Various terms have been used to describe AI and human partnerships in attempting to accomplish a goal defined by the human. For example, a recent book by Ethan Mollick was titled “Co-intelligence”. Recently, I have been reading about the work of several authors who have described different ways in which learners might interact with AI some being more successful than others. I will return to this material after my preliminary remarks. These analyses also consider various types of collaboration.
As I read the most recent set of papers, I flashed back on research I encountered in the early 1990s. In this case, there was no AI partner, but educational researchers and instructional designers were investigating ways in which student peers could collaborate (e.g., Johnson, et al., 1991; Slavin, 1995). If you were a preservice or practicing teacher at that time, you likely heard a lot about cooperative learning. There were multiple proposed benefits of cooperative learning. Social interactions were motivating. Multiple individuals have unique experiences and skills and combining these resources benefits all who are exposed. Interaction, whether it be a form of teaching, working through differences of opinion or error-checking each other, or simply sharing experiences, augments an individual’s cognitive activity. More individuals, theoretically, allow more to be accomplished in less time.
There were also concerns about cooperative learning activities, and if the connection to learning with an AI partner is not obvious, identifying educators’ and researchers’ concerns should clarify the similarities. The first was called the freeloader effect – in a pair version of cooperative learning, one participant might do all of the work and the other would do and learn very little. Perhaps one student would simply be more motivated or more capable and find that working alone was more efficient. Other concerns included a lack of experience and skill in cooperative planning or in effectively using the talents of multiple individuals. A few remedies from that era I recall included individual accountability (e.g., individual tests on content), positive interdependence (e.g., clear identification of tasks whose outputs will be combined in the final product), and a structured or scaffolded process.
Scaffolding will come up again when discussing the potentially effective use of AI, so perhaps an example of what this might look like would be helpful. In building construction, a scaffold provides temporary support for workers. In education, a scaffold provides a structure that supports a learning task by guiding how work is done. Consider a version of a common strategy for a cooperative task – i.e., think, pair, share. Students are given a task. Then, each student writes (or just thinks) of a proposed solution. Finally, students consider each other’s proposals (pair) and integrate them to arrive at a solution (share). The imposed structure in this case ensures that each individual participates and provides a record of their work, should the teacher want to hold each accountable for participating. The jigsaw cooperative learning technique offers another method for scaffolding a cooperative learning activity to promote participation and individual accountability. A project is selected that requires identifiable roles or tasks. For example, when creating a brochure describing the butterflies one might most likely encounter in a local garden, individual students could be assigned to research different butterflies and then asked to combine their research into a single document.
Before moving on to the cooperation between a student and AI, I propose that a body of research examines peer interaction in educational settings, and that its identified issues and remedies might be useful to those now focused on AI.
Differentiating Ways Learners Use AI In Attempting to Identify Productive Approaches
First, I would refer readers to a previous post in which I examined the classification scheme of AI learning strategies. The following is my previous explanation of the levels in this classification scheme.
The three levels have the following characteristics:
Zone 1: No AI Involvement
In this level, learning occurs without any AI assistance. While learning happens, it is often “capacity-constrained” because the learner must spend significant time and effort on execution and task completion, leaving less bandwidth for higher-order reflection.
Zone 2: Scattered, Half-Hearted Use
This is characterized by using AI for minor tasks like fixing sentences, checking facts, or tidying paragraphs. It often produces the worst learning outcomes. The learner still carries nearly the full cognitive load but adds the overhead of managing AI interactions without gaining significant cognitive savings. Note: this summary paraphrases the description of the authors. My version would add having the student using the AI tool to perform the task based on simplistic instructions.
Zone 3: Committed, Strategic Delegation
This level involves offloading entire categories of substantive work to AI to free up genuine cognitive capacity. This freed bandwidth is then redirected toward tasks AI cannot do, such as critiquing frameworks, questioning assumptions, and making complex judgment calls. This zone is where “transformative learning” is thought to live, provided the course design is intentional about how and why tasks are delegated.
My attempt to summarize this scheme suggested that the use or nonuse of AI and its relationship to successful learning experiences could be explained by investigating how certain conditions of student motivation, metacognitive proficiency, and working memory interact. For motivation, was the learner focused on developing personal knowledge and/or skills beyond task completion and on receiving a positive evaluation? Working memory reflects the capacity for meaningful learning beyond underlying task demands. The notion here is that AI might, in some situations, handle nonessential or untargeted tasks, allowing the learner to devote their attention and processing capacity to accomplishing targeted goals. Metacognitive proficiency suggests that more sophisticated learners with sufficient available attentional capacity are more likely to make sound decisions about when to use AI to free cognitive capacity to accomplish goal-related knowledge and skills.
I hope it makes some sense how these factors might interact. Allow an argument based on a personal perspective. I would suggest that my own learning offers motivational advantages. I learn to accomplish personal goals rather than be subject to external goals and reward structures in a classroom setting. I am also more metacognitively sophisticated than secondary school students. I understand the tasks I want to accomplish well having explored them countless times over the decades and such experiences offer me useful insights, but also means that I have background knowledge and cognitive skills that require less working memory capacity on my part. My use of AI could enable Zone 3 processing. I am not saying that this is always the case, but it makes sense I have a greater opportunity to function at this level.
Does this analysis then suggest that less-experienced learners, even secondary students, should not work with an AI partner? Not necessarily. This is where the scaffolding found to be important in cooperative learning and proposed as a way for Zone 3 functioning to be practical. Scaffolding bridges the gap, allowing tasks to be accomplished before all necessary conditions are in place and offering a mechanism for introducing required skills.
The mention of design in the Zone Three descriptions is another way of suggesting the value of scaffolding. I have included multiple references I found that explain the three-zone model and offer suggestions for scaffolding. My typical reaction is that the examples never seem to include the tasks educators might most need assistance in addressing. For example, writing tasks are commonly described, but not general “homework” tasks.
I think the best advice is to focus on processes and discuss assignment goals with students, differentiating those processes students are free to use AI to accomplish and which are expected to be completed by students. Consider how students might document their activity in completing each. For example, again using a writing example, use AI to identify content you will use in your writing project (available for submission). Submit your draft based on this content. Have AI evaluate the quality of your draft (available for submission). Submit your final version. Students might also be given a general topic on which a written product will be required. Students will bring the relevant content they have found to class and then receive specific instructions on the product to be fashioned during that class period. Both resources and product to be submitted.
General Resources:
Mollick, E. (2024). Co-intelligence: Living and working with AI. Penguin. Penguin.
Johnson, D., Johnson, R., & Holubec, E. (1991). Cooperation in the classroom (rev. ed.). Edina, MN: Interaction.
Slavin, R. (1995). Cooperative learning: Theory, research, and practice (2nd ed.). New York: Merrill.
Sources for AI Level Analysis
Hardman – The cognitive offloading paradox
Lodge and Lobel. Artificial intelligence, cognitive offloading and implications for education
Means. Strategic Cognitive Offloading: What the Research Says, and Why Higher Education Isn’t Ready for It
Wang, S., Zhang, H. Pedagogical partnerships with generative AI in higher education: how dual cognitive pathways paradoxically enable transformative learning. Int J Educ Technol High Educ 23, 11 (2026). https://doi.org/10.1186/s41239-026-00585-x
