The IRIS Scale: disclosing AI use without mistaking it for learning

Two students can submit the same polished text: one wrote it and requested a spelling check; the other accepted a complete chatbot draft. “I used AI” does not distinguish them. On 21 July 2026, Fundació Escola Cristiana de Catalunya introduced the IRIS Scale to describe that difference. It proposes six levels, from work without AI to automated production under supervision, plus a short declaration at the end of a task.

Its useful contribution is not a claim that more AI is better. The full framework written by FECC’s digital group stresses that this is neither a rigid ranking nor a control instrument. Level zero may be the best choice when the objective is to practise an unassisted capability. The question it teaches is different: what cognitive process must the person perform, and what can be delegated without emptying the learning from the task?

Six levels describing a relationship

At IRIS-0, the task is completed without AI. At IRIS-1, the tool corrects formal aspects of human material—spelling, grammar, translation or formatting—without contributing ideas or transforming content. At IRIS-2, it proposes options from a one-off request, while the person selects and develops them. The distinction between correction and inspiration matters: a topic suggestion already influences intellectual direction even when it does not draft the product.

At IRIS-3, AI generates an outline or first draft that the user must check, transform and complete with independent sources. IRIS-4 describes iterative co-creation: person and system engage in several rounds of dialogue, reformulation and decision-making. At IRIS-5, the tool is configured to produce a complete artefact—a presentation, visualisation, report or process—and a person supervises and validates the result.

The one-page IRIS infographic makes the sequence memorable, but the long document supplies the fine print. Levels are not marks of merit or a mandatory progression. The fifth is aspirational for students: configuring autonomous agents requires digital maturity and teacher supervision that cannot be assumed. For students, it means supervised use of advanced tools, not the design of autonomous systems.

The framework is not confined to students. It applies the same levels to teachers preparing exams, families supporting homework, leadership teams analysing results and administrative staff managing enrolment. That breadth avoids a double standard in which students must disclose AI while adults conceal that a test, message or preliminary diagnosis was generated with automated assistance.

Start with the objective, then set permission

A scale guides practice only if the teacher begins with the capability to be observed. If a task assesses writing, allowing AI to draft the first version changes the object of assessment. If it assesses verification, the same draft can become ideal material: the student must find weak claims, locate sources and explain corrections. There is no correct level outside the pedagogical purpose.

The Catalan education department’s guidance on AI in education reaches a compatible conclusion: teachers need to redesign activities that applications solve easily and choose use according to the competence being developed. It also warns that the field changes too quickly for guidance to be presented as definitive truth. IRIS provides vocabulary for the decision; it does not make the decision for a teacher.

A clear design can be written in three lines before work starts. First: “you will learn to compare sources and build your own conclusion”. Second: “you may use IRIS-1 for language correction, but not for ideas or paragraphs”. Third: “you will submit reading notes, sources and a short explanation of two decisions”. Permission, activity and evidence now point to the same capability.

The contrast appears when IRIS-3 is allowed. The objective might be “evaluate and rebuild a generated argument”. A draft is then permitted, but the initial version, consulted sources, substantive changes and an oral defence are required. The final product is no longer the only evidence. Assessment examines what the person detected, rejected and rebuilt.

Disclosure is not proof

IRIS recommends a note on every AI-assisted document stating the level and briefly describing the delegated tasks. That is better than yes or no: “IRIS-2 for possible approaches; selection, sources and writing by the student” provides information that can be discussed. It also avoids guessing use from style, an unreliable practice that mistakes linguistic uniformity for misconduct.

But the label is an author’s claim, not evidence of learning. A student may declare IRIS-3 and alter a draft only superficially. The document acknowledges this limit: the scale identifies AI integration but does not determine whether meaningful learning occurred. Its later guidance suggests comparing versions, requesting independent sources, reviewing decision histories or asking for an oral explanation.

Three layers should remain separate. Disclosure describes tools and tasks. Process evidence shows drafts, decisions and checks. Assessment of understanding shows that the student can explain and transfer the learning without relying on the submitted artefact. A school may use IRIS for the first layer and still fail at the other two; the scale should therefore change the rubric, not become an administrative checkbox.

Transparency has limits too

Requiring a complete chatbot history may expose personal questions, classmates’ data or information that should never have been entered. The Catalan guidance notes that many applications require accounts and retain data capable of tracking users’ interests and activity. Before requesting a conversation as evidence, a school should define authorised tools, prohibited data, who will see the record and how long it will be retained.

Less intrusive alternatives exist. A student can select two important decisions, copy only the necessary excerpts, describe the check against a source and remove personal data. A teacher can observe part of the process in class or request a short oral defence. Traceability should be proportionate: enough to support assessment, not an invitation to collect an entire digital life.

The issue grows in level-five administrative examples. The document imagines agents supplied with rubrics, school plans, prior results or enrolment rules, while also calling for anonymised data and final human responsibility. “Supervised” does not settle purpose, access, security or legality. Before automation, a school needs a data inventory, permissions, genuine human review and a procedure for correcting errors.

A framework built on earlier work

IRIS did not appear from nowhere. Its cover says it was inspired by AIAS, MIAE and SEIA, other assessment and integrity frameworks. The AI Assessment Scale already organises AI permissions around assessment design. FECC adds examples for several participants, a cognitive-delegation lens and a declaration model. Acknowledging that lineage enables comparison and avoids presenting an adaptation as an isolated invention.

It also continues FECC’s 2025 responsible-use protocol, which covered values, privacy, security, critical thinking, social impact and a roadmap for schools. The new scale narrows the focus to AI intervention in a task. It does not replace the broader protocol. Knowing that work is IRIS-4 says nothing by itself about a tool’s age requirement, data processing or bias.

The FECC expert-group page identifies the coordinator and members and says materials are shared so each school can adapt them. That scope matters: IRIS is a proposal from an education network, not an official rule for every Catalan school or a validated test of improved outcomes. Its effectiveness will need to be observed in use, with agreed criteria and review after implementation.

Turning the scale into a useful conversation

Before every task, a teacher can answer four questions: what the student should learn, which assistance level preserves that capability, what process evidence will be collected and how data will be protected. The student then declares actual use and explains one decision. Finally, the teacher assesses understanding rather than compliance with a label. If use exceeded permission, the response can identify which evidence became invalid rather than assume the whole piece has no value.

The transferable capability is to align objective, permission, evidence and assessment. IRIS supplies language for permission and disclosure; the teacher must complete the other pieces. When all four fit, the question stops being “did you use AI?” and becomes “what did you learn, what did you delegate, and how can you demonstrate the difference?”.