Before You Click Generate: Teaching Special Educators to Us…

Dec 22/Jan 23 Closing The Gap Solutions - Collaborating With Augmentative and Alternative Communication (AAC) Users Gains A New Perspective To Best Support Clients By Lydia Dawley

? accessibility & UDL

Before You Click Generate: Teaching Special Educators to Use AI Well, Not Just Fast A look at PAUSE and PBNJ, two practical frameworks helping special education teachers close the gap between AI adoption and AI training Summary: This article will introduce readers to two practical frameworks for special education teachers navigating AI: PAUSE, a five-step ethical check covering privacy, accuracy, bias, independence, and equity, and PBNJ, a research-grounded prompting method for generating classroom-ready, differentiated materials. It will explain how these frameworks connect to FERPA, IDEA, and UDL 3.0, extend the same privacy caution to families, and offer teachers and parents concrete, low-lift starting points for using AI responsibly with students who have disabilities.

THE REALITY TEACHERS ARE ALREADY LIVING IN The case for structured training starts with a blunt set of numbers. Data drawn from Center for Democracy and Technol- ogy research published in October 2025 found that 57 percent of special education teachers had already used an AI tool on an IEP or 504 plan task in the 2024 to 2025 school year, while only 22 percent had received any training on the risks involved. No federal laws currently address AI use in special education, leav- ing individual teachers to make high-stakes decisions with few guardrails. Naming that gap is not meant to frighten teachers away from AI. It is meant to make the case for informed use rather than either blanket avoidance or unexamined trust. The underlying message of most current AI ethics training for educators is not that teachers should stop using AI, but that they should use it with their eyes open. WHAT HAPPENS THE MOMENT YOU HIT GENERATE A scenario many teachers recognize immediately illustrates the risk clearly. A teacher, hoping to save time, pastes a real stu-

Across the country, special education teachers are already us- ing artificial intelligence to draft progress notes, adapt reading passages, and brainstorm behavior supports. National surveys put regular AI use among special education teachers above half, yet fewer than a quarter of those same teachers report receiv- ing any formal training on the risks. That gap, between how fast teachers have adopted AI and how little guidance they have been given for using it responsibly, is the starting point for a growing body of professional development work centered on two practi- cal, classroom-ready frameworks: PAUSE, a five-question ethical check for any AI tool used with or for a student, and PBNJ, a four- part prompting structure grounded in peer-reviewed research that turns a vague AI request into classroom-ready material. The frameworks, developed by Dr. Tiffanie Zaugg, a profes- sional development lead at Central Rivers Area Education Agen- cy in Iowa and a former state assistive technology lead, treat AI use as a professional skill that has to be taught deliberately, the same way teachers are taught to write a behavior intervention plan or interpret an evaluation report. (See YouTube video next page)

DR. TIFFANIE ZAUGG is the assistive technology and artificial intelligence specialist at Central Rivers Area Education Agency in Iowa, a former Iowa state Area Education Agency assistive technology lead, and principal investigator of Project RAISE, a federally funded robotics and AI intervention for students in special education. She is a co-author of the cited 2025 study on AI-assisted lesson planning (Faith, Zaugg, Stolys, Szabo, Haghi, Baldis, & Olmedo, published in the journal AI).

BACK TO CONTENTS

YouTube Video: AI in Classrooms A 12-Page Guide for Educators. https://www.youtube.com/watch?v=_xtf8ShkKII

dent's first name, disability category, and specific performance data into a consumer chatbot to draft progress-monitoring lan- guage. Within seconds, the identity of a child with a disability, along with sensitive academic data, has left the classroom and entered a system that may store it, use it to train future models, or expose it to a data breach or legal subpoena. The scenario illustrates a point that is easy to miss in the rush of a busy school day: under the Family Educational Rights and Privacy Act, personally identifiable information is not limited to a student's name. Any combination of details, such as a disability label paired with a grade level, a school name, and specific per- formance data, can be enough to identify a student even with- out a name attached. Many consumer AI tools' terms of service explicitly allow user input to be used for model training, and few include the kind of data-processing agreement that a school dis- trict's use of student information requires. The safeguard is straightforward: treat every AI prompt like a public billboard. If a detail would not be comfortable posted in a school hallway, it should not be pasted into an AI tool. In prac- tice, that means stripping out names, school names, and family circumstances, and replacing disability labels with functional descriptions. A prompt asking for help with a behavior interven-

tion plan for a named student with a specific diagnosis and a difficult family situation becomes a request for a plan addressing physical aggression during academic tasks, with no identifying details attached. Teachers who make this shift consistently find the output just as useful without exposing a real child's identity. THE PRIVACY MESSAGE BELONGS TO FAMILIES TOO The de-identification habit is usually framed as guidance for teachers, but the same risk exists on the other side of the IEP table. Parents and guardians increasingly turn to consumer AI tools for their own help, drafting a letter to a school team, trying to understand evaluation language, or researching a diagno- sis, and in doing so, many paste in the same kind of identifying details schools are working to keep out of AI prompts: a child's name, school, disability label, and specific performance or be- havior data. That information is just as exposed when a parent enters it as when a teacher does. Districts already send home guidance on topics like internet safety and social media use. Plain-language guidance on AI is becoming an equally basic literacy need, and it fits naturally into channels schools already use: a line in the family handbook, a handout at an IEP meeting, or a brief mention during a family

BACK TO CONTENTS

AI literacy night. The core message translates easily for a home audience: before typing anything about a child into a public AI tool, leave out the name, the school, and the specific diagnosis, and describe the situation in general terms instead. Families do not need to become privacy experts to protect their own chil- dren. They need the same billboard test teachers are learning, phrased for a kitchen table instead of a classroom. WHERE BIAS HIDES IN THE MACHINE A second pressure point in AI ethics training surprises more teachers: bias. Artificial intelligence systems learn patterns from historical data, and the historical record in American education includes decades of documented inequity, including the dispro- portionate placement of Black and Latino students in restrictive settings. When an AI tool recommends a placement, drafts a goal, or interprets an assessment, it can inherit and repeat those patterns instead of correcting them. Three concrete ways this shows up in practice include placement recommendations that skew more restrictive for student profiles mentioning behavioral challenges, IEP goals that reflect lower expectations for students with certain disability labels, and screening algorithms that gen- erate more false positives for particular demographic groups. Representation gaps compound the problem. AI models tend to know less about low-incidence disabilities, culturally di- verse learners, and multilingual students with disabilities, simply because there is less training data describing their experiences well. And left unchecked, AI-generated descriptions of disabil- ity tend to default to deficit-based language rather than the strengths-based, person-first framing the field has worked for decades to normalize. The response to that risk is not to abandon AI over bias con- cerns but to interrogate any new tool before it reaches a class- room. Five questions are worth asking before adopting any AI tool: what data was it trained on, who was involved in building it, does it perform equally well across different student profiles, does its output use strengths-based language, and can a teach- er override or edit its recommendations? That last question car- ries particular weight. If a teacher cannot modify or correct an AI tool's output, the tool is effectively making decisions for the teacher rather than with the teacher, which is a very different and much riskier relationship. THE PAUSE FRAMEWORK Every thread from the ethics discussion- privacy, accuracy, bias, independence, and equity- converges into a single tool teachers can carry with them: PAUSE. The framework is designed to be printed, kept next to a computer, and run through in full before using any AI tool with or for a student.

Letter

The Question to Ask Before Clicking Generate

I s any student identifiable in this prompt? Has all data been de-identified? Can the output be verified? Is professional judgment still driving the decision? Could this output reflect or amplify bias against any group of students?

P - Privacy

A - Accuracy

U - Underlying Bias

S - Independence Does the tool build a skill, or replace it? Is it a scaffold or a crutch?

Does every student benefit equally? Who might be left out or harmed?

E - Equity

Table 1: PAUSE Framework

The Independence letter draws a distinction central to the framework: the difference between a scaffold and a crutch. An AI tool that reads text aloud for a student with dyslexia is a scaffold, temporary support that keeps a skill accessible while it devel- ops. An AI tool that generates all of a student's written responses when that student is capable of typing is a crutch, something that replaces a skill rather than building toward it. The same dis- tinction applies to organizing a graphic organizer for a student to fill in, which supports independence, versus letting AI write IEP goals without a teacher's clinical judgment behind them, which erodes it. For each AI use with a given student, the practi- cal question is whether the tool builds a bridge toward indepen- dence or a wall around it. AI INSIDE THE UDL FRAMEWORK Alongside the ethical framework, special education increas- ingly discusses AI use through the lens of Universal Design for Learning. UDL's most recent version, UDL 3.0, organizes instruc- tion around three principles: engagement, or the why of learn- ing; representation, or the what of learning; and action and expression, or the how of learning. Specific AI uses map onto each principle. On the engagement side, AI can generate inter- est-based word problems, culturally responsive scenarios, and student choice menus. On the representation side, it can pro- duce text-to-speech output, image descriptions, reading-level adjustments, and language translations. On the action and ex- pression side, it can support speech-to-text, writing scaffolds, multimodal output options, and organizational supports such as visual schedules. A related question carries legal weight: can generative AI be written into a student's Individualized Education Program as as- sistive technology? The answer is yes, when the tool functions as AT for that particular student. The federal definition of assistive technology under the Individuals with Disabilities Education Act is broad enough to encompass many AI tools, provided the tool addresses a documented need, the student uses it consistently, it is named in the IEP's assistive technology section, and staff is

(See Table 1: PAUSE Framework)

BACK TO CONTENTS

trained to support it. That framing sets up a live legal question many districts have not yet resolved: when a student uses an AI tool as a documented accommodation, a classroom-wide ban on AI use can create real tension with that student's rights under Section 504 and the ADA. Districts need a clear answer before a dispute forces one. FROM ETHICS TO EXECUTION: THE PBNJ FRAMEWORK Ethical grounding answers whether and when to use AI. It does not answer the separate question of how to get an AI tool to produce something actually useful for a specific student. That second question is the subject of research Zaugg co-authored with a team of educator-researchers at the University of Toronto. The study, published in 2025 in the journal AI and led by Faith, Zaugg, Stolys, Szabo, Haghi, Baldis, and Olmedo, spent fourteen weeks examining how pre-service teachers used AI for differentiated lesson planning. Its central finding runs against the intuition many teachers bring to AI: writing one long, highly detailed prompt and expecting a finished, usable lesson back does not work well. Even thorough single prompts tended to produce material teachers described as impersonal and, in the researchers' words, alien. What worked instead was iterative back-and-forth dialogue with the AI tool, an approach the re- searchers describe as jamming, in which a teacher builds a re- sult gradually, adds detail in stages, and keeps critical judgment engaged throughout the exchange rather than accepting a first draft wholesale. The study also found that prompting AI to act as a critical friend, explicitly asking it to challenge assumptions and surface blind spots, helped teachers notice gaps in their own planning, particularly for students with diverse needs. Underly- ing it all was a finding the research team returns to repeated- ly: teachers perform a kind of contextual alchemy, drawing on knowledge of an individual child, the physical classroom, timing, and relationships that no AI system can replicate. That research carries its own name: PBNJ, short for Persona, Break Glass, Name Plan, Jam. Rather than a separate teaching shorthand layered on top of the study, PBNJ is the four-step workflow the research itself produced, and any teacher can ap- ply it in under two minutes.

Letter

What It Prompts the Teacher to Do

Assign the AI a specific role suited to the task, such as an ex- perienced special education teacher creating differentiated materials. Start with the simple, direct version of the request, the basic ask before any refinement.

P - Persona

B - Break Glass

N - Name Plan Name the plan for the output: what it should cover, what format it should take, and what the student's spe- cific needs are.

Iterate through back-and-forth dialogue, treating the AI as a critical friend and refining the result gradually rather than accepting the first draft.

J - Jam

Table 2: PBNJ Framework The PBNJ prompting framework, from Faith, Zaugg, Stolys, Szabo, Haghi, Baldis, & Olmedo (2025), AI, 6(12), 310.

diately. The AI system is not smarter in the second example. The prompt is. MATCHING TOOLS TO STUDENT NEED Zaugg maintains a continuously updated, publicly available spreadsheet cataloging AI tools by student need, organized into more than twenty tabs and rated for privacy and compliance, us- ing a color code of green for compliant, yellow for use with cau- tion, and red for tools that do not meet standards. The resource reflects her broader teaching philosophy: organize tools around what a student needs rather than which AI product is trending. The available tools fall into four practical categories. For reading and text access, tools such as Diffit can instantly adapt a pasted article or text into multiple reading levels with vo- cabulary support, while general-purpose tools like ChatGPT or Claude can generate leveled passages from scratch when given a specific Lexile level, grade, and topic. Education-specific plat- forms such as MagicSchool AI include special-education-specific generators for goals and accommodations, and interactive tools such as Curipod build in differentiation for inclusive, whole-class lessons. For writing, IEP planning, and executive function sup- port, general-purpose AI tools handle sentence starters, graphic organizers, and rubric adaptation, while a free tool called Goblin Tools, built by a neurodivergent developer, breaks a complex as- signment into step-by-step chunks and has become a particular favorite among teachers for its intuitive grasp of executive func- tion challenges. Teachers generally do not need a different specialized tool for every task. A general-purpose AI assistant paired with the PBNJ framework can handle roughly 80 percent of what a special ed- ucation teacher needs day to day. The specialized, purpose-built tools are useful additions for specific recurring workflows, not a replacement for learning to prompt well.

(See Table 2: PBNJ Framework)

The framework makes a clear difference. A vague prompt asking simply for a reading worksheet about animals produc- es generic, grade-level material a teacher then spends twenty minutes editing to fit actual student needs. A PBNJ-structured prompt that assigns the AI a persona, opens with the basic ask, names a clear plan for a 150-word passage about ocean animals at a second-grade reading level with a defined Lexile score, three comprehension questions of specified types, and a five- word vocabulary list, and then jams with the AI to refine the result, produces a passage a teacher can print and use imme-

BACK TO CONTENTS

FIVE-MINUTE MONDAY WINS Structured AI training for teachers tends to close the same way: with an insistence on one small, immediate action rather than an overwhelming overhaul of practice. Five-minute start- ing points include running tomorrow's reading passage through a leveling tool to create versions matched to each student's read- ing level, pasting a complex assignment into Goblin Tools for an instant step-by-step breakdown, asking an AI tool to generate a visual schedule or social story for a student who needs one that week, creating an accommodated version of an upcoming quiz, and using a PBNJ-structured prompt to draft a progress moni- toring template tied to a specific IEP goal. Saying a chosen commitment out loud to a colleague before leaving the room is a small step backed by research showing that spoken commitments are roughly twice as likely to be fol- lowed through on as silent intentions. The goal is not mastery in a single session. It is one small win on a Monday that builds enough momentum to try a second thing, and then a third, over the course of a school year. THE BIGGER PICTURE Taken together, the ethics framework and the prompting framework sketch out a coherent answer to the adoption-train- ing gap described at the start of this article. Teachers do not need to be talked out of using AI, and they do not need to be handed a tool with no instructions and left to sort out the risks them- selves. What the research and the classroom experience both point to is the same kind of structured professional preparation the field already provides for any other high-stakes instructional decision: a clear ethical framework, an understanding of the rel- evant legal landscape, and a concrete method for getting use- ful results without cutting corners on student privacy, accuracy, bias, independence, or equity. AI does not replace the contextual alchemy that makes a teacher irreplaceable, the knowledge of an individual child, a classroom, and a relationship that no algorithm can access. Used well, and used within a clear framework, it gives a teacher a little more time and a few more tools to practice that judgment, not less reason to exercise it.

BACK TO CONTENTS

Page 1 Page 2 Page 3 Page 4 Page 5

www.closingthegap.com

Made with FlippingBook Ebook Creator