out, and would anyone know if it never caused a visible incident?
Is the absence of guidance in your institution a decision, or a gap that individual educators are filling on your behalf?
What would you have to publish for a cautious educator to choose the sanctioned path without losing a week of preparation?

Introduction

The adoption literature keeps returning to the same gap. In EDUCAUSE's 2026 survey of 1,960 staff and faculty, 94% had used AI tools for work in the past six months, while only 54% were aware of the policies meant to guide that use. Almost all respondents' institutions, 92%, reported a work-related AI strategy, and just 13% were measuring return on investment for it. More than half, 56%, said they used tools their institution does not provide.

That combination describes a governance surface that exists on paper and a practice that exists elsewhere. People are not waiting for permission, and in many institutions they are not being asked to. Shadow AI is the name for what fills the space between a strategy document and a classroom.

The literature complicates the moral framing. Rameskumar and colleagues interviewed 20 academic educators in Germany and found that informal use was not hidden cynicism but a practice educators justified to themselves in professional terms. Kelso and colleagues surveyed 375 educators who named 1,373 apps and services, 494 of them unique. Dangol and colleagues followed 22 teachers in a U.S. district that had endorsed AI before writing rules, and found them drawing their own boundaries. Shadow AI, on this evidence, is what a provision gap looks like from the inside.

Why it happens

The proximate cause is usually missing guidance. Educators in the German study described rules that were absent, too general, fragmented, or disconnected from concrete teaching situations. One said plainly that nobody had told them what was or was not allowed.

The structural cause is speed. Procurement cycles and policy committees move at the pace of institutions; model releases move at the pace of vendors. A tool that a colleague demonstrates on Monday can be in use by Friday, and the approval process that would have caught it has not opened yet.

The motivational cause is utility, and it is not close. In the ASU survey, pedagogical utility accounted for 213 of the selections educators gave for choosing a tool, functional convenience 155, accessibility 187 and cost 164. Security was named by 24 of 375. Discovery ran through people rather than procurement: 290 learned about tools from professional peers, 141 from administrators, 100 from search, and 100 from advertising.

Cost and access deserve their own line. Free tiers are the reason many educators can try a tool at all, and the reason the institution cannot see the trial. When the sanctioned alternative is a licensed seat with a waiting list, the free tool wins on the only criterion the teacher is measuring that afternoon.

How educators justify it

The German study's contribution is a sequence rather than a count. Educators first classified a use by whether it fit the professional role: preparing materials, organizing content and generating examples counted as role-supportive; wording an email sat outside the role; producing work that displaced student thinking was problematic. They then invoked efficiency, with time savings treated as sufficient justification on their own. They added value claims, that the material was better or more varied, not merely faster. And they closed with a responsibility move: the tool is neutral, the educator decides, and the educator remains accountable for what reaches students.

That last step is the one institutions should study. It is also the one that makes a ban feel irrelevant, because the person applying it already believes they exercised judgment. The paper reads it as moral legitimation rather than as evasion, which is a more uncomfortable finding than simple noncompliance.

What it costs

The ASU survey gives the harm side concrete numbers. Of 375 educators, 167 encountered privacy- or security-invasive behavior and uninstalled the app, 108 of them in K-12. But 62 who worried their data was being sold or shared kept the tools anyway, citing workflow dependence. Administrators named 16 serious concerns, led by data leakage and privacy invasion.

Integration is where exposure compounds. K-12 educators connected 341 apps to institutional systems and higher education educators 116, and most of those apps required account provisioning. Nearly a third of K-12 educators, 30.3%, knew of an institutional policy covering unsanctioned use; 22 admitted violating one.

The privacy problem is structural rather than individual. A tool that receives a class list, an IEP or a disciplinary note has created a record the institution did not authorize, cannot audit, and may only learn about when a breach notification arrives.

What teachers actually do about it

The district study is the most useful corrective to the image of the careless adopter. Teachers there had a district-wide platform, MagicSchool AI, and still used general systems such as ChatGPT and Gemini. Their boundaries were pedagogical: one allowed AI during growth work and withdrew it at mastery, so students had to demonstrate the skill unaided. A special education teacher taught fourth graders to interrogate authorship, treating rephrasing as acceptable and idea generation as not.

Voice and privacy followed. Teachers rewrote outputs so materials still sounded like them, cross-checked facts, and treated generated text as a majority source rather than an authority. They redacted student names and used district accounts deliberately, so that their usage could be audited if it were challenged. That is collaborative sensemaking with an audit trail, and it is closer to what a governance office says it wants than to what a compliance report usually captures.

The cost of this improvisation falls unevenly. Teachers who had time, confidence and collegial support built the practices; teachers in schools without them did not. Equity gaps that begin as staffing gaps end as differences in what students are asked to do.

What institutions can do

The first move is to publish the sanctioned path, not to close the unofficial one. The teachers in the district study asked for rules, vetted-tool lists, examples and time, and none of those requests require a new platform.

The second is to shorten the path. MindRouter and the Fontys gateway both show a model in which approval is a routing decision rather than a procurement event: an approved model appears in a gateway with a model card naming region, cost and alternatives, and a teacher can use it the same day. Institutional AI infrastructure is what makes that possible; the absence of it is what shadow AI is made of.

The third is to measure what is being spent and used, including the spend that never reached a purchase order. EDUCAUSE's 13% ROI figure suggests that most institutions cannot currently answer whether the sanctioned tools are earning their seats, which leaves the argument for them resting on authority alone.

The fourth is to treat AI literacy as a control, not a curriculum add-on. A teacher who can read a data policy, recognize a vendor's retention terms and redact a student name is doing risk management that no approval form replaces.

Implications for Practice

Faculty developers can publish the sanctioned path with worked examples, in the language of teaching rather than procurement. The teachers in the district study asked for rules, a vetted list, examples and time; none of those requires a new platform. An approval route that takes a week is long enough for a new model to ship, so the useful target is days.

Instructors can adopt the habits the careful users in that study already had. Redact student names before pasting anything, use an institutional account so the record can be audited, and treat generated text as a draft to check rather than an authority to cite.

Administrators can measure the spend that never reached a purchase order and read unapproved use as a report on an unmet need. If a free tool wins because the licensed one has a waiting list, the adoption problem is the queue.

IT and policy staff can turn approval into a routing decision. A gateway that lists approved models with region, cost and alternatives lets a teacher start the same day, which is the difference between governance that shapes use and governance that documents it afterward.

Department chairs can build the collegial support that let some teachers develop boundaries and left others without them. The gap that begins as staffing becomes a difference in what students are asked to do.

Common Misconceptions

"Shadow AI is just shadow IT with a new name." The German study draws the distinction: shadow IT risks attach to infrastructure, while shadow AI risks attach to content, judgment and the record of what was produced.

"A ban will end it." In the ASU survey, 30.3% of K-12 educators knew of a policy and 22 admitted breaking one. Prohibition without a sanctioned alternative changes where the use happens, not whether it does.

"This is a student problem." Every study here is about educators: preparation, grading, communication, translation. The student story is real, but it is not the one with administrative credentials attached.

"Awareness training solves it." 94% used the tools and 54% knew the policy, and the 40-point gap did not close when people were told. What closes it is a usable alternative with a short path to approval.

Connected Concepts

Connected Articles

The Impact of AI on Work in Higher Education — EDUCAUSE: 94% used AI for work, 54% knew the policy, 13% measured ROI
How Is Shadow AI Legitimized in Higher Education? A Qualitative Study Among Academic Educators in the Absence of Institutional Guidance — a three-step account of how educators justify unsanctioned use
'It's not approved, but many, like myself, ignore the rule': Investigating the Landscape and Consequences of Unauthorized Technology Use in Educational Institutes — 375 educators, 1,373 apps, 24 naming security as a reason
Relief or displacement? How teachers are negotiating generative AI's role in their professional practice — teachers drawing pedagogical boundaries in a policy vacuum
AI Intrapreneurship: Educators as Frontline Innovators in the Age of Generative AI — educator-led adoption as the alternative to waiting for institutional guidance

Metadata

Discipline	Information Technology
Level	Higher Ed, K 12
Intended audience	Administrators, Instructors, Faculty Developers, Educational Technology Developers
Technologies and techniques	Generative AI, Large Language Models (LLMs)
Equity, ethics, and responsible use	Privacy, Equity

Source History Edit Chat