This is not hypothetical. These are the actual tools in deployment at US universities in 2025–2026.

Accuracy is measured differently by different tests, and adversarial conditions (prompt engineering to evade detection) reduce it significantly.

The practical implication: in a 650-word SOP, a 4% sentence-level false positive rate means a human-written essay can generate 2–3 false flags. This is why detection tool output is treated as grounds for closer scrutiny, not automatic rejection. Carnegie Mellon has formally stated that detection output is "insufficient evidence to conclude that a violation occurred." A conversation with the applicant is required.

This is consistent with what we cover in how admissions officers actually read SOPs — detection is a triage tool, not a verdict.

This is the most serious fairness issue in AI detection right now.

Why does this happen? AI-generated text and ESL writing share some statistical patterns: simpler, more predictable sentence structures; less idiomatic variation; lower entropy in word choice. Detection tools struggle to distinguish between "written by an AI" and "written by someone whose second language is English."

Turnitin's bias is documented at 2–3× the baseline for ESL writers. Documents under 300 words show elevated false positive rates across all tools.

This concern is one reason why major universities like Johns Hopkins and Vanderbilt have disabled AI detection entirely. It is also a direct argument against over-reliance on detection software in admissions decisions.

It is worth noting that some of the most prominent universities have explicitly rejected AI detection tools.

The fact that these institutions have disabled detection software does not mean they are unconcerned about AI use. It means their human readers are the primary detection mechanism. This is arguably harder to fool than software, because a faculty reader who reads 50+ applications per week develops a genuine sensitivity to the AI writing pattern — particularly the absence of technical specificity that characterises real research experience. This is explored further in our article on what universities actually do about AI-written SOPs.

Yes. This is less widely known, but AI detection is actively applied to recommendation letters as well as personal statements.

Slate Technolutions' AI Reader feature — used across many US college recruitment CRM systems — explicitly scans LORs alongside essays. A 2025 peer-reviewed study analysed 600,000+ Common Application letters and found that letter specificity (individualized details about the applicant) strongly correlates with admissions outcomes. AI-generated letters lack this specificity by definition.

Human detection of AI LORs is also reliable. Generic recommendation letters — those that could apply to any applicant — are recognisable without software. A letter that says "She showed excellent leadership qualities and academic diligence" could have been written about anyone in any field. A letter that says "Her redesign of our lab's data pipeline reduced processing time by 40% and directly contributed to our Nature paper" is specific, verifiable, and impossible to fake credibly.

The contrast between a genuine and generic LOR is also explored in our article on the biggest lies students tell in SOPs, which covers the cross-referencing between application components.

Software accuracy ranges from 84–94%. Human pattern recognition in a specific academic field is harder to quantify — but the signals that trigger it are well-documented.