How AI Detectors Actually Work
Most AI detectors use two primary signals:
Perplexity measures how "surprising" the text is. If a language model would have easily predicted each word given the words before it, perplexity is low. AI text has characteristically low perplexity because the model produces the statistically expected output. Human text is less predictable — we use unexpected words, make unusual connections, and sometimes go in directions a model wouldn't predict.
Burstiness measures how much sentence length varies. Human writing has high burstiness — we naturally mix short punchy sentences with long flowing ones. AI produces more uniform sentence lengths because it's optimizing for clarity and completeness at each step.
Some detectors additionally measure:
- Repetition of specific phrases associated with AI training data
- Syntactic patterns (passive voice frequency, subordinate clause structure)
- Semantic coherence (AI text is often "too" coherent — humans digress)
Understanding this tells you immediately which approaches will fail.
What We Tested
We ran 50 text samples (academic essays, blog articles, and emails) through 7 different humanization approaches, then tested each output against GPTZero, Originality.ai, Copyleaks, Winston AI, and ZeroGPT.
Approach 1: Synonym replacement only
Average detection rate: 78% still flagged as AI. The sentence structure and perplexity patterns are unchanged by synonym swapping. Doesn't work.
Approach 2: Manual paraphrasing, same structure
Average detection rate: 61% flagged. Better, but still high. Rephrasing without restructuring preserves the underlying patterns.
Approach 3: Adding transitional variety
Average detection rate: 52% flagged. Swapping "Furthermore" for "Also" and varying transitions helps modestly. Not enough on its own.
Approach 4: Sentence length variation (manual)
Average detection rate: 31% flagged. Aggressively varying sentence length — the single most impactful individual change — drops detection substantially.
Approach 5: Full structural rewrite (manual)
Average detection rate: 8% flagged. When you restructure at the paragraph level, vary rhythm, add specific details, and remove AI phrasing patterns, results are very good. Time cost: 15-20 minutes per 500 words.
Approach 6: Dedicated humanizer tool (basic)
Average detection rate: 22% flagged. Basic AI humanizers that mainly swap synonyms and paraphrase. Not enough.
Approach 7: Dedicated humanizer tool (structural)
Average detection rate: 4% flagged. Humanizers that rewrite at the structural level — restructuring sentences, varying rhythm, changing syntactic patterns — produce results comparable to full manual rewrites at a fraction of the time.
The Approaches That Actually Work
Based on our testing, these are the changes with the highest impact on AI detection scores:
1. Sentence length variation (highest impact)
Mix very short sentences (3-7 words) with longer ones (20-35 words). Don't let three consecutive sentences be the same length. This is the single change with the biggest effect on burstiness scores.
2. Structural reorganization at paragraph level
Don't just rephrase sentences. Restructure how information flows through the paragraph. Merge two sentences into one, split one long sentence into three short ones, move the main point from the end to the beginning of a paragraph.
3. Remove AI phrasing signatures
Certain phrases are statistically overrepresented in AI output. Removing them reduces the "fingerprint": "It is important to note", "Furthermore", "In conclusion", "It is worth mentioning", "This demonstrates", "It can be seen that".
4. Increase lexical unpredictability
Use a word that's technically less common but perfectly appropriate. Write "the data doesn't add up" instead of "the results are inconclusive." The unexpected but correct word choice increases perplexity.
5. Add human markers
Contractions, rhetorical questions, short fragmentary sentences for emphasis. "Honestly." "Not great." "This matters more than people realize." These patterns are rare in AI output and common in human writing.
Detector-Specific Notes
GPTZero is most sensitive to perplexity and burstiness. Sentence length variation and structural rewrites have the biggest effect here. Currently the most widely-used detector in academic settings.
Originality.ai v3 is more sophisticated — it uses its own trained model rather than just perplexity scoring. It's harder to bypass with simple changes. Structural rewriting at the paragraph level is necessary, not optional.
Copyleaks focuses heavily on syntactic patterns and is particularly good at detecting passive voice overuse and formal transition phrases.
Winston AI is commonly used in professional content contexts. It's more lenient overall, and sentence variation alone often reduces scores substantially.
ZeroGPT is the least reliable of the major detectors and the easiest to bypass. Don't treat a clean ZeroGPT score as meaningful.
Sapling focuses on grammatical patterns associated with language model output. Contractions and informal phrasing have an outsized effect here.
The best strategy is to aim for results that pass the two hardest detectors — GPTZero and Originality.ai — since anything that passes those will pass everything else.
The Honest Limitation
No method is 100% reliable, and detector technology is improving continuously. What works today may need adjustment in 12 months.
The more sustainable approach is to focus on writing that is genuinely readable and natural for humans — not text engineered specifically to fool detectors. When text reads naturally, it passes detectors as a byproduct. When it's engineered specifically to manipulate scores, it often reads oddly to human readers even when it fools the algorithm.
The goal should be: would a human reader believe a human wrote this? If yes, the detectors will almost certainly agree.
Try it yourself
Test your own text with Humanified — see your AI detection score drop in real time.
Try Humanified free →