One of the most damaging assumptions in the AI-in-education conversation is that giving students AI tools means giving up human control. Either we ban AI and preserve rigor, or we allow it and accept the risks. I’ve argued against this binary for years on this blog, and I keep running into it in institutional policy, faculty discussions, and even in research.
A 2020 paper by Ben Shneiderman, published in the International Journal of Human-Computer Interaction, offers a theoretical rejection of that binary. Shneiderman proposes a Human-Centered AI (HCAI) framework that says high levels of human control and high levels of computer automation can and should coexist. The goal is to design for both.
This paper was published before ChatGPT and before the generative AI panic. Shneiderman was writing about self-driving cars, medical devices, and aviation. But his framework maps onto education so precisely that reading it in 2026 feels like finding a blueprint someone drew before the building went up.
The One-Dimensional Trap in Human-Centered AI Design
Shneiderman’s starting point is a critique of the Sheridan and Verplank (1978) ten-level scale of automation, which runs from full human control at one end to full computer autonomy at the other. This one-dimensional model has shaped decades of AI design thinking. The assumption buried in it is that automation and human control are inversely related: the more the machine does, the less the human controls. Shneiderman calls this a false trade-off, and I think he’s right.
He proposes a two-dimensional framework with human control on one axis and computer automation on the other, creating four quadrants. The upper-right quadrant, where both human control and automation are high, is the target zone. Shneiderman calls systems in this quadrant Reliable, Safe & Trustworthy (RST).
The upper-left is human mastery territory (high control, low automation, like a bicycle or piano). The lower-right is computer control territory (high automation, low human control, like a pacemaker or airbag). And the lower-left is simple or low-tech devices.
I find this reframing powerful because it breaks the assumption that educators face a binary choice with AI. The AI Assessment Scale I covered from Perkins, Roe, and Furze (2024) tries to do something similar for assessment design: calibrate the level of AI use per task, acknowledging that different assignments call for different levels of human and AI involvement. Shneiderman’s framework provides the theoretical backbone for that kind of thinking.
What Happens at the Extremes
Shneiderman identifies two danger zones design conversations ignore: excessive automation and excessive human control. The Boeing 737 MAX disaster is his central example of excessive automation. The MCAS system was built to override pilot input based on a single sensor, and when that sensor failed, the pilots couldn’t override the override because the system’s existence wasn’t described in the manual and they hadn’t been trained on it. A catastrophic design failure, not a pilot error.
On the other side, Shneiderman argues that many “human mistakes” are actually design failures. If the interface doesn’t make the system’s state clear, if the user can’t predict what the machine will do next, and if there’s no way to intervene quickly, the problem is the design, not the operator.
He flags the Tesla “Autopilot” naming itself as a design failure: the word suggests the car drives itself, which lowers driver vigilance. The National Transportation Safety Board’s 2016 crash investigation found exactly that.
This connects directly to what I see happening in education. When we give students AI tools with no pedagogical scaffolding, no guidance on when to use them and when to think independently, we’re creating the educational equivalent of excessive automation.
I wrote about Shaw and Nave’s (2026) concept of cognitive surrender, the idea that students stop engaging their own reasoning when AI is available, and Shneiderman’s framework explains why: the system wasn’t designed to keep the human in the loop. The automation was there, but the human control wasn’t.
Reliability, Safety, and Trustworthiness as Design Goals
Shneiderman breaks down what it takes to build RST systems across three pillars. Reliability comes from technical practices: audit trails, benchmark testing, continuous data quality and bias review, explainable interfaces. The safety pillar is organizational, built on leadership commitment, open reporting of failures, internal oversight boards, and public accountability. Trustworthiness requires a third layer entirely: independent oversight from professional organizations, government regulators, NGOs, and third-party auditors.
This three-layer architecture resonates with me because it mirrors what I think education needs from AI. We’ve been talking almost exclusively about the technical layer, which tools work, which ones hallucinate, which ones detect AI text. The safety and trustworthiness layers have barely been touched. Open reporting of AI failures in the classroom? Internal oversight of how AI-generated feedback affects student learning? Independent review of AI tools before adoption? Those conversations are just beginning in most institutions.
Shneiderman also organizes applications into three tiers based on consequence level. Recommender systems occupy the lowest tier, where errors are annoying but not harmful. Consequential applications, things like medical diagnosis, financial advising, and legal tools, occupy the middle. Life-critical systems, aviation, military, implantable medical devices, are the highest tier. Education falls somewhere in the middle. The stakes of a bad AI-generated assessment or a flawed feedback algorithm aren’t life-or-death, but they can shape a student’s academic trajectory, their confidence, their relationship to learning. That’s consequential.
The Prometheus Principles and What They Mean for Educators
Shneiderman calls his design guidelines the “Prometheus Principles,” drawn from commercial, academic, and government standards. These include: consistent interfaces that let users form and express intent, continuous visual display of the system’s state, rapid and reversible actions, error prevention, informative feedback for every action, and progress indicators. The principles are meant to produce systems that are comprehensible, predictable, and controllable.
There’s one line in the paper I want to highlight. Shneiderman writes that “computers are not teammates, collaborators, or co-active partners, as many suggest. Humans are responsible for actions of the technology that they use.” I find this a useful corrective to a lot of the language I hear in edtech circles, where AI is described as a “co-pilot,” a “thinking partner,” or a “collaborator.”
That language obscures accountability. If an AI tool gives a student inaccurate feedback on an essay, the student didn’t make the error and neither did the AI. The instructor who chose the tool and designed the assignment bears the responsibility, and the design of the tool determined whether they had the control needed to exercise it well.
I’ve written about cognitive offloading research from Gerlich (2025) and the neural engagement findings from Kosmyna et al. (2025), both of which document what happens when AI does the thinking and the human disengages. Shneiderman’s framework names the design failure behind those outcomes: systems that increase automation without increasing human control. The answer is designing for both dimensions at once, not dialing back the automation.
Reading Shneiderman in 2026
This paper is six years old, and it was written for HCI researchers and product designers, not educators. Shneiderman’s examples are thermostats, elevators, self-driving cars, and digital cameras. He doesn’t mention classrooms, students, or learning outcomes. But the framework translates directly.
Every time an educator decides how much AI to allow on an assignment, or an institution writes a policy on AI use, the question is the same one Shneiderman posed in 2020: are we designing for high automation and high human control, or are we stuck in one-dimensional thinking?
The technology has changed enormously since 2020. The design question hasn’t changed at all.
References
- Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. _Societies_, 15(1), Article 6. https://doi.org/10.3390/soc15010006.
- Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing tasks. MIT Media Lab. https://www.media.mit.edu/publications/your-brain-on-chatgpt/.
- Perkins, M., Roe, J., & Furze, L. (2024). The AI assessment scale revisited: A framework for educational assessment. arXiv preprint arXiv:2412.09029. https://medkharbach.com/ai-assessment-scale-framework/
- Shaw, S. D., & Nave, G. (2026). Thinking fast, slow, and artificial: How AI is reshaping human reasoning and the rise of cognitive surrender. Working paper, The Wharton School, University of Pennsylvania. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646
- Shneiderman, B. (2020). Human-centered artificial intelligence: Reliable, safe & trustworthy. International Journal of Human-Computer Interaction, 36(6), 495-504. https://doi.org/10.1080/10447318.2020.1741118