AI is especially well suited for work that is generative, iterative, variable, data-heavy, or repetitive. It can produce and revise content quickly, generate many alternatives with little additional effort, process large amounts of information, transform content from one form to another, and repeat well-defined tasks consistently. These affordances make AI particularly useful for the following ID tasks.
Generating Design Artifacts and Media. AI can significantly accelerate the development of instructional materials by helping designers draft design artifacts and create media. For example, AI can generate first drafts of learning objectives, storyboards, lesson outlines, assessment items, scenarios, scripts, job aids, and feedback, giving instructional designers a strong starting point that they can refine based on learner needs and design goals. AI can also support media production by creating or assisting with images, illustrations, voiceovers, animations, video scripts, captions, and other multimedia elements. This can reduce the time spent on routine production tasks. However, the instructional designer still plays the critical role of reviewing the output for accuracy, alignment, relevance, and overall instructional quality.
Ideation, Variation, and Rapid Prototyping. AI is especially useful for brainstorming, generating variations, and rapid prototyping. It can quickly suggest alternative activities, examples, scenarios, explanations, and assessment ideas; represent the same concept in multiple ways; and turn ideas into draft storyboards or prototypes. It is also very effective at producing many practice opportunities, such as case studies, role-plays, quiz questions, examples, non-examples, and feedback. This allows instructional designers to explore more options quickly and then select and refine the strongest ones.
Data Management, Processing, and Initial Analysis. AI can also help instructional designers manage and make sense of large amounts of data. It can organize learner feedback, clean and categorize information, summarize survey responses or interview notes, and compare results across multiple sources. It is particularly useful for initial analysis, where it can quickly scan large amounts of information to identify recurring themes, preliminary patterns, unusual responses, or areas that may deserve closer attention. This gives the instructional designer a useful first pass through the data and helps identify where deeper analysis may be needed, and can save considerable time, especially when working with large datasets or multiple sources of feedback. However, instructional designers still need to question and interpret the results cautiously and avoid treating AI-generated analyses as automatically accurate or conclusive.
Automating Routine and Administrative Tasks. AI can take on many repetitive and administrative tasks that consume instructional designers’ time, such as formatting content, summarizing meeting notes, organizing feedback, drafting routine emails, creating documentation, tagging or categorizing materials, checking consistency, and converting content into different formats. It can also help automate established workflows and routine project managing and tracking tasks, especially when the steps are already clearly defined and repeatable. By reducing this workload, AI gives instructional designers more time to focus on higher-value work such as learner analysis, design decisions, collaboration, evaluation, and improving the learning experience.
AI is not naturally suited for work that depends heavily on deep local context, nuanced professional judgment, determining what really matters in a specific situation, human relationships, or ethical accountability. It can generate plausible recommendations, but it does not fully understand the people, history, culture, constraints, and consequences surrounding a design problem. These limitations mean that AI may support but not replace instructional designers in the following ID work.
Identifying the Real Problem in Context. AI can work only with the information that is available to it, but the real instructional or performance problem is often hidden in the local context. A request for “more training,” for example, may actually reflect unclear expectations, poor feedback, limited access to tools, weak incentives, low motivation, or a mismatch between what people are taught and what they are expected to do. Identifying the real problem requires talking with learners, subject-matter experts (SMEs), instructors, managers, and other stakeholders, observing how work is actually done, and understanding the environment in which learning and performance take place. This contextual investigation is difficult to automate because much of the most important information is tacit, informal, and never written down. AI can help organize and analyze what is collected, but the instructional designer still has to ask the right questions, interpret what people say, notice contradictions, and determine what problem is truly worth solving.
Judging Pedagogical Fit. AI can generate activities, assessments, and design recommendations that sound convincing, but it cannot be fully trusted to determine whether they are the right choices for a particular learning situation. Good design judgment requires understanding the learners, the instructional goals, the context, the constraints, and the tradeoffs among different approaches. An activity may look engaging but fail to support the objective, or an assessment may appear rigorous while measuring the wrong thing. AI can suggest options, but the instructional designer must decide whether those options are appropriate, aligned, and likely to improve learning.
Guaranteeing Trustworthy Output. AI can produce answers that sound clear, confident, and well written even when the information is incomplete, outdated, or incorrect. It may invent facts, misrepresent sources, or overlook important details. For this reason, instructional designers should not treat AI output as automatically trustworthy. Content, references, and recommendations still need to be checked against reliable sources and reviewed by SMEs when appropriate. The same caution applies to data processed or analyzed by AI. Designers need not only to assess whether the patterns AI identifies are meaningful, but also to look for what it may have overlooked, excluded, or failed to recognize.
Taking Accountability or Responsibility. AI can recommend actions, generate content, and support decision-making, but it cannot be held accountable for the consequences of a design. It does not take responsibility when information is wrong, when an assessment is unfair, or when a design fails to meet organizational or ethical standards. The instructional designer must remain responsible for reviewing AI output, making final decisions, documenting the rationale for those decisions, and ensuring that the design is accurate, appropriate, and aligned with learner and organizational needs.
As AI takes on more drafting, production, analysis, and routine work, the instructional designer’s role shifts toward the skills that require stronger human expertise. The most important capabilities are no longer just about creating instructional materials, but about understanding how people learn, making sound design judgments, working effectively with others, evaluating AI-supported work, and knowing where human oversight is essential. In this environment, several skills become especially important.
Learning Science as the Foundation. As AI makes it easier to generate lessons, activities, assessments, feedback, and media, instructional designers need an even stronger foundation in learning science to judge whether those outputs will actually support learning. AI can produce something that looks polished and plausible without understanding whether it matches how people learn, whether the cognitive load is appropriate, whether practice is meaningful, or whether feedback will help learners improve. A deeper knowledge of learning science helps instructional designers move beyond asking, “Can AI create this?” to asking, “Will this actually help learners learn, remember, transfer, and perform?” In the AI age, learning science becomes the basis for directing, evaluating, refining, and sometimes rejecting AI-generated design ideas.
AI Literacy as the Essential Competence. AI literacy is now a core competency for instructional designers because AI is becoming part of both the ID process and the learning experience itself. A good instructional designer needs to understand, at a conceptual level at least, how AI works, what generative AI can and cannot do, and where its limitations and risks lie. They also need to know how to effectively communicate with AI, critically evaluate its output, create tailored AI tools, and design human-AI workflows that keep appropriate human judgment in the loop. Just as importantly, instructional designers must understand issues related to bias, ethics, privacy, copyright, and responsible use. Without this foundation, it is easy to overtrust AI, use it inappropriately, or design learning experiences that are technically impressive but not necessarily accurate, fair, or pedagogically sound.
Data-Informed Decision Making. Data-informed decision making becomes even more important in the age of AI because instructional designers now have access to more learner data, faster analysis, and a growing number of AI-generated recommendations. The challenge is no longer simply gathering information. It is deciding which data are meaningful, how much confidence to place in the findings, and what those findings actually mean for design. Instructional designers therefore need to do more than organize and analyze data. They must interpret evidence through the lens of learning science, contextual knowledge, and professional judgment, while also questioning the assumptions and limitations behind the results. The goal is not to follow the data blindly, but to use evidence thoughtfully to guide decisions, test ideas, and continuously improve the learning experience.
Communication and Collaboration as the Key Human Skill. AI can generate content quickly, but it cannot build shared understanding among people. Instructional designers still need to work closely with learners, SMEs, instructors, managers, and stakeholders to uncover needs, clarify goals, gather context, resolve differences, and make design decisions. They also need to explain how and why AI is being used, communicate risks and limitations, and help others interpret results and evidence. As AI speeds up production, the instructional designer’s value increasingly comes from connecting people, asking the right questions, translating between different perspectives, and building trust around the design process.
Other Core ID Skills as a Necessity. A pilot can use autopilot, but still needs to know how to fly. It is important for instructional designers to retain traditional core ID skills even when AI can perform parts of them, because once a skill is fully offloaded, the designer can lose the ability to judge the quality of the AI’s work. If instructional designers repeatedly let AI do the difficult cognitive work, they may gradually lose diagnostic ability, design intuition, ability to notice poor alignment, ability to challenge SMEs, ability to explain design decisions, and ability to work when AI is wrong or unavailable. This means instructional designers need to intentionally keep doing certain tasks themselves even when AI can do them, especially while their design expertise is still developing.
In sum, AI is increasingly becoming a design partner and co-worker in the ID process. It can generate alternatives, accelerate production, and help designers explore possibilities that might otherwise take much longer to do. The instructional designer, however, remains the director of the design process: the person who defines the problem, understands the context, decides what matters, exercises professional judgment, and takes responsibility for the final design. The most valuable instructional designer in the AI era will probably not be the person who can get AI to do the most work. It will be the person who knows what can safely be offloaded to AI, where human judgment must remain, and how to make the most of both.
Association for Talent Development. (2025). AI in instructional design: Transforming workflows and content creation. https://www.td.org/product/research-report–ai-in-instructional-design-transforming-workflows-and-content-creation/792512
Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8
Dell’Acqua, F., McFowland, E., III, Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality (Working Paper No. 24-013). Harvard Business School. https://ssrn.com/abstract=4573321
Goh, E., Gallo, R., Hom, J., Strong, E., Weng, Y., Kerman, H., Cool, J. A., Kanjee, Z., Parsons, A. S., Ahuja, N., Horvitz, E., Yang, D., Milstein, A., Olson, A. P. J., Rodman, A., & Chen, J. H. (2024). Large language model influence on diagnostic reasoning: A randomized clinical trial. JAMA Network Open, 7(10), Article e2440969. https://doi.org/10.1001/jamanetworkopen.2024.40969
Nuhoğlu Kibar, P., & Ilgaz, H. (2026). The intersection of artificial intelligence and instructional design practice: A systematic review. Educational Technology Research and Development, 74, 1989–2015. https://doi.org/10.1007/s11423-026-10624-z
Rowland, G. (1992). What do instructional designers actually do? An initial investigation of expert practice. Performance Improvement Quarterly, 5(2), 65–86. https://doi.org/10.1111/j.1937-8327.1992.tb00546.x
U.S. Copyright Office. (2025). Copyright and artificial intelligence, Part 2: Copyrightability. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf
Wang, X., Chen, Y., Ritzhaupt, A. D., & Martin, F. (2021). Examining competencies for the instructional design professional: An exploratory job announcement analysis. International Journal of Training and Development, 25(2), 95–123. https://doi.org/10.1111/ijtd.12209
]]>Imagine a learner logging into an online course. They open the assignment, read the instructions, and close the tab without trying it. They have not answered a question or received a failing grade. Nothing seems to have gone wrong. Yet they may already be giving up because they have already reached one conclusion: “I cannot do this.”
Something similar happened in competitive running before 1954, when the four-minute mile was widely considered beyond human reach. Then, on May 6, Roger Bannister ran the mile in 3 minutes and 59.4 seconds (Roger Bannister, n.d.). Other runners soon followed, crossing a barrier that had stood there for decades. John Landy broke Bannister’s record 46 days later (Sport Australia Hall of Fame, n.d.). Human bodies had not suddenly changed. What changed was the belief about what was possible.
Albert Bandura called this belief self-efficacy: your judgment of whether you can perform the actions needed to achieve a particular result (Bandura, 1977).1 According to expectancy-value theory, people’s “choice, persistence, and performance can be explained by their beliefs about how well they will do on the activity and the extent to which they value the activity” (Wigfield & Eccles, 2000, p. 68). Self-efficacy is closely related to the first belief: whether people believe they can perform the activity successfully.2 A learner may care deeply about a subject and still make no attempt if they expect to fail. From the outside, that inaction can look like laziness or a lack of skill. We are likely to respond with more practice or stricter accountability, yet neither will necessarily change the belief holding the learner back. Before choosing a solution, we need to understand what helps learners strengthen their self-efficacy.
Bandura identified four sources from which learners form beliefs about their capabilities: performance accomplishments, vicarious experience, verbal persuasion, and physiological states.3 Each of them points to a different way learning design can strengthen or weaken self-efficacy.
The first is to help the learner succeed at least once (i.e., performance accomplishment). The task should be challenging enough to count, but manageable enough to complete. As Bandura noted, succeeding at an easy task “provides no new information” (1977, p. 201). A learner can revise a perceived ceiling only by completing work that once seemed beyond reach. This raises a practical design question: Is the task designed to give a learner the right level of challenge as well as a genuine chance to succeed?
The second is watching someone like them succeed, which Bandura calls a vicarious experience. The choice of model seems to matter. Bandura argued, from studies of people overcoming fears, that observers gain more from watching someone overcome difficulty through determined effort than from the “facile performances” of adept models (Bandura, 1977, p. 197). Later classroom experiments found this specifically true for learners who are struggling with something new (Schunk, Hanson, & Cox, 1987). The same principle should guide our design. Seeing the polished work of someone who always succeeds may not make the goal feel more attainable to a struggling learner. Seeing a peer struggle, adjust, and eventually succeed may make the goal feel within reach.
Verbal persuasion is on Bandura’s list too. However, as he suggested, it is “widely used because of its ease and ready availability” (Bandura, 1977, p. 198), not because of its effectiveness. The confidence created by encouragement alone is fragile. A single experience of failure can readily extinguish it through “disconfirming experiences” (p. 198). Therefore, encouragement is the easiest form of support to offer, but its effects may also be the quickest to fade.
As you design instruction, consider what your learners believe is beyond their reach, and ask whether your design reinforces or challenges that belief. Teaching does more than present content. It also shapes what learners believe they can do.
Bandura, A. (1977). Self-efficacy: Toward a unifying theory of behavioral change. Psychological Review, 84(2), 191–215.
Roger Bannister. (n.d.). Chronology: Key moments in eight years as an athlete. Retrieved September 12, 2026, from https://www.sirrogerbannister.com/athletics-career
Schunk, D. H., Hanson, A. R., & Cox, P. D. (1987). Peer-model attributes and children’s achievement behaviors. Journal of Educational Psychology, 79(1), 54–61.
Sport Australia Hall of Fame. (n.d.). John Landy. Retrieved September 12, 2026, from https://sahof.org.au/hall-of-fame-member/john-landy/
Wigfield, A., & Eccles, J. S. (2000). Expectancy–value theory of achievement motivation. Contemporary Educational Psychology, 25(1), 68–81. https://doi.org/10.1006/ceps.1999.1015
Bandura also treated self-efficacy as specific, not as a general supply of confidence that people either have or lack. We can feel capable of explaining a concept to our students and completely incapable of repairing our car. Even within one subject, we may feel ready to describe a study but not to conduct its statistical analysis. The task, its difficulty, and the setting all matter. ↩
Bandura separated efficacy expectations, the belief that one can accomplish a task, from outcome expectancies, the belief that a given action will lead to a given outcome, and argued that expectancy-value theorists had modeled the outcome expectancies. The expectancy-value theorists stated that they measure “individuals’ own expectations for success, rather than their outcome expectations,” so their expectancy construct “is more similar to Bandura’s efficacy expectation construct than it is to the outcome expectancy construct” (2000, p. 71). ↩
Learners also use physical sensations as information about their capabilities (Bandura, 1977). What matters is how they interpret those sensations. A racing heart before a difficult task might be read as excitement, or as evidence that the task is beyond them. The same physical response can therefore strengthen or weaken self-efficacy. ↩
Researchers have asked similar questions in the past. In 1983, Richard Clark reviewed decades of studies that compared one teaching medium with another and wrote one of the most widely quoted sentences in the field. Media are “mere vehicles that deliver instruction but do not influence student achievement any more than the truck that delivers our groceries causes changes in our nutrition” (Clark, 1983, p. 445). When performance improved after a new medium was introduced, Clark usually attributed the gain to the instructional method that came with it, or to novelty, not to the medium itself.
So, is AI just another truck? I think there is another question to ask this time: Will the truck change the types of groceries delivered? Clark’s vehicles delivered instruction without altering it. Generative AI, however, can change what learners receive each time learners ask a question, write a prompt, or share something about themselves.
Bastani et al. (2025) provide some of the answers. In fall 2023, they conducted a preregistered field experiment with nearly 1,000 students in grades 9 through 11 across about 50 mathematics classes at one Turkish high school. All students received the same teacher-led instruction first. The researchers then assigned the classes to three conditions for the practice session.
One group used GPT Base, a chat interface similar to ChatGPT. Its prompt asked GPT-4 to act as a tutor and help the student solve the problem. Another group used GPT Tutor. It had the same interface, but its prompt included a set of safeguards such as teacher-written solutions, common student mistakes, and hints. The prompt also instructed the model not to give away the full solution. Students in the control group used only their notes and textbook.
After practice, everyone took a closed-book exam without AI. Each exam problem tested the same concept as a problem the students had just practiced.
During practice, students assigned to GPT Base scored 48% higher than students in the control group. Students assigned to GPT Tutor scored 127% higher than students in the control group. Those numbers make both tools look promising, right?
But here is the twist.
On the unassisted exam, students who had practiced with GPT Base scored 17% lower than students in the control group. Students who had practiced with GPT Tutor did not score significantly differently from the control group. So, the safeguards prevented the decline seen with students using GPT Base, but students using GPT Tutor did not do better than students using notes and a textbook.
The researchers then considered two explanations for the lower GPT Base exam scores. First, GPT-4 might have given students incorrect answers. When the researchers repeatedly asked GPT Base for answers to the 57 practice problems, it returned a correct answer in 51% of trials. Its logical errors predicted lower scores on assisted practice problems, but the researchers found no statistically significant spillover to the paired exam problems. So, incorrect answers may not explain the lower exam performance.
Second, students might have used it as a crutch instead of engaging with the mathematics. The researchers classified two types of messages: superficial message which repeated the problem or asked for an answer and nonsuperficial message which attempted an answer or asked for help. For students working with GPT Base, superficial first messages rose from 56% to 67% across the session. In the GPT Tutor condition, they fell from 42% to 37%. It is clear that the two designs elicited different ways of interacting with the AI.
Fan et al. (2025) found a related pattern in a study on teaching university writing with AI. They asked 117 Chinese university students who used English as a second language to revise an essay. The researchers randomly assigned students to four conditions: ChatGPT, a human expert, checklist tools, or no additional support. They restricted ChatGPT to giving advice rather than generating the essay, although they observed some students copying example sentences that it supplied!
Students in the ChatGPT group improved their essay scores by 3.60 points on average. The mean improvements were 1.63 points in the no-support group, 1.48 in the human-expert group, and 1.40 in the checklist group. The improvement in the ChatGPT group was significantly greater than the improvement in each of the other three groups.
However, the ChatGPT group’s advantage in essay revision did not extend to other measures. The four groups did not differ significantly in knowledge gain, on a transfer test about a related topic, or on any post-task motivation dimension: interest and enjoyment, perceived competence, effort and importance, or pressure and tension.
And here is the most interesting finding. During revision, students in the ChatGPT group repeatedly moved between writing and ChatGPT. Students working with the human expert did not form that loop with their expert. Their revising stayed tied to the reading materials, and they kept moving between the task instructions and their own checking of the draft. The authors read the pattern shown by the students using ChatGPT as potential metacognitive laziness. That is, students offloaded some of the work of regulating their learning to AI.1
Please do not get me wrong. We do have evidence showing that AI can improve learning measured after the lesson, not just performance during it.
Kestin et al. (2025), for example, developed a GPT-4 tutor around research-based instructional practices. The tutor prompted active engagement, managed cognitive load, promoted a growth mindset, guided students through each problem in sequence, and used instructor-written step-by-step solutions.
The researchers then conducted a crossover experiment with 194 students in an introductory physics course at Harvard during fall 2023. Each student completed one lesson with the AI tutor at home and another through in-class active learning. The median posttest score was 4.5 after the AI-tutored lesson and 3.5 after the active-learning lesson, against a combined pretest median of 2.75. Students spent a median of 49 minutes with the tutor, compared with 60 minutes allocated to learning in class. They also reported feeling more engaged in the AI condition.
The authors set clear limits around this result. They write that they “do not presume that structured AI tutoring will always outperform in-class active learning in all contexts, for example, those requiring complex synthesis of multiple concepts and higher-order critical thinking” (Kestin et al., 2025, p. 6). They also argue that AI tutors should not replace in-person teaching.
Taken together, these studies do not tell a simple story about AI. GPT Base offered help without safeguards and was associated with lower unassisted exam scores than the control condition. GPT Tutor offered hints, withheld answers, and produced no significant difference from the control condition on the exam. Kestin’s tutor paired instructor-written step-by-step solutions with a platform developed over months to structure students’ progress. The same students learned more from the AI-tutored lesson than from the in-class lesson.
Three studies cannot settle the question. The participants ranged from ninth graders at one Turkish high school to undergraduates at Harvard. The AI was used to support practice after a lesson, help students revise an essay, or deliver the lesson itself. The comparison conditions also differed. We cannot attribute the different results to one design feature alone.2
However, one pattern is clear. Someone has to decide what an AI tutor will provide and/or hold back, when it will offer a hint, how it will sequence a problem, and which mistakes it will anticipate. Each choice is a learning design decision. Each depends on an understanding of how people learn. The design is what makes the difference.
That said, our work may change a bit when we design learning with AI. Designing the prompts and structures that guide how AI interacts with students is becoming part of what we do. The work can be challenging because we need to understand AI’s instincts. A teacher who improvises in the moment draws on years of teaching and years of being a learner. Those instincts are usually worth following. An AI has instincts too (and this is where the truck may change the groceries): Because AI is trained to be helpful and fluent, it may “happily” hand over an answer to students when learning requires a hint, a question, or time to struggle. Learning designers must know how to guide AI to support learning. Doing this well requires a good understanding of how AI works as well as how learning works.
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), Article e2422633122. https://doi.org/10.1073/pnas.2422633122
Clark, R. E. (1983). Reconsidering research on learning from media. Review of Educational Research, 53(4), 445–459.
Deng, R., Jiang, M., Yu, X., Lu, Y., & Liu, S. (2025). Does ChatGPT enhance student learning? A systematic review and meta-analysis of experimental studies. Computers & Education, 227, Article 105224. https://doi.org/10.1016/j.compedu.2024.105224
Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544
Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, Article 17458. https://doi.org/10.1038/s41598-025-97652-6
Wu, X., Zhu, P., Zhang, J., Yin, M., & Wang, Y. (2026). ChatGPT’s impact on student learning outcomes: A meta-analysis of 35 experimental studies. Humanities and Social Sciences Communications, 13, Article 684. https://doi.org/10.1057/s41599-026-07019-z
The term describes an inference from trace patterns, not a disposition that the researchers measured directly. ↩
Two recent meta-analyses (Deng et al., 2025; Wu et al., 2026) found that students who used ChatGPT generally performed better than those in the comparison groups. However, these findings do not necessarily tell us how much students learned. Deng et al. (2025) found that nine studies allowed students to use ChatGPT during the final assessment. Another 33 did not report whether ChatGPT was allowed. In the first nine studies, the assessment measured what students could produce with ChatGPT, much like the practice scores in the Turkish high school study. In the other 33, we do not know whether the students worked independently or not. Wu et al. (2026) did not examine this issue. The pooled results therefore mix different kinds of performance and cannot tell us clearly what students could do on their own. ↩
Which of the following short paragraphs would you remember best? They are ranked by how closely the two sentences are related, from strongest to weakest, and they all end the same way:
Level 1. “Joey’s big brother punched him again and again. The next day his body was covered with bruises.”
Level 2. “Racing down the hill, Joey fell off his bike. The next day his body was covered with bruises.”
Level 3. “Joey’s crazy mother became furiously angry with him. The next day his body was covered with bruises.”
Level 4. “Joey went to a neighbor’s house to play. The next day his body was covered with bruises.”
You might choose the first one (Level 1) because the punches lead directly to the bruises. Or you might choose the fourth (Level 4) because it leaves a mystery.
In 1984, three memory researchers asked 36 students at the University of Denver to read short paragraphs like these (Keenan et al., 1984). Each student saw only one version of Joey. Then, after about half an hour of other reading tasks, the researchers asked: how did Joey get his bruises?
The results? Across the paragraphs, average recall was 69% for the strongest causal links (Level 1 punching), 83% for Level 2 (falling off the bike), 75% for Level 3 (angry mom), and 50% for the weakest links (Level 4 going to a neighbor’s house).
The weakest links (Level 4) were recalled significantly less often than every other level (p < .01). The trend is clear: recall rose and then fell across the four levels (p < .001). And the gap between Level 2, the bike, and Level 1, the punching, was marginally significant (p < .10).1
You may wonder why the fall from a bike (Level 2) was remembered so well.
The researchers offered one explanation. When one event almost inevitably leads to the next event, the reader does not need to do much mental processing. Repeated punches, for example, already bring bruises to mind. However, when the connection is less direct but still plausible, the reader has to actively search for the missing link. A fall from a bike does not always cause bruises, but you can imagine how this can happen. The small effort required to make that inference may lead the reader to process the information more deeply.
You might argue that the Level 4 paragraph required even more mental work and should therefore have produced the best recall. So what happened there? The researchers proposed that the extra processing of the paragraph did not produce a firm causal connection. A visit to a neighbor could lead to bruises in many ways, and the two sentences did not provide a firm answer. “If increased processing only results in a tenuous causal connection”, the researchers concluded, “then it cannot facilitate memory” (Keenan et al., 1984, p. 125).
Daniel Willingham, a cognitive psychologist who writes for teachers, stated the same principle in five words: “Memory is as thinking does” (2003, “Memory Is as Thinking Does” section). Learners do not necessarily remember the material we give them. What they remember is what they think about as they process it. The two do not always line up. Much of the craft of teaching lies in bringing them together.
Willingham illustrates this mismatch with an example. A teacher asked his students to bake biscuits so they could appreciate what escaped slaves ate on the Underground Railroad. These students had probably spent 30 seconds thinking about how the biscuits were related to the course material, and then 30 minutes thinking about measuring flour and mixing dough. If memory follows thought, Willingham suggested, baking biscuits is likely to be what they remembered!
Together, the 1984 study and Willingham’s article raise a design question that neither of them directly addresses: How much of the connection should we reveal when designing instruction?
If we explain every step before learners have a chance to think, they may follow easily but do little thinking themselves. But if we leave out a crucial step, they may miss the intended connection or construct a plausible but incorrect one. A useful approach is to leave a gap that learners can close by thinking with what they already know.
In practice, we can present an outcome first and ask what might have caused it, or show two events and ask learners to explain the connection. Then we can provide the missing information and introduce the concept that makes sense of the relationship. Consider a lesson on confirmation bias. Instead of beginning with a definition and a list of common mistakes, we could introduce a hiring manager who believes a new interview method works. She points to every successful hire who passed the interview but explains away every unsuccessful one. We could ask learners what evidence she accepted, what she dismissed, and whether her conclusion was justified. After they examine her reasoning, we would introduce confirmation bias as the concept that explains the pattern.
Finally, before designing an activity, follow Willingham’s own rule for teachers: “Anticipate what your lesson will lead students to think about” (2003, “What Does This Mean for Teachers?” section). The activity supports learning when that thinking focuses on the concept or idea you want them to learn. If it does not, learners may remember an enjoyable experience while forgetting the lesson it was meant to teach.
Keenan, J. M., Baillet, S. D., & Brown, P. (1984). The effects of causal cohesion on comprehension and memory. Journal of Verbal Learning and Verbal Behavior, 23(1), 115–126.
Willingham, D. T. (2003). Ask the cognitive scientist: Students remember… what they think about. American Educator, 27(2), 37–41.
The authors noted that the middle-level cause sentences were longer on average than the others, which may have influenced recall. They described the finding as tentative and called for further research to verify it (Keenan et al., 1984, pp. 124–125). ↩