How It Works
A look inside the GetSolved AI Detector — how three independent analysis layers work together to determine whether a text was written by a human or generated by AI.
Architecture Overview
The GetSolved AI Detector uses a three-stage classification architecture to distinguish human-authored text from AI-generated content — including text produced by ChatGPT, Claude, Gemini, and other large language models. It is not locked to any specific AI model.
Every submitted text passes through three independent analysis layers: a predictability scorer that measures how "surprising" the text reads, a linguistic feature extractor that captures writing style patterns, and a neural network classifier that combines both signals into a final AI probability — along with model attribution (GPT, Claude, Gemini, or "Other AI" for less common models).
Layer 1 — Predictability Scoring
An algorithm evaluates how predictable the text is. For each token, it asks: "given the words so far, how surprising is this next word?" AI-generated text picks high-probability continuations and produces a low score. Human writing is naturally more varied and produces a higher score.
- →Outputs a single overall score for the document
- →Per-sentence scores show which sentences read as most AI-like
- →Per-token scores power the color-coded sentence heatmap
- →Calibrated against a reference baseline so scores are comparable across different texts
Layer 2 — Linguistic Feature Analysis
Over 50 linguistic features are extracted from the text in a single pass. Each feature captures a stylistic signal that tends to differ between human writing and AI output — together they form a high-dimensional fingerprint of the text's style.
- →Readability — Flesch Reading Ease, SMOG, Coleman-Liau and other standard indices
- →Part-of-speech distribution — proportions of nouns, verbs, adjectives, adverbs, function words
- →Syntactic complexity — average sentence length, dependency depth, clause density
- →Lexical diversity — vocabulary richness and repeated-word patterns
- →N-gram repetition — bigram/trigram overlap (AI text tends to recycle short phrases)
Layer 3 — Neural Network Classifier
A neural network combines the Layer 1 predictability score with every linguistic feature from Layer 2 and produces a calibrated probability across five output categories — so the final verdict is informed by every signal at once, not just one.
- →Inputs: Layer 1 predictability score + all Layer 2 linguistic features
- →Outputs: Five class probabilities (Human, GPT, Claude, Gemini, Other AI) that sum to 100%
- →Overall AI probability = 100% − P(Human)
- →Probabilities are calibrated, so the predicted confidence reflects real-world frequency
How Signals Combine
Each layer contributes a different type of evidence. The predictability score captures statistical patterns; the linguistic features capture writing style; the classifier fuses both into a single calibrated probability.
The final output includes an overall AI probability for the entire document, a per-sentence breakdown showing which parts are likely AI-generated, and model attribution indicating which AI model was most likely used.
What you see in the results
A single percentage showing how likely the text is AI-generated. Values above 70% strongly suggest AI content; 30–70% is the borderline zone.
Which AI model most likely generated the text — GPT, Claude, Gemini, or "Other AI" — each with its own confidence level.
A color-coded breakdown showing AI probability for each sentence, so you can see exactly which parts are flagged.
For mixed texts, independent analysis of each section — identifying where human writing ends and AI content begins.
Output Categories
The neural network classifier sorts every text — or every detected segment within a mixed text — into one of five output categories. Each category gets its own probability, and these probabilities always sum to 100%.
The "AI probability" you see in the report is simply 100% − P(Human). The remaining four categories tell you which AI tool most likely produced the text.
Five-class classifier output
Worked Example
Here is a simplified walkthrough of how the detector would analyze a mixed document — one where a student wrote the introduction themselves, then used an AI tool to generate the body paragraphs.
Each layer scores the text independently. The classifier then combines all signals to produce a final result with per-segment analysis.
"In my experience studying environmental policy, I've noticed that local communities often bear the greatest impact of industrial decisions. This essay explores three case studies from my home region..."
"Environmental policy frameworks have evolved significantly over the past two decades. The implementation of regulatory mechanisms has demonstrated measurable outcomes across multiple jurisdictions..."
Final Result
The Borderline Zone
Not every result is a clear "human" or "AI" verdict. Some texts fall in the borderline zone — the range where the detector's confidence is lower and the result should be interpreted with more nuance.
This typically happens with heavily edited AI text, formulaic human writing (such as legal or medical templates), or very short submissions. In these cases, we recommend reviewing the sentence-level heatmap and the per-model confidence breakdown rather than relying solely on the overall percentage.
What to do with borderline results:
- →Check the sentence-level heatmap for specific flagged passages
- →Review the model attribution confidence for each AI model
- →Consider the text length — shorter texts are less reliable
- →Look at the segment analysis if the document is long enough
- →Use the result as one input alongside your own judgment
Confidence Spectrum
The text shows strong human writing characteristics. Low risk of false positive.
The detector is less certain. Review the sentence heatmap and model confidence for more detail.
The text has strong AI patterns. Model attribution indicates which AI most likely generated it.
Mixed-Text Detection
Many documents are not purely human or purely AI — they contain a mix of both. The detector automatically locates the boundaries between sections written by different authors, including stretches of AI-generated content embedded inside an otherwise human-written document.
Each detected segment is then classified independently and gets its own AI probability and model attribution — so an AI-polished conclusion or a manually-rewritten paragraph in the middle of an AI draft is flagged accurately, instead of being averaged out across the whole document. This is especially valuable for academic submissions, essays, and reports where AI assistance is often partial.
How segmentation works
Sentences 1–3 identified as human-written with high confidence.
Sentences 4–8 flagged as AI-generated (GPT) with high confidence.
Each segment receives its own AI probability and model attribution. This ensures that human-written sections are not penalized by the presence of AI content elsewhere in the document.