In 2025 it feels as though everyone—teachers, editors, HR teams, even casual bloggers—has an opinion on AI text detectors. Some swear the tools are flawless lie-detectors; others dismiss them as gimmicks. To cut through the noise, the Merlin AI research team reviewed 2.3 million real-world scans collected over the past year. Their dataset covers school essays, corporate white papers, social posts, and 21 languages. In this FAQ-style post we’ll use that huge sample to debunk six myths we keep hearing. Plain language, concrete numbers, no spin.

Are AI Detectors Always Right?

Myth: “If a detector says ‘AI-generated,’ it must be true.” Reality: Even the best tools miss some AI text and sometimes flag humans. Merlin’s 2024 benchmark puts its overall accuracy at 91 %—excellent, but not perfect. Across 100,000 English essays, Merlin produced: True positives: 46,000 (AI found correctly)

True negatives: 45,000 (human text cleared)

False positives: 4,500 (human text flagged)

False negatives: 4,500 (AI text missed)

That breakdown shows why results arrive as probabilities (“85 % likely AI”) instead of yes/no verdicts. Merlin’s engineers recommend treating any single scan as a strong clue, not legal proof. Always read the highlighted lines and context before making a call.

Will Rephrasing Fool the System?

Myth: “Run the paragraph through a paraphraser and detectors give up.” Reality: Paraphrasing can blur obvious AI fingerprints, but it rarely erases them completely. Merlin retested 50,000 AI-written paragraphs after users put them through popular paraphrase tools. Outcomes: Direct copy: 94 % flagged

Light paraphrase (word swaps): 78 % flagged

Heavy paraphrase (sentence rewrites): 62 % flagged

Why the drop? Paraphrasers raise perplexity and burstiness, two signals detectors track. Yet statistical quirks—repetition patterns, improbable phrase clusters, “too-neat” grammar—often remain. Merlin’s advice: if content integrity matters, run both versions through a detector. Drastic rewriting may lower the score, but it seldom turns a solid “AI” into a firm “Human.”

How Short Is Too Short to Detect?

Myth: “Anything under 100 words is invisible to AI detectors.” ** Reality:** Length matters, but short text can still raise red flags. Merlin’s lab fed the detector tightly sliced samples from the same 500-word AI essay:

Word Count vs. Confidence Score

The rule of thumb Merlin recommends: ≥ 75 words for a “useful” signal,** ≥ 150 words** for high confidence. Below 50 words, scores wobble because there just aren’t enough patterns to judge. Detectors still flag some tweets and email snippets, but treat those results cautiously.

Do AI Detectors Store My Text?

Myth: “When I paste a document, the company keeps it forever.” Reality: Merlin’s policy deletes plain-text content within 24 hours of scanning. During that window the text is encrypted, hashed, and used only to compute metrics. No sentences enter public databases or search indexes. The system logs anonymous feature vectors—statistical fingerprints stripped of readable words—so engineers can improve accuracy without seeing user content. Merlin also passed an independent GDPR audit in April 2025. Translation: the tool can’t be used later to accuse you of copying your own work, and it won’t leak sensitive data. Always read a vendor’s privacy page, but blanket “AI detectors steal your text” claims don’t match Merlin’s documented practice.

Is Detection Limited to English?

Myth: “AI detectors break down outside English.” Reality: Language coverage varies by vendor, yet Merlin’s 2025 multilingual benchmark shows solid performance in 21 languages. The detector’s F1-score (a balance of precision and recall) stays above 0.80 in Spanish, French, German, Hindi, and Indonesian. Accuracy does dip in lower-resource tongues like Zulu or Icelandic, where training data are scarce. Still, the tool rarely defaults to wild guesses. Merlin flags results with a language-confidence badge—green for well-supported languages, yellow for limited-data ones—so users know when to double-check. If you write in a niche language, expect a wider margin of error, but the myth that detectors are “English-only toys” is outdated.

What Does Merlin’s Data Actually Show?

Here are five headline stats drawn from Merlin’s 2.3 million-document corpus:

Combined, these numbers debunk the loudest myths. Detectors aren’t perfect, but neither are they guess-machines or privacy traps. They work across languages, spot many paraphrases, and publish clear confidence scores you can evaluate at a glance.

Conclusion: Separating Hype from Hard Numbers

AI text detectors inspire bold claims—some glowing, others dire. Merlin’s multi-million-scan dataset shows the truth is comfortably in the middle. Accuracy hovers around 90 %, paraphrases still leave footprints, short text yields weaker but usable signals, and robust privacy policies stop your words from lingering online. Next time you see sensational headlines about detectors being “always right” or “easily fooled,” check for real metrics. Ask for the AUC, look at perplexity and burstiness gaps, and read the vendor’s deletion policy. Myths fade fast when data walks in the door.