How to Detect AI-Generated Writing

AI detection has become important for publishers, educators, and content teams — but it’s widely misunderstood. Most guides oversell what detectors can do. This one doesn’t.

What AI detectors actually measure

AI content detectors don’t “know” whether a human or AI wrote something. They measure the statistical probability that a piece of text was generated by an AI language model — based on patterns like:

• Perplexity: how predictable the word choices are (AI tends toward high-probability word sequences)
• Burstiness: how much variation exists in sentence length and structure (humans vary more)
• Entropy: the overall randomness and surprise in the text

This means detectors can be wrong in both directions: they produce false positives (flagging human content as AI) and false negatives (missing heavily edited or prompted AI content).

What detectors can and can’t do

Can do	Cannot do
Flag statistically likely AI-generated text	Definitively prove who wrote something
Identify unedited AI drafts with high accuracy	Detect well-edited AI content reliably
Provide sentence-level probability signals	Detect all AI models equally well
Serve as one signal in a content quality review	Serve as sole proof for punitive action

Critical point: No detector should be used as the sole basis for academic penalties, employment decisions, or content rejection. Use them as one signal among several.

The three main tools

GPTZero

Best for: educators and academic integrity

• Sentence-level AI probability highlighting
• Batch document checking
• Free plan: 10,000 words/month
• Strong in academic use cases

Originality.ai

Best for: web publishers and content agencies

• Combines AI detection + plagiarism in one scan
• Credit-based pricing (flexible for agencies)
• Team management features
• No free plan

Winston AI

Best for: teams needing OCR detection (PDFs, images)

• Supports image and PDF uploads via OCR
• Shareable detection reports
• Free plan: 2,000 words/month

How to use detectors correctly

Step 1: Run the scan Paste the text or upload the document into your tool of choice. For GPTZero, the sentence-level highlighting is more useful than the overall score.

Step 2: Look at sentence patterns, not just the score A 65% AI probability score tells you less than seeing which specific sentences are flagged. AI-written paragraphs tend to cluster; human writing with occasional AI assistance shows mixed signals.

Step 3: Read the content manually Detectors miss context. Read the flagged content for:

• Generic phrasing that could apply to anything
• Missing specific examples, numbers, or anecdotes
• Overly uniform sentence rhythm
• Claims without attribution

Step 4: Apply judgment, not automation Detection results are inputs to a human decision, not the decision itself.

How to spot AI writing manually (without tools)

Once you’ve read enough AI-generated text, you develop a feel for the patterns:

Common AI writing tells:

• “In today’s fast-paced world…” type openings
• Conclusions that summarize what was just said in the same paragraph
• Phrases like “it’s worth noting,” “it’s important to highlight,” “in conclusion, it’s clear that”
• Lists that start with the same grammatical structure every time
• Absence of specific facts, dates, or named sources
• No personal perspective or experience — all claims are hedged and generic
• Transitions that feel mechanical: “Furthermore,” “Additionally,” “In addition to”

This isn’t conclusive — good writers sometimes use these patterns, and good AI prompting avoids them. But clusters of these signals warrant closer inspection.

What actually matters for publishers

If you’re a publisher or content manager, what you actually want to know isn’t “was this AI-written?” but rather: Is this content good enough to publish?

Those are different questions. A human can write bad content; an AI-assisted workflow can produce excellent content.

The more useful quality checklist:

• Does this content contain specific, verifiable information?
• Is there original insight or perspective that I couldn’t get from any other source?
• Are all factual claims accurate and sourced?
• Is the tone consistent with our brand voice?
• Would a reader find this genuinely useful?

If the answer to these is yes, the question of AI involvement matters less than it might seem.

For educators: a more honest framing

Detection tools are more useful as a deterrent than as a punishment mechanism. The existence of AI detectors discourages the most casual AI misuse — submitting an unedited ChatGPT response.

But any student willing to spend 20 minutes editing and personalizing AI output will likely evade most detection. The better response is redesigning assessments to require demonstration of genuine understanding — presentations, in-class writing, project defenses — rather than policing submittable text.

See our GPTZero vs Originality.ai comparison for a detailed breakdown of the two most popular tools.