I’ve spent the last two years arguing that the real question about AI in education isn’t the technology itself. It’s the pedagogy around it. Good teaching design makes AI a powerful tool. Bad teaching design makes it a shortcut machine. But a new editorial by Bozkurt, Crompton, and Fell Kurban (2026) in Open Praxis forced me to reconsider the scale of what we’re dealing with. The problem they describe isn’t students using AI to write essays. It’s AI systems that can attend school on a student’s behalf, and nobody in the system can tell the difference.
From Generative AI to Agentic AI: What Changed
The shift the authors describe is categorical. Generative AI tools like ChatGPT and Gemini are reactive. You prompt them, they respond. Agentic AI is different. These systems combine a large language model (the reasoning engine) with an autonomous browser (the execution layer) to create what Bozkurt et al. call a “ghost student,” a digital surrogate that can log into a learning management system, open course modules, watch embedded videos, respond to quizzes, submit assignments, and move between platforms without a single human keystroke.
That’s the part that should alarm anyone who works in education. We’ve built entire digital learning ecosystems on the assumption that a valid login equals a present human being. Bozkurt et al. call this “the mirage of the login,” and they’re right. A password is not proof of presence. It never was, but it didn’t matter much until now.
The Cost of Invisible Learning: Cognitive Debt
The authors frame the deeper harm through what they call cognitive debt. Learning science has long established that productive struggle, the effort, the confusion, the active engagement with difficult material, is what builds durable understanding. When an AI browser completes coursework on a student’s behalf, none of that cognitive work happens. The grade gets recorded. The module gets checked off. But the student hasn’t changed.
I’ve covered this pattern from different angles on this blog. Gerlich (2025) documented how cognitive offloading weakens critical thinking when students lean too heavily on AI for answers. Fan et al. (2025) described it as metacognitive laziness, a gradual erosion of the self-monitoring that makes learning stick. Shaw and Nave (2026) went further and called it cognitive surrender, the point where students stop trying to think independently altogether. Kosmyna et al. (2025) at MIT measured it with brain imaging and found measurably lower cognitive effort during AI-assisted writing.
What Bozkurt et al. add to this conversation is a structural escalation. Those earlier studies examined students who were still present, still reading outputs, still making some decisions. Agentic AI browsers remove the student from the loop entirely. There’s no struggle, no engagement, and no opportunity for learning to occur. The system records success. The student records nothing.
Why Blocking, Proctoring, and Waiting Won’t Work
The authors systematically dismantle the three most common institutional reflexes. Start with blocking. You can’t block a tool you can’t see. Students install AI browsers on personal devices. The institution has no mechanism to detect it, no authority to remove it, and no technical means to prevent it from accessing web-based systems using valid credentials. The agent is invisible on the network.
Proctoring doesn’t fare any better. Digital proctoring tools like Proctorio are playing a reactive game, blocking specific AI browsers by name only to face new ones the following week. Browser fingerprinting can be spoofed. Surveillance technology, the authors note, was designed to watch a human. It’s fundamentally unprepared to detect a machine that has successfully mimicked human browser-level behavior.
And then there’s the hope that technology companies will self-regulate. Bozkurt et al. are direct about this: tech companies have engineered agentic capabilities for maximum efficiency and market reach, “often with a reckless disregard for the educational ecosystems they disrupt.” No major AI browser company has implemented restrictions for educational sites. In a competitive market, autonomy is the product. Any company that cripples its agent’s capabilities for education loses users to a competitor that won’t.
I find this section of the paper compelling. The authors are being accurate about the fact that we can’t outsource the solution to the companies that built the problem.
Credentials at Risk: When a Degree Proves Nothing
Bozkurt et al. make a sharp argument about what ghost students do to the value of credentials. Degrees and certificates are social contracts. They attest that a specific person mastered specific knowledge and developed specific capabilities. When institutions can’t verify who completed the work, that contract breaks for everyone, including the students who actually did the learning.
The paper warns that we’re “approaching a threshold where the only thing a degree proves is that the holder has a sufficiently powerful AI subscription.” That’s a provocation, but it’s not far off for programs built entirely around output-based assessments submitted through digital platforms. Online and distance education institutions, the authors argue, face an existential version of this crisis because they’ve always operated under a legitimacy gap. If they can’t demonstrate that humans were present for the learning, the degree becomes a receipt for a transaction the student was never involved in.
I’ve written about this erosion of assessment validity from multiple angles. Dawson, Bearman, Dollinger, and Boud (2024) argued that validity should take priority over catching cheaters, and that resonates here. Perkins and Roe (2025) declared it the end of assessment as we know it. Bozkurt et al. are saying something even more unsettling: it’s not just the assessment that’s at risk. It’s the entire credentialing system.
Designing for Human Presence
The authors’ core recommendation is a radical shift from output-based to process-based assessment. Oral examinations, live defenses, spontaneous questioning, in-person demonstrations, dialogic assessment, reflective portfolios. These methods resist automation because they require the messy, real-time process of human thinking that no agent can simulate. The agentic surrogate can produce a polished product, but it can’t explain the evolution of a thought, pivot during a live defense, or respond to divergent questioning.
I agree with this direction, and it aligns with what I’ve been advocating on this blog about building assessment around learning goals. But the authors also raise an equity concern I want to flag. Oral defenses and sustained reflection can disadvantage students with disabilities, language barriers, or neurodivergence. The challenge, as they put it, is to “find ways to verify the human without creating new barriers to entry” (p.8). That tension won’t resolve easily, and any institution racing to implement presence-based assessment needs to take it seriously. Faculty development is non-negotiable here. Instructors can’t design assessments that resist agentic AI if they don’t understand how AI browsers and autonomous agents actually work.
What Comes Next
The authors close with a provocation that I think every educator should grapple with: “The ghosts are already in the machine. It is up to us to decide if there is still a human in the room” (p. 9). That’s a description of where we are right now. The tools exist, they’re commercially available, and they work.
The response can’t be technical. It has to be pedagogical. If the system can’t tell a person from a proxy, we need to change the system so that only a person can succeed. That means rethinking how we design courses, how we assess learning, and what we actually value when we hand someone a credential. The technology will keep advancing. The pedagogical redesign can’t wait for it to slow down.
References
- Bozkurt, A., Crompton, H., & Fell Kurban, C. (2026). The devil is in the det[ai]ls: AI agents, ghost students, and the crisis of verified presence in an agentic AI world. Open Praxis, 18(1), 1–12. https://doi.org/10.55982/openpraxis.18.1.1145
- 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
- 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/
- 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