Beyond the "AI-Powered" Label: What Changes in the Design of an EdTech Product, and How to See It Part 1

EdTech Tulna

October 7th, 2026

Share:

Download PDF
Executive Summary

The rapid mainstreaming of Artificial Intelligence (AI) has sharply expanded what EdTech products can do. EdTech products today generate instructional content, conduct assessments, deliver feedback, and tutor learners in real time. For India, this lands on a problem the country was already trying to solve. The National Strategy for Artificial Intelligence (NITI Aayog, 2018) names education as a priority sector and pins its two hardest constraints as scale, the sheer number of learners to reach, and distance, the gap between quality instruction and the learner who needs it. AI is attractive in India precisely because it can work on both at once, a system that explains, questions, and responds in a learner's own language. The opportunity is real, and it is arriving faster than the tools for telling good AI-powered products from weak ones.

That turns product selection into an evaluation challenge. A stakeholder used to be able to inspect an EdTech product before adopting it, its content and pathways were fixed and visible. An AI product cannot be judged that way: its instruction is generated live, so its quality shows only in how it behaves with a learner, never in the claim or the demo a policymaker, a school, or a parent is handed.

To address this, Tulna, housed at IIT Madras, traced how AI can be implemented in the design of learning products, what it now enables for learners, and what it puts at risk.

What AI Now Does in a Learning Product

 

  • Automation: AI applied to well-bounded, repetitive tasks where a correct outcome is checkable.
  • Adaptive and predictive personalisation: AI that uses learner data to sequence content, adjust difficulty, and recommend the next step.
  • Generative AI:  AI that produces explanations, worked examples, hints, feedback, even whole assessment items, composed live.
  • Conversational AI:  AI that engages a learner in multi-turn, natural-language dialogue, where the conversation itself is the medium of instruction.

Risks of AI-Mediated Learning

  • In what the product puts in front of the learner: Covers pedagogical and interactional risks, where generated content, feedback, scoring, and dialogue can fail a learner.
  • In the learner, over sustained use: Covers cognitive and psychosocial risks, where steady reliance on the system begins to affect how a learner thinks and develops.
  • In the model underneath: Covers model- and training-level risks, where bias and error carried in the model surface in what a learner sees.
  • In how the product is deployed: Covers implementation and governance risks, where data handling, silent updates, and unprepared oversight become the risk.
The Four Faces of AI in EdTech: Capability and Opportunity

AI integration in EdTech spans four functionally distinct capabilities, each with its own behavioural signature and its own risk profile.

 

Automation

AI applied to well-bounded, repetitive tasks where a correct outcome is checkable. 

Opportunity:  AI reduces the administrative burden of repetitive, rule-bound tasks. It also solves for fairness and volume, since the same criteria get applied consistently across large volumes of work rather than varying from one instance to the next. For example, an AI-powered exam grader offers the potential for evaluations that are more impartial and consistent by prioritizing comprehension and methodology rather than simply correct or incorrect responses, while also processing volumes of student work at considerable scale and speed, freeing instructor time for direct teaching.

Adaptive and Predictive Personalisation

AI that uses learner data to sequence content, adjust difficulty, and recommend the next step.

Opportunity:  AI brings a new degree of dynamism to something that was already dynamic. The genuine opportunity lies in the capacity to run algorithms sophisticated enough to draw on a broader spectrum of learner data, including background and multimodal inputs such as visual and textual responses, and to personalize not merely what a learner encounters next but how it is explained, calibrated to their linguistic, cultural, and educational background. For example, in a personalised adaptive learning solution, a learner's mastery is re-evaluated mid-lesson rather than only at the outset. The system reconstructs the next set of problems from wherever the actual gap resides, renders that material in a manner suited to the learner's background, and advances the learner into a different knowledge tier the moment their understanding shifts, all within a single session.

Generative AI

AI that produces explanations, worked examples, hints, feedback, even whole assessment items, composed live rather than drawn from a fixed library.

Opportunity: Generative AI extends learning beyond what a fixed instructional framework can offer. The genuine opportunity lies in interactivity: unlike a fixed question-and-answer exchange, generative systems require the user to construct their own queries to get targeted output, cultivating initiative in how a user explores a topic. This interactivity, paired with feedback shaped to the specific question asked, offers the potential to foster critical thinking rather than simply deliver information, with the AI able to act as either a collaborator alongside the learner or a lead the learner directs. For example, in a lesson plan generator, a full lesson plan is drafted live around the brief a teacher provides, rather than adapted from an existing template.

Conversational AI

AI that engages a learner in multi-turn, natural-language dialogue, where the conversation itself is the medium of instruction.

Opportunity: Conversational AI carries the potential to cultivate a learner's autonomy, interactivity, and metacognition through the exchange itself. The genuine opportunity lies in scaffolding that responds to a learner's reasoning as it unfolds, rather than only their final answer. Dialogue-based tutoring systems anticipate both the steps toward an ideal answer and the specific misconceptions a learner is likely to hold, and steer the conversation to address whichever of these the learner's response reveals, offering hints, counterexamples, or probing questions rather than the answer itself.  For example, an AI tutor that mimics how a tutor works through a problem with a student, using nudges and questions to guide them to the answer rather than handing it over.

These four categories can also be layered within a single product. It can combine automated grading support for instructors, adaptive sequencing drawn from a mastery-tracked content library, generative feedback composed in the moment, and turn-by-turn dialogue as the primary learner interface, all at once.

Risks Of AI-Mediated Learning

The expansion of AI capabilities in EdTech carries a corresponding expansion of risk. The risks show up in four places: in what the product puts in front of the learner, in the learning behaviour of a learner over time, in the model underneath, and in how the product is implemented. They trace back to the same four AI capabilities, each surfacing differently across these four places, generation and dialogue mostly in what the learner is shown, adaptivity in decisions no one can see, automation in scoring that is inconsistent or unexplained.

 

Risks in What The Product Puts in Front of the Learner:
  • Generative systems produce instructional material of variable factual quality and are prone to fabrication. They can reproduce representational bias and stereotyping in the examples, contexts, and imagery they present to learners. Generated material may also diverge from curricular expectations and from the learner's linguistic and cultural context, even where it is factually sound.

  • Dialogic engagement yields critical-thinking and metacognitive benefits under deliberately designed interaction. Accordingly, conversational systems that provide answers where the pedagogical moment calls for learner reasoning risk reducing the dialogic engagement and cognitive effort through which learning is supported. 

  • AI-generated feedback often aligns only partially with the principles of effective feedback, offering generic comments rather than directing a learner to the specific error and the next step. Automated scoring of open responses can be inconsistent and difficult to validate, since the same response can be scored differently from one instance to the next or diverge from the intended rubric. Generative AI integrated into assessment environments can also facilitate unauthorized assistance, challenging the integrity of both formative and summative assessment.

  • Model-determined pathways and recommendations, when neither transparent nor explainable, limit the ability of learners and teachers to understand and critically evaluate the instructional decisions made by adaptive systems.

Risks to Learners:
  • Over-reliance on AI dialogue systems may impair critical thinking, decision-making, and analytical reasoning, the outcome of cognitive offloading in which the system performs the reasoning the learner should be practising. 

  • Generative systems place a novel interpretive burden on learners, who, as novices, often lack the expertise to critically evaluate the explanations and solutions generated for them. Additionally, continuous, information-dense conversational output that is not calibrated to the learner's level may overwhelm rather than support comprehension. 

  • AI models can shift their answers toward whatever option a user suggests, improving accuracy when the user proposes the correct answer but degrading it by a comparable margin when the user proposes an incorrect one. 

  • Beyond these cognitive risks, the policy literature also identifies psychosocial risks associated with sustained AI interaction, including psychological dependency, the displacement of the human interaction on which socio-emotional development relies, and the exposure of learners to systems inappropriate to their developmental stage.

Risks in The Model Underneath:
  • Machine-learning components in education can encode demographic and linguistic bias acquired during training.

  • Hallucination remains a fundamental limitation of current generative models rather than an isolated implementation error.

  • AI systems are frequently deployed without adequate transparency regarding the models used, the data on which they are trained, the rationale underlying their recommendations, their confidence or uncertainty, and the role of human-in-the-loop mechanisms in generating, reviewing, or overriding AI outputs, limiting users' ability to understand or scrutinize AI-generated decisions.

Risks in How the Product is Implemented:
  • Data protection acquires a specifically conversational character, since free-text dialogue invites disclosures of sensitive personal information that structured interfaces never elicited.

  • Products are also routinely updated after deployment, through retraining or version changes that can silently alter system behaviour without evaluators, teachers, or procurement bodies being informed, a problem termed update opacity, which binds any product evaluation to the version assessed. 

  • Oversight is similarly unmatched by readiness among stakeholders: teachers report insufficient preparation to supervise AI-mediated learning.

What This Means for Each Stakeholder

Read together, the capabilities and risks indicate what AI brings to learning and what must be addressed before it reaches classrooms. These capabilities offer value to learners, and they carry risks that can be designed for rather than left to chance. Understood this way, all stakeholders share a role in realising this value and addressing these risks.

  • For product design, the capabilities and risks define what to build. They show where an AI-powered product can support a learner, when it explains a concept in the learner's own context, responds in their language, or offers a hint at the moment of difficulty. They also show where it tends to fail one, when it gets an answer wrong, lets attention drift, or invites too much reliance. Both are best addressed before a product ships, rather than after a class has already used it.

  • For policy, this understanding shows where standards are needed. National policies already endorse technology-enabled personalisation, but the behaviours that decide whether a product helps or harms a learner, protection of a learner’s data and conversations, developmental-stage appropriateness, and the understanding of systems being used, have almost no quality standard to be held to. Closing that gap is the task.

  • For researchers, the potential and limitations of AI raise a set of open questions. How much reasoning a learner hands over through sustained use, when dialogue supports learning and when it replaces it, and which learners a system helps and which it leaves further behind, are still asserted more often than they are measured.

  • For evaluation bodies, these opportunities and challenges indicate what to test. They define the behaviours to assess before a product reaches a learner, not the claims made in a demo. Evaluators need to see whether the product adapts, contextualises, and scaffolds as intended, and how it behaves when a real learner errs, challenges, or struggles.

 

Across all four, the starting point is the same. The AI capabilities and their risks have to be understood together before an AI-powered product is placed in front of a learner.

 

EdTech Tulna, as an evaluation index, has extended its framework to capture what AI capabilities newly offer learners and to address the risks they carry. One new cluster and eleven new criteria are introduced, and existing criteria are re-founded on what a product generates in use rather than on what it was authored to contain. Part 2 elaborates the extension in full.