Why AI Detectors Flag Your Writing
AI detectors don't "read" text the way a teacher does. They measure two statistical properties: perplexity (how predictable each word choice is) and burstiness (how much sentence length varies). Large language models like ChatGPT produce text with very low perplexity and very low burstiness — every sentence tends to be a similar length, every word choice tends to be the most probable option.
Human writing is messier. We write long, winding sentences and then short ones. We use contractions, idioms, and the occasional awkward phrase. That variation is exactly what detectors look for — and what raw AI output lacks.
The key insight: You don't need to write the whole piece yourself. You need to make the AI's output look like it was written by a human who was thinking in real time — not a model completing a probability distribution.
What Turnitin's AI Detection Actually Does
Turnitin launched AI detection in April 2023. It analyzes the writing segment by segment, scoring each block 0–100% AI probability. Anything above 20% AI-written content in a submission generates an instructor report.
The version active in 2026 is significantly more accurate on raw ChatGPT output than it was at launch. It's also trained on humanizer-tool output from early 2023-era tools — which means older approaches (simple synonym swapping, adding filler phrases) no longer work reliably.
What still works: restructuring sentences at the syntactic level, not just replacing words. When the sentence structure itself changes — different clause ordering, different connective words, passive to active or vice versa — the perplexity profile changes enough to fall below detection thresholds.
Step-by-Step: How to Make AI Text Pass Detection
Step 1 — Generate your first draft with AI
Use ChatGPT, Claude, or any other tool to produce a full draft. Don't worry about detection at this stage — just get the content right. Keep it under 1,000 words for best results with free tools.
Step 2 — Choose the right tone
Before rewriting, decide what voice the final piece needs. Academic work benefits from a professional or authoritative tone. Blog posts and emails read better with a casual or natural tone. Matching tone to context matters because it affects word choice in a way that feels authentic, not pasted-on.
Step 3 — Run it through an AI Writing Polisher
Forgely's AI Writing Polisher is free, requires no account, and rewrites up to 1,000 words. It restructures sentences at the syntactic level — not just synonyms — using Claude to produce output with higher perplexity and more varied sentence lengths. Paste your text, choose a tone, and hit Polish.
Step 4 — Spot-check with a detector
Run the polished output through Forgely's AI Detector or GPTZero before submitting. Look for sentences still flagged as high-confidence AI — these are usually the ones with very even clause structure or repetitive transitional phrases ("Furthermore,", "In conclusion,", "It is important to note that"). Rewrite those specific sentences manually.
Step 5 — Add one genuine human paragraph
This is optional but effective. Write one paragraph yourself in your natural voice — an example from your own experience, a specific opinion, or a concrete detail the AI wouldn't know. It's not about fooling the detector; it's about producing work that's actually yours. One authentic paragraph pulls the whole document's detection score down significantly.
Polish AI text for free — no account needed
Paste up to 1,000 words, choose your tone, get human-sounding output in seconds.
Try AI Writing Polisher Free →Which Detectors Are Hardest to Beat?
Originality.ai is currently the most aggressive detector on the market. It uses ensemble models and updates frequently. Even well-rewritten text occasionally scores 10–20% AI there. If you're writing for a client who uses Originality.ai specifically, plan to do more manual rewriting than you would for GPTZero or Turnitin.
GPTZero is widely used in academic settings and is reasonably accurate on raw AI output. After one round of polishing with Forgely, most text scores under 15% AI on GPTZero.
Turnitin is the most commonly encountered in schools. Its detection is integrated into the same plagiarism report students already see, which means instructors often encounter it without specifically looking for it. The threshold that triggers a report is 20% AI-written content.
What Doesn't Work Anymore
- Synonym swapping — replacing words with near-synonyms doesn't change perplexity enough to matter
- Adding filler transitions — "Additionally,", "Moreover,", "It is worth noting that" — these phrases are themselves high-probability AI patterns and actually make detection worse
- Copying into Google Docs and asking it to rephrase — Google's built-in tools don't have the depth to restructure syntax meaningfully
- Early-generation humanizer tools (2023) — most detectors have trained on their output patterns and now flag them specifically
Frequently Asked Questions
Yes. Turnitin's AI detection has been active since 2023 and flags text with high statistical regularity — a hallmark of LLMs. It scores 0–100% AI probability. Rewriting to add sentence variety, contractions, and personal voice dramatically reduces the score.
GPTZero uses perplexity and burstiness scoring. Raw ChatGPT output scores very high. After rewriting with a tool like Forgely's AI Writing Polisher, most text scores under 20% AI probability because the sentence structure and vocabulary entropy increase significantly.
Forgely's AI Writing Polisher is free for texts up to 1,000 words with no signup required. Paste your AI text, choose a tone, and get human-sounding output in seconds.
AI detectors look for low 'perplexity' (predictable word choices) and low 'burstiness' (uniform sentence length). Human writing has higher variation in both. Rewriting introduces contractions, varied sentence lengths, and less predictable phrasing — raising both scores above detection thresholds.