Discovery Learning: Bruner, Examples and When It Works

Go from storyboard to an interactive course in minutes

Start Free

Picture a customer service officer at a credit union branch, three years on the counter, who passed the annual fraud awareness quiz last month. It listed the common scam types and she got every question right. Then a member of 30 years asks to send $18,400 to a new payee "for a term deposit with a better rate", while still on the phone to her "adviser". The new payee, the size of the payment and the person on the phone each look ordinary on their own, so the officer checks them one at a time and processes it.

I'm an instructional designer at Coursebox, where we make instructional design software that turns a brief into a course outline, so I'll mention our product where I tested it. Discovery learning offers one answer to that officer's problem: let staff work out the pattern from realistic cases instead of reading another list. It is also one of the most argued-over methods in education research, because the unguided version teaches beginners less. This guide covers what discovery learning is, what Jerome Bruner did and did not say, what the evidence found, two real studies with adults, and how to design guided discovery for the officer's course, with two free templates.

What is discovery learning?

Discovery learning is a teaching method in which learners work out a principle, rule or solution from the problems, cases or materials they are given, instead of being told it first, with anything from no guidance to hints and feedback at each step.

View the full-size diagram

The researchers behind the largest review of the evidence, Alfieri and colleagues in the Journal of Educational Psychology in 2011, used a working definition: discovery learning "occurs whenever the learner is not provided with the target information or conceptual understanding and must find it independently and with only the provided materials" (p. 2). They added that the guidance along the way can be "intensive or, conversely, minimal".

That range is the key to the whole topic. The basics first:

• Scope. It has been studied mostly in school maths and science, and also in university courses, software training and medical education.
• Purpose. Its advocates argue that what people work out for themselves is easier to use on new problems than what they are told.
• Structure. It is not a fixed sequence of stages. A designer makes two decisions: how much help learners get while they try, and when the explicit teaching arrives.
• Current use. Researchers moved away from the pure, unguided version long ago. The forms with current evidence behind them are guided discovery and designs that teach straight after the attempt.

Here are the main forms at a glance.

Our guide to constructivism learning theory covers the theory side, and why constructivist courses still need guidance.

Pure vs guided discovery: what is the difference?

Pure discovery leaves learners to find the answer with little or no help. Guided discovery gives them a problem plus hints, prompts or feedback along the way, so they still do the thinking but are less likely to get lost or settle on a wrong answer. The evidence favours support over leaving learners to it; it does not show that guided discovery beats every well-designed explicit lesson.

The sharpest case against pure discovery came from Richard Mayer in a 2004 article in American Psychologist titled "Should there be a three-strikes rule against pure discovery learning?". He reviewed studies in which learners left to themselves failed to discover problem-solving rules, conservation strategies and programming concepts, and concluded: "Pure discovery did not work in the 1960s, it did not work in the 1970s, and it did not work in the 1980s, so after these three strikes, there is little reason to believe that pure discovery will somehow work today" (p. 18).

Mayer's target was the unguided version, not discovery as such. He argued that effective methods involve "cognitive activity rather than behavioral activity, instructional guidance rather than pure discovery, and curricular focus rather than unstructured exploration" (p. 14), and that learners need "enough freedom to become cognitively active" and "enough guidance so that their cognitive activity results in the construction of useful knowledge" (p. 16). Reading the same studies, Manu Kapur later noted that in Mayer's review "guided discovery outperformed both unguided discovery and direct instruction conditions."

The research community moved the same way. Alfieri and colleagues defined guided discovery as "some form of instructional guidance (i.e., scaffolding) or regular feedback to assist the learner at each stage of the learning tasks" (p. 5). Even the strongest critics of discovery describe this shift. John Sweller, Paul Kirschner and Richard Clark wrote in a 2007 reply to their critics that "discovery learning in its pure form was slowly jettisoned and replaced by 'guided discovery' as the full disaster of pure discovery became apparent" (p. 117).

An IBM study from the 1980s shows both sides in one project: guided exploration working for office workers, and a pure discovery step that had to be abandoned.

Case study: Office workers learning a word processor at IBM in the US

In the 1980s, IBM researchers studied how new users learned a widely used office word processor. Its self-instruction manual asked users to "proceed step-by-step through sequences of drill and practice exercises", but new users wanted to start on real tasks, like typing a letter, and skipped the steps that seemed irrelevant.

• What they did: A first study replaced the manual with a set of unordered cards that gave only goals, hints, checkpoints and help with recovering from errors. The team then built the Minimal Manual: real tasks first, far less text, error recovery and some steps left incomplete on purpose. They had hoped to let users discover document revision on their own, but "This did not work out; the subskill had to be further decomposed in order to be reliably executed by learners."
• Evidence and result: In Carroll and colleagues' 1987 paper in Human-Computer Interaction, the card users spent less than half as long learning as manual users and made half as many errors, yet "sometimes voiced a desire for a more structured training tool". In the main Minimal Manual experiment, 19 temporary office workers, most of them secretary-typists, spent up to three days in a simulated office at IBM's Austin laboratory. The 10 using the Minimal Manual needed 40% less learning time than the 9 using the standard manual, and completed 2.7 times as many of the test's subtasks (18 people scored).
• Limitation: The samples were small (19, then 32 in a second experiment), the office was simulated and the software dates from the 1980s. IBM's own researchers evaluated their own design. The Minimal Manual changed several things at once, so the result cannot be credited to discovery alone, and part of the 2.7 times gap came from standard-manual users running out of time before the advanced tasks. Everything was measured during or straight after training, not on the job.
• Lesson: Adults learning a work tool did better starting on real tasks with brief hints, checkpoints and help with errors than following a long manual. Where learners could not reliably work a skill out, the designers had to teach it. Guide the exploration, and teach what people cannot discover.

Who is the father of discovery learning?

Jerome Bruner is the name most associated with discovery learning, through his 1961 essay "The act of discovery" in the Harvard Educational Review. He did not invent it. Experiments comparing discovery with telling were running in the 1940s and 1950s, and John Dewey had argued decades earlier for learning through inquiry into real problems. Mayer described the essay as "Bruner's (1961) eloquent call for discovery methods" (2004, p. 15).

Bruner's essay argued for four benefits, usually summarised as greater skill at solving problems, a move from external rewards to the satisfaction of discovering, learning the methods of discovery itself, and better memory. The line most often quoted from it is: "Practice in discovering for oneself teaches one to acquire information in a way that makes that information more readily viable in problem solving." These were his arguments, not findings.

He also set limits. Alfieri and colleagues pointed out that Bruner cautioned that discovery "could not be made a priori or without at least some base of knowledge in the domain in question", and that "many have seemingly ignored his warnings" (p. 2). In their account, Bruner "also warned that the learner's mind has to be prepared for discovery" (p. 13). Bruner himself wrote that "left to himself, the child will go about discovering things for himself within limits" (1961, p. 22, as quoted by Alfieri and colleagues). So the essay is a weak citation for leaving learners to it.

Kittel's 1957 result, as Kapur summarises it, found "guided discovery to be the best for retention and transfer", which means the case for guidance is older than Bruner's essay. The workplace root is older than most people assume too. In the 1960s an OECD project tested a "discovery method" for retraining older workers, including 83 British locomotive drivers moving from steam to diesel, and its summary says the method was useful "if tuned to a given group". I could read only that summary, not the results.

Which discovery learning papers to read

• Mayer (2004), "Should there be a three-strikes rule against pure discovery learning?" Read this one first. It is six pages long and makes the case against pure discovery and for guided discovery in plain language.
• Alfieri, Brooks, Aldrich and Tenenbaum (2011), "Does discovery-based instruction enhance learning?" Read it for the evidence. Its tables show which kinds of support helped, for which ages and in which subjects.
• Kapur (2015), "Learning from productive failure". Read it before you design a problem-first lesson. It sets out the design principles and the conditions under which the struggle pays off.

Bruner's 1961 essay is the best-known early statement, but read it alongside Alfieri's account of his warnings.

Principles of discovery learning

Bruner's principles were claims about benefits. The research since then has added principles about design: how much help to give, and when. These are the ones with evidence behind them.

Alfieri and colleagues drew the practical lesson themselves: "Although direct teaching is better than unassisted discovery, providing learners with worked examples or timely feedback is preferable" (p. 12). Teaching, they wrote, "should employ scaffolded tasks that have support in place as learners attempt to reach some objective, and/or activities that require learners to explain their own ideas."

What does the research say about discovery learning?

On average, unassisted discovery taught less than explicit instruction, and discovery with support did better than other methods. The kind of support decided the result.

Alfieri and colleagues ran two meta-analyses covering 164 studies. Across 108 studies, unassisted discovery did worse than explicit instruction (d = -0.38). Across 56 studies, enhanced discovery did better than other forms of instruction (d = 0.30). In their words, "unassisted discovery does not benefit learners, whereas feedback, worked examples, scaffolding, and elicited explanations do" (p. 1). Our comparison of behaviourism, cognitivism and constructivism uses the first of those figures to show why constructivist design needs guidance.

"Enhanced" covered three kinds of support: generation, elicited explanations and guided discovery (p. 5). They did not work equally well.

An effect size (d) of 0.50 means the average guided-discovery learner scored about half a standard deviation higher than the average learner taught another way. The finding a designer should act on sits next to it: asking people to come up with the answer, with nothing else, did slightly worse than other methods. The gains came with the hints, the feedback and the request to explain, not with discovering on its own. Against worked examples, no statistically significant difference was detected (d = 0.06), so a good worked example remains a strong choice for beginners.

The breakdowns add detail that a single average hides.

Age made a difference as well. Against explicit instruction, unassisted discovery cost adults less (d = -0.26) than adolescents (d = -0.53), and "the adolescents tended to benefit least and the adults tended to benefit most from both unassisted-discovery tasks and enhanced-discovery tasks" (p. 12). The authors suggest adults' "greater domain-relevant knowledge" as the reason, which fits Bruner's warning about prior knowledge.

Read these numbers as a direction. The results varied widely between studies, the disadvantage of unassisted discovery was larger in first-tier journals (d = -0.67) than in second-tier ones (d = -0.24), and enhanced discovery averaged d = -0.54 in unpublished studies against 0.35 to 0.40 in journals. The samples were mostly school and university students, so none of these figures is a workplace result.

Discovery learning vs problem-based, inquiry and direct instruction

Problem-based and inquiry learning share discovery's starting point, a problem before the answer, but as designed they are heavily scaffolded. Whether they count as discovery learning is still argued.

The argument came to a head in 2006. Paul Kirschner, John Sweller and Richard Clark grouped several of these approaches together in a 2006 article in Educational Psychologist, writing that "The minimally guided approach has been called by various names including discovery learning" (p. 75). They argued that minimally guided instruction "is less effective and less efficient than instructional approaches that place a strong emphasis on guidance of the student learning process", and that guidance becomes less important "only when learners have sufficiently high prior knowledge to provide 'internal' guidance" (p. 75).

Two replies rejected the grouping. Cindy Hmelo-Silver, Ravit Duncan and Clark Chinn answered in 2007 that problem-based and inquiry learning "are highly scaffolded" (p. 99), that they "are not discovery approaches and are not instances of minimally guided instruction", and that direct instruction can be given "on a just-in-time basis" (p. 100). Schmidt and colleagues, in a second reply, agreed "about the failure of minimally guided instruction for novices learning in structured domains" (p. 91), but pointed out that in problem-based learning, a tutor shares missing knowledge with the group when a task "turns out to be too complex" (p. 95).

Both sides agree that pure discovery fails beginners. They disagree on whether a scaffolded, problem-first design beats full explanation. Sweller, Kirschner and Clark replied that they were "still unable to detect the differences between guided discovery and PBL and/or IL with scaffolding", and that providing learners with a complete solution was "better still" (2007, p. 117). Hmelo-Silver and her co-authors wrote that "'Does it work?' is the wrong question" (p. 105), and asked instead under what circumstances guided inquiry works and for which outcomes. That is the more useful question for a designer. The comparison that follows is my synthesis of the sources in this guide, not a ranking.

Several of our guides cover the guided end of this range. The cognitive load theory guide explains why worked examples suit beginners and how to fade them. In the cognitive apprenticeship model, exploration comes last, once coaching and other support have faded. Merrill's principles of instruction put a real task at the centre and add demonstration and guidance around it.

Problem first, then teach: productive failure

Productive failure is a design in which learners attempt a hard problem before they are taught, fail to find the full answer, and then get explicit teaching that builds on their attempts. It is one of the most studied modern forms of discovery, and the teaching is not optional.

Manu Kapur, whose studies were mostly with school students in Singapore, first published it in 2008. In his 2015 summary, the design has "a generation and exploration phase (Phase 1) followed by a consolidation phase (Phase 2)". In the first phase, learners "are asked to generate solutions without any cognitive support or scaffolds", so they "will typically be unable to generate or discover the correct solution(s) by themselves". In the second, the teacher compares their solutions with each other and with the correct one, then teaches it.

In his best-known study, 133 ninth graders in Singapore worked for two lessons on a problem about which soccer player was most consistent, before they had been taught standard deviation. The teacher then compared their solutions and worked through the standard method. The productive failure classes "significantly outperformed their DI counterparts on conceptual understanding and transfer without compromising procedural fluency", where DI means direct instruction.

Kapur's design principles explain why this is not "leave them to it". The problems should "challenge but do not frustrate, rely on prior mathematical resources, and admit multiple solutions", learners need "opportunities for explanation and elaboration", and they need "opportunities to compare and contrast" failed solutions with the correct one. His tasks were pilot-tested and refined before use.

A 2021 meta-analysis by Tanmay Sinha and Manu Kapur pooled 53 studies and found that problem solving followed by instruction beat instruction followed by problem solving on average (g = 0.36, from the abstract), with stronger effects when designs stuck closely to productive failure. The same abstract reports the reverse for younger primary school children and for general skills, and Kapur wrote that "Thus far, the efficacy of PF has been demonstrated in the domains of math and science."

The adult evidence is thinner but points the same way. With 40 first-year pharmacy students in Toronto, Steenhof and colleagues found in 2019 no difference on tests of what students had learned and how they applied it, but a better result for productive failure on a test of preparation for future learning (abstract only). A related line of training research, error management training, encourages people to explore and make errors during training. A 2008 meta-analysis by Keith and Frese of 24 studies with 2,183 participants found an average effect of d = 0.44, rising to d = 0.80 on transfer tasks that differed in structure from the training (abstract only). The clearest recent trial with working clinicians tested hard cases before the teaching.

Case study: Emergency doctors learning to read head CT scans in the US

Emergency doctors in residency training need to spot time-sensitive findings on head CT scans, a skill the trial's authors say "is not well covered in residency training". Handling unfamiliar cases, which they call adaptive expertise, is especially hard to teach. In 2022 and 2023, seven US emergency medicine residency programs ran a one-hour online session in their weekly teaching time.

• What they did: Residents worked through 9 head CT cases on a real radiology viewer and were randomly assigned to one of three versions with the same cases and teaching content. One group answered difficult questions about each case before seeing the explanation, a design meant to produce errors. A second group answered easy questions first. The third group was shown how to find the critical findings before looking at the cases.
• Evidence and result: In the randomised trial published in JAMA Network Open in 2024, 212 residents were randomised and 150 completed the posttest. On new kinds of cases, the hard-questions-first group scored 60.6%, against 45.2% for easy questions first and 40.9% for taught first. On familiar cases no statistically significant difference was detected (75.1%, 73.6% and 74.4%). The study's statistical mediation analysis linked much of the hard-first group's advantage to training errors; errors themselves were not randomly assigned, and first-year residents gained the most.
• Limitation: The outcome was an unsupervised online test straight after one hour of training, not patient care or later retention. Fewer residents finished in the two problem-first groups (64.3% and 67.6%, against 80.3%), so those groups may have held more motivated residents. Most programs were university-based, and the authors say the study cannot show how difficult the cases need to be.
• Lesson: Hard, realistic cases before the explanation helped doctors who already knew the basics handle unfamiliar cases, with no statistically significant difference on familiar ones. Easy questions first showed no detected advantage over teaching first. Make the attempt hard enough to produce errors, follow it with specific teaching, and test on new cases, not repeats.

A smaller Danish study points the same way: medical students in Copenhagen who were told to make errors during ultrasound simulation did better on a transfer test with real patients 7 to 10 days later, in a 2017 trial of 60 students (abstract only).

When does discovery learning suit adults at work?

Discovery suits a principle that learners can work out from cases they already know how to read, where a wrong first attempt is safe and there is time for the teaching that follows. Procedures, safety steps, legal duties, rules that no case reveals, and anything a beginner cannot yet reason about are better told and then practised. This rule is my synthesis of the studies in this guide, not a validated checklist.

The evidence behind each part:

• Prior knowledge. Kirschner, Sweller and Clark argue that guidance becomes less important only when learners have enough prior knowledge to guide themselves. Slava Kalyuga and colleagues described the expertise reversal effect: "Instructional techniques that are highly effective with inexperienced learners can lose their effectiveness and even have negative consequences when used with more experienced learners" (2003, p. 23). The residents in the CT trial were already "familiar" with head CT.
• Adults and structure. Adults lost less from unassisted discovery and gained more from guided discovery in Alfieri's analysis. In a 2010 study of 161 working-age adults with an average age of 39.7, Carter and Beier found that instructions encouraging errors "in high structure training were beneficial for older learners, especially when higher in ability" (abstract only). The error-friendly approach helped inside a structured course.
• What can be discovered. IBM's designers could not get learners to discover document revision reliably and had to teach it in smaller steps.
• The goal. In the CT trial the problem-first benefit showed on new cases, not familiar ones. If the goal is a fixed routine, discovery adds little.

Discovery learning examples

Discovery learning shows up in classrooms, universities and training rooms. These examples come from studies, apart from the last one.

• Secondary maths in Singapore. Ninth graders tried to rank soccer players by consistency before being taught standard deviation, then the teacher built on their attempts.
• Primary science in the US. In a 2004 study by Klahr and Nigam, 112 third and fourth graders learned how to design a fair experiment, and "many more children learned from direct instruction than from discovery learning" (abstract). It is a useful counterexample for children learning a procedure.
• Office software in the US. IBM's office workers learned a word processor from task cards with hints instead of a step-by-step manual.
• Emergency medicine in the US. Residents attempted hard head CT cases before the explanation.
• A credit union branch. A hypothetical course for counter staff, used for the rest of this guide.

A worked example: spotting scam-driven payments at a credit union

Here is the example I'll design with, carried through the lesson plan, the product test and both templates. It is hypothetical.

A member-owned Australian credit union with about 20 branches wants a short course for its branch customer service officers. Every officer has at least 6 months on the counter. They already process payments, know the core banking screen and can see each member's payment history. All of them have done the annual fraud awareness module, which listed scam types and ended with a quiz. Passing that quiz does not show they can spot a scam at the counter.

The problem is the one from the opening. Officers passed the quiz but still processed scam-driven payments, because each request looked ordinary on one dimension and they checked the signs one at a time.

The task is to decide, at the counter, whether to process a payment request or pause it. The credit union's own rule (also hypothetical) uses four warning signals:

1. A new payee, or changed payee details.
2. An amount or pattern unlike the member's history.
3. Urgency or secrecy pressure, such as "today" or "don't tell the bank".
4. A third party directing the payment, by phone, text or email, or a story the member has been told to give.

Two or more signals together means pause before processing. The procedure after a pause is fixed: do not process; ask the open questions in the credit union's scam conversation guide; never complete a payment while the member is on a call with the person asking for it; never accuse the member; and refer to the fraud team before anything is sent. An officer cannot refuse a member's own money on their own judgement, and the fraud team decides on any longer hold.

The course gives officers four case files, read alongside each member's payment history, with every fact the decision needs (amounts in AUD):

• Request A. A member aged 71, with the credit union for 30 years, asks to send $18,400 to a new payee "for a term deposit with a better rate". Her payment history shows nothing of that size. She is on the phone to "her adviser", who told her the branch might try to talk her out of it. All four signals. Pause.
• Request B. A member aged 34 pays $2,150 to the payee she has paid every month for two years. The rent went up after a lease renewal, and she shows the notice. Only a changed amount, which she explains. Process.
• Request C. A member aged 45 wants $6,900 sent today to a new payee for "bail money", after a text from an unknown number saying it is her son, overseas, with a new phone. Her history shows no payments like it. All four signals. Pause.
• Request D. A member aged 52 pays $3,000 to a new payee for a used car he inspected in person. The seller's bank details are on a signed sale contract, the amount is within his usual spending range, and there is no pressure or third party directing him. One signal. Process with normal checks.

A transfer case comes at the end, with a different surface and the same principle. A small-business member, a landscaping business, asks to pay a regular supplier's $12,480 invoice to new bank details that arrived by email: "we've changed banks, please update before Friday". For this transfer case, the credit union also requires independent verification of any emailed supplier bank-detail change, even without a second signal. Pause and call the supplier on a number already on file, not one in the email. This is an additional supplied procedure, not a second application of the two-signal threshold.

By the end, officers should be able to sort payment requests into process and pause, say which signals they saw, have the conversation and refer correctly. The course does not certify fraud competence or give anyone authority to refuse a payment.

This example suits discovery learning for four reasons. The principle, that no single sign decides and the combination does, can be worked out by comparing a handful of requests. The officers have the prior knowledge to compare them. The course is low risk, because nobody's real money moves. And there is time, because it is a 45-minute module. It also shows the limit: the procedure and the officer's duties are told, not discovered, because guessing them wrong is costly and no case would reveal them.

How to design a guided discovery lesson

This workflow is my own, built on Alfieri's evidence for guided discovery, Kapur's order of attempt then teaching, and Mayer's balance of freedom and guidance. It follows the credit union course.

1. Split the objectives. Decide which ones learners will discover and which you will tell. Here, officers discover the two-or-more pattern and are told the procedure.
2. Check the prior knowledge. Confirm learners can already read the cases. Officers know the banking screen and payment histories, so they can compare requests.
3. Choose cases that differ on the principle. Include cases that should not trigger the rule, so learners have to compare. Requests B and D are ordinary payments that share one surface feature with the scams.
4. Give learners every fact the decision needs. Put the amounts, history and circumstances in the case files. Do not ask for a judgement the information cannot support.
5. Write a hint ladder. Write three hints, from general to specific, that the learner opens one at a time after a first answer, with only the last one pointing at the pattern.
6. Ask for an explanation. After each decision, ask what made them decide. Explanation was one of the forms of support that helped in Alfieri's analysis.
7. Give feedback on every case. Name what the learner should have noticed, not only whether they were right.
8. Teach the principle, then the procedure. State the rule explicitly, link it back to the learner's own answers, then teach the procedure directly.
9. Test on a new case. Use a different surface and no hints at first, then give feedback.

The credit union lesson, step by step

The three hints for Case 1 narrow step by step:

1. "Compare each request with what you know about the member's usual payments."
2. "Look at who is asking for the payment, and how much time the member is being given."
3. "Requests A and C share several things that B and D do not. What are they?"

Treat the first answers and the pattern officers write as formative assessment: they show what each officer noticed before the teaching, and keeping them beside the final answers shows what changed. None of this shows that officers pause the right payments at work. That needs a check on the job, such as a review of the payments each officer referred over the following weeks.

What course requests ask for

To see whether people brief discovery into their courses, I looked at course requests typed into Coursebox in the 90 days to 7 October 2026: one per person, from English-language browsers, with no uploaded file. Of 266 prompts, 104 were course requests. Two independent AI readers each classified all 104 against written definitions, quoting the words behind every positive call, and every positive request was re-read in full by hand. The readers agreed on every zero. Counts are classification estimates from the union of the two readers. Positive-call checks do not establish an upper bound because missed cases remain possible.

People asked for practice often. The classification identified 21 of the 104 as asking for hands-on practice, exercises, activities or simulations.

"Hands-on" in these briefs meant practice after the content. One asked for "practical step-by-step lessons, examples, exercises", and another for "step-by-step explanations, practical examples and exercises". The only request asking for less telling, a school health course with "Minimal teacher lecture", paired it straight away with guided practice, which is guided discovery's shape rather than pure discovery. These are 104 English requests over 90 days, many of them not for workplaces, and people may add discovery activities later by editing. Still, in this sample, the move that defines discovery learning, a problem before the explanation, never appeared in a brief. If you want it, write the problem, the hints, the feedback and where the explanation lands into the brief yourself.

Limitations of discovery learning

• Unassisted discovery teaches beginners less. It did worse than explicit instruction on average across 108 studies in Alfieri's analysis, and Mayer found it failed across three decades of studies.
• Discovery without support is the weak version. Generation alone did slightly worse than other methods.
• It takes more time. Kapur's ninth graders spent two lessons on the problem before any teaching. Sweller, Kirschner and Clark point out that "Requiring that a learner discover a problem solution always reduces guidance compared to presenting the solution" (2007, pp. 115 to 116).
• Wrong ideas can stick. Without the teaching that follows, learners may leave with their own wrong answer. Kapur's design adds support specifically to keep students from "unproductive failure experiences".
• The help has to match the learner. By the expertise reversal effect, help that suits a beginner can hold back an experienced colleague, and too little help overloads a beginner. One course may need two levels of hints.
• The evidence is mostly from schools and universities. Productive failure has been shown mainly in maths and science, and reversed for younger children and general skills. Social and verbal skills showed the biggest loss from unassisted discovery.
• Adult evidence measures training tests, not work. The IBM office workers and the emergency residents were tested during or straight after training. I found no study that tracked a discovery-based workplace program through to job performance.

Using discovery learning with an AI course tool

On 9 October 2026 I briefed Coursebox to build the credit union course. Coursebox is my product, so weigh my account with that in mind. The brief asked for a 45-minute self-paced course called "Pause or process? Spotting scam-driven payments at the counter", with seven sections in the order of the credit union lesson. It gave the four requests, the signals, the procedure and the transfer case as fixed facts. It said not to state the warning signs or the rule anywhere before section 4, to put three hints in section 2 that open one at a time, to give feedback on each request, to use open-text answers in sections 2, 3 and 6, and to put the wrap-up last.

The discovery half was in place. In the first outline, the four requests came before any rule, and "What did you notice?" asked learners to describe the dividing pattern in their own words.

The issue was the other half. The first outline had 4 sections instead of 7. "The pattern", "What the credit union requires" and the transfer case were squeezed into one final page, "From pattern to safe action", there was no separate wrap-up, and two recall quizzes had been added. The outline compressed the teaching and transfer test instead of giving each the planned space. Reading that outline alone could not establish whether the condensed page taught the rule adequately.

I sent one follow-up message: keep the discovery part, restore the seven sections with the pattern and the procedure taught explicitly and the transfer case in place, and drop the quizzes. After reloading, the outline had 7 sections, 14 pages, no quizzes and 4 assignments. I then opened the generated pages. "Make your four calls" is an open-text assignment with requests A to D and no warning signals or rule stated. The next page, "Hints and feedback", comes after the attempt, gives the three hints exactly as briefed and keeps each hint panel closed until the learner opens it, which I checked in the page itself. It then gives feedback on each request. "The four signals" and "Why the calls differ" teach the rule after the attempt, "Required counter procedure" and "Adviser on the line" teach the procedure, and the transfer case ends with feedback. The independent verification rule for emailed supplier bank changes is clarified in the plan here. Before using this course, confirm that the teaching and transfer feedback state that rule; the page check reported above did not verify this clarification.

I checked the outline and the generated pages, not how officers respond to them, because the course has not been run with branch staff. Two things need trimming before use: the course came to about 1 hour 8 minutes against the 45 minutes I asked for, and the first assignment asks for 150 to 200 words per request, more than I briefed. The hints sit on the page after the attempt, above the feedback, so learners open them once their first calls are written, not while they decide. In the pages I checked, hints opened when the learner chose; I did not test any system-triggered support based on difficulty. With the missing sections back, the course follows the guided discovery order: cases first, then hints and feedback, then the rule taught against the officers' own answers and a new case to try it on.

Before you brief a course like this, decide which objectives people should discover and which you should tell.

Putting discovery learning to work

Discovery learning works when it is guided. Learners who already know the basics attempt realistic cases, get hints and feedback as they go, explain what they decided, and are then taught the principle against their own answers. Left to work things out alone, beginners learn less, and procedures, legal duties and rules no case reveals should be told whatever the learners' experience. Use discovery for patterns people keep missing at work and for handling new cases. Use explanation and worked examples for steps that have one right answer.

Pick one objective in a course you are building where staff know the rule but still miss the pattern on the job. Write three or four case files that differ on that pattern, including at least one that should not trigger it, a three-step hint ladder and the explanation that will follow. Then try it with a few learners and read their first answers before you build the rest.