ECD, UDL, and The Cow Path
Educational measurement is at a crossroads. Multimodal artificial intelligence (AI) finally opens a pathway to move beyond the industrial-era, one-size-fits-all summative test in favor of unobtrusive, agentic, and adaptive assessment.
But directing raw computing power at children without guardrails is ill-advised. Using AI to churn out traditional items at scale risks “paving old cow paths”—automating subpar measurement, amplifying biases, and compounding inefficiency.
To architect a SAFE–Safe, Accountable, Fair, and Efficious–future of assessment, we must weave together two foundational frameworks: the late Dr. Robert J. Mislevy’s Evidence-Centered Design (ECD) and Universal Design for Learning (UDL). To bring this architecture to life, we must invest in technologies purpose-built to break down barriers for students with disabilities and learning differences, proving that designing for the edges enhances design for all learners..
Courtrooms and the Unless Clause
Educational assessment is an exercise in evidentiary reasoning. ECD introduces this by comparing the test-maker to a lawyer in a courtroom making a Claim (e.g., “This student understands kinematics”), providing Data (a plotted graph), and relying on a Warrant (the rationale: graphing indicates understanding).
However, every score carries an invisible caveat: the student understands… unless they couldn’t read the tiny font, were overwhelmed, or confused by syntax. These are construct-irrelevant barriers. If a student hits a perceptual or executive-functioning hurdle they cannot clear, they are shut out. Their performance drops to guessing, yielding zero information about their capability. ECD asks developers to proactively engineer these out of the assessment from day one.
Signal, Noise, and the Microscope
The testing industry long relied on “Marginal Inference”—the assertion that if everyone faces the exact same surface (e.g., font and time limits), fairness is achieved. However, fairness requires Conditional Inference: standardizing the validity of the construct while actively varying the delivery.
A standardized test is like an unadjustable microscope. If the eyepiece isn’t focused for the unique user, they can’t see the slide. The failure isn’t the student’s vision or the academic content (the slide)—the failure is the instrument itself.
Consider an eye exam versus a letter recognition test. Standardizing distance is mandatory for visual acuity, but forcing a low-vision student to stand back to read the alphabet confounds the test.
Every task demands a web of Knowledge, Skills, and Abilities (KSAs). We must separate Focal KSAs (the Signal) from Additional KSAs (the Noise). In a spelling bee, forcing a spoken response potentially introduces an Additional KSA. For a student with a speech impairment, this measures the impairment, not spelling. Typing might preserve rigor while removing the barrier. UDL is an engine informing these adjustments.
Proof from the Margins
Realizing this architecture requires investing in technologies purpose-built for students with learning differences. Programs like the US Department of Education and Institute of Education Sciences’ ED/IES SBIR lead the way by underwriting needed high-risk, high-social-impact R&D. Seeded to tackle complex access challenges, these companies have built a portfolio of field-tested tools that prove a core thesis: designing from the edges removes construct-irrelevant barriers, enables students to show what they know, and ultimately strengthens the accuracy and fairness of assessment for all learners.
Sensory Hurdles
When sensory access is inhibited, tools must adapt to the medium. Alchemie (Kasi) uses computer vision and tactile manipulatives so visually impaired students can interact with complex chemistry concepts. IDRT’s myASL Quizmaker assesses natively in American Sign Language, eliminating the written English barrier for Deaf students. Nimble Tools integrated custom overlays, read-aloud supports, and magnification into standard test administrations.
Linguistic and Speech Hurdles
Alternative modalities unlock expression for students with developmental and language delays. IQ Sonics (Sing and Speak 4 Kids) uses music-based games to assess and strengthen expressive language skills in children with speech delays and autism. Attainment Company provides adapted texts and prompting via its Early Reading Skills Builder, unlocking foundational literacy for students with intellectual disabilities. CAPTI (ReadBasix) leverages AI to create scenario-based assessments localized to students’ lived experiences, reducing cultural bias.
Cognitive and Executive Functioning Hurdles
Where traditional tests trigger cognitive overload, adaptive platforms clear the path. ObjectiveEd (BuddyBooks) embeds visual models, personalized pacing, and multimodal supports to reduce working-memory load for students with dyslexia. Teachley builds conceptual math understanding through virtual manipulatives, bypassing procedural practice barriers for students with ADHD. Filament Games (PLEx Life Science) accommodates reading difficulties in science through universally designed gameplay. OKO Labs uses collaborative, game-based AI facilitation and voice interactions, ensuring students aren’t penalized by written-response barriers. Handhold Adaptive (iPrompts) provided individualized digital visual schedules and just-in-time scaffolding to support executive functioning for students with autism.
Systemic Hurdles
To ensure these principled designs reach the learner, infrastructure is required. Presence built a teletherapy platform to ensure secure, remote access to speech, occupational, and behavioral therapy. Education Modified employs AI to translate Individualized Education Programs (IEPs) directly into day-to-day instructional workflows.
Shearing Layers and TV Captions
Historically, creating universally designed, conditional task variants was a manual, expensive bottleneck. Today, multimodal AI has the potential to position products to dynamically adjust Variable Task Features in real-time.
But AI needs guardrails. We must view assessment through the architectural metaphor of “Shearing Layers.” Modifying an assessment’s UI (the “Skin”) using AI must not dismantle the underlying construct (the “Structure”). ECD’s Design Patterns must act as the “system prompts” so AI doesn’t hallucinate validity.
Policy, procurement, and investment must demand the ability to enable architecture upfront. Retrofitting accommodations onto finalized AI tools is scientifically sloppy and financially exorbitant. Like closed captioning—once a retrofit for the deaf community, now a universal feature—accessible design must be baked in from day one.
AI does not change the fundamental science of evidentiary reasoning; it just might give us the computational scale to achieve it. By anchoring our investments in ECD and UDL, we ensure every student is provided an unclouded, focused lens through which to show us their brilliance.
