Abstract

Research and policy have become good at naming the cognitive risks that generative AI (GenAI) poses to learning. Namely, detrimental cognitive offloading, metacognitive laziness, and the illusion of competence. What this literature does much less well is tell teachers what to do with that knowledge in their everyday practice. This article argues that alongside AI literacy, regulation and safeguards protecting educational stakeholders, one of the central challenges facing education in the age of GenAI is related to didactic design. When a finished text or answer no longer reveals what a student has understood, struggled with, revised, verified or learned, teaching and assessment have to be redesigned so that the process of thinking is made visible and necessary again. We argue that this redesign should be approached through the classic didactic questions: why, what, how, when and for whom. We connect these questions with Question Zero (Q0), an approach to responsible technology adoption, through the dimensions of why, who, what, how and where. Both didactic and Q0 perspectives and underlying questions move the starting point from the technology to educational purpose, stakeholders and the conditions under which the use of selected technology is well-designed and justified. We argue that current AI strategies, guidelines and institutional frameworks in education leave a conspicuous gap precisely here. We translate these didactic and Q0 perspective questions into a concrete, phased design for teaching and assessment, whose guiding principle is not to forbid cognitive offloading but to regulate it didactically. We also argue that making the learning process visible is not in itself a neutral act, and that its design must attend to who carries its costs. This challenge recurs across education systems globally and cannot be solved by top-down regulation alone. Instead, translating high-profile principles of ethical AI in education into pedagogically sound use of GenAI requires the contextual and professional judgement of educators.

From cognitive risk to didactic response

Recent reports synthesise the evidence on AI and learning. A particularly useful contribution is the distinction between beneficial and detrimental cognitive offloading (Dunn & Risko, 2016; Risko & Gilbert, 2016; Lodge & Loble, 2026). AI can manage extraneous load and free working memory for the intrinsic work of learning, or it can bypass that intrinsic work altogether. Equally useful is the distinction between a task-completion goal and a learn-the-concept goal (Kalyuga & Plass, 2025). These distinctions give us a more precise language for what is at stake.

The problem is not simply that students might cheat, or that they might produce better answers than they could have produced alone. The more serious issue is that generative AI (GenAI) can interfere with the cognitive and metacognitive processes through which students learn (Fan et al., 2025). Students may complete the task without developing the knowledge, judgement or self-regulated learning capacity the task was meant to support. The most recent international synthesis reaches the same conclusion. The OECD (2026) finds that when students lean too heavily on GenAI, metacognitive engagement falls and task performance becomes misaligned with genuine learning.

The current empirical data and preliminary findings describe the risk well and give much less guidance on what teachers are supposed to do with this knowledge. How should they design teaching and assessment when the finished product no longer tells us very much about what the student has understood, revised, verified or learned? This is the question we want to place at the centre.

The pedagogical problem beyond AI literacy

Much of current education policy responds to GenAI by emphasising AI literacy and competence. We agree with their importance. But they do not constitute a holistic pedagogical response to cognitive outsourcing. Treating them as such risks diverting attention and resources from the difficult didactic work required (Weidlich et al., 2025). Knowing what a large language model is, and how to prompt it, does not by itself help a teacher decide how to assess understanding when fluent, confident output is available on demand. A recent article makes a parallel point: conventional, skills-based AI literacy frameworks remain anchored in functional competence and say little about the metacognitive and epistemic demands of probabilistic, opaque systems (Shapiro et al., 2026). That argument moves toward expanding literacy itself, into a critical, metacognitive and civic practice, whereas we argue the missing piece is the appropriate redesign of teaching and assessment.

In this opinion piece, we argue that the challenge of AI in education is fundamentally pedagogical rather than solely technological. The same tool produces opposite effects depending on the instructional design and the learner’s goals. A systematic review of 67 empirical studies confirms the pattern. Generative AI’s cognitive effects are not intrinsic to the tool but dependent on pedagogical framing, with scaffolded, inquiry-oriented use supporting higher-order thinking and unstructured use triggering cognitive offloading (Li et al., 2026). A risk taxonomy, however accurate, does not contain its own remedy. Naming “metacognitive laziness” tells a teacher what to fear, not how to handle it in the classroom.

The tension is structural rather than rhetorical. Cognitive offloading is, at its core, a way of reducing mental effort (Risko & Gilbert, 2016), and generative AI lowers the cost of such offloading further than any earlier technologies. Education rests on people working their way through uncertainty and complexity. This is not a superficial tension but a conflict between what the tool is built to optimise and what learning requires. Decades of research on desirable difficulties suggest that effort, struggle, and even the unpleasantness of thinking are often what produce durable learning (Bjork & Bjork, 2015; David et al., 2024), which is precisely the work that a tool designed to minimise effort invites students to skip. Early empirical work points the same way. In a study of 666 participants, more frequent use of AI tools was associated with weaker critical thinking, an association largely mediated by cognitive offloading (Gerlich, 2025).

If AI offers the fastest route to a finished product, then teaching has to make the process of thinking (i.e., explaining, revising, verifying, judging) more explicit. This is not a small adjustment. It touches the design of assignments, the role of feedback, assessment practices, the relationship between teacher and student, and students’ motivation to engage in the difficult and sometimes frustrating work that learning requires.

Several strands of assessment scholarship that predate generative AI are especially relevant here. Work on assessment reform for an age of GenAI argues for distinguishing between assessment that secures and validates individual achievement and assessment that develops capability through authentic, process-rich tasks (Lodge et al., 2023; Dawson, 2021). Work on formative assessment and feedback shows how assessment can be designed to drive self-regulated learning rather than merely certify a product (Nicol & Macfarlane-Dick, 2006; Carless & Boud, 2018; Hathcoat et al., 2026). And the notion of evaluative judgement (i.e., the capacity to judge the quality of work, one’s own and others’) names precisely the capability that detrimental offloading erodes and that teaching now has to cultivate deliberately (Tai et al., 2018; Bearman et al., 2024). The OECD’s synthesis points in the same direction. Effective integration, it argues, may require teachers to emphasise process, how students think and learn, rather than student output, and finds that hybrid systems pairing GenAI with explicit pedagogical models, such as structured tutoring or evidence-centred assessment design, show more promise than general-purpose chatbots (OECD, 2026). A 2025 UNESCO anthology reaches a similar conclusion, devoting a major section to assessment and arguing that standardised, output-focused testing is being destabilised by generative AI in favour of more human-centred, process-oriented forms of knowing (UNESCO, 2025b).

Making learning visible, but for whom?

There is a cautionary precedent worth recalling. Earlier efforts to make learning visible through digital tools did not automatically translate into changed teaching. A study of the introduction of mandatory personal development plans and digital documentation in Swedish compulsory schools found that, although the reform was meant to centre each student’s learning, students’ own goals had little effect on actual teaching. Responsibility for reaching them was largely devolved to the individual student, and the process drifted towards documentation, behaviour management and the production of the “model student” (Mårell-Olsson, 2012). Teachers, working under pressure, rationalised the task. For instance, teachers copied and adapted written assessments across students. The lesson is relevant to the current stage of GenAI adoption in education. Mandating a digital tool without redesigning teaching and assessment tends to create an administrative routine rather than reshape the learning process it was meant to support.

It is also important to be careful with the word ‘visible’. Making the learning process visible is not a neutral or automatically benign act. The study showed how documentation practices can turn into surveillance, control and the disciplining of students towards a prevailing norm, a dynamic that Bernstein’s distinction between visible and invisible pedagogy helps to name (Mårell-Olsson, 2012). Later studies of one-to-one computing classrooms found a similar dynamic, showing how teachers’ enacted didactical design decisions carry symbolic power and control over students (Bergström et al., 2019). Thus, designing for process-visible learning has to attend to the ‘for whom’. Visibility can support metacognition and self-regulation, or it can become a mechanism of compliance that weighs most heavily on the students with the least resources to navigate it.

The aim, therefore, is not visibility for its own sake but tasks designed so that the thinking serves learning rather than control, and so that students have a reason to understand the concept and apply it in the real world, rather than complete the task (see Lodge et al., 2023).

A didactic reframing: why, what, how, when, for whom

A further observation sharpens the design task. The evidence suggests that GenAI works best as a scaffold, not as a substitute. When it is used for Socratic questioning, feedback, alternative examples, concept explanations and support for self-regulation, it can support learning. When it is used to produce finished texts, solutions or analyses before the student has done their own cognitive work, the risk of shallow learning rises. A synthesis of research on curiosity and metacognition draws this distinction directly, warning that large language models’ default interaction modes can act as a cognitive shortcut but when redesigned, they can instead become a partner for sustained epistemic development (Desvaux et al., 2026). The design question, therefore, is not AI or no AI, but when in the learning process AI is used, for what purpose, and with what degree of student responsibility. That reframing is  what the didactic questions are built to handle.

We propose that AI in education be approached through the questions educators have long asked of any teaching decision. Why something should be taught, what learners need to know, how it should be taught, when in the course of learning, and for whom. These are the didactic questions of the Didaktik tradition (Klafki, 2000; Hopmann, 2007; Uljens, 1997). This tradition has already been extended to digital tools through, for example, the concept of digital didactical design, which provides an empirical framework for analysing how teachers’ design decisions shape technology-mediated learning (Jahnke et al., 2017).

Approached this way, GenAI becomes one more factor in teaching design, to be shaped by the same questions rather than settled by a blanket rule:

• The why recovers the purpose of an assignment, which is often what makes offloading either harmless or harmful. This mirrors, at the level of the classroom, a broader argument that Europe’s strategy should move from an AI first to a purpose first posture, placing societal benefit and human-centric values ahead of acceleration for its own sake (Dignum et al., 2025). What purpose first is to AI policy, the didactic why is to teaching.
• The what forces a distinction between knowledge that must be built in the student and knowledge that may legitimately be offloaded.
• The how and when concern where in a learning sequence friction is desirable and where it is merely unnecessary.
• The for whom is not incidental. The cognitive risks of AI fall hardest on novices and on students with weaker self-regulation, raising the prospect of a widening equity gap (Bastani et al., 2025). It is the same population, and the same equity concern, raised by the caution about visibility above. Making the process visible can support these students or, if designed carelessly, weigh most heavily on them. This is also why frameworks aimed only at upgrading teachers’ technical competence, important as they are (Redecker, 2017; UNESCO, 2024a, 2024b), need to be complemented by frameworks for pedagogical and content-specific judgement (Mishra & Koehler, 2006; Mishra et al., 2023).

Next, we introduce the Question Zero tool from the policy world with five overarching questions-dimensions related to thoughtful and responsible AI adoption and in what ways they overlap and complement the didactic questions.

Question Zero: why, who, what, how, where

Question Zero (Q0) from the policy world asks: “Under what conditions should an AI system be adopted, if at all?” (Dignum et al., 2026; Titareva et al., 2026).  Its purpose is to resist an “AI First” logic in which the technology is perceived as the only next step. And problems are identified later for the technology to solve. Instead, the problem and available alternatives (why?), affected stakeholders (who?), type of technology to solve a specific problem at hand (what?), and conditions of adoption (how and where?) should be considered before an AI tool is selected as the solution.

In education, a similar mistake is to begin with the assumption that generative AI will be used and then impose the decision on teachers and students to use it without pedagogically sound guidance. A didactic approach requires an earlier question: why we are using the tool and what educational problem are we trying to solve?

This is where Q0 and the classical didactic questions, coming from different disciplines and angles, become connected. Q0 is a governance question about responsible and thoughtful pre-adoption stage; the didactic why is a question about educational purpose. Both questions move the focus from technology as their starting point to the purpose. Q0 requires organisations to describe honestly the existing problem (not the solution/tool), consider human and non-AI alternatives, and justify why AI is appropriate before adopting it. The didactic why for teaching asks what an activity is intended to achieve (desirable teachers and learners’ goals) before the stakeholders decide what tools to use.

The connection between didactic questions and Q0 dimensions moves beyond why. Q0 operationalises responsible adoption through five iterative dimensions: why, who, what, how and where. These dimensions include: motivation for using the tool starting from the real problem identification and alternatives, affected stakeholders, the type and implications of the AI system, the adoption process, and infrastructure and monitoring. Didactic questions ask what happens when technology is imposed on teaching practices: what knowledge has to remain with the learner, when assistance should be provided, how much cognitive work may be offloaded, and for whom a particular design encourages or limits learning. The two approaches and questions are complementary and help make informed and weighted decisions prior to technology adoption in core activities, including in teaching and assessment.

Table 1

Didactic Questions vs. Q0 Dimensions

Didactic Questions	Question Zero dimensions
Why?	Why?
What?	Who?
How?	What?
When?	How?
For whom?	Where?

Why top-down regulation is not enough

It is tempting to respond to the resulting uncertainty with detailed, centralised rules, which is important for such complex systems and environments as education.  Indeed, the lack of concrete guidance does create a large space for interpretation, which will likely produce considerable variation in how AI is understood and used across classrooms and institutions (Mårell-Olsson, in press). However, education cannot be governed as if teaching was mainly a compliance task. Teachers need pedagogical freedom and the exercise of professional evaluated judgement, because the right use of AI is contextual. It depends on the subject, the stage, the goal and the student.

The current high-profile policy documents vividly demonstrate it. High-level, rights-based guiding documents such as UNICEF’s “Guidance on AI and Children” (2025) and UNESCO’s “AI and education: protecting the rights of learners” (2025a) are valuable precisely because they are general enough to apply across very different contexts. However, their authors are explicit that these guidelines must remain high-level, and they do not and cannot be easily applied at the level of each classroom. Even the European Commission’s dedicated, recently updated “Ethical guidelines on the use of AI and data in teaching and learning for educators” (European Commission, 2026), which now incorporates the AI Act and GDPR and offers core principles, guiding questions and classroom scenarios, is framed around ethical and responsible use and risk awareness rather than around the teaching redesign that the cognitive-offloading problem demands. The same is true of most national AI strategies and sectoral guidelines. They identify principles and risks. They do not, and arguably should not, prescribe teaching design. That work has to be done with and by teachers, not to them.

From high-profile principles to practice: designing teaching and assessment in GenAI times

Assessment has to begin by measuring process, judgement and independence, not only the finished product. The research and policy debate on AI and assessment points towards combinations of AI-free and AI-integrated components (Lodge et al., 2023; Dawson, 2021). AI-free components are needed to secure basic individual understanding. AI-integrated components are needed to assess how a student uses AI critically, ethically and in a subject-responsible way. These are complementary, not alternatives. Below we propose a framework for putting didactic/Q0 questions into practice in pedagogical settings in the age of generative AI.  

We propose designing teaching in phases:

Stage 1. Own understanding, without AI (i.e., securing initial cognitive work): The purpose is to protect the cognitive work through which initial understanding is built. The student must first read, think, formulate and test ideas on their own.

Stage 2. AI as dialogue partner, critic or feedback resource (i.e., extending and challenging thinking):  GenAI is used only to extend thinking rather than replace it. AI is introduced to question, challenge, exemplify and provide feedback.

Stage 3. Verification (i.e., developing disciplinary and epistemic judgement): AI use is conditional. The student must compare GenAI tool’s output against course literature, empirical evidence, theory or their own data, thereby turning passive acceptance into active judgement.

Stage 4. Accountability (i.e., making the use of AI transparent and reflective): The student reports what AI was used for, what was accepted, what was rejected, and why.

Stage 5. Monitoring: Assessment can then ask a different set of questions: Can the student explain this without AI? Can the student use AI without losing subject-matter judgement? Can the student identify errors, uncertainties and limitations? Can the student defend their choices, orally or in writing? Can the student show how their understanding developed over time?

Table 2

5-Stage AI Pedagogy Framework (Mårell-Olsson & Titareva, 2026)

Stage	Focus & Purpose	AI Status	Key Student & Instructor Actions
1. Own Understanding	Secure initial cognitive work	No AI	Student reads, thinks, formulates, and tests initial concepts independently.
2. Dialogue & Challenge	Extend thinking	Active AI Partner	AI acts as critic / Socratic dialogue partner to question and exemplify.
3. Verification	Develop epistemic judgement	Conditional AI	Student cross-references AI outputs against peer-reviewed literature, theory, and data.
4. Accountability	Transparency & reflection	Transparent AI	Student documents AI prompts, accepted/rejected output, and rationale.
5. Monitoring	Assessment of core mastery	Evaluated Mastery	Assessment tests independent defence, error detection, and student growth over time.

We are not claiming to offer a novel solution to designing teaching and assessment for GenAI reality. A recent framework for higher education proposes similar principles, preserving cognitive friction, positioning AI as a provisional thinking partner, embedding evaluation throughout, and sequencing AI-free with AI-mediated phases (Vendrell & Johnston, 2026). Wazan (2026) builds an assessment design and examination system around the same logic. Students must commit to an initial answer before consulting GenAI, then interrogate the model’s output (including deliberately flawed responses) and justify a refined answer, scored by a rubric that assesses reasoning as a process rather than a product. That two such proposals, developed from cognitive psychology, critical pedagogy and epistemology tradition, converge on the same sequence is significant in itself.  What we add is the didactic account of why it works, the why, what, how, when and for whom, its extension across schooling as well as higher education, and the caution developed above that making the process visible is not automatically benign.

To illustrate our framework in practice, consider a written assignment that requires students to analyse a problem using course material. In the first phase, students draft their analysis in class, without AI, so that the initial reading and reasoning are their own. In the second stage, students use AI to interrogate their draft, asking it to challenge the argument, surface counterexamples or offer alternative framings, and then deciding what to keep. In the third phase, students check every AI-generated claim, source and example against the literature and the evidence, and correct what does not hold. In the fourth stage, students submit a short account of how AI was used, what they accepted or rejected and why, and are examined on whether they can defend the analysis without it. In the fifth stage, the educator assesses how students can explain their analysis without AI, show critical thinking and demonstrate the growth of their own thought and knowledge over the process of completing the assignment.

These phases are a design logic rather than a fixed timetable. They may recur within a longer task, and where friction is most valuable depends on the students and the subject. A novice may need a longer, more protected first phase, while an advanced student may begin closer to verification.

We believe that, rather than forbidding cognitive offloading, education in the future must regulate it pedagogically. AI may reduce extraneous load. It must not be allowed to take over the load that builds understanding. This principle, drawn from the distinction between beneficial and detrimental offloading, is helpful in guiding teaching and assessment design in GenAI times.

Implications for professional development and a different kind of guidance

We offer two crucial implications. First, education systems face a very large need for professional development, from schools to higher education globally. This goes beyond AI literacy. Teachers need practical pedagogical support in designing teaching and assessment that reduces detrimental offloading and metacognitive laziness, while still allowing beneficial offloading, and in helping students understand why the goal is to learn the concept and its applicability to the real world, not merely to complete the task (see Lodge et al., 2023). Policy recommendations are starting to reflect this need. The Council of the European Union’s 2026 “Conclusions on teachers in the era of artificial intelligence (AI)” place teachers at the centre, recognising that AI may reshape how they design, deliver and assess instruction and calling for guidance and supportive frameworks for their continuous development (Council of the European Union, 2026). Characteristically, though, such instruments invite and recognise rather than specify what that development should contain, which is  the gap we attempt to cover in this article.

Second, the kind of guidance that policy produces needs to pivot. The most useful instruments for teachers and classrooms will not be longer lists of risks or prohibitions, but resources that help teachers translate risks into concrete decisions about assignment design, feedback, and assessment (working examples, design patterns, and exemplars of process-visible tasks) while preserving professional judgement. Better AI tools may help at the margins, such as Socratic tutoring models, tools designed to scaffold metacognition rather than provide complete answers, and teacher-facing applications that augment rather than replace an expert (Wang et al., 2024; OECD, 2026). The OECD frames this as a choice between replacement, complementarity and augmentation, and argues that augmentation (e.g., teacher and AI refining each other’s work, while professional judgement is preserved) holds the most promise (OECD, 2026). But tools are not the fundamental issue. The fundamental issue is how generative AI changes the design of teaching and the assessment of students’ knowledge.

Conclusion

Many current AI policies and reports have grown adept at identifying the risks generative AI poses to learning. They are useful but give teachers very little practical support in translating those risks into teaching and assessment design. We argue that closing that gap is the central educational task of this moment, and it is a shared task across education systems globally. Just as AI policy is being urged to put purpose before acceleration (Dignum et al., 2025), education should put didactics before tools. The argument is not merely that didactics matter, but that the didactic questions can be turned into a concrete, phased design for teaching and assessment, one whose guiding principle is to regulate cognitive offloading rather than forbid it. Closing the gap, therefore, depends a lot on the design work, and on investing, at scale, in the professional judgement of educators.

In conclusion, earlier we introduced Question Zero (Q0) as a way of asking the right questions before adoption: “Under what conditions should an AI system be adopted, if at all?” Q0 includes dimensions of why, who, what, how and where. What real problem are we trying to solve? Who might be affected? What kind of technology, if any, is appropriate to solve a problem at hand? How should it be adopted and governed? And where, and under what conditions should it operate? The didactic questions ask: why, what, how, when and for whom? The two approaches converge on one overarching premise: the adoption of the technology does not start with the tool, instead it starts with the purpose, i.e., answering why and what do we want to achieve? Both invite us to think from a perspective of purpose, context, people and the goals and conditions in which the tool is planned to be applied.

References

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), Article e2422633122. https://doi.org/10.1073/pnas.2422633122

Bearman, M., Tai, J., Dawson, P., Boud, D., & Ajjawi, R. (2024). Developing evaluative judgement for a time of generative artificial intelligence. Assessment & Evaluation in Higher Education, 49(6), 893–905. https://doi.org/10.1080/02602938.2024.2335321

Bergström, P., Mårell-Olsson, E., & Jahnke, I. (2019). Variations of symbolic power and control in the one-to-one computing classroom: Swedish teachers’ enacted didactical design decisions. Scandinavian Journal of Educational Research, 63(1), 38–52. https://doi.org/10.1080/00313831.2017.1324902

Bjork, E. L., & Bjork, R. A. (2015). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M. A. Gernsbacher & J. R. Pomerantz (Eds.), Psychology and the real world: Essays illustrating fundamental contributions to society (2nd ed., pp. 59–68). Worth.

Carless, D., & Boud, D. (2018). The development of student feedback literacy: Enabling uptake of feedback. Assessment & Evaluation in Higher Education, 43(8), 1315–1325. https://doi.org/10.1080/02602938.2018.1463354

Council of the European Union. (2026). Council conclusions on teachers in the era of artificial intelligence (AI) (9003/26). https://data.consilium.europa.eu/doc/document/ST-9003-2026-INIT/en/pdf

David, L., Vassena, E., & Bijleveld, E. (2024). The unpleasantness of thinking: A meta-analytic review of the association between mental effort and negative affect. Psychological Bulletin, 150(9), 1070–1093. https://doi.org/10.1037/bul0000443

Dawson, P. (2021). Defending assessment security in a digital world: Preventing e-cheating and supporting academic integrity in higher education. Routledge. https://doi.org/10.4324/9780429324178

Desvaux, C., Abdelghani, R., Oudeyer, P.-Y., & Sauzéon, H. (2026). Curiosity and metacognition: Towards a unified framework for learning and education in the age of AI [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2604.25648

Dignum, V., Carli, R., Ericson, P., Titareva, T., & Tucker, J. (2025). “AI first” to “purpose first”: Rethinking Europe’s AI strategy. AI Policy Lab (AIPEX), Umeå University. https://doi.org/10.63439/LPOU6506

Dignum, V., Carli, R., Dahlgren Lindström, A., Ericson, P., Titareva, T., & Tucker, J. (2026 June). Question Zero for Explainability and Vice Versa: The Case of the EU’s AI First Strategy. In International Conference on Human-Computer Interaction (pp. 17-31). Cham: Springer Nature Switzerland.

Dunn, T. L., & Risko, E. F. (2016). Toward a metacognitive account of cognitive offloading. Cognitive Science, 40(5), 1080–1127. https://doi.org/10.1111/cogs.12273

European Commission. (2026, May 12). Guidelines on the ethical use of artificial intelligence and data in teaching and learning. European School Education Platform. https://school-education.ec.europa.eu/en/discover/publications/guidelines-ethical-use-ai-and-data-teaching-and-learning

Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544

Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), Article 6. https://doi.org/10.3390/soc15010006

Hathcoat, J. D., Slotnick, R., & Miller, W. (2026). The End of Assessment? Disruption and Transformation in the Age of AI. Research & Practice in Assessment, 21(3), 5-26 (Special Issue–AI in Assessment).

Hopmann, S. (2007). Restrained teaching: The common core of Didaktik. European Educational Research Journal, 6(2), 109–124. https://doi.org/10.2304/eerj.2007.6.2.109

Jahnke, I., Bergström, P., Mårell-Olsson, E., Häll, L., & Kumar, S. (2017). Digital didactical designs as research framework: iPad integration in Nordic schools. Computers & Education, 113, 1–15. https://doi.org/10.1016/j.compedu.2017.05.006

Kalyuga, S., & Plass, J. L. (2025). Rethinking cognitive load theory. Oxford University Press. https://doi.org/10.1093/9780190078539.001.0001

Klafki, W. (2000). Didaktik analysis as the core of preparation of instruction. In I. Westbury, S. Hopmann, & K. Riquarts (Eds.), Teaching as a reflective practice: The German Didaktik tradition (pp. 139–159). Lawrence Erlbaum Associates.

Li, C., Cui, H., & Hagedorn, L. S. (2026). The cognitive impact of ChatGPT in higher education: A systematic review of critical and creative thinking outcomes. Computers and Education: Artificial Intelligence, 10, Article 100571. https://doi.org/10.1016/j.caeai.2026.100571

Lodge, J. M., Howard, S., Bearman, M., Dawson, P., & Associates. (2023). Assessment reform for the age of artificial intelligence. Tertiary Education Quality and Standards Agency (TEQSA). https://www.teqsa.gov.au/sites/default/files/2023-09/assessment-reform-age-artificial-intelligence-discussion-paper.pdf

Lodge, J.M., Loble, L. (2026). Artificial intelligence, cognitive offloading and implications for education. University of Technology Sydney. https://doi.org/10.71741/4pyxmbnjaq.31302475

Mårell-Olsson, E. (2012). Att göra lärandet synligt? Individuella utvecklingsplaner och digital dokumentation [Making learning visible? Personal development plans and digital documentation] [Doctoral dissertation, Umeå University]. DiVA. http://urn.kb.se/resolve?urn=urn:nbn:se:umu:diva-59420

Mårell-Olsson, E. (in press). AI literacy, governance, and equity in Swedish education policy. In T. Titareva, A. Csonska-Stambekova, & E. Mårell-Olsson (Eds.), AI ethics in global education: Building a future grounded in local realities. CRC Press.

Mishra, P., & Koehler, M. J. (2006). Technological pedagogical content knowledge: A framework for teacher knowledge. Teachers College Record, 108(6), 1017–1054. https://doi.org/10.1111/j.1467-9620.2006.00684.x

Mishra, P., Warr, M., & Islam, R. (2023). TPACK in the age of ChatGPT and generative AI. Journal of Digital Learning in Teacher Education, 39(4), 235–251. https://doi.org/10.1080/21532974.2023.2247480

Nicol, D. J., & Macfarlane-Dick, D. (2006). Formative assessment and self-regulated learning: A model and seven principles of good feedback practice. Studies in Higher Education, 31(2), 199–218. https://doi.org/10.1080/03075070600572090

OECD. (2026). OECD digital education outlook 2026: Exploring effective uses of generative AI in education. OECD Publishing. https://doi.org/10.1787/062a7394-en

Redecker, C. (2017). European framework for the digital competence of educators: DigCompEdu (Y. Punie, Ed.). Publications Office of the European Union. https://doi.org/10.2760/159770

Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002

Shapiro, H., Souto-Otero, M., & Watermeyer, R. (2026). Metacognitive AI literacy: Going beyond the AI skills gap agenda. Learning, Media and Technology. https://doi.org/10.1080/17439884.2026.2652638

Tai, J., Ajjawi, R., Boud, D., Dawson, P., & Panadero, E. (2018). Developing evaluative judgement: Enabling students to make decisions about the quality of work. Higher Education, 76(3), 467–481. https://doi.org/10.1007/s10734-017-0220-3

Titareva T., Tucker J., Carli R., Movchan V., Dignum V. (2026). Question Zero: Why Responsible AI Begins Before AI Adoption. AI Policy Exchange Forum (AIPEX). https://doi.org/10.63439/WCEZ7337

Uljens, M. (1997). School didactics and learning: A school didactic model framing an analysis of pedagogical implications of learning theory. Psychology Press. https://doi.org/10.4324/9780203304778

UNESCO. (2024a). AI competency framework for teachers. https://doi.org/10.54675/ZJTE2084

UNESCO. (2024b). AI competency framework for students. https://doi.org/10.54675/JKJB9835

UNESCO. (2025a). AI and the future of education: Disruptions, dilemmas and directions. https://doi.org/10.54675/KECK1261

UNESCO (2025). AI and education: protecting the rights of learners.  https://doi.org/10.54675/ROQH4287

UNICEF. (2025b). Policy guidance on AI for children (Version 3.0). https://www.unicef.org/innocenti/reports/policy-guidance-ai-children

Vendrell, M., & Johnston, S.-K. (2026). Scaffolding critical thinking with generative AI: Design principles for integrating large language models in higher education. Computers and Education: Artificial Intelligence, 10, Article 100572. https://doi.org/10.1016/j.caeai.2026.100572

Wang, R. E., Ribeiro, A. T., Robinson, C. D., Loeb, S., & Demszky, D. (2024). Tutor CoPilot: A human-AI approach for scaling real-time expertise [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2410.03017

Wazan, A. S. (2026). Strategies for creating uncertainty in the AI era to trigger students’ critical thinking: Pedagogical design, assessment rubric, and exam system [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2602.00026

Weidlich, J., Gašević, D., Drachsler, H., & Kirschner, P. (2025). ChatGPT in education: An effect in search of a cause. Journal of Computer Assisted Learning, 41(5), Article e70105. https://doi.org/10.1111/jcal.70105

Abstract

The Question Zero (Q0) Self-Assessment Tool for Responsible AI, developed by the AI Policy Lab at Umeå University, supports organisations in posing foundational questions before adopting AI. Grounded in the concept of Question Zero: “Under what conditions should an AI system be adopted, if at all?”, the tool offers support for cross-functional team discussions covering motivation, stakeholder mapping, system type, adoption process and infrastructure. It is designed for public institutions, civil society organisations, policymakers and other actors looking for responsible AI decision-making approaches. With the help of the tool, we argue that purposeful, problem-led assessment must precede any procurement and deployment decisions.

Keywords: Question Zero, Q0, Responsible AI, AI Self-Assessment, AI Governance, EU AI First Strategy, AI Policy, AI Adoption Motivation, Stakeholder Mapping, Explainability, AI Policy Lab, Umeå University

 

The Problem with Starting from the Answer

Public and private organisations face growing pressure to adopt artificial intelligence (AI). Governments across the EU and beyond are committing vast public resources to AI acceleration. Organisations, large and small, feel urged to integrate AI into their operations for fear of being perceived as falling behind. The European Commission’s Apply AI Strategy (European Commission, 2025) has formalised this pressure into policy direction, framing AI as the default first response to organisational and societal challenges. The same trend can be seen at a national level, as well as across sectors and organisations.

However, this urgency leads to a fundamental disruption of sound decision-making. Rather than beginning with a problem and asking what solutions are available, an “AI First” logic begins with the technology and asks where it can be applied. As we have argued elsewhere, the result is that the question of whether AI is appropriate in a given context is hardly ever properly posed (Dignum et al., 2025). The consequences of skipping this step are not abstract.

The Dutch childcare benefit scandal, in which an algorithmic system wrongly flagged tens of thousands of parents as committing fraud, disproportionately affecting ethnic minority and low-income families (Amnesty International, 2021), is one of the examples of what might happen when AI is deployed without adequate prior deliberation about necessity, appropriateness and who might be harmed. Similarly, the UK’s A-level grading algorithm reinforced existing inequalities, putting students from smaller schools and lower socio-economic backgrounds at a disadvantage by relying on historical data to determine results (Kolkman, 2020). In the United States, ProPublica’s investigation into the COMPAS recidivism algorithm revealed significant racial biases, with Black defendants nearly twice as likely as their white counterparts to be wrongly flagged as high risk (Larson et al., 2016). Australia’s Robodebt scandal exposed the dangers of deploying flawed automated decision-making systems at scale, showing how vulnerable citizens can be unjustly penalised without adequate oversight, transparency and safeguards (Royal Commission into the Robodebt Scheme, 2023).

Together, these cases illustrate how without robust AI governance, such systems can deepen discrimination against marginalised groups and undermine public trust in key institutions. While these cases differ in context and technical form, they share an important feature: technology was introduced without adequately addressing Question Zero (Q0), meaning the prior scrutiny of necessity, appropriateness, accountability and social impact was weak or absent.

The question Q0 (“Under what conditions should an AI system be adopted, if at all?”) we pose, is a call for rigour and strategic foresight. It asks organisations to define the problem before considering technological solutions, to assess alternatives, to map who benefits and who bears risks and to ensure that any decision on how to proceed is genuinely justified. Only sometimes, not always, is AI the right answer. Thus, the purpose of the Q0 Self-Assessment Tool is to facilitate a structured pause to allow for better decision making about responsible AI.

What is the Q0 Self-Assessment Tool?

The Q0 Self-Assessment Tool is a free assessment tool designed not as a compliance checklist or technical audit, but as a structured self-reflection instrument for cross-functional groups within organisations. Its primary goal is to support decision-makers in working through the foundational questions of responsible AI governance before procurement or deployment decisions are finalised (Titareva et al., 2026).

The tool is organised into five sections: Why, Who, What, How and Where, each targeting a distinct dimension of responsible AI governance. These sections are not a linear sequence; organisations are encouraged to move between them iteratively, as answers in one area frequently raise additional questions in others.

Section A: Why (Motivation) asks organisations’ decision-makers and employees to articulate the problem they are attempting to solve, the reasons they are considering AI, the alternatives they have considered (including human and non-AI technical solutions) and why and whether AI is the best solution for the existing problem. This is the core of Q0: encouraging explicit justification before any further investment of time or resources into AI procurement and deployment.

Section B: Who (Stakeholder Mapping) asks organisations to identify the stakeholders who may be directly or indirectly affected by the adoption of an AI system. This includes colleagues, employees, customers, and members of marginalised or underrepresented groups who are often overlooked in initial assessments (as well as at later stages). The section also prompts organisations to consider who currently performs the tasks the AI system is intended to undertake, who is likely to benefit, who could be harmed and whether meaningful opt-out options exist. By mapping the full stakeholder landscape, organisations can assess how different groups may experience the adoption of AI and ensure that no group is disproportionately burdened, excluded or deprived of meaningful choice.

Section C: What (Type of AI System) asks organisations to define and describe the specific AI system, method or tool under consideration (e.g., predictive, generative, categorising or hybrid), as well as whether it will be procured off the shelf, developed in-house or built for specific needs of an organisation. This section encourages organisations to ensure that the chosen system fits the problem they want to solve and that the system’s complexity, development approach and flexibility match its intended use.

Section D: How (Adoption Process) asks organisations to consider how the AI system would be deployed, how existing workflows would change, how outputs and performance would be monitored, how data security would be ensured and how affected users would be able to understand the basis of the system’s conclusions. It also prompts organisations to define how complaints and errors would be handled and who would be accountable for oversight throughout the system’s lifecycle. This section emphasises that accountability must be embedded from the beginning rather than added retrospectively, and that explainability should be treated as an ongoing organisational practice rather than merely a compliance requirement.

Section E: Where (Infrastructure and Control) asks organisations to consider where the AI system will operate, where data will be stored, where the provider is based and where the training data originates. These questions address issues of digital sovereignty, data jurisdiction and the extent to which organisations can maintain meaningful control over the AI systems they adopt. This section emphasises that effective governance and accountability depend on understanding where the system operates, who controls it and how data is collected, stored and managed. It therefore encourages organisations to ensure transparency regarding infrastructure, models and training data in order to meet legal, security and ethical responsibilities and requirements.

Each section concludes with a confidence rating on a five-point scale, allowing organisations to reflect on how confidently they can answer the questions within that dimension. Rather than measuring compliance or assigning a level of AI maturity, the ratings are intended to stimulate discussion, identify areas of uncertainty and highlight where further information, stakeholder engagement or organisational preparation may be needed. The aggregated scores are visualised in a radar diagram, providing an overview of confidence across all five dimensions. This visualisation enables organisations to compare results across teams, over time or between different AI use cases, thus supporting internal learning, reflection and continuous improvement. The ratings are not intended for external accreditation, benchmarking or audit, but as a self-assessment tool that helps organisations recognise strengths, identify gaps and guide responsible AI governance.

A final section of concluding questions invites decision-makers and employees of organisations to reflect on the insights gained throughout the self-assessment process by evaluating any pilot experiences, considering the expected benefits of the proposed AI system alongside its possible risks and harms and identifying appropriate next steps. Rather than prescribing decisions or producing a pass-fail outcome, the tool is designed to support informed, context-sensitive judgement, recognising that responsible AI adoption depends on an organisation’s objectives, values, legal obligations and operational context. As a decision-support instrument, the Q0 tool helps structure deliberation, surface assumptions and strengthen transparency, but responsibility for interpreting the findings and deciding how to act upon them remains with the organisation and its employees as part of its internal AI governance.

                          Figure 1. Question Zero Conceptual Operationalisation

Positioning the Q0 Self-Assessment Tool Among Existing AI Assessment Frameworks

The Q0 self-assessment tool builds on a growing ecosystem of AI governance and assessment instruments. At the same time, it addresses a different stage of the AI lifecycle. Existing frameworks such as the European Commission’s Assessment List for Trustworthy Artificial Intelligence (ALTAI), UNESCO’s Ethical Impact Assessment (EIA), the UNESCO Readiness Assessment Methodology (RAM), the UNDP Human Rights Impact Assessment Toolkit and similar international and national frameworks support responsible AI through structured reflection on legal, ethical, technical and organisational issues. Although these tools have different purposes and target different audiences, they share a common goal: They help governments or organisations identify, manage and govern the risks associated with AI systems and promote transparency, accountability, human rights and public trust.

The Q0 tool shares many of these principles. Like ALTAI, it addresses human oversight, transparency, accountability, fairness and data governance. Similar to UNESCO’s Ethical Impact Assessment, it encourages multidisciplinary participation and considers the wider social impacts of AI alongside technical issues. The UNDP Human Rights Impact Assessment Toolkit also emphasises the importance of identifying rights holders and affected communities. UNESCO’s Readiness Assessment Methodology takes a broader perspective by examining institutional capacity, governance, infrastructure and organisational readiness. Q0 tool follows a similar approach by treating responsible AI governance as an institutional challenge rather than only a technical one.

The main difference is not the questions themselves, but when they are asked. The assessment tools mentioned above often assume that an organisation has already decided to develop, procure or deploy an AI system. Their purpose is to assess whether that system is trustworthy, legally compliant or ethically governed. Some frameworks do include questions about necessity or proportionality during the pre-adoption stage. However, these questions usually form one part of a broader impact or compliance assessment when the AI system’s adoption project has already been initiated.

The Q0 tool starts one step earlier. It asks whether AI should be adopted at all, and if yes, under what conditions. This is the central purpose of Question Zero. The tool encourages organisations to define the problem first, consider non-AI alternatives and examine governance, stakeholder impacts and organisational readiness before resources are committed or procurement begins. It also places greater emphasis on collective deliberation and participation. It recommends cross-functional participation, the inclusion of affected stakeholders where possible, iterative use of the tool, documentation of disagreements and integration with existing governance processes. The goal is not a single assessment but an ongoing process of organisational learning and reflection.

A second difference is the intended use of the tool. Many existing assessment instruments support regulatory compliance, formal impact assessment or risk management. The Q0 tool has a different purpose. It is a self-assessment instrument. It does not generate a compliance score or certify that an AI system or case is trustworthy. The confidence ratings are intended to stimulate discussion, reveal uncertainty and identify gaps in knowledge. The radar diagram allows organisations to compare discussions across teams, projects or points in time. This helps identify differences in understanding and areas that require further attention.

The Q0 Self-Assessment Tool is intended to complement, not replace, existing governance frameworks. Organisations that decide AI is appropriate can then carry out more detailed assessments, such as ethical impact assessments, human rights impact assessments, data protection impact assessments or assessments required under the national or regional legislation (e.g., the EU AI Act and others). Q0 tool therefore supports an earlier stage of decision-making. It aims to improve the quality of the decision that determines whether these later assessments will be necessary.

The Q0 tool also has limitations. It is a self-assessment tool. Its value depends on organisations’ employees engaging with the questions and including a diversity of perspectives. It cannot remove organisational bias or guarantee that all relevant stakeholders are represented. It cannot ensure that recommendations will be implemented. It is also not a substitute for legal, technical or human rights assessments where these are required. Finally, no questionnaire can capture the full complexity of every organisational context. The Q0 tool should therefore be seen as a governance aid rather than a decision-making mechanism. It supports better judgement, but responsibility for the final decision always remains with the organisations’ leadership and employees.

Example of the Q0 Self-Assessment Tool in Practice

A university department is considering introducing a generative AI assistant to answer students’ questions about courses, deadlines and administrative procedures. Instead of immediately assuming that an AI chatbot is the right solution, the department first uses the Q0 Self-Assessment Tool. Addressing the questions in the section “Why”, university employees from diverse departments and functional areas define the problem, for example, long response times during busy periods. They also consider other options, such as improving online guidance, extending office hours or employing more student assistants. AI is treated as one possible solution, not the starting point.

The remaining Q0 tool’s sections help to broaden the discussion. Under the section “Who”, the participants identify as many categories of stakeholders as possible, who may be affected, including students, lecturers, administrative staff, IT services and others. It also considers students with disabilities, international students and others who may be disadvantaged by digital-only support. Under “What”, participants compare different AI systems and ask whether a simpler non-AI or rule-based solution could meet the need. Under the sections “How” and “Where”, they discuss governance, accountability, transparency, data security, hosting arrangements and whether the university keeps appropriate control over its data.

At the end of the process, the group gives a confidence rating for each section. The radar diagram shows where participants feel confident and where more work is needed. In this example, confidence could be high for the motivation behind the project but lower for infrastructure, data governance and long-term accountability. These results do not determine whether AI should be adopted. Instead, they help the department identify where further discussion, expert advice or pilot testing is needed before making a final decision.

Conclusion

AI is not inevitable. Its adoption is not inherently beneficial. Responsible AI governance begins before procurement, development or deployment. It begins by asking whether AI is the best solution for your current problem, if so, under what conditions. This is the purpose of Question Zero.

The Q0 Self-Assessment Tool provides organisations with a practical way to support this early reflection. It does not replace existing governance frameworks or assessments. Instead, it complements them by addressing a stage of decision-making that is often overlooked. It focuses on motivations, stakeholders, system choice, adoption, infrastructure and organisational control.

The value of the Q0 Self-Assessment Tool lies not in providing answers, but in improving the quality of the questions organisations ask before adopting AI systems. It creates space for discussion, makes assumptions explicit and helps identify issues that require further attention. In the end, the responsibility for deciding whether to adopt an AI system remains with the organisations’ employees. The Q0 process helps ensure that this decision is more informed, transparent and accountable.

Access the Tool

The Q0 Self-Assessment Tool (Version 3, April 2026) is available: HERE 

 

References

Ala-Pietilä, P., Bonnet, Y., Bergmann, U., Bielikova, M., Bonefeld-Dahl, C., Bauer, W., … & Van Wynsberghe, A. (2020). The assessment list for trustworthy artificial intelligence (ALTAI). European Commission.

Amnesty International (2021, October 2025). Xenophobic machines: Discrimination through unregulated use of algorithms in the Dutch childcare benefits scandal.  https://www.amnesty.org/en/documents/eur35/4686/2021/en/

Dignum, V., Carli, R., Ericson, P., Titareva, T., & Tucker, J. (2025). ‘AI First’ to ‘Purpose First’: Rethinking Europe’s AI Strategy. AI Policy Exchange Forum (AIPEX).  https://doi.org/10.63439/LPOU6506

Dignum, V., Carli, R., Dahlgren Lindström, A., Ericson, P., Titareva, T., & Tucker, J. (2026, June). Question Zero for Explainability and Vice Versa: The Case of the EU’s AI First Strategy. In International Conference on Human-Computer Interaction (pp. 17-31). Cham: Springer Nature Switzerland. https://doi.org/10.1007/978-3-032-29456-2_2 

European Commission (2026, March 27). Apply AI Strategy. Shaping Europe’s Digital Future. https://digital-strategy.ec.europa.eu/en/policies/apply-ai

Kolkman, D. (2020, August 26). “f**k the algorithm?”: What the world can learn from the UK’s A-level grading fiasco. The London School of Economics and Political Science (LSE) Impact.  https://blogs.lse.ac.uk/impactofsocialsciences/2020/08/26/fk-the-algorithm-what-the-world-can-learn-from-the-uks-a-level-grading-fiasco/#:~:text=%E2%80%9CF**k%20the%20algorithm%E2%80%9D?:%20What%20the%20world%20can,address%20through%20making%20their%20algorithms%20more%20explainable.

Larson, J., Mattu, S., Kirchner, L., & Angwin, J. (2016, May 23). How we analyzed the COMPAS recidivism algorithm. ProPublica. https://www.propublica.org/article/how-we-analyzed-the-compas-recidivism-algorithm

Royal Commission into the Robodebt Scheme (2023, July 7). The Report of the Royal Commission into the Robodebt Scheme to the Governor-General, His Excellency, General the Honourable David Hurley AC DSC (Retd). https://robodebt.royalcommission.gov.au/publications/report

Titareva, T., Carli, R., Dahlgren Lindström, A., Dignum, V., Fabris, B., Ericson, P., & Tucker, J. (2026). Q0 Self-Assessment Tool with Guiding Questions for Responsible AI Approach by AI Policy Lab (AIPL). Test version 3 from April 2026. Umeå University.

UNESCO. (2023). Ethical Impact Assessment: A tool of the Recommendation on the Ethics of Artificial Intelligence. https://doi.org/10.54678/YTSA7796

UNESCO. (2023). Readiness Assessment Methodology: a Tool of the Recommendation on the Ethics of Artificial Intelligence. https://doi.org/10.54678/YHAA4429

United Nations Development Programme. (2025). Human Rights Impact of AI Assessment Tool.  https://www.undp.org/eurasia/publications/human-rights-impact-ai-assessment-toolkit