Artificial Intelligence Assessment Scale (AIAS)
- AIAS is a five-level framework defining acceptable GenAI use in academic assessments, ranging from no AI to full co-creation.
- It integrates ethical standards, constructivist learning theory, and adaptable rubrics to align tool use with learning outcomes.
- Empirical findings and revisions demonstrate enhanced academic integrity and innovative assessment practices across disciplines.
The Artificial Intelligence Assessment Scale (AIAS) is, in the educational literature, a five-level framework for specifying acceptable use of Generative Artificial Intelligence (GenAI) in assessment. It was developed to help instructors clarify and communicate appropriate AI use, reduce misconduct, preserve academic integrity, and integrate AI ethically into teaching and assessment (Perkins et al., 2023). Subsequent work revised the framework to support open dialogue between educators and students, redesign assessments in an era of expanding AI capabilities, and foreground assessment validity rather than compliance with brittle technological rules (Perkins et al., 2024). Domain-specific adaptations, most notably the EAP-AIAS for English for Academic Purposes, retain the five-level structure while reworking descriptors around language development, genre literacy, and critical evaluation of AI outputs (Roe et al., 2024).
1. Emergence and conceptual basis
The AIAS emerged against a background in which higher-education institutions were attempting to reconcile two imperatives: preserving academic integrity and preparing graduates for an AI-rich workplace. Early responses to GenAI often took the form of bans, blocking, or simplified attempts to “AI-proof” assessment, while the AIAS reframed the issue as one of structured integration rather than punitive detection (Furze et al., 2024). In the original educational formulation, the scale was presented as a practical, simple, and sufficiently comprehensive tool for integrating GenAI into educational assessment, with an explicit ethical rationale centred on academic integrity, skill development, equity, and transparency (Perkins et al., 2023).
Its theoretical underpinnings were described in several closely related ways across the literature. The original and EAP-adapted accounts grounded the framework in principles of academic integrity, ethical technology use, and constructivist assessment theory, with the assumption that transparent criteria guide learners’ responsible tool use (Roe et al., 2024). The revised framework later made the social constructivist basis explicit, drawing on Vygotsky’s view of learning as socially mediated and presenting GenAI as a scaffolding technology within learners’ Zone of Proximal Development. In that account, the AIAS preserves the social dimension of learning by encouraging peer and teacher dialogue about when and how to use AI, while assessment validity is defined as ensuring that an assessment genuinely measures intended knowledge, skills, or competencies rather than mere compliance with technological restrictions (Perkins et al., 2024).
A central conceptual shift in the AIAS literature is therefore from “AI = cheating” to a more differentiated account in which AI can be prohibited, bounded, scaffolded, critically evaluated, or strategically integrated depending on the learning outcome being assessed (Perkins et al., 2023).
2. Level structure and framework evolution
Across its educational formulations, the AIAS remains a five-level ordinal framework. The original version organised assessment from “No AI” to “Full AI Co-Creation,” while the revised version retained five distinct levels but relabelled them as “No AI,” “AI Planning,” “AI Collaboration,” “Full AI,” and “AI Exploration” (Perkins et al., 2023, Perkins et al., 2024).
| Earlier educational AIAS | Revised AIAS | Typical role of GenAI |
|---|---|---|
| No AI | No AI | Zero access to GenAI |
| AI-Assisted Idea Generation and Structuring | AI Planning | Pre-task ideation only |
| AI-Assisted Editing | AI Collaboration | Drafting, feedback, refinement |
| AI Task Completion with Human Evaluation | Full AI | Extensive strategic use |
| Full AI Co-Creation | AI Exploration | Co-designed experimentation |
At Level 1, students complete the task in a controlled, supervised environment with zero access to any AI tool; assistive technologies for accessibility are explicitly distinguished from GenAI (Perkins et al., 2024). At the second level, AI is limited to pre-task stages such as brainstorming, outlining, preliminary research, or locating potential sources, and no AI-generated content may appear in the final submission (Perkins et al., 2023). The third level permits drafting support, editing assistance, or other collaborative functions, but requires critical evaluation, editing, and integration in the student’s own voice (Perkins et al., 2024). The fourth level allows extensive or required AI use, with the emphasis shifting to prompting, evaluating, and directing AI toward discipline-specific goals (Perkins et al., 2024). The fifth level, “AI Exploration,” is reserved for co-designed, open-ended tasks that test the boundaries of current GenAI, including synthetic media experiments or bespoke applications (Perkins et al., 2024).
A notable revision concerns visual representation. Earlier versions used a traffic-light logic; the revised framework replaced this with a neutral colour palette and circular diagram because red–amber–green coding implied prohibition and endorsement, and therefore conflicted with the principle that no single level is inherently superior and that level choice is context-dependent (Perkins et al., 2024). Accessibility considerations, including contrast ratios and colour-blind-friendly palettes, were also built into the redesign.
3. Adaptation to language education: EAP-AIAS and EFL use
The EAP-AIAS adapts the educational AIAS specifically for English for Academic Purposes, where assessment targets not only content but discourse-level features such as vocabulary range, cohesion, and genre conventions (Roe et al., 2024). The rationale for adaptation is twofold. First, EAP uniquely blends language proficiency with academic acculturation. Second, GenAI’s ability to produce near-native-level text creates an acute challenge for measuring genuine EAP outcomes, especially in take-home and formative tasks such as essays, annotated bibliographies, and presentations, where misuse is difficult to detect (Roe et al., 2024).
The adaptation process mapped the original AIAS levels onto EAP-specific tasks and objectives, redesigned descriptors to foreground language development, genre literacy, and critical evaluation of AI outputs, and consulted EAP literature on assessment validity and academic literacies to ensure alignment with disciplinary norms (Roe et al., 2024). The resulting five levels are descriptive rubrics rather than numerical scoring formulas. Level 1 requires all language and skill tasks to be completed without any AI assistance and focuses on independent demonstration of core macro-skills. Level 2 permits AI only to generate or adapt input materials for learners, so that students respond to AI-generated input but produce final answers entirely unaided. Level 3 allows controlled practice of discrete language features or academic micro-skills, with AI suggestions treated as provisional and subject to student justification or correction. Level 4 permits brainstorming, outlines, or rough text, but requires critical evaluation, substantial revision, and annotation of AI contributions. Level 5 permits extensive, strategic AI use, for example in data synthesis or statistical analysis, provided that final work reflects independent insight and that AI use is explicitly cited (Roe et al., 2024).
Applications span writing tasks, oral presentations, and research projects. The paper gives examples including in-class timed essays at Level 1, paragraph-level practice with AI grammar feedback at Level 3, literature reviews and full essay drafting with annotated revisions at Levels 4–5, AI-supported slide design for presentations at Level 5, and AI-assisted data analysis with methodological and ethical reflection in research projects (Roe et al., 2024). Best practices include pairing AI-generated drafts with peer- and self-assessment focused on AI reliability and bias, embedding explicit training sessions on prompt engineering and AI source evaluation, and using reflective journals in Levels 4–5 to document decision-making (Roe et al., 2024).
A related 2025 paper extends the updated AIAS into EFL teaching and learning, arguing that few formalised frameworks are available for developing AI literacy skills in EFL and demonstrating how three levels of the updated AIAS can structure EFL writing instruction in ways that promote academic literacy and transparency (Roe et al., 1 Jan 2025).
4. Implementation and assessment redesign
The AIAS literature presents the framework not merely as a label set but as an instrument for redesigning assessment around learning outcomes. A pilot implementation at British University Vietnam described a four-step process: mapping existing assessments to AIAS levels; documenting permissible AI functions and required artefacts for Levels 2–3; embedding explicit reflection or co-design prompts for Levels 4–5; and providing faculty training, policy documentation, and student guides around the principles of Ethics & Transparency, Security & Privacy, and Limitations & Bias (Furze et al., 2024). The same report states that policy materials and assessment briefs were released in October 2023 together with workshops on prompt engineering, citation of AI sources, and rubric alignment (Furze et al., 2024).
The revised AIAS supplements this procedural guidance with composite vignettes from K-12 and higher education. These include a first-year rhetoric course using handwritten in-class writing at Level 1, a documentary pre-production storyboard task using sketch-to-image systems and text-based GenAI at Level 2, a chemistry lab report permitting Copilot for data analysis and Claude or ChatGPT for initial drafting at Level 3, a third-year software design course using GitHub Copilot, Cursor, or OpenAI o1 at Level 4, and a specialist dance-technology lab in which students and instructors co-design real-time movement-analysis systems at Level 5 (Perkins et al., 2024).
Across these examples, the pedagogical logic is consistent: AI permission is specified in relation to what evidence of learning the assessment must produce. The framework therefore functions as a mechanism for aligning tool use, task design, and intended outcomes rather than as a generic approval or prohibition device.
5. Empirical findings and institutional experience
The most explicit quantitative evidence in the educational AIAS literature comes from the British University Vietnam pilot. That study reported GenAI-related academic misconduct cases falling from 116 out of 4,080 submissions in April 2023 to 0 reported cases out of 4,125 submissions in October 2023; a mean grade increase of 5.9% comparing September 2022 and September 2023 semesters; a module pass-rate increase from 66.7% to 100%; and an increase in faculty using AIAS Levels 3–5 from approximately 20% to approximately 80% (Furze et al., 2024). The same paper also reported anecdotal uptake of multimodal projects and qualitative feedback indicating that non-native English speakers found AI tools beneficial for rendering ideas in accessible formats (Furze et al., 2024).
Those findings are accompanied by substantial methodological cautions. The pilot used a pre–post comparison with no concurrent control group, reported aggregated outcomes rather than individual-level sampling, and relied on descriptive analyses without formal inferential models or reported -values. The paper also notes that policy changes allowed staff to downgrade rather than report offences, which may partly explain the drop in misconduct reports (Furze et al., 2024). The evidence therefore supports institutional change and pedagogical uptake, but not a clean causal attribution.
A later qualitative study examined how staff experienced AIAS implementation at a private international university in Vietnam and a public university in the United Kingdom. Using five focus groups with 30 academic staff and hybrid thematic analysis, the study developed six themes: recognising and integrating AI, facilitating conditions, building capacity, pathways to adoption, ethics in practice, and reframing pedagogy (Perkins et al., 25 Jun 2026). Staff valued the AIAS as a shared language for legitimising GenAI use, clarifying boundaries, and prompting reflection on assessment design. At the same time, implementation was shaped by governance, tool access, staff confidence, workload, integrity concerns, disciplinary context, and alignment with learning outcomes (Perkins et al., 25 Jun 2026). The study concluded that the AIAS could prompt authentic assessment design and student engagement, but may become a compliance layer when disconnected from learning outcomes, disciplinary context, and staff capacity.
6. Limitations, misconceptions, and scope of the term
A recurring misconception is that the AIAS is a psychometric scale in the narrow sense of a numerical instrument with formal scoring equations. The revised AIAS paper explicitly states that it does not introduce formal mathematical models or LaTeX-formatted formulas for scoring rubrics or validity measures, and the EAP-AIAS similarly states that it offers descriptive rubrics rather than numerical scoring formulas (Perkins et al., 2024, Roe et al., 2024). In the educational literature, AIAS is therefore better understood as a structured rubric or framework for assessment design.
Another misconception is that higher levels are intrinsically preferable. The revised framework rejects this by emphasising that no single level is inherently superior and that appropriate level choice depends on context, learning outcomes, and discipline (Perkins et al., 2024). This suggests that the key evaluative question is not how much AI is permitted, but whether the permitted use still yields valid evidence of student capability.
The framework’s current limitations are also clearly stated. Equity of access remains imperfect even where institutions provide tools; some implementations have been superficial, applying level labels without redesigning tasks; and rapid GenAI development, including multimodal and BCI-enabled interfaces, may outpace static frameworks (Perkins et al., 2024). Proposed future directions include integrating prompt-engineering competencies and detailed evaluative-judgement rubrics, developing modular guidance for emerging modalities such as vision, audio, and video, empirically validating the AIAS across diverse cultural and disciplinary settings, exploring tiered micro-levels within each stage, and revisiting the relation between AI tool sustainability and pedagogical benefit (Perkins et al., 2024).
Finally, the acronym “AIAS” is not unique to the educational framework associated with Perkins, Furze, Roe, MacVaugh, and related work. Earlier papers used closely related terminology for a standard intelligent machine model and artificial intelligence IQ or intelligence grade taxonomy, and another paper used AIAS in connection with adapting the SQuaRE quality model for AI systems (Liu et al., 2015, Liu et al., 2017, Kuwajima et al., 2019). In current educational discourse, however, AIAS most commonly denotes the five-level GenAI-in-assessment framework and its domain adaptations.