Everything you need to know about our AI detection engine — how it works, how accurate it is, and what has changed in each update.
Three sections — one complete picture.
Understand our three-layer detection engine: predictability scoring, linguistic analysis, and neural classification.
See how the detector performs across AI models, text lengths, and content types — with honest limitations documented.
Track improvements to the detection engine — from the latest accuracy tuning to new features like model attribution.
Each layer targets a different aspect of how AI-generated text differs from human writing. All three signals are combined into a single calibrated probability.
An algorithm evaluates how predictable the text is. AI text follows high-probability patterns; human writing is naturally more varied and surprising.
Over 50 features are extracted from the text — readability, sentence structure, vocabulary diversity, phrase repetition, and writing style patterns that differ between human and AI writing.
A neural network combines both signals into five output categories — Human, GPT, Claude, Gemini, or 'Other AI' — plus a sentence-level heatmap and overall AI probability.
Common questions about the GetSolved AI Detector.
Submitted text is processed in memory and is not permanently stored after the scan completes. The downloadable PDF report is generated on the fly and is not retained on our servers.
The engine is updated regularly — typically every 1–2 months. Each update is documented on the Engine Updates page with a full changelog of what changed and why.
Heavily paraphrased AI text may lower the overall AI probability, but the sentence-level heatmap can still identify suspicious patterns. No AI detector is immune to adversarial rewriting, and we document this transparently in our Known Limitations section.
When the AI probability falls between approximately 30% and 70%, the result is in the borderline zone — the detector is less certain. We recommend reviewing the sentence heatmap and model attribution for more context rather than relying on the headline number alone.
Every engine update ships with updated documentation. Accuracy figures are measured on held-out evaluation sets. Limitations are documented honestly — not buried.
