## The screen can look busy while the mind learns very little
A student is tapping quickly. Correct-answer animations appear. A progress bar moves. Streaks are growing. The app says 92 per cent mastery. The student is smiling and has completed twenty-seven activities without complaint.
It looks like learning.
But what happens ten minutes after the device is closed?
Can the student explain the idea without the hints? Can they solve a differently worded problem? Can they remember the key relationship tomorrow? Can they use the knowledge in a book, laboratory, conversation or examination where the app’s interface is absent? Can they detect when the tool has guided them toward an answer they did not truly understand?
Those questions matter because educational technology can produce several kinds of success that are not identical. It can increase attention. It can make practice more convenient. It can give immediate feedback. It can improve task completion. It can raise scores *inside the tool*. It can also, under good conditions, improve durable knowledge and transferable skill.
The mistake is assuming these outcomes are the same thing.
In September 2026, the American Psychological Association released an expert report on children’s and adolescents’ learning with educational technology. Its message is unusually practical: time spent, enjoyment, participation and completion are often mistaken for learning; students may perform well within a tool yet fail to retain or apply the learning elsewhere; and AI can improve performance while reducing the mental work needed for learning. The OECD’s 2026 Digital Education Outlook makes a related distinction for generative AI: the technology can support learning when it is built around sound pedagogy, but outsourcing a task can improve visible performance without producing real learning gains.
That gives schools and families a better question than “Is this a good app?”
Ask: **What evidence would convince us that this tool changes the learner after the tool is gone?**
This article builds an EdTech evidence test around that question.
## 1. Engagement is valuable, but it is not a learning outcome
Engagement matters. A learner who refuses to participate cannot benefit from even the best-designed explanation. A game, simulation, animation or interactive tutor can make practice easier to start and easier to sustain. That is not trivial.
The problem begins when engagement becomes the endpoint.
An app can optimise for actions that look educational: taps, streaks, badges, time on task, levels completed, daily logins. These metrics are easy to count. Durable understanding is harder to measure. So products and users can drift toward the visible metric.
Imagine two vocabulary tools. Tool A keeps a student busy for twenty minutes with colourful recognition tasks. Tool B is less entertaining but requires the learner to retrieve the word from memory, use it in a new sentence and distinguish it from a near-synonym. Tool A may win on session time and smiles. Tool B may create stronger learning.
The correct conclusion is not “fun is bad”. It is that **engagement should serve the learning mechanism**.
A useful audit question is: *What is the student doing mentally while the screen is engaging them?*
Are they retrieving? Comparing? Explaining? Predicting? Solving? Generating? Checking? Or are they mainly navigating rewards and reacting to surface cues?
The same visible attention can contain very different cognition.
## 2. Separate four outcomes: engagement, performance, retention and transfer
A school can avoid many EdTech mistakes by naming four separate outcomes.
### Engagement
Will the student begin, attend and continue?
### Performance
Can the student succeed on the task while using the tool?
### Retention
Can the student still recall or execute the relevant knowledge later?
### Transfer
Can the student use the learning in a different context, format or problem?
These outcomes can move together, but they do not have to.
A student might be highly engaged yet learn little. A student might perform well because the interface supplies cues, then forget quickly. A student might remember facts but fail to apply them to a novel problem. A tool might produce lower immediate scores because it makes practice harder, yet create better retention because the learner had to retrieve rather than recognise.
The four-part model is useful because it changes procurement and classroom questions. Instead of asking for “evidence of engagement”, ask which outcome was measured. Instead of accepting an improvement in in-app scores, ask whether a delayed or independent assessment was used. Instead of assuming a mathematics game builds mathematics generally, ask whether students transferred the skill to paper problems with different surface features.
Educational claims become easier to inspect when the outcome is precise.
## 3. The most important test happens after support disappears
Educational technology is often a form of scaffold. It can sequence tasks, highlight relevant information, provide hints, show worked steps, read text aloud, translate language, adapt difficulty or point out errors.
Scaffolds can be extremely useful. But a scaffold has an educational obligation: if independence is the target, support must eventually be reduced or removed.
Suppose a fractions app shades the correct denominator region after every wrong attempt. Students improve quickly. Have they learned fraction comparison, or have they learned how to use the app’s visual cue?
We cannot tell until the cue is gone.
This suggests a simple **support-removal test**:
1. Let the student practise with the tool.
2. Remove the tool or its strongest hint.
3. Give a fresh task that requires the same underlying idea.
4. Observe what survives.
The test is not hostile to technology. It is how we distinguish a useful scaffold from a dependency.
If performance falls slightly but remains substantially better than before, the tool may have helped build skill. If performance collapses to baseline, the interface may have been carrying more of the task than the learner realised.
## 4. Beware the recognition trap
Many digital tasks are built around recognition because it is efficient. Tap the correct answer. Drag the label. Match the pairs. Select the definition. Choose one of four options.
Recognition is a real cognitive ability, but it is easier than generating an answer from memory. A student who recognises “photosynthesis” among four options may not be able to produce the term when asked, explain it in their own words or apply it to a plant investigation.
The recognition trap occurs when repeated success inside the app creates a feeling of mastery that does not survive open recall.
One quick test is to change the response format.
If the app teaches spelling through multiple choice, later ask the student to spell the word without options. If it teaches science concepts through matching, ask for a short explanation. If it teaches equation solving through guided steps, present a clean problem on paper. If it teaches vocabulary through word-definition pairing, ask the learner to use the word precisely in an unfamiliar context.
A strong tool should not merely improve performance in its own interface. It should help the learner function when the interface stops giving clues.
## 5. Productive friction is not a design failure
Digital products are usually designed to reduce friction. Buttons should be obvious. Feedback should be quick. Navigation should be smooth. Those are sensible usability principles.
Learning, however, has a complication: some forms of mental friction are the work.
Retrieving an answer before seeing it is effortful. Explaining why a method works is slower than clicking the method name. Reading a difficult paragraph and building a coherent meaning takes time. Comparing two plausible answers requires more thought than following a highlighted path.
An EdTech tool can therefore become *too helpful*. If it supplies the next step before the learner has tried, reveals the answer after a shallow attempt or turns every complex problem into a sequence of obvious choices, it may improve completion while reducing the cognitive activity that creates learning.
This is especially relevant to generative AI. The answer can arrive before the learner has generated a question, attempted an explanation or retrieved prior knowledge.
The design question is not “How do we make the task effortless?” It is “Which effort is unnecessary, and which effort is educationally valuable?”
Good technology removes logistical friction while preserving intellectual friction.
## 6. Hints should diagnose, not simply rescue
Adaptive learning systems often promise personalised hints. The quality of those hints matters.
A weak hint gives away the next step. A stronger hint helps the learner locate the problem while leaving enough work to do.
Consider an algebra question. A student makes a sign error while moving a term across the equals sign.
A rescue hint says: “Change +5 to -5.”
A diagnostic hint says: “What operation keeps both sides balanced? Write that operation on both sides before simplifying.”
The second hint targets the mathematical structure. It may feel less efficient, but it creates a better chance of transfer.
A useful hint ladder is:
**Prompt attention → remind a principle → suggest a representation → reveal one step → show a worked example.**
The system should begin as lightly as possible and escalate only when needed. If the learner always receives the full route, the app becomes a solution dispenser.
Schools evaluating adaptive platforms should therefore ask not only whether “personalised support” exists but what the support actually does to the learner’s thinking.
## 7. Feedback is only useful if it changes the next attempt
Digital tools can deliver feedback faster than a teacher could manually, which is one of their genuine strengths. But speed alone does not make feedback educational.
“Incorrect” is immediate but thin. A red cross may tell the learner that something went wrong without telling them what kind of repair is needed. At the opposite extreme, a long generated explanation may be so complete that the learner reads it passively and moves on.
Good feedback should answer a practical question: **What should the learner do differently on the next attempt?**
For a grammar task, that might mean identifying the boundary where agreement broke. For mathematics, it might mean checking the base quantity before calculating percentage change. For reading, it could mean returning to the sentence that establishes the pronoun reference. For science, it may ask the learner to distinguish observation from inference.
The strongest systems then give another opportunity to use the feedback. Without a new attempt, feedback remains information rather than learning.
When evaluating a tool, inspect the loop: error → feedback → retry → changed performance. If the loop ends at the explanation screen, the learning opportunity may be incomplete.
## 8. Delayed tests tell us more than immediate post-tests
Many interventions look good immediately after use. The information is still active, the interface is familiar and the learner remembers the procedure.
Education cares about tomorrow, next week and next term.
A delayed test asks whether the learning survived time. This is important because some methods optimise short-term fluency rather than durable memory.
Schools do not need laboratory conditions to use the principle. If a class uses a digital tool for a topic on Monday, include a few clean retrieval items on Thursday. If an app claims mastery of multiplication facts, test a sample a week later without the app. If students use AI feedback to revise a paragraph, give them a fresh paragraph problem later and see whether the improvement reappears.
Delayed checks can be small. Their value is diagnostic.
The question changes from “Did the student finish the unit?” to “What survived the unit?”
## 9. Transfer needs a change in surface form
A student can learn the visual pattern of an app rather than the underlying concept. Transfer tests whether the idea survives a new surface.
For example, a ratio app may always use recipes. A transfer problem might use map scales or speed. A coding platform may teach loops through a game character; transfer could require writing a short program in a text editor. A vocabulary tool may teach a word in one sentence; transfer requires interpreting it in a new subject context.
The change should be meaningful but fair. If the transfer task introduces entirely new prerequisite knowledge, failure tells us little. We want the same deep idea in a different wrapping.
A practical transfer ladder is:
1. Same structure, different numbers.
2. Same structure, different wording.
3. Same concept, different representation.
4. Same principle, different domain or context.
5. Novel problem requiring the learner to select the principle independently.
The farther learning travels, the more confident we can be that the student owns it rather than the interface.
## 10. AI tools require an extra distinction: product quality versus learner quality
Generative AI makes the EdTech evidence problem more urgent because it can improve the quality of an output instantly.
A student writes a weak paragraph. AI rewrites it. The paragraph improves.
What improved: the product or the writer?
Those are separate questions.
The tool may be valuable if the student compares versions, identifies why the revision is stronger, applies the principle to a new paragraph and later writes better independently. It may contribute little to learning if the student accepts the rewrite and submits it.
The same applies in coding, mathematics, science explanation and research.
A useful AI protocol is **generate less, diagnose more**. Instead of asking the system to produce the final answer, ask it to:
– identify ambiguity without rewriting;
– generate a counterexample;
– pose a question that tests the reasoning;
– classify likely error types;
– provide two competing explanations for the student to evaluate;
– give a hint capped at one step;
– check a result after the learner has produced a route.
The technology becomes a training partner rather than a performance substitute.
## 11. Ask what the app is replacing
“Screen time” by itself is often an unhelpful metric. The educational effect of thirty minutes on a tablet depends on what the activity is and what it displaces.
A high-quality interactive reading activity might replace passive waiting and increase useful practice. A low-quality reward loop might replace outdoor play, sustained reading or teacher feedback. A video explanation may help a student revisit a concept; autoplaying short clips may fragment attention.
The APA’s 2026 report makes this replacement principle explicit: the amount of screen time matters less than the activity being replaced.
Schools and families can therefore ask:
**What would the learner be doing if this tool were not here?**
If the answer is “nothing useful”, even modest benefits may matter. If the answer is “a richer discussion with a teacher”, the threshold should be higher.
Technology should be compared with the realistic alternative, not with an imaginary world where every non-digital minute is perfect instruction.
## 12. Independent evidence should outrank marketing metrics
EdTech marketing often presents impressive numbers: millions of questions answered, hours saved, completion rates, satisfaction percentages or gains among active users.
These metrics may be true and still not answer the learning question.
A stronger evidence hierarchy is:
**Level 1: Usage evidence.** Students use the product.
**Level 2: Satisfaction evidence.** Students or teachers like the product.
**Level 3: In-tool performance evidence.** Scores improve inside the product.
**Level 4: Independent learning evidence.** Students improve on an external measure of the intended knowledge or skill.
**Level 5: Durable and transferable evidence.** Gains persist and generalise beyond the original context.
**Level 6: Independent replication.** Results are reproduced by researchers or settings not financially tied to the product.
A product does not need Level 6 evidence before any classroom use. Teachers try new resources all the time. But claims should match evidence. A tool with usage data can claim adoption, not proven learning. A tool with an immediate internal pre-post test can claim something about in-tool improvement, not necessarily durable transfer.
Precision protects schools from buying a learning claim that was never actually measured.
## 13. Run a small local trial before a large commitment
Even strong research evidence does not guarantee the same effect in every school. Implementation matters: device access, teacher training, curriculum alignment, class size, language, age, student needs and how the tool fits into lessons.
Before a large rollout, schools can run a bounded local trial.
Define the learning target. Establish a baseline. Use the tool for a specified period. Track implementation: did students actually use it as intended? Then test retention and transfer with measures that are not produced by the vendor. Compare patterns across learners. Gather teacher observations about workload and instructional usefulness.
Do not change ten things at once. If the trial includes new curriculum, new devices, new grouping, extra teaching time and a new app, it becomes hard to know what caused the result.
The purpose is not to create a perfect randomised controlled trial in every school. It is to replace impression with disciplined observation.
A small pilot can reveal practical problems a glossy demonstration cannot: students gaming the hint system, teachers spending too long on setup, weak learners clicking through, strong learners being held back, or an adaptive sequence drifting away from the school syllabus.
## 14. Measure the learners who disappear inside averages
A tool may help on average while helping some students much more than others. It may also disadvantage a group whose needs are hidden by the overall result.
Schools should therefore inspect subgroup patterns where ethically and statistically sensible. Does the tool work differently for learners with weak reading proficiency? Does heavy visual design overload some pupils? Are accessibility features compatible with screen readers? Does speech recognition handle different accents reliably? Does adaptive difficulty trap a learner at low-level items after an early poor performance?
The question is not only “Does it work?” but “**For whom, under what conditions, and at what cost?**”
This is particularly important when algorithms make placement or recommendation decisions. A student should not become permanently defined by an early data point.
Human review remains important because educational data are incomplete representations of learners.
## 15. Teacher involvement is not evidence that technology failed
Some EdTech narratives imply that a successful tool should remove the need for teacher involvement. That is a poor standard.
A book does not fail because a teacher uses it well. A laboratory does not fail because a teacher frames the experiment. Likewise, technology can be most effective as part of a teaching system.
Teachers decide when to introduce a tool, how to connect it to prior knowledge, which errors matter, when to pause, what discussion should follow, and when support should be removed. They also notice affective and social information that a platform may miss: fatigue, embarrassment, avoidance, overconfidence, copying or confusion hidden behind correct clicks.
The APA’s 2026 recommendations explicitly preserve a significant role for adult involvement. That should be treated as design reality, not inefficiency.
The better question is: does the technology amplify valuable teacher work, or merely add another dashboard to manage?
## 16. Beware dashboard certainty
Dashboards are persuasive because they turn messy learning into clean numbers. Mastery 78%. Engagement high. Topic red. Topic green.
These signals can help. They can also create false precision.
A mastery score is produced by a model with assumptions. It may reflect question difficulty, response history, hint use, time, confidence estimates or proprietary algorithms. Teachers should know enough to ask what the number means.
Does “mastered” mean correct once, correct repeatedly, correct after delay, correct without hints or predicted probability of future success? Are easy and hard items weighted differently? Does the score reset after forgetting? Can a learner reach mastery through repeated guessing?
If nobody can explain the metric, it should not make high-stakes decisions by itself.
Use dashboards as signals for investigation. “This student appears weak in fractions” should lead to a short human diagnostic, not an automatic identity label.
## 17. A practical EdTech evidence test for teachers and parents
Before adopting or trusting a tool, ask ten questions.
### 1. What exact capability is this supposed to improve?
Avoid “critical thinking” or “brain training” as vague claims. Name the knowledge or skill.
### 2. What does the learner actually do?
Describe the mental activity, not the interface feature.
### 3. What help does the tool provide?
Hints, cues, multiple-choice options, worked steps and AI generation may inflate in-tool performance.
### 4. Is there a clean attempt without the strongest support?
Independence should be tested if independence is the target.
### 5. Is learning checked after a delay?
Immediate success is not enough for durable learning claims.
### 6. Is transfer checked in a new format or context?
The learner should not need the original interface.
### 7. What evidence is independent of the company?
Vendor research can be useful, but independent evaluation adds confidence.
### 8. What does the tool replace?
Compare against the realistic alternative.
### 9. Which learners benefit less or experience new barriers?
Averages can hide important variation.
### 10. Can the teacher explain when to stop using it?
A tool without an exit condition can create dependency.
If a product cannot answer all ten, that does not automatically make it bad. But the unanswered questions tell you how cautious the claim should be.
## 18. A classroom protocol: Use, remove, vary, revisit
Teachers need something simpler than a research report. One four-step cycle can reveal a lot.
**Use:** Let students practise with the technology under normal conditions.
**Remove:** Give a short independent item without the tool or its major cues.
**Vary:** Change the surface form, representation or context.
**Revisit:** Test a small sample again after time has passed.
This takes the four outcomes seriously: performance, independence, transfer and retention.
Suppose a Primary 5 class uses an app to learn percentage change. After practice, students solve one paper problem with no hints. Then they solve a differently worded scenario where the base quantity is less obvious. Three days later, the teacher includes another item in a warm-up.
That sequence provides more information than a dashboard’s 90 per cent mastery label.
It also sends a message to students: learning is what you can carry with you.
## 19. Good EdTech should make itself less necessary
This is a demanding principle, but it is worth keeping.
If the learning target is independent capability, successful technology should gradually reduce the amount of support a learner needs. The student may continue to use the tool for efficiency, extension or practice, but the core skill should become more portable.
A reading app succeeds when the child reads better beyond the app. A mathematics tutor succeeds when the learner can solve and explain mathematics without its prompts. A writing assistant succeeds educationally when better decisions appear in fresh writing, not only in AI-revised drafts.
This does not mean every digital tool should disappear. Calculators, spreadsheets, coding environments and reference tools can become legitimate parts of expert practice. The question is whether the student understands the underlying capability expected at their stage.
Experts use tools differently from novices because they can often detect when the tool is wrong.
Education should build that independence before dependence becomes invisible.
## 20. What high-quality digital learning often has in common
No single feature guarantees learning, but stronger tools often share several characteristics.
They keep the learning objective visible.
They require meaningful learner responses rather than constant recognition.
They provide feedback that points toward repair.
They allow retrieval and reattempts.
They space or revisit important knowledge.
They vary examples enough to discourage surface memorisation.
They do not drown essential information in decorative stimulation.
They give teachers usable evidence without pretending the dashboard is the whole learner.
They support accessibility without unnecessarily lowering the intellectual target.
They make the learner explain, choose, predict or generate when those actions matter.
And they make it possible to see whether learning survives away from the screen.
These features do not make technology magical. They align technology with principles of learning.
## 21. The evidence test should include cost, privacy and opportunity cost
A tool can produce a learning benefit and still be the wrong choice for a particular school. Educational decisions are never made in a vacuum. Money, teacher time, device management, technical support, student data and timetable space all have alternative uses.
This is why the EdTech evidence test should end with three system questions.
**What does the gain cost?** A small improvement may be worthwhile for a low-cost tool that fits existing routines. The same improvement may be unattractive if it requires expensive licences, new hardware, extensive training and constant troubleshooting. Cost-effectiveness is not the same as effectiveness.
**What data are collected to produce the gain?** A platform may record names, clicks, response times, writing samples, voice, location, behavioural patterns or inferred ability. Schools need to know what is necessary, how long information is retained, who can access it, whether it trains commercial models, and what happens when a family leaves the platform. “Personalisation” should not become a blank cheque for unnecessary data collection.
**What better activity might be displaced?** Ten minutes of adaptive practice may be valuable. Forty minutes may crowd out teacher explanation, collaborative problem solving, extended reading, handwriting, practical work, physical movement or a restorative break. The opportunity cost can change by age, subject and learner.
These questions protect against a common technology bias: evaluating a product only by what it adds. Every intervention also consumes attention, time and resources.
A simple procurement sheet can therefore include five columns: intended learning outcome, independent evidence, implementation burden, data risk, and displaced activity. The final decision becomes a comparison rather than a product demonstration.
This also encourages an important habit: define an exit condition before adoption. What result would tell the school to continue, modify or stop the programme? Perhaps independent mathematics scores should improve by a specified amount, teacher workload should remain within a limit, or the platform must demonstrate that hint dependence falls over time. If no success criterion exists, weak tools can continue indefinitely because they have become part of the furniture.
Technology deserves neither automatic suspicion nor automatic enthusiasm. It deserves the same disciplined question we ask of any educational intervention: compared with the realistic alternative, does it produce enough durable learning to justify the costs and risks?
## 22. Three common demonstrations that look impressive but prove less than they seem
A disciplined evidence test becomes easier when we recognise three recurring demonstration patterns.
**The perfect live demo.** A vendor shows a student moving smoothly through a personalised sequence. The system identifies an error, gives a hint and produces a correct answer. This proves that the interface can perform the sequence under demonstration conditions. It does not yet prove that learners retain the skill, that the algorithm diagnoses errors accurately across a diverse population, or that the tool works when teachers have thirty students and unreliable Wi-Fi. Ask for independent outcomes and implementation evidence.
**The dramatic before-and-after product.** A weak paragraph becomes polished after AI feedback. A rough slide deck becomes professional. A coding error disappears. The difference is visible, so the improvement feels educational. But product improvement is not the same as learner improvement. Ask the student to explain the changes, reproduce the principle on a new task and perform part of the work without assistance. If the improvement transfers, the demonstration becomes more convincing.
**The rising mastery dashboard.** Students begin at 43 per cent and end at 91 per cent. The graph is compelling. But what generated the score? Were questions repeated? Did hint use count? Did difficulty adapt downward? Was the final assessment independent of the training items? Was there a delayed check? Without these details, the graph may describe progress through a platform rather than progress in the underlying subject.
None of these demonstrations should be dismissed automatically. They are useful starting evidence. The error is treating them as the end of the argument.
A strong school response is curious rather than cynical. “Show us what students can do later, elsewhere and without the strongest support.” A strong developer should welcome that question because durable transfer is exactly what an educational product ought to help create.
The same standard should apply to teachers’ favourite tools. Familiarity can create its own marketing effect. If a platform has been used for years, that history does not remove the need to ask what students are learning from it.
## 23. Teacher workload belongs inside the learning calculation
A digital tool can save student time while quietly consuming teacher time. Accounts need creating, classes need syncing, forgotten passwords need resolving, dashboards need interpreting, assignments need configuring, technical problems need troubleshooting and data need exporting. If those costs are ignored, a product can look efficient only because the labour has moved.
Schools should therefore measure implementation burden during pilots. How many minutes does setup take each week? How often does a lesson stall because the platform fails? Does the dashboard reduce marking, or does it add another stream of information that teachers must reconcile with exercise books, classroom observations and assessments? Can teachers act on the data, or are they simply collecting it?
The same principle applies to student attention. Every new platform has an interface to learn. Tool-switching has a cognitive cost. If pupils move from textbook to browser to quiz app to shared document to AI assistant in one lesson, the technology stack itself may become part of the workload.
A high-value tool often does one job exceptionally well and fits the existing learning sequence. It may be less visually impressive than an all-in-one platform, but the teacher can explain exactly why it is there.
This is another reason local trials matter. Independent research can estimate learning effects; only the school can see whether the product fits its timetable, staff capacity, devices and curriculum. Implementation is not an administrative detail added after effectiveness. It is part of whether effectiveness can appear at all.
## Conclusion: ask what travels beyond the screen
Educational technology can be excellent. It can expand access, accelerate feedback, support practice, personalise pacing, simulate environments, reduce administrative friction and give learners opportunities that would otherwise be difficult or expensive.
But the glow of the screen can make weak evidence look strong.
A busy student is not automatically learning. A high in-app score is not automatically durable knowledge. A polished AI-assisted product is not automatically stronger human capability. A mastery dashboard is not automatically mastery.
The test is what travels.
Does the knowledge survive time? Does the skill survive the removal of hints? Does the concept survive a new representation? Does the learner make better decisions when the app is absent?
That is the EdTech evidence test.
The most important metric may be the one the product cannot display while the student is using it: **what the learner can still do after the product is closed.**
—
## Related eduKateSG reading
– **How Education Technology Makes Education Better, Or Does It?**
How Education Technology Makes Education Better, Or Does It?
– **How Student Engagement Works | Attention Is Not Enough**
How Student Engagement Works | Attention Is Not Enough
– **How Studying Works | Learning Conversion Efficiency — How Much Study Actually Becomes Independent Capability**
How Studying Works | Learning Conversion Efficiency — How Much Study Actually Becomes Independent Capability
– **When Does Classroom Technology Cost More Attention Than It Saves? | How EdTech Tool Friction Works**
When Does Classroom Technology Cost More Attention Than It Saves? | How EdTech Tool Friction Works
## Sources and further reading
– American Psychological Association (3 September 2026), *With education technology, engagement is not the same as learning*.
https://www.apa.org/news/press/releases/2026/09/education-technology-engagement
– American Psychological Association (2026), *Children’s and adolescents’ learning with educational technology*.
https://www.apa.org/pubs/reports/children-adolescent-learning-educational-technology
– OECD (19 January 2026), *OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education*.
https://www.oecd.org/en/publications/oecd-digital-education-outlook-2026_062a7394-en.html
– American Psychological Association (1 April 2025), *Psychological science helps build digital educational media that capitalizes on how human brains learn*.
https://www.apa.org/monitor/2025/04-05/kids-digital-educational-media
