How AI detection works

What the AI score measures, how it was trained, how accurate it is, and why it is always an estimate that can be wrong about who wrote a text.

SerpPal shows an AI score in many places: the AI checker, the Humanizer, the essay writer, the paraphraser's statistics and the Recent lists. All of them use the same checker. This page explains, in plain words, what that number is and what it is not.

#What the checker looks at

AI models leave patterns in their writing. People leave different ones. The checker measures about 140 of these style patterns, then a trained model weighs them into one number. The patterns fall into a few groups:

• Word choice. Words models favour (like "seamless", "robust", "leverage", "delve") and everyday words people use that models tend to avoid.
• Sentence rhythm. People write sentences of very uneven length. Models often write runs of sentences that are about the same length, or many short "punchy" ones.
• Linking words. Models lean on words like however, these, each, overall and additionally.
• Stiff transitions. Sentences that open with Furthermore, Moreover or Additionally.
• Participle tails. Endings like ", ensuring that..." and ", allowing them to...".
• Stock phrases. Phrases like "in today's fast-paced world" or "it is important to note".
• Punctuation habits. People use brackets, semicolons, questions and loose punctuation more.
• Contractions. How often "don't", "it's" and "we're" appear.
• Polish. Human text usually has the odd typo, slang word or casual sentence. Text with none at all reads more machine-made.
• Vocabulary variety. How many different words are used, and how long they are.

In the AI checker, the four cards (Word choice, Sentence rhythm, Predictability, Structure) group these measurements. The i button under the cards lists each one with its value.

#How it was trained

The model learned from about 13,000 texts that were labelled as human-written or AI-written: student essays, news, stories, how-to guides, forum answers and encyclopedia articles, written by people and by several popular AI models, plus output from humanizer tools. Older blog posts (written before AI tools were common) and current AI-written blog articles were added so it handles web content well.

The checker runs inside SerpPal. Your text is not sent to an outside detection service.

#How accurate it is

On texts it had never seen during training, the model separates human and AI writing well. At the Likely AI line (60% and up), it flags about 91% of AI-written texts and about 7% of human-written texts.

That second number matters. Roughly 1 in 14 human texts can be flagged as Likely AI. And about 1 in 11 AI texts can slip through as lower. That is why the score is a signal to look closer, not a verdict.

#How the score is made

1. The model works out the probability that the whole text is AI-written.
2. That probability is turned into the 0 to 100% score you see. Clearly human text reads near 0, and a confident AI call reads near 100.
3. The label follows the score: Likely human below 25%, Mixed from 25%, Likely AI from 60%.

#Sentence highlights

Each sentence's color starts from the verdict for the whole text, then moves up or down with that sentence's own signs:

• up for a stock phrase, an AI-typical word, a stiff opening, a participle tail, a list of three, a "Label: text" opening, or the same length as the sentences around it;
• down for words people use more than models, very short sentences, and casual touches like "yeah", "kinda" or a double question mark.

So a sentence in a text that reads as human overall rarely shows red, and a plain sentence in a very AI-like text can still show yellow.

#Other languages

The word-based measurements were learned from English. For other languages, the checker uses only rhythm and punctuation, which is a weaker signal, and it holds the score back to avoid false alarms. Scores for non-English texts are a rougher guide. Always pick the right language in the checker.

#Why a score can be wrong

The score can run high for human writing when the text is:

• very formal, polished or edited for consistency (legal, academic, corporate);
• written in a textbook style, as non-native writers are often taught;
• mostly lists, short definitions or template-like sections;
• very short (under 40 words the checker won't score it at all, and short texts are still less reliable).

The score can run low for AI writing when the text:

• has been edited by a person, or humanized;
• copies a casual voice with slang, asides and questions;
• is in another language.

#How to use the score well

• Compare versions of the same text, not texts from different writers. A drop from 70% to 10% after editing is meaningful.
• Look at the highlights, not only the number. They show you which sentences to work on.
• Use the AI editor's opinion as a second view. If both agree, you can be more confident. If they don't, read the text yourself.
• Different detectors disagree. Each uses its own model, so a text can score differently elsewhere.

#Common questions

Does checking a text use AI words? No. The main check is free. Only AI editor's opinion uses AI words.

Does the Humanizer use the same checker? Yes. The Humanizer checks its own output with this checker and works on the parts that still read as AI. See Humanizer options.

Can I prove to my teacher that I wrote something? A low score helps, but it is not proof either way. Your drafts, notes and version history are much better evidence.