糖衣本身就在说：药是苦的

The Sugar Coating Says the Pill Is Bitter

The Honey on the Rim#

Lucretius, in the first book of De Rerum Natura​, explains why he sets out his doctrine in verse:

But as physicians, when they seek to give Young boys the nauseous wormwood, first do touch The brim around the cup with the sweet juice …1

The comparison is to his own poem. His doctrine is the wormwood and his verse is the honey. Lucretius does not hide the bitterness; he states it in the same passage that offers the sweetener. A coating is a confession: whoever sweetens the rim has already decided what the drinker will taste.

Horace supplied the respectable descendant. Omne tulit punctum qui miscuit utile dulci​: in a plain rendering, he has won every point who has mixed the useful with the sweet.2 The line is about how a poem wins its adult readers, some of whom dismiss what lacks profit and some of whom pass over austere poems. It is not a method for teaching children. The line's modern descendant is edutainment, teaching wrapped in entertainment,3 and gamification, the subject of this article, is edutainment's industrial form.

Gamification's own literature defines it. Sailer and Homner's 2020 meta-analysis takes gamification to be "the use of game design elements in non-game contexts" and describes gamified learning as "augmenting or altering an existing learning process."4 By that definition the subject comes first and the game is added on top. That is a coating.

This article takes a position in two parts. First, learning has a pleasure of its own when difficulty and the sense of progress are right, so the coating is superfluous; that part rests on evidence gathered near the end. Second, the coating misleads: it tells the learner that the pill is bitter. Gamification says it with points, streaks, ranks, bars and locks, to every learner who opens the app. The second part is my judgment, and no study measures it. Each section below ends on what its device makes of learning; those lines are my reading of the device, not a finding. I find the second part the more frightening, and the sections below try to earn that word.

The image is old and not mine. Dewey used it in 1902,5 and Kohn, in a 2018 piece for the New York Times (quoted here from the expanded version on his site), wrote that many teachers, parents and managers persist in using versions of what has been called "sugar-coated control."6

Psychologists who study rewards sort them by what they are paid for: engaging in a task, finishing it, or doing it well. Deci, Koestner and Ryan's meta-analysis of 128 experiments uses those three classes.7 Applied to gamification's devices, the same question becomes a field guide: what is this one paying for, and what does paying for that teach? The devices are ordered by how deep the answer reaches, from what the learner does, to why, to against whom, to how learning feels, to what the subject itself is.

The Streak: Paid for Showing Up#

In its basic form a streak is a counter of consecutive days that returns to zero after one missed day. It does not count what was understood, what was hard, or whether anything was learned. It counts that a lesson was finished, and in the company's own write-up of a 2020 test, one lesson a day was enough to extend the streak.8

Read as a reward (here and in each device below, the mapping of the meta-analysis's reward classes onto game devices is mine), a streak is completion-contingent. In Deci and colleagues' meta-analysis, that class lowered later free-choice interest: on initially interesting tasks, measured after the reward was gone, d = -0.36, against -0.40 for rewards merely for engaging in a task and -0.28 for rewards for performing well.7

Experiments on streaks show what the count does once it exists. Silverman and Barasch built a mock language app "modeled after Duolingo" (Study 3, N = 602). After a few Portuguese questions, participants chose between continuing and switching to Hawaiian. With a streak log on screen, 92.47% continued after an intact streak and 45.21% after a broken one; with no log, the figures were 64.94% and 60.90%.9 A broken count pushed continuation below where it stood with no count at all.

Figure: Share of participants who chose to continue with Portuguese rather than switch to Hawaiian. Study 3, a mock language app "modeled after Duolingo," online participants, streaks three or four questions long. No learning was measured.

The authors report that the effect holds "independent of actual past behavior," and read it as showing that people "consider maintaining a logged streak to be a meaningful goal in and of itself."9 The step from this to a learner defending day thirty is mine. The participants were recruited online, the streaks were three or four questions long, and no learning was measured.

The product's maker names the mechanism in its own words. A Duolingo blog post of 31 January 2022 reads: "As you grow your streak to new heights, the novelty of extending your streak may begin to wear, so the feature begins to motivate you in a different way. This is when we tap into 'loss aversion,' an internal bias in your brain that makes you particularly averse to losing something, like a learning streak."10

The Streak Freeze, which pauses the count for a day, is the company's answer, and allowing a second one, the post says, "actually increased the relative number of active learners on Duolingo every day by +0.38%."10 The figure is the company's, and the outcome it reports is active learners. The remedy for a lost streak has also been for sale: the company's registration statement of June 2021 lists "streak freezes" among the in-app purchases that made up about 1% of its 2020 revenue.11

The company also says what it counts as success. In the earnings release of 10 November 2022 its co-founder and CEO is quoted: "I'm particularly proud that our products are encouraging more engagement, which can lead to better learning outcomes."12 Learning is what engagement "can lead to."

The one randomized field experiment on streaks in a learning setting deserves its numbers in full. Aulagnon, Cristia, Cueto and Malamud call theirs, to their knowledge, "the first experimental study examining the effects of highlighting streaks in an educational context."13 In the working-paper version, 60,000 Peruvian children in grades four to six who had used a free maths platform were randomly assigned, for six summer weeks, to one of four groups: messages highlighting their streaks (defined as consecutive weeks of completed problem sets), personalized reminders with positive reinforcement, generic reminders, or no messages. In the streak group use rose: the chance of connecting at all by 2.8 percentage points, and the share of weeks connected, among those who connected, by 9.4 points. Among the 1,503 children who took an endline test, the streak group scored 0.13 to 0.17 standard deviations above control. The authors: "We do not observe a discouragement effect from highlighting streaks."13

The streak bought what a streak pays for, which is showing up, and against a personalized reminder the point estimates split. The personalized-reminder arm, whose messages were not about streaks, had the higher estimate for getting children to connect at all (+3.8 percentage points, against +2.8 for the streak messages); the streak arm's was higher for the share of weeks connected (+9.4 points, against +6.9). Because both arms received two messages a week, those contrasts carry no message-count confound. The streak's edge in use over generic reminders does: two messages a week for the streak arm, one for the generic arm.

The test was aligned with the platform's own exercises and, in the authors' words, "does not represent a broad-based measure of academic achievement."13 Only 1,503 of the 60,000 took it, the differences between treatment arms on the test were not significant, and the authors add that the effect on learning for the overall population would be much smaller. No outcome was interest or belief, and the streaks in the study were weekly and delivered as messages: a notification about a weekly streak, not a daily counter. The authors' own summary names the prize: highlighting streaks "is most effective when the objective is to generate greater attachment among individuals who are already engaged in a desired behavior."13 In my reading that is attachment to the platform, scored by the platform's own test.

The latest company figure shows what a day's learning becomes. In June 2026, in what Duolingo's shareholder letter calls "a one-time event," eligible learners could "restore their longest-ever streak by opting in and completing three lessons"; by the company's count, "15.4 million learners revived their streaks."14 The day's learning is done when the count moves, and the count is what is owed: an obligation to a counter.

Points: Meaning Converted into Currency#

Points, usually called XP, are awarded per lesson, per correct answer, per unit finished, and they buy levels. One currency covers everything. A grammar lesson and a vocabulary drill become comparable by what they pay, and, in my reading, a task that pays nothing stops registering as a task.

The experiment that shows the economy in miniature is older than any app. Lepper, Greene and Nisbett (1973) took fifty-one preschoolers who already liked drawing with markers and assigned them to three groups: one expected an award for drawing, one received the same award unexpectedly, and one received none. Later, in free play with no award in sight, the children who had expected the award spent 8.59% of their time drawing, against 16.73% for the no-award group and 18.09% for the unexpected-award group. The pictures they drew for the award, in the session itself, were rated lower than either other group's: 2.18, against 2.85 for the unexpected-award group and 2.69 for the no-award group.15 The unexpected award did no harm, so the mechanism is expectation plus contingency. And the children who expected the award did not only draw less; they drew worse. Points in a learning app are expected and contingent by design. The experiment used a certificate for drawing, not points for a lesson. Points belong to the same class the streak did, rewards for finishing, with the same direction on later free-choice interest.7 The standard objection is that points are feedback, and the same meta-analysis found that positive verbal feedback raised free-choice behavior (d = 0.33). A point is not that feedback. Feedback says what was done well; a point says what it was worth. The closer cousin is Lepper's certificate.

Asked whether rewards motivate, Kohn answered in his 1994 summary of the research: "Absolutely: they motivate students to get rewards."16

What learning becomes under a points economy is labor paid in currency. The learner's question moves from "what is this?" to "what is this worth?" A currency also invites comparison: who has more?

The Leaderboard: Learning Converted into Rank#

Badges certify milestones. A leaderboard ranks learners. What it displays is position, and a ranked board has a property that needs no study: most of its learners are below the top by construction, and one learner's rise is another's fall. Each time the board is consulted, it tells most of the people on it that others are ahead of them. That last step is my reading of the device. The arithmetic is not.

The semester-long field study of a badge-and-leaderboard course on which this section rests is Hanus and Fox (2015), reported here as the authors report it. Two sections of one elective communication course at a large Midwestern university, taught by one of the authors, were compared over a semester. One section had a system of 22 badges, required as part of the grade, and a leaderboard visible to the class under pseudonyms that ordered students by badge count; the authors write that "all of our game mechanics were aligned with learning objectives."17 On seven-point scales, intrinsic motivation in the gamified section ran 4.74, 4.57, 4.66 across the three surveys after the badge system began (Times 1 to 3), against 4.78, 4.92, 4.95 in the other section. Satisfaction ran 4.10, 3.51, 3.57, against 4.18, 4.15, 4.00. On the final exam there was no significant difference. The only exam link the authors found ran through lower intrinsic motivation, and they read it as harm. The sections were intact classes, not randomized, and 71 students completed all four surveys.17

The authors offer their own account of why. "When a reward system is imposed on top of a class students already find interesting, it may feel constraining and forced."17 That is their suggestion, not a tested mechanism, and the design cannot separate the badges from the leaderboard, or either from differences between the two sections.

The meta-analysis offers the nearest measured pattern: its class of rewards for performing well lowered later free-choice interest, though not self-reported interest.7 Hanus and Fox measured social comparison and found that the sections differed only marginally, at the last survey (Time 3, p = .06),17 so the study does not show the board working through comparison. The case for that mechanism is the arithmetic above; the case for the harm is the motivation curve.

What learning becomes here is a contest.

Ranks and points measure the learner against something, a price or a rival. The progress bar measures the learner against nothing at all, and it feels like learning.

The Progress Bar: Ease Mistaken for Learning#

Levels, rings and bars fill with every completed item. The bars in a lesson path only rise.

Learning scientists draw a line that this device erases. Soderstrom and Bjork's review says that performance during instruction "is often an unreliable index of whether the relatively long-term changes that constitute learning have taken place," and that people "often mistakenly interpret their performance during acquisition as a reliable guide to long-term learning."18 A progress bar draws a count of items completed, which is less even than performance during acquisition. Every bridge in this section from a study to the device is mine.

The feeling can run backwards. In one calculus-based introductory mechanics course at Harvard, in a randomized crossover design (N = 149), students scored 0.46 SD higher on a test of learning after active classes and rated their own feeling of learning 0.56 SD lower than after passive lectures. The authors write that students "initially take that effort to signify poorer learning."19 The study concerns lectures and college physics, not gamification. The misreading is already there without any bar. A bar that rises smoothly feeds it: it confirms, at every step, that real learning should feel like the bar.

The adults read it the same way. Infants aged 12 to 18 months watched a popular vocabulary DVD at home several times a week for about four weeks and learned no more words than controls; "parents who liked the DVD tended to overestimate how much their children had learned from it."20 The overestimate is the parents', not the infants'. Liking the product was read as learning. The DVD had no progress bar.

The point is not that learning should be made hard. A difficulty helps only when the learner can meet it.

What learning becomes under a bar is a bar that fills. Confusion, the ordinary feel of understanding being built, now reads as malfunction, because the interface has taught the learner what progress looks like.

The Unlock: The Lesson as Toll#

The fifth device gates the game behind the content: answer the questions to unlock the next level, finish the quiz to get back to the game, play the minigame after the worksheet. The content is positioned as the price of admission.

Habgood and Ainsworth built the cleanest test of this arrangement. They made two versions of the same maths game, Zombie Division, for children of about seven to eleven. In the extrinsic version the game ran as normal and division appeared as a quiz at the end of each level. The researchers' own description of that tradition, in their published abstract, is games used "as a sugar coating for learning content."21

The interview table in the study records this remark about the extrinsic version, from children who had played both: "you think: oh I've had the fun part, now I have to do a test – I'm just going to turn it off and not bother."21

The design announced which part was the pill. These are sixteen children in one group interview, with quotes selected for a table; the remark shows what children said, not what learners in general conclude. The same table records children imagining that teachers would object to the intrinsic version as "too much fun – and hasn't got a test."21 What it shows, in my reading, is that the children expect adults to hold that learning must contain a bitter part.

Dewey described the method in 1902. Leave the subject as it is, he wrote, and "by trick of method" arouse interest: "to cover it with sugar-coating," and so, "as it were, to get the child to swallow and digest the unpalatable morsel while he is enjoying tasting something quite different." He called this the "easier and simpler" course.5 The unlock is that sentence turned into a game loop.

Sweetness that is not the subject also has a measured price, though a small one. In a multilevel meta-analysis of 177 effect sizes from 50 studies, interesting but irrelevant additions to lessons, called seductive details, lowered learning by g = -0.16, working "primarily by increasing extraneous cognitive load."22 Those studies concern irrelevant material inside a lesson, not a game wrapped around one, and the bridge to the minigame is mine.

What learning becomes behind a lock is the price of the fun. The lock trains the learner to wait for the sugar: the subject turns into the interval between rewards.

What the Best Numbers Measure#

The meta-analytic case for gamification starts with Sailer and Homner's 2020 synthesis. Across all the studies it includes, the estimates are as follows. For cognitive outcomes, g = .49 (95% CI 0.30 to 0.69; k = 19 study samples, N = 1,686). For motivational outcomes, g = .36 (0.18 to 0.54; k = 16 study samples, N = 2,246). For behavioral outcomes, g = .25 (0.04 to 0.46; k = 9 study samples, N = 951). The synthesis also examined a subsplit of studies with high methodological rigor: experimental designs, or quasi-experimental designs with pre- and posttests. There the cognitive effect stayed significant (g = .42, p < .01; k = 9 study samples). The motivational effect did not (g = .22, p = .20; k = 7 study samples), and neither did the behavioral (g = .27, p = .22; k = 5 study samples).4 The intervals around the two nonsignificant estimates are wide enough to include the full-sample values. The authors write that the factors behind success "are still somewhat unresolved, especially for cognitive learning outcomes."4

Figure: Sailer and Homner's meta-analysis, effect size g. The high-rigor subsplit is experimental designs, or quasi-experimental designs with pre- and posttests. Rounded node: significant. Hexagon: not significant. The authors call the subsplits underpowered (k = 7 motivational, k = 5 behavioral, where k is the number of study samples), so a non-significant result here means the effect could not be established, not that it was measured as zero.

The numbers are accurate.

First, the effect that holds under rigor is the one a coating predicts. A sweetened rim gets the cup drunk. In a classroom study that followed scores across three tests, 473 university students prepared with ordinary quizzes or with gamified ones ("a wager option, a progress bar, encouraging messages"). The authors report in their abstract that the gamified students "had significantly better scores on the first test" and that "the beneficial effect of gamification did not persist for subsequent tests"; they read this as gamification that "might work through a novelty effect."23 It is evidence that the pill went down once. It is silent on what the learner thinks the pill is.

Second, g = .49 is the effect of adding a layer to an existing learning process, which is the definition quoted at the start of this article. It is not a comparison between a layer and a subject that is itself the action. Where that comparison was run, in the randomized Zombie Division study with seven- and eight-year-olds, the delayed test gave 41% for the coated (extrinsic) version, 38% for the control version with the maths removed from the game, and 58% for the version in which the maths was the play. At the immediate post-test the groups did not differ.21 The paper reports the intrinsic version ahead of the other two; it does not report the three-point gap between the coated version and the game with no maths as a difference, so against a game with no maths in it at all the coating showed no gain the paper could report. The version in which the maths was the play stood 17 to 20 points ahead of both. That is one school and 51 children analyzed.

Figure: Delayed test, two weeks after the post-test. Seven- and eight-year-olds, randomly assigned; 51 children analyzed, one school. "Coated" is the extrinsic version (division as a quiz at the end of each level); "maths removed" is the control; "maths is the play" is the intrinsic version. At the immediate post-test the groups did not differ. The paper does not report the 41% versus 38% gap as a difference.

I also read two studies together here, and the reading is mine: a gamified lift present at the first test and gone by the later ones (Sanchez and colleagues),23 and an intrinsic advantage absent at once and present two weeks later (Habgood and Ainsworth). A score taken at the moment of the game says little about what stays, which is the learning-versus-performance distinction of the progress-bar section working in both directions.18 Sailer and Homner's own coding bears on the same point. Cognitive and motivational outcomes were "almost exclusively measured after interventions," whereas behavioral outcomes were "almost exclusively measured during interventions (i.e., the measurement was to some extent confounded with the intervention itself)."4 The authors' hedged reading of their moderator results, which were mostly found for behavioral outcomes, is that certain variables may affect behavior in the immediate situation, which does not necessarily transfer outside the gamified context.4

Third, the selling point is the unstable part. Gamification is sold on motivation and engagement, and the authors' own verdict on their high-rigor subsplit is that the summary effects on motivational and behavioral outcomes "are not robust according to the subsplit analyses."4 They give it with the caveat that the subsplits were underpowered (k = 7 and k = 5), so the verdict is that these effects could not be established in the rigorous studies, not a measurement of no effect. Rodrigues and colleagues show what the instability looks like over time. They compared gamified cohorts of an introductory programming course (2017 and 2018) with a non-gamified cohort from 2016: 756 undergraduates in all. In logged behavior the effect "started to decrease after four weeks," turned upward "without any intervention," and "did not restore the positive effect to the level of the initial contribution."24 The authors' reading: "gamification likely suffers from the novelty effect, but also benefits from the familiarization effect."24 The U shape is their exploratory, visual reading. What was logged was attempts, minutes in the coding environment, and system access: activity in the system, not learning. The game was never taken away.

Fourth, the measure that matters for the charge is what remains once the rewards stop. The experiments that measure it are the free-choice experiments, Lepper's drawing study among them,15 and the meta-analysis pools the pattern.7 Gamification's own removal literature is small and measures something else. Seaborn's rapid review found eight removal studies from 2012 to 2020: "Findings on gamification removal were mixed but lean negative. Most confirm earlier fears, with even initial positive results declining on removal. Given the small pool and noted issues, firm conclusions cannot be drawn."25 Six of the eight, she reports, focused on "measures related to evaluating the worth of gamification for the platform creators or purveyors."25

One study did survey students after gamification stopped, and its result cuts against the expected story. García Iruela and colleagues ran a four-week course, gamified for two weeks and then not, against the reverse order, and concluded that "no significant differences were found in the 4 aspects analyzed when students stop gamifying."26 The four aspects were pressure, perceived choice, perceived competence, and effort, asked two weeks after the switch. Of the 190 registered students, 113 completed the second survey. The authors attribute what they saw to "the loss of the novelty effect" of the course itself.26 I found no study that asks what a learner believes learning is once the points stop.

Figure: The three studies in this article that report a time course. The final box is the author's report of the search, not a finding of any of the studies.

Read on its own terms, gamification's best evidence is an indictment. It shows that the pill can be made to go down. Only studies that follow the same learners through a course can show what happens to the sweetness over time, and the three here that report a time course of a gamification effect (Sanchez's three tests, Rodrigues's fourteen weeks, Hanus and Fox's semester) record the lift not persisting past the first test (as Sanchez's abstract reports it), the effect never regaining its starting level, and motivation slipping behind the other section's, the last two with the game still running. And the evidence leaves unmeasured the learner facing a cup without honey.

If the coating does not produce the wanting, something else does. That is the cell the field guide leaves out.

Where the Sweetness Cannot Be Peeled Off#

One test separates a coating from a pleasure that belongs to the subject: can the sweetness be peeled off without changing what is learned? The distinction is not new. Malone drew it in 1981, between fantasies that "depend only weakly on the skill used in a game" and fantasies "intimately related to the use of the skill."27

Habgood and Ainsworth's other version of Zombie Division passes the test. In the intrinsic version, division is how the player fights the skeletons: there is no quiz, and the maths is the move. Given free choice between the two versions over about 135 minutes of available play, 16 children aged 9 to 11, the same children whose interview is quoted above, spent 75.7 minutes on the intrinsic version and 10.28 on the extrinsic one, over seven times as long.21 Strip the skeletons' costumes and a game is still left, because the move the player makes, moment to moment, is the division. What cannot be peeled is the play.

The same table holds a remark about the intrinsic version: "it's like mixing paint [...] the maths in the game with the fun [...] you don't really think you're doing that much."21 Paint, once mixed, cannot be unmixed. That is the peel test in a child's words.

Papert quotes a child who described programming as "It's fun. It's hard. It's Logo." His comment: "I have no doubt that this kid called the work fun because it was hard rather than in spite of being hard." His proviso is that the challenges "have to be the right things matched to the individual."28

Willingham, a cognitive scientist, gives the condition. "Solving problems brings pleasure," and "the pleasure is in the solving of the problem." Problems at the right level of difficulty are rewarding, while problems "too easy or too difficult" are unpleasant.29 That is the true form of the first leg of this argument: the subject offers pleasure when difficulty and progress are right. Getting them right is a teaching problem, and no coating solves it.

Chinese readers inherit the same ranking from the Analects: knowing it ranks below liking it, and liking it below delighting in it (6.20 in the modern numbering, 6.18 in Legge).30 The object of the three verbs is left implicit, so reading it as learning is an interpretive step, and I take it.

The consequence for gamification comes from the same meta-analysis that supplied the reward numbers. Among the 11 studies with a free-choice measure that varied how interesting the task was, rewards lowered later free-choice interest on interesting tasks (d = -0.68) and had no significant effect on dull ones (d = +0.18, not significant). The authors conclude that the undermining by tangible rewards "does not extend to dull-boring tasks."7 Read both halves. Where the subject already delights, the reward spends the delight; where it does not, the reward creates none that the analysis could detect. Hanus and Fox drew the same line about their own course. Their students were in an elective, "presumably of their own free will and because it was somewhat interesting to them," and for bored students they allow that rewards "might make the material more engaging."17

The standard reply is to gamify only the dull parts: a reward need not create interest in a dull task, only get the practice done. On a dull task, then, the coating buys compliance, practice delivered at a price per unit, which is the strict camp's goal reached by a sweeter road. And "dull" is seldom a fact about the subject. In my reading, a part that bores is more often a part that is badly calibrated in the sense Willingham gives,29 and the coating paints over the calibration and certifies the failure as a property of the subject. It also accepts the coating's premise that the part is dull. The scope finding fits that reading: an announcement that a task is a chore can change the mind only of someone who thought it a pleasure, and the measured damage falls on interesting tasks.

None of this rehabilitates the game layer. The intrinsic version is not gamification succeeding; it is the subject being the action, and the test excludes it from the field guide.

If the pleasure was in the cup all along, who decided it was wormwood?

The Same Sentence, Twice#

Every device in the field guide says one sentence about the subject: it would not be chosen without help. The streak says it with a count, points with a price, the board with a rank, the bar with a fill, the lock with a gate. That is what the coating teaches. I state it as my judgment: no study measures what learners conclude about learning. What frightens me is the dose. The sentence is said to anyone who opens such an app, as often as it is opened, in the same motion as the reward, and the company reports in its own count that "Over 20% of our DAUs now have streaks longer than a year" (release of 7 August 2024).31

Figure: The reward classes are Deci, Koestner and Ryan's;7 applying the question to each device is the author's. Left-hand boxes describe the devices as defined in the sections above. The right-hand column is the author's reading of each device. No study measures what learners conclude about learning.

A judgment that no study measures can still stake a prediction. The test would measure how much free time learners give the subject outside the app in the month after a streak ends, against learners who never held one. I predict that the first group gives less.

The strict tradition says the same sentence with a whip instead of honey. The Three Character Classic​, the primer that Chinese children recite, holds up students who tied their hair to a beam or pricked a thigh with an awl: "One hung his head from a beam, one pricked his thigh with an awl; unbidden, they toiled hard of their own accord."32 Those lines praise effort taken up unprompted, and it would be a caricature to hear a cruelty manual in them. The primer closes on the other half: "Diligence earns merit; play brings no benefit."32 Diogenes Laertius reports Aristotle saying: "The roots of education, he said, are bitter, but the fruit is sweet."33 The pairing is my reading of two texts. The strict camp puts the sweetness at the end of the labor; the coating puts it on the rim; both agree that the root is bitter. Rejecting the coating is therefore not a vote for the awl.

Dewey saw the shared premise in 1913, in Interest and Effort in Education​, and this article applies his diagnosis without having discovered it. "It is not unusual, though somewhat surprising, that there is generally a common principle unconsciously assumed at the basis of two theories which to all outward appearances are the extreme opposites of each other."34 The principle he named was "the externality of the object, idea, or end to be mastered to the self." Because the object is assumed to lie outside the self, it "has to be made interesting; to be surrounded with artificial stimuli and with fictitious inducements to attention," or else "the sheer power of 'will'" must be summoned.34

Inside his critique of the interest theory's excess, Dewey wrote of the sugar-coated child that "he soon learns to turn from everything that is not artificially surrounded with diverting circumstances."34 His way out was identification: once the learner is joined to the subject, "we have neither to appeal to sheer strength of will, nor to occupy ourselves with making things interesting."34 What I add to Dewey is one sentence. The coating is not only a habit-former; it is an announcement.

Figure: The shared premise and the quotations from both sides and the way out are Dewey's (1913). The pairing with Aristotle's saying, and the closing remark that rejecting the coating is not a vote for the awl, are the author's.

Postman made the attitude point about television in 1985. "We now know that Sesame Street encourages children to love school only if school is like Sesame Street​."35 A few sentences later: "As a television show, and a good one, Sesame Street does not encourage children to love school or anything about school. It encourages them to love television."35 That is an observation, not a measurement, and Postman reports no effect size. Gamification extends the line by my reading: the learner can come to love the streak.

The industry has had Dewey's diagnosis for over a century and has built on the premise anyway. In August 2026 Duolingo's shareholder letter reported that its one-time event restoring lost streaks "became one of our most successful campaigns," and drew the lesson in its own words: "It shows how much the streak matters to our learners."14 The company states it as a result.

Return to the cup. Lucretius says of his doctrine that it seems tristior​, rather grim, to those who have not handled it (​quibus non est tractata​).1 Even he placed the bitterness in the impression of people who had not tasted the thing, and he sweetened the rim for them anyway. He expected the child at the cup to be tricked (​ludificetur​, decepta​) and not harmed. The children in the extrinsic version of Zombie Division, as they described it, were not tricked. They read the rim. And what the rim told them, in my reading, is the thesis at its sharpest: the coating manufactures the bitterness it presumes.

References#

1. Lucretius. De Rerum Natura 1.936–950 (the passage recurs at about 4.11–25). Latin: The Latin Library, https://www.thelatinlibrary.com/lucretius/lucretius1.shtml. English verse, Book 1 lines quoted: Leonard, W. E., trans. (1916), Book I; Project Gutenberg ebook #785, https://www.gutenberg.org/ebooks/785. The Book 4 repetition of the passage (Leonard) is at https://fgh.perseids.org/read/latinLit/phi0550/phi001/perseus-eng1/4.1-4.26. The English rendering of lines 943–945 in the final section is the author's plain gloss of the Latin, not a published translation.
2. Horace. Ars Poetica 333–344, esp. 341–344. Latin: The Latin Library, https://www.thelatinlibrary.com/horace/arspoet.shtml. English glosses are the author's.
3. Chinese renders the idea as 寓教于乐 (yù jiào yú lè, literally "lodging teaching in pleasure"). Chinese reference sources trace it to Horace's Ars Poetica (《诗艺》); its first Chinese use was not found.
4. Sailer, M., & Homner, L. "The gamification of learning: A meta-analysis." Educational Psychology Review​, 32(1), 2020, pp. 77–112. doi:10.1007/s10648-019-09498-w. https://link.springer.com/content/pdf/10.1007/s10648-019-09498-w.pdf. Full-sample figures and the "still somewhat unresolved" quotation are from the abstract, p. 77; subsplit figures are from p. 98; the "not robust" verdict and the "underpowered" qualifier from p. 101; the measurement-timing passage from p. 102; the hedged reading of the moderator results from p. 107.
5. Dewey, J. The Child and the Curriculum​. University of Chicago Press, 1902. Text: Project Gutenberg #29259, https://gutenberg.org/cache/epub/29259/pg29259-images.html
6. Kohn, A. "Science Confirms It: People Are Not Pets." New York Times​, 28 October 2018. The quotation is from the slightly expanded version of the article, "Rewards Are Still Bad News (25 Years Later)," at https://www.alfiekohn.org/article/rewards-25-years-later/; the sentence was checked there, not in the Times text.
7. Deci, E. L., Koestner, R., & Ryan, R. M. "A meta-analytic review of experiments examining the effects of extrinsic rewards on intrinsic motivation." Psychological Bulletin​, 125(6), 1999, pp. 627–668. doi:10.1037/0033-2909.125.6.627. Full-text copy: https://depts.washington.edu/techdocs/papers/deciExtrinsicRewardsAndIntrinsicMotivation99.pdf. The interesting-versus-boring comparison (d = -0.68 versus +0.18, ns) is in the section "Interesting Versus Boring Tasks."
8. Yu, A. "Improving the streak: Forming habits one lesson at a time." Duolingo Blog​, 19 November 2020. https://blog.duolingo.com/improving-the-streak. A company statement describing a company A/B test in which one lesson a day extended the streak; results are reported as relative changes, with no denominators or intervals.
9. Silverman, J., & Barasch, A. "On or off track: How (broken) streaks affect consumer decisions." Journal of Consumer Research​, 49(6), 2023, pp. 1095–1117. doi:10.1093/jcr/ucac029. Accepted manuscript: https://udspace.udel.edu/bitstreams/7e3e3b64-8bf2-4527-ac93-94a160358165/download (Study 3, manuscript pp. 24–25; abstract, p. 3).
10. Mansur, O. "The Duolingo Streak Uses Habit Research to Keep You Motivated." Duolingo Blog​, 31 January 2022. https://blog.duolingo.com/how-duolingo-streak-builds-habit. A company statement; the +0.38% figure is the company's own and carries no denominator or interval.
11. Duolingo, Inc. Form S-1, SEC, filed 28 June 2021. https://www.sec.gov/Archives/edgar/data/1562088/000162828021013065/duolingos-1.htm. A company statement in a registration filing; the in-app purchases it lists ("streak freezes" and "timer boosts") are described as about 1% of 2020 revenue. The final prospectus wording was not compared.
12. Duolingo, Inc. Q3 FY2022 earnings press release (SEC exhibit 99.1), 10 November 2022. https://www.sec.gov/Archives/edgar/data/1562088/000156208822000163/q3fy22pressrelease.htm. A company statement, quoting co-founder and CEO Luis von Ahn.
13. Aulagnon, R., Cristia, J., Cueto, S., & Malamud, O. "Streaks to success?" NBER Working Paper 34173, August 2025 (revised September 2025). https://www.nber.org/system/files/working_papers/w34173/w34173.pdf. Figures are the working paper's; the journal version (​Economics of Education Review 109, article 102721, 2025, doi:10.1016/j.econedurev.2025.102721) was not read. The streak-versus-generic contrast is confounded with message count (two messages a week for the streak arm, one for the generic arm); the personalized-reminder arm also received two a week.
14. Duolingo, Inc. Q2 2026 shareholder letter (SEC exhibit 99.2), announced 5 August 2026. https://www.sec.gov/Archives/edgar/data/0001562088/000162828026053299/q2fy26duolingo6-30x26share.htm. A company statement; the 15.4 million count is the company's, with no method given.
15. Lepper, M. R., Greene, D., & Nisbett, R. E. "Undermining children's intrinsic interest with extrinsic reward: A test of the 'overjustification' hypothesis." Journal of Personality and Social Psychology​, 28(1), 1973, pp. 129–137. doi:10.1037/h0035519. https://www.heartofcharacter.org/wp-content/uploads/Undermining_Childrens_Intrinsic_Interest_with_Ext-1.pdf
16. Kohn, A. "The Risks of Rewards." ERIC Digest ED376990, 1994. https://www.alfiekohn.org/article/risks-rewards/
17. Hanus, M. D., & Fox, J. "Assessing the effects of gamification in the classroom: A longitudinal study on intrinsic motivation, social comparison, satisfaction, effort, and academic performance." Computers & Education​, 80, 2015, pp. 152–161. doi:10.1016/j.compedu.2014.08.019. Figures and quotations are from the article as first published online (15 September 2014); the 2018 corrigendum (​Computers & Education 127, p. 298, doi:10.1016/j.compedu.2018.09.019) was not retrieved.
18. Soderstrom, N. C., & Bjork, R. A. "Learning versus performance: An integrative review." Perspectives on Psychological Science​, 10(2), 2015, pp. 176–199. doi:10.1177/1745691615569000. Abstract-level.
19. Deslauriers, L., McCarty, L. S., Miller, K., Callaghan, K., & Kestin, G. "Measuring actual learning versus feeling of learning in response to being actively engaged in the classroom." PNAS​, 116(39), 2019, pp. 19251–19257. doi:10.1073/pnas.1821936116 (PMC6765278).
20. DeLoache, J. S., Chiong, C., Sherman, K., Islam, N., Vanderborght, M., Troseth, G. L., Strouse, G. A., & O'Doherty, K. "Do babies learn from baby media?" Psychological Science​, 21(11), 2010, pp. 1570–1574. doi:10.1177/0956797610384145. Abstract-level.
21. Habgood, M. P. J., & Ainsworth, S. E. "Motivating children to learn effectively: Exploring the value of intrinsic integration in educational games." Journal of the Learning Sciences​, 20(2), 2011, pp. 169–206. doi:10.1080/10508406.2010.508029. Figures and interview quotations are from the author accepted manuscript, https://shura.shu.ac.uk/3556/1/Habgood_Ainsworth_final.pdf (Study 1, pp. 12–21, 29; Study 2, pp. 26–28; interview table, p. 28). The phrase "a sugar coating for learning content" is from the published abstract as indexed by OpenAlex; the manuscript spells it "sugar-coating."
22. Cheng, C., Wu, Y., Wang, R., & Wang, Z. "Seductive details, cognitive load, and learning outcomes: A multi-level meta-analysis and MASEM." Educational Psychology Review​, 38(1), article 28, 2026. doi:10.1007/s10648-025-10099-z. Abstract page.
23. Sanchez, D. R., Langer, M., & Kaur, R. "Gamification in the classroom: Examining the impact of gamified quizzes on student learning." Computers & Education​, 144, article 103666, 2020. doi:10.1016/j.compedu.2019.103666. Abstract only; the full text was not read.
24. Rodrigues, L., Pereira, F. D., Toda, A. M., Palomino, P. T., Pessoa, M., Carvalho, L. S. G., Fernandes, D., Oliveira, E. H. T., Cristea, A. I., & Isotani, S. "Gamification suffers from the novelty effect but benefits from the familiarization effect: Findings from a longitudinal study." International Journal of Educational Technology in Higher Education​, 19, article 13, 2022. doi:10.1186/s41239-021-00314-6
25. Seaborn, K. "Removing gamification: A research agenda." Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems (CHI EA '21), 2021. doi:10.1145/3411763.3451695. Read in the arXiv version, https://arxiv.org/pdf/2103.05862v1
26. García Iruela, M., Fonseca, M. J., Hijón-Neira, R., & Chambel, T. "What happens when gamification ends?" In Bittencourt, I. I., et al. (Eds.), Artificial Intelligence in Education (AIED 2020), LNAI 12164, Springer, 2020, pp. 98–102. doi:10.1007/978-3-030-52240-7_18
27. Malone, T. W. "Toward a theory of intrinsically motivating instruction." Cognitive Science​, 5(4), 1981, pp. 333–369. doi:10.1207/s15516709cog0504_2. Abstract-level.
28. Papert, S. "Hard Fun." Column, Bangor Daily News​, 2002 (outlet and year from the archive listing at https://dailypapert.com/?p=1441). Text: https://dailypapert.com/hard-fun/
29. Willingham, D. T. "Why Don't Students Like School? Because the Mind Is Not Designed for Thinking." American Educator​, Spring 2009, pp. 4–13 (excerpt of ch. 1 of Why Don't Students Like School?, Jossey-Bass, 2009); quotations on pp. 6–7. https://www.aft.org/sites/default/files/WILLINGHAM(2).pdf
30. 《论语·雍也》, 6.20 in the modern (Yang Bojun) numbering; chapter XVIII (6.18) in Legge and on zh.wikisource (雍也第六). Text: https://zh.wikisource.org (論語/雍也第六). Original, given here in simplified characters: 「知之者不如好之者，好之者不如乐之者。」 The English in the text is the author's rendering.
31. Duolingo, Inc. Q2 FY2024 earnings press release (SEC exhibit), 7 August 2024. https://www.sec.gov/Archives/edgar/data/1562088/000156208824000203/q2fy24duolingo06-30x24pres.htm. A company statement; "DAUs" are the company's daily active users. This figure is not combined with the company's separate 2022 statement about streaks longer than 7 days.
32. 《三字经》 (Three Character Classic), anonymous; traditionally credited to Wang Yinglin (王应麟), with authorship and dating disputed. Text: Project Gutenberg #12479, http://www.gutenberg.org/cache/epub/12479/pg12479.txt (traditional characters). Originals, given here in simplified characters: 「头悬梁，锥刺股。彼不教，自勤苦。」 and 「勤有功，戏无益。」 The English renderings in the text are the author's.
33. Diogenes Laertius. Lives of Eminent Philosophers 5 (Aristotle). English: Perseids edition (R. D. Hicks, ed.), https://fgh.perseids.org/read/greekLit/tlg0004/tlg001/perseus-eng2/5.1.15-5.1.19. The section number is unresolved (5.18 or 5.21). The saying is Diogenes's report, compiled about five centuries after Aristotle.
34. Dewey, J. Interest and Effort in Education​, ch. 1, section I ("Unified Versus Divided Activity"), 1913. English Wikisource transcription: https://en.wikisource.org/wiki/Interest_and_Effort_in_Education/Chapter_1 (print imprint not verified against a catalogue record). The sugar-coating sentence appears inside Dewey's critique of the excess of the interest theory.
35. Postman, N. Amusing Ourselves to Death: Public Discourse in the Age of Show Business​. Viking, 1985, ch. 10, "Teaching as an Amusing Activity." Wording confirmed in an Internet Archive OCR text of the Penguin reissue; print pagination not verified.
36. Chinese titles: 《孟子·告子上》 6A7; 王艮《乐学歌》; 韩愈《进学解》. Texts: zh.wikisource (孟子/告子上); 王艮's poem as quoted by 谢青松 in 《光明日报》, reprinted at https://www.rujiazg.com/article/25751; https://zh.wikisource.org/wiki/進學解.

Further Reading#

• Harp, S. F., & Mayer, R. E. "How seductive details do their damage: A theory of cognitive interest in science learning." Journal of Educational Psychology​, 90(3), 1998, pp. 414–434. doi:10.1037/0022-0663.90.3.414. — Read it for the original seductive-details experiments: interesting but irrelevant additions to a science lesson, and what they did to recall and transfer.
• Bjork, E. L., & Bjork, R. A. "Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning." In Psychology and the Real World​, Worth, 2011, pp. 56–64. — Read it for the condition attached to desirable difficulties: a difficulty helps learning only when the learner can meet it.
• Kearney, M. S., & Levine, P. B. "Early childhood education by television: Lessons from Sesame Street​." American Economic Journal: Applied Economics​, 11(1), 2019, pp. 318–350. doi:10.1257/app.20170300. — Worth reading for the evidence that access to Sesame Street was linked to children being at the age-appropriate grade level. One caveat: the study's design cannot separate the programme's content from its presentation.
• Wouters, P., van Nimwegen, C., van Oostendorp, H., & van der Spek, E. D. "A meta-analysis of the cognitive and motivational effects of serious games." Journal of Educational Psychology​, 105(2), 2013, pp. 249–265. doi:10.1037/a0031311. — Read it for the case of serious games, which are not the same as gamification: the meta-analysis found effects on learning and retention, but did not show higher motivation.
• Hamari, J., Koivisto, J., & Sarsa, H. "Does gamification work? A literature review of empirical studies on gamification." Proceedings of HICSS-47​, 2014. — A widely cited review of empirical gamification studies, worth reading for its own verdict: gamification "does work, but some caveats exist."
• Kohn, A. "Dewey, Piaget, and Frosted Mini Wheats." 8 March 2021. https://www.alfiekohn.org/blogs/sugarcoating/ — The modern use of the sugar-coating image, with Dewey credited as its source.
• Moldon, L., Strohmaier, M., & Wachs, J. "How gamification affects software developers: Cautionary evidence from a natural experiment on GitHub." ICSE 2021​, pp. 549–561. doi:10.1109/ICSE43902.2021.00058. — The subjects are software developers, not learners, so it works as an analogy rather than a classroom result. Still worth reading for what happened to their behavior after streak counters were removed.
• Mogavi, R. H., et al. "When gamification spoils your learning: A qualitative case study of gamification misuse in a language-learning app." Proceedings of Learning @ Scale 2022​, pp. 175–188. doi:10.1145/3491140.3528274. — Read it for learners' own words about chasing points. The sample was selected for prior misuse, so it shows what misuse looks like, not how common it is.
• Mencius 6A7 ("Gaozi I"); Wang Gen, Le xue ge ("Song of Delight in Learning"); Han Yu, Jinxue jie ("An Explanation of Advancing in Learning").36 — Chinese texts on delight, from the mind's pleasure in what is right (Mencius) to delight as learning itself (Wang Gen), and on diligence set against play (Han Yu); read together, they show the old argument from both sides.