A content manager submits a 1,500-word article. The client's editor runs it through Originality.ai. It comes back 94% AI. The piece gets rejected, even though a human edited every paragraph, rewrote the introduction twice, and added three original case references.
This happens every week in B2B content workflows. The problem is not that AI was used. The problem is that the writing still carries the statistical fingerprint of AI output, and detection tools are trained to find exactly that fingerprint.
This article covers how those detectors work, which writing patterns trigger them, and five techniques that address the actual signals, not the surface symptoms. The goal is not to deceive anyone. The goal is to write content that reads and scores like human writing, because it has been genuinely rewritten to that standard.
How AI detectors actually work
Two signals drive most detection scores. Understanding them changes how you approach the fix.
Most AI detectors run on two core measurements: perplexity and burstiness. Every other signal is secondary to these two.
Perplexity measures how predictable word choices are. Language models predict the most probable next token at each step. When a model writes, it tends to choose high-probability words in high-probability sequences. Detectors flag text where word choices are consistently predictable, because that pattern correlates with model output rather than human writing.
Burstiness measures sentence length variation. Human writers naturally shift between very short sentences and long, clause-heavy ones. AI output tends toward uniform sentence length, clustering around 18 to 22 words per sentence. Detectors score low burstiness as a strong AI signal.
Secondary signals include repetitive transitional phrases, symmetrical paragraph structure, and the absence of hedging or contradiction. These patterns reinforce the primary scores but rarely drive a flag on their own.
How a detection score is calculated
Input Text
raw content submitted for analysis
Perplexity Scorer
measures word choice predictability
Burstiness Analyzer
measures sentence length variation
Pattern Classifier
detects structural and phrase-level signals
Combined Confidence Score
weighted output from all three analyzers
Flag / Pass
threshold varies by platform and setting
What detectors actually measure
Detectors do not detect AI. They detect statistical patterns that correlate with AI output. That distinction changes how you approach the fix. You are not trying to hide AI use. You are trying to produce text whose statistical properties match human writing.
Why "humanizing" tools mostly fail
Tools like Undetectable.ai address surface patterns. Enterprise detectors are now trained on their output.
A category of tools claims to rewrite AI text so it passes detection. Undetectable.ai, Quillbot, and similar products work by substituting synonyms and shuffling sentence order. This changes perplexity scores slightly, because different word choices produce different token probabilities.
The problem is structural. Synonym substitution does not fix uniform sentence length. It does not add genuine opinion or contradiction. It does not break paragraph symmetry. The underlying statistical fingerprint survives the surface rewrite.
Enterprise-grade detectors have adapted. Originality.ai and Winston AI now include classifiers trained specifically on humanizer output. Running content through Undetectable.ai and then through Originality.ai often produces a higher AI score than the original, because the humanizer output matches a known pattern.
Humanizer tools introduce factual errors
Some humanizer tools paraphrase claims during synonym substitution and change the meaning of specific facts, statistics, or named references. Always verify every factual claim after running content through any rewriting tool. Do not assume the output says what the input said.
The structural patterns that trigger detection
These are the exact patterns AI models default to. Detectors are trained on all of them.
Before you can fix the problem, you need to recognize it in your own drafts. These patterns appear in GPT-4, Claude, and Gemini output at high frequency. They are not bugs. They are the result of training that rewards clarity, completeness, and predictability.
Run any AI-generated draft through this list before you run it through a detector. If you find five or more of these patterns, the draft will flag on most enterprise tools.
Why GPT-4 defaults to these patterns
GPT-4 and similar models are trained using reinforcement learning from human feedback (RLHF). Human raters reward responses that are clear, complete, and easy to follow. That feedback loop trains the model to produce text that is optimized for comprehension, not for sounding like a specific human voice.
The result is a writing style that is maximally legible. Sentences are a consistent length because consistent length is easy to read. Transitional phrases appear because they signal logical connection explicitly. Paragraphs follow a predictable structure because predictable structure is easier for raters to evaluate.
The model is not trying to sound like AI. It is trying to be understood. Those two things produce the same statistical output, and detectors are trained on exactly that output.
This is why prompting the model to "write more naturally" rarely works. The model's definition of natural is its training distribution. You need to give it explicit structural constraints that force it outside that distribution.
The five writing techniques that actually work
Each technique maps to a specific detection signal. Apply all five to a draft before testing.
Introduce sentence length chaos
Deliberately vary sentence length across every paragraph. Write one two-word sentence. Then write a longer one that builds on a specific detail or example you just introduced. Break the rhythm. Burstiness scores improve when the standard deviation of sentence lengths increases. Aim for at least three sentences under eight words and two sentences over 30 words per 500-word section.
Remove transitional scaffolding
Delete phrases like "This means," "As a result," and "In other words." Let the logic connect without signposting. Readers follow cause and effect without labels. Removing these phrases also raises perplexity scores because the model defaults to them at high probability. Their absence makes the text statistically less predictable.
Insert a genuine opinion or contradiction
State something you actually believe, or acknowledge a case where your advice fails. Write it in first person. Detectors score low on content that takes a specific position, because models trained on RLHF are optimized to avoid controversy. A sentence like "I think this approach fails for technical content above 2,000 words" introduces a pattern the model was not rewarded for producing.
Use specific, non-round numbers
"73% of content managers" reads differently than "most content managers." Specific numbers raise perplexity scores because they are less predictable than round figures or vague quantifiers. They also signal that a human did reporting or research. Use specific numbers for statistics, time estimates, word counts, and any other quantifiable claim in the piece.
Break paragraph symmetry
Write one two-sentence paragraph. Follow it with a seven-sentence one. Vary the rhythm at the structural level, not just the sentence level. Paragraph length variation is a secondary burstiness signal that many practitioners miss. A piece where every paragraph is four to six sentences long will flag even if sentence length varies within each paragraph.
Sentence length variation
The original AI output clusters sentences at similar lengths. The rewrite introduces short and long sentences in the same paragraph.
Removing transitional scaffolding
The original uses three transitional phrases in four sentences. The rewrite removes all of them and lets the logic connect directly.
Inserting opinion and contradiction
The original avoids taking a position. The rewrite states a specific opinion and acknowledges where the advice breaks down.
Prompting AI to write less detectable content from the start
Build anti-detection constraints into the prompt. Reduce the editing work before the draft exists.
The five techniques in the previous section apply to editing an existing draft. This section covers how to generate a draft that requires less editing, by building the structural constraints directly into the prompt.
The prompt below forces the model to vary sentence length, avoid transitional scaffolding, include specific numbers, and break paragraph symmetry from the first output. It does not guarantee a passing score on every detector. It reduces the editing time required to reach one.
Prompt: Generate low-perplexity-resistant content
Claude / GPT-4You are writing a [content type] about [topic] for [audience]. Follow these constraints exactly: 1. Vary sentence length aggressively. Include at least three sentences under eight words. Include at least two sentences over 30 words. Do not cluster similar-length sentences together. 2. Do not use these transitional phrases: "It's worth noting," "This means that," "In other words," "As a result," "Additionally," "Furthermore," "It's important to." 3. Include at least one specific opinion stated in first person. It should be a position a cautious writer might avoid. 4. Use at least two specific numbers that are not round figures. 5. Write one paragraph that is two sentences or fewer. Write one paragraph that is six sentences or more. 6. Do not summarize at the end of each section. End sections by extending the idea, not restating it. 7. Contradict or qualify one claim you make earlier in the piece. Topic: [INSERT TOPIC] Audience: [INSERT AUDIENCE] Tone: [INSERT TONE] Word count: [INSERT COUNT]
Why these constraints work
These constraints force the model outside its default optimization target. You are not asking it to sound human. You are giving it structural rules that produce higher perplexity output. The model follows explicit instructions more reliably than it follows vague style directions like "write naturally" or "vary your sentences."
How different detectors score differently
Originality.ai, GPTZero, Winston AI, and Copyleaks use different methods. Your client probably uses one of these four.
Not all detectors are the same. A piece that passes GPTZero may flag on Originality.ai. A piece that scores 30% AI on Winston may score 70% on Copyleaks. Before you test, find out which detector your client or platform uses. Then calibrate against that specific tool.
Originality.ai
Primary use
SEO agencies, content publishers
Detection method
Perplexity + pattern classifier
False positive rate
Higher on technical content
API available
Yes
Used by
Content agencies, SEO platforms
Key behavior
Now includes classifier trained on humanizer tool output
GPTZero
Primary use
Education, editorial review
Detection method
Perplexity + burstiness
False positive rate
Moderate; higher for non-native English
API available
Yes
Used by
Academic institutions, some media outlets
Key behavior
Sentence-level highlighting shows which passages triggered the score
Winston AI
Primary use
Enterprise compliance
Detection method
Multi-model ensemble
False positive rate
Lower than single-model tools
API available
Yes
Used by
Legal teams, compliance departments
Key behavior
Scores multiple AI models separately, not just a combined figure
Copyleaks
Primary use
Plagiarism and AI combined check
Detection method
Hybrid: plagiarism + AI pattern
False positive rate
Variable; depends on content type
API available
Yes
Used by
Publishers, HR teams screening applicants
Key behavior
Runs AI detection alongside source matching in a single pass
No detector has published a peer-reviewed validation of its accuracy claims. Treat scores as signals, not verdicts. A 78% AI score on Originality.ai does not mean 78% of the content is AI-generated. It means the content's statistical properties match AI output at a confidence level the tool's model assigns to that score range.
The tools themselves acknowledge this. GPTZero's documentation states explicitly that its output should not be used as the sole basis for a decision about content authenticity.
Testing your content before delivery
A repeatable process for testing against multiple detectors and interpreting conflicting scores.
Run every AI-assisted piece through at least two detectors before delivery. Single-detector testing misses the variance between tools. A piece that passes one tool and fails another tells you something specific: you are close to the threshold, and the client's specific tool matters.
Pre-delivery testing workflow
Draft content
AI-assisted or fully AI-generated
Structural edit
Apply all five techniques from Section 5
Test: Originality.ai
Primary detector, most common in agency workflows
Test: GPTZero
Secondary detector, common in editorial and media
Compare scores
Both above 60%? Return to structural edit
Scores conflict?
Check which detector your client uses specifically
Deliver
With disclosure if required by contract terms
Pre-delivery content review
The false positive problem and how to handle client disputes
Human-written content gets flagged regularly. Here is how to dispute a detection result professionally.
False positives are not rare. A 2023 study from Stanford's NLP group found that GPTZero flagged human-written text as AI-generated at rates between 10% and 30% depending on the writing style and domain. Non-native English speakers face a higher false positive rate. One study found GPTZero flagged non-native English academic writing as AI at significantly higher rates than native English writing with equivalent content quality.
This matters for client disputes. If a client rejects a piece based solely on a detector score, they are acting on a probabilistic signal that the tool itself does not claim is definitive proof.
How to respond to a detection-based rejection
If a client disputes a piece based on a detector score alone, ask them to identify the specific passages they believe are AI-generated. Detectors produce scores, not evidence. A score is not admissible proof of anything. Ask for the passages. If they cannot identify specific passages, the dispute is about a number, not about the content.
Template language for disputing a detection-based rejection
Use this language as a starting point when responding to a client who has rejected content based on a detector score. Adjust the specifics to match your contract terms and the detector they used.
Subject: Re: Content review on [piece title]
Thank you for flagging this. I want to address the detection score directly.
[Detector name] produces a probabilistic confidence score, not a binary determination of AI authorship. The tool's own documentation states that scores should not be used as the sole basis for a content decision. False positive rates on human-edited content range from 10% to 30% depending on writing style and subject matter, based on independent testing by Stanford NLP researchers and others.
I am happy to walk through the specific passages that triggered the score. If you can share the sentence-level breakdown from the tool, I can address each flagged section directly and show you the editing decisions behind them.
If your contract requires a specific score threshold, I will revise the piece to meet that threshold. I want to be transparent: I used AI assistance in the drafting phase and human editing throughout. If your contract prohibits any AI involvement in the drafting process, please let me know and we can discuss scope and timeline for a fully human-drafted version.
I am available to discuss this by phone or video if that would be faster than email.
