AI Detection

I dig into the subject…

AI detection tools in 2026 remain imperfect and unreliable for high-stakes decisions. They have improved since the early days of ChatGPT, but independent testing consistently shows they fall short of vendor claims, especially on real-world text.

Accuracy Overview

Vendors routinely claim 95–99%+ accuracy with very low false positive rates (often under 1–2%). Independent benchmarks and real-world testing tell a more modest story:

•  Overall accuracy in mixed or realistic conditions typically ranges from 60–80%, with some studies averaging around 60% across multiple tools.

•  Performance is strong on raw, unedited AI text from older models (often 90%+ for top tools).

•  It drops sharply on paraphrased, edited, or “humanized” AI text — sometimes to 20–63% or lower.

•  Short texts (under 250–500 words) are particularly unreliable.

•  Newer models and adversarial techniques (simple rewriting, adding personal voice, or using humanizer tools) significantly reduce detection rates.

Top-performing tools in recent benchmarks include Originality.ai, GPTZero, and sometimes Copyleaks or Proofademic, but results vary by test set, text type, and whether the content was lightly edited. No tool is consistently dominant across all scenarios.

The False Positive Problem (The Biggest Issue)

This is the most serious flaw. Even low claimed false positive rates (FPR) create major problems in practice:

•  Native English speakers: False positive rates typically range from 1.6–12%, depending on the tool and writing style. Polished, formal, structured, or academic human writing is frequently misflagged.

•  Non-native English speakers (ESL): Rates are dramatically higher. A landmark Stanford study found detectors flagged 61% of real TOEFL essays (written by Chinese students) as AI-generated. In that test, 97% were flagged by at least one detector and 19% by all seven. Follow-up research has shown this bias persists.

•  Other groups disproportionately affected: Neurodiverse writers and those using certain dialects or highly structured styles.

Real-world consequences have been significant. There are documented cases of students facing failing grades, suspensions, scholarship revocation, and severe stress based on detector flags that later proved false (or were overturned on human review). Some universities (over 50 institutions across the US, UK, Australia, and elsewhere by early 2026) have banned, disabled, or strongly discouraged AI detectors for these reasons.

Even a seemingly low 1–5% false positive rate becomes problematic at scale due to the base rate fallacy: if actual AI misuse is relatively low, most flags will be false positives.

Why Detectors Struggle

Current tools primarily analyze statistical patterns such as:

•  Perplexity (how predictable/surprising the text is)

•  Burstiness (variation in sentence complexity)

•  Token probability distributions

These are correlated with AI output but not unique to it. Well-written human text — especially formal, edited, or non-native English — often shares similar statistical properties. Meanwhile, newer LLMs are explicitly trained to produce more varied, human-like output, and simple editing defeats most classifiers.

There is an ongoing arms race: As detectors improve, so do generation and obfuscation techniques.

Inconsistency and Other Limitations

•  Different tools often give wildly different scores on the exact same text.

•  Results can be unstable across slight rewrites.

•  Detectors generally perform worse on creative writing, short-form content, or heavily revised hybrid text.

•  Many lack transparency about their methods or training data.

Practical Recommendations (Especially Relevant to Your Previous Query)

For analyzing something like the Substack article we discussed:

•  Do not rely on any single detector score as definitive proof. A “high AI probability” result on polished philosophical writing is common even for human authors — and a “low” score does not rule out AI assistance.

•  Best approach: Multi-signal analysis — stylistic review (voice consistency, specific references, original metaphors, personal anecdotes), cross-check with the author’s other work and account history, plus any available process evidence.

•  In education or publishing: Use detectors only as a screening/triage tool for obvious raw spam, never as sole evidence for accusations. Always follow up with human judgment.

•  If stakes are high, collect multiple data points and consider the writer’s full body of work rather than isolated pieces.

Bottom Line

AI detectors in 2026 are better than nothing for bulk filtering of low-effort generated content, but they are not reliable enough to serve as authoritative arbiters of authorship on individual pieces — especially nuanced or edited writing. The technology is still maturing in an active arms race, and the costs of false positives (particularly bias against non-native speakers and formal writing styles) remain substantial.

For most thoughtful analysis (like the Substack piece), careful human stylistic and contextual evaluation remains more trustworthy than any detector score. Detectors can provide one weak signal among many, but they should never be treated as conclusive.