Homoglyph-based Adversarial Perturbation of Introductory Computer Science Theory Problems

TEXT ANALYSIS: arXiv cs.CY — Homoglyph-based Adversarial Perturbation of Introductory Computer Science Theory Problems

A. THE DISSECTION

This paper presents itself as an academic integrity measure — a method for perturbing CS homework questions with homoglyphs (e.g., replacing Latin characters with visually identical Cyrillic ones) to prevent AI tools from solving them while preserving semantic meaning for human students.

What it is actually doing: Engineering a last-mile CAPTCHA for homework — a desperate procedural patch on a fundamentally broken assessment model. The paper treats AI-assisted cheating as a technical problem solvable by text mangling, while studiously avoiding the systemic question of what "introductory CS theory" coursework even means when AI can execute algorithmic reasoning at superhuman speed and zero cost.

B. THE CORE FALLACY

The paper operates on the implicit premise that the learning objective of introductory CS theory is for students to perform computations that AI can no longer perform better.

This premise died in 2022. The "lazy student model" the paper references — students using AI to skip homework — is not a behavioral pathology requiring correction. It is the rational response to a structurally obsolete assignment. The students who skip CS theory homework to use Claude are making the correct economic calculation, just as someone who owns a forklift doesn't feel guilty about not hand-loading freight.

The fallacy is confusing the medium of assessment (problem-solving by humans) with the value being assessed (comprehension of computational theory). Homoglyph perturbation solves for the former while leaving the latter — whether CS theory education produces anything economically defensible — completely unaddressed.

C. HIDDEN ASSUMPTIONS

1. Introductory CS theory homework is worth saving. No argument is offered for why students need to manually trace Turing machines or prove pumping lemmas by hand when AI systems demonstrate, execute, and extend this knowledge at scale.

2. Homoglyph perturbation is a durable defense. This is explicitly acknowledged as a lazy student model countermeasure, but provides no model of AI capability trajectory. Current frontier models already handle homoglyph injection — this is a moat measured in months, not years.

3. The relevant threat is student laziness, not institutional obsolescence. The framing positions AI use as individual moral failure (laziness) rather than a structural signal that the work product being assessed has been automated away.

4. Human readability is the bottleneck. The paper assumes human legibility is preserved by homoglyph perturbation. It does not model the scenario where the bottleneck shifts entirely to AI interpretation — where human-readable homework is simply irrelevant to any defensible learning outcome.

D. SOCIAL FUNCTION

This paper occupies the "transition management" and "prestige signaling" register simultaneously.

• Transition Management: It provides faculty and administrators with a procedural response to AI cheating that lets them continue existing course structures without confronting whether those structures should exist. It's institutional anesthesia.

• Prestige Signaling: Publishing on arXiv demonstrates that researchers are engaging with "the AI problem," lending academic cover to a solution that is functionally theater.

The "interactive tool" proposed at the end is the tell. A tool to conveniently apply homoglyph perturbations to homework is the academic equivalent of a company issuing branded fidget spinners during economic collapse — it treats symptom management as strategy.

E. THE VERDICT

This paper is a diagnostic artifact of institutional collapse, not a solution to it. Its existence proves that the assessment layer of higher education has recognized AI capability but is structurally incapable of responding with anything other than resistance to the old model. Homoglyph perturbation is to CS education what the floppy disk icon is to software — a visual relic of a workflow that no longer reflects the underlying reality.

The question the paper refuses to ask: What is the surviving economic purpose of a human being proving pumping lemmas in a world where AI proves novel theorems?

The honest answer this paper cannot deliver is that the purpose is disappearing, and procedural obfuscation of homework problems does not change that.