The Devil is in the Details: AI Agents, Ghost Students, and the Crisis of Verified Presence in an Agentic AI World

Synthesis: Bozkurt, Crompton, and Fell Kurban (2026) argue that the shift from reactive generative AI to agentic AI marks a categorical break in digital education, enabling "ghost students"—digital surrogates formed by coupling Large Language Models (LLMs) "minds" with agentic AI browser "bodies" that navigate LMSs, engage content, and complete assessments with human-like mimicry. This creates a verification gap that traditional proctoring and detection tools are structurally unable to close, with consequences including cognitive debt, erosion of credential trust, and legal vulnerabilities. The authors argue institutions must move away from output-based assessments and "design for human presence."

Key Findings

1. The ghost student is real, not speculative. Agentic browsers (e.g., Perplexity Comet, OpenAI Atlas) can log into an LMS with valid authenticated credentials, navigate modules, watch videos, take quizzes, and submit work—rendering the human learner's presence optional.
2. Detection is structurally impossible. Anthology (Blackboard) conceded AI agents "look identical to normal student activity"; browser fingerprinting can be spoofed, and proctoring tools are trapped in a reactive whack-a-mole against each new agent.
3. Blocking, proctoring, and waiting all fail. BYOD culture defeats system-level blocking; proctoring cannot distinguish a mimicked machine; and vendors have no commercial incentive to cripple agentic utility for education.
4. Cognitive debt accumulates invisibly. Because the "productive struggle" of learning is offloaded, the agent bypasses learning itself—metrics look flawless while the student remains unchanged (echoing evidence that GPT-4 access raised practice scores 48% but lowered exam performance 17%).
5. Credentials erode for everyone. When institutions cannot verify who completed the work, degree value falls for all students, including those who engaged fully; O-ODE programs face existential legitimacy risk.
6. Severe data-protection and security risks. BYOA ("bring your own agent") triggers catastrophic failure of frameworks like FERPA, GDPR, and Türkiye's KVKK, which assume institutional control that no longer exists; prompt-injection attacks let malicious pages hijack the agent's privileged access.
7. The solution is pedagogical, not technical. Institutions must shift from product-oriented to process-oriented, dialogic, and reflective assessments—presence-required formats such as oral defenses, live presentations, and reflective portfolios that value "human becoming" over algorithmic proxies.

What this means for practice

• Instructors. Move assessment to presence-required formats — oral defenses, live spontaneous questioning, in-person demonstrations, dialogic Assessment and reflective portfolios — because agentic browsers with valid LMS credentials look identical to normal student activity.
• Instructors. Learn how agentic browsers actually work (authenticated LMS access, content navigation, prompt-injection exposure) before designing against them; the authors treat faculty development at massive scale as a prerequisite, not an add-on.
• Administrators. Write explicit, operationally grounded governance that names agentic browsers as distinct entities, communicates risk transparently, and addresses the loss of the "private session" under data-protection mandates, since BYOA breaks the institutional-control assumptions behind FERPA, GDPR, and KVKK.
• Institutions. Design inclusive presence deliberately: presence-required formats can disadvantage students with disabilities, language barriers, or neurodivergence, so equity vigilance is part of the format decision.
• Instructors. Ask students to articulate what they learned, not merely what they submitted, and embed that reflection in the curriculum so the choice between delegating and learning is named explicitly.

Limitations

• This is a theoretical analysis with no empirical study: there is no sample, no measured prevalence of ghost students, and no field test of the proposed assessment redesign.
• The claim that detection is structurally impossible rests on a vendor acknowledgement (Anthology/Blackboard's statement that AI agents "look identical to normal student activity"), browser-fingerprinting arguments, and platform capability descriptions rather than independent validation.
• The cognitive-debt and credential-trust arguments rest partly on cited external evidence — for instance the result that GPT-4 access raised practice scores 48% while lowering exam performance 17% — rather than new data collected for this paper.
• The authors themselves acknowledge that designing for presence is inherently difficult and does not scale like a quiz: oral defenses and portfolios require time, funding, and professional support that many institutions have cut, and the paper offers no cost or workload evidence.

Citation

Bozkurt, A., Crompton, H., & Fell Kurban, C. (2026). The Devil is in the Details: AI agents, ghost students, and the crisis of verified presence in an agentic AI world. Open Praxis, 18(1).