LUDIA: A Design and Evidence Statement

Synthesis: LUDIA is a no-cost, private, multilingual AI thought partner that connects educators with the Universal Design for Learning (UDL) framework. The August 2026 relaunch was rebuilt from the ground up for privacy (no accounts, no cookies, no data collection), accessibility (WCAG 2.2 Level AA), and scale (13 languages, public-good architecture). Rather than a solution engine, LUDIA is deliberately positioned as a thought partner, and its authors read it against four 2026 guides — UNICEF's EdTech for Good Framework, the European Commission's ethical AI guidelines, the EdTech Quality Indicators, and the European EdTech Alliance's Needs-Based Evidence Mapping — while explicitly disclaiming any evidence that LUDIA improves learning. The document's central move is to replace a claim of impact with a culture of "proof of trust."

Key Findings

1. LUDIA is a no-cost, private, multilingual AI thought partner connecting educators, instructional designers, and learning leaders with the UDL framework, relaunched in August 2026 on its own purpose-built platform after running on Poe from 2023.
2. The rebuild is grounded in three architectural commitments — privacy by design (no accounts, no cookies, no server-side chat storage, no data collected for training), accessibility (WCAG 2.2 Level AA), and multilingual reach (13 machine-translated languages) — so the tool "measures up" by design rather than by promise.
3. LUDIA is engineered as a thought partner rather than a solution engine, structured around the Four T's (Tell, Tinker, Tweak, Transfer), and guards against over-reliance by framing every response as an option to weigh and by pointing back to CAST's published Guidelines.
4. The authors hold no evidence that LUDIA improves learning and say so throughout, reading the tool against four 2026 evaluation guides as a statement of intent made before the evidence exists rather than as a scoring exercise after the fact.
5. Following the Needs-Based Evidence Mapping, LUDIA organizes its future inquiry around "proof of trust" rather than proof of impact, because trust is treated as a relational property earned between a person, a tool, and a setting rather than a certifiable feature.

A thought partner, not a solution engine

LUDIA exists to connect educators with the Universal Design for Learning framework at the moment a design decision is still open — before a learner meets a barrier, not after. Where most AI agents are engineered as solution engines that deliver an answer, LUDIA is purpose-built to help educators and instructional designers work through the UDL principles of Engagement, Representation, and Action and Expression, so that the reasoning connecting a particular situation to the relevant Guidelines is visible and not only the suggestion. The Scaffolding intent is explicit: responses are "scaffolds and entry points" that leave the educator's own thinking intact, since "AI can certainly enhance but not replace our thinking as educators." The tool's use is structured around the Four T's — Tell LUDIA about the learners and challenges, Tinker with follow-up questions, Tweak the provisional outcomes, and Transfer by reflecting on how reducing a barrier deepened learning, often through Project Zero's Visible Thinking Routines. An exchange that stops before Transfer is LUDIA serving a purpose, but not yet being used fully as a thought partner.

A privacy-first, accessible rebuild

The relaunched LUDIA requires no account, sets no cookies, and writes no chat content to its own servers; an exchange lives in the browser for the session and can be downloaded and re-uploaded, and nothing is used for model training. This is Privacy by architecture rather than by policy, a deliberate break from the third-party commercial platform that hosted the earlier version and could not guarantee the same commitments. The tool is designed to Accessibility standards — WCAG 2.2 Level AA, keyboard-only operation, screen reader support, zoom to 400 percent, and light and dark schemes — because an interface inherited from a host platform was "the wrong compromise" for a tool built on UDL. Launching in 13 machine-translated languages widens reach, though the authors disclose that non-English versions may carry errors and the assumptions of the English they came from. The build is an open-ethos, public-good architecture with no premium tier, no advertising, and no commercial relationship with curriculum vendors, aimed at reducing the access barriers that fall hardest on educators with the least institutional backing.

The problem: barriers in the design, not in the learner

LUDIA is built to address a structural gap: despite a projected global EdTech market of hundreds of billions of dollars, roughly half of the 240 million children worldwide experiencing disability in low- and middle-income countries are out of school, and across OECD systems a third of teachers report lacking competencies to support students with specific needs. The paper's stance is that the barrier sits in the design, not in the learner — the core premise of Inclusive Learning and UDL. The framework carries legal standing through interpretation rather than treaty text: the Convention on the Rights of Persons with Disabilities (CRPD) establishes the right to inclusive education in Article 24, and its General Comment No. 4 directs states to adopt the universal design for learning approach. UDL is also named in national law across the United States, Chile, Colombia, and Spain, and in Professional Development policy. The authors name the "knowing-doing divide" as the obstacle — UDL guidance exists in abundance but is written in general terms, while a barrier is always particular to this engagement, these learners, this room, this week — and position LUDIA as the bridge across it.

Reading LUDIA against four 2026 guides

LUDIA is published as a design and evidence statement read against four guides: UNICEF's EdTech for Good Framework, the European Commission's guidelines on the ethical use of AI in teaching and learning, the EdTech Quality Indicators Guide (co-developed by CAST), and the European EdTech Alliance's Needs-Based Evidence Mapping. The authors use them as a set of questions to keep answering rather than a rubric to be scored against. The European Commission's guidelines place LUDIA within the EU AI Act and GDPR — LUDIA performs no emotion recognition, produces no scores or rankings, and is read as falling outside the high-risk categories, carrying instead the transparency duty to identify itself as an AI system. The paper is candid about its own AI Governance gaps: a Data Processing Agreement with the model provider is requested but not concluded, no independent accessibility audit exists yet, and the reading of the AI Act is the authors' own, not a legal determination. It also confronts the deeper sycophancy risk — a model trained on human feedback that drifts toward validating the user's view — which is named as "the failure that would empty LUDIA out," and which the design seeks to counter by refusing engagement mechanics and grounding every response in the published Guidelines.

An honest evidence stance: proof of trust over proof of impact

The paper's most distinctive move is its refusal to claim evidence of effectiveness. "We hold no evidence that LUDIA improves learning, and we say so throughout" — there is no trial, no pre-post measure, and no claim of learning impact. Instead, following the Needs-Based Evidence Mapping, LUDIA adopts a culture of "proof of trust": evidence that is credible, context-rich, explainable, and socially responsible. Trust is treated not as a property a tool possesses and can be certified as holding, but as something that "sits between a person, a tool, and a setting" and must be earned again whenever any of the three changes — a framing echoed in the calibrated trust literature. Every output passes through an educator before it can affect anyone, a human-in-the-loop safeguard the authors themselves call weaker than it sounds, since automation bias means the educator most likely to accept a poor suggestion is the one who trusts that LUDIA knows UDL. The Bias Mitigation gaps are named plainly: nobody has tested LUDIA's outputs for cultural or linguistic bias, and the model may skew toward well-resourced, English-medium, Global North schooling.

What this means for practice

• Designers. Engineer AI tools as thought partners, not solution engines: frame every response as an option to weigh, make the reasoning against a published framework visible, and point back to the authoritative source.
• Designers. Make privacy and accessibility architectural commitments — no account, no cookies, no server-side chat storage, no data used for training, and conformance to WCAG 2.2 Level AA with keyboard-only operation and screen reader support — and treat the substance of replies, such as register and length, as an Accessibility question rather than a matter of tone.
• Instructors. Use a structured cycle such as the Four T's and complete the Transfer step by reflecting on how reducing a barrier deepened learning; an exchange that stops before that reflection is not yet full thought-partner use.
• Instructors. Keep your own knowledge of your learners as the decisive check on every output and treat the tool as one option among several, because automation bias means the educator who trusts the tool most is the least likely to catch a poor suggestion.
• Institutions. Organize evidence for early-stage public-good tools by the purpose it serves rather than methodological prestige, and specify in advance what counts as "proof of trust"; never treat the tool as a safeguarding route, since a privacy architecture that holds nothing cannot act on a disclosure.

Limitations

• No learning-impact evidence: the authors state plainly that "we hold no evidence that LUDIA improves learning," with no trial, no pre-post measure, and no claim of impact.
• Nothing is independently verified: no external audit of accessibility, security, or reliability; no independent researcher review of the evidence claims; usability testing was informal and undocumented; and the Accessibility Conformance Report is self-authored and still in preparation.
• Outputs have never been tested for cultural or linguistic bias, and the model may skew toward well-resourced, English-medium, Global North schooling; the 13 languages are machine translations that carry the assumptions of the English source.
• The comparison guides were self-selected after the tool was largely built, one of them was co-developed by an organization that gave the authors an award, and the authors call reading yourself against guides you picked "the weakest form of assessment there is."

Citation

Stark, B., & Rostan, J. (2026). LUDIA: A Design and Evidence Statement. EdArXiv preprint.