A 10-method ensemble analyzer that detects AI-generated content through perplexity analysis, burstiness scoring, linguistic patterns, hallucination detection, and even adversarial evasion attempts, with detailed explanations for every verdict.
AI writing tools got good fast. Students discovered that ChatGPT could produce passable essays, and overnight, faculty lost the ability to distinguish student work from AI output. Commercial detectors emerged but they’re black boxes, they give a percentage and no explanation. Faculty can’t tell a student “this was flagged as AI-generated” if they can’t explain why.
False positives ruin trust. False negatives enable cheating. And students quickly learned evasion techniques: paraphrasing tools, strategic typos, thesaurus substitution, even mixing Cyrillic characters into Latin text. A single detection method is easy to fool.
The arms race between generation and detection needed a different approach, not one better detector, but many independent detectors working together.
Commercial detectors say ‘87% AI’ with no explanation. Faculty can’t justify academic integrity decisions with an opaque percentage.
Any one detection technique can be fooled. Paraphrase tools, humanizers, and strategic editing beat simple detectors.
Students use homoglyph substitution, thesaurus abuse, strategic typos, and filler injection to dodge detection tools.
Wrongly accusing a student of AI use is devastating. Faculty need confidence levels and explanations, not just scores.
Built 10 independent detection methods, each analyzing text from a completely different angle. No single method is reliable alone, but their combined signal is strong. Weighted ensemble scoring produces a final classification.
Two methods use GPT-2 as a reference model: Perplexity Detection measures how predictable the text is (AI text is more predictable), and Zero-Shot Detection analyzes log-likelihood distributions without any training data.
Four methods examine writing patterns: N-gram frequency, POS distribution, function word usage, sentence complexity, syntactic structure, voice patterns, and readability metrics across 5 scales (Flesch, Gunning Fog, Coleman-Liau, SMOG, ARI).
Two specialized methods detect evasion attempts (homoglyphs, strategic typos, thesaurus abuse, filler injection, copy-paste boundaries) and AI hallucination patterns (fabricated citations, anachronisms, AI self-revelation phrases, unsourced statistics).
Let’s talk about how a multi-method AI detection system could work for your institution.
Start a ConversationPerplexity, Burstiness, N-gram, Linguistic, Stylometric, Readability, Pangram, Hallucination, Zero-Shot, and Adversarial, each analyzing text from a unique angle.
Every verdict comes with a multi-section report: overall assessment, per-detector findings, key insights, and specific recommendations. Faculty can understand and explain every flag.
Catches fabricated citations (future dates, generic authors, mixed styles), anachronisms, AI self-revelation phrases, and unsourced precise statistics.
Identifies homoglyph substitution (Cyrillic in Latin text), known ‘humanizing’ typos, thesaurus abuse density, filler injection, and copy-paste boundaries.
Enable or disable individual detectors, adjust weights, and create custom detection profiles tuned to specific use cases or document types.
AI (0-35%), Uncertain (35-65%), Human (65-100%). The ‘Uncertain’ zone is deliberate, it prevents false positives and tells faculty when they need to investigate further.
The comprehensive analyzer running all 10 methods on a text sample, showing individual detector scores and the weighted ensemble verdict.
A multi-section explanation showing per-detector findings, key insights, and the reasoning behind the final classification.
The adversarial detector flagging homoglyph substitution and thesaurus abuse patterns in a tampered text sample.
Other tools just give us a percentage. This one tells us why it thinks something is AI-generated, which specific patterns it found, which methods agree, and where the text is suspicious. That’s the difference between ‘I think you cheated’ and ‘here’s what the evidence shows.’
Binary AI/Human classification is irresponsible. The 35-65% ‘Uncertain’ band explicitly tells faculty ‘we don’t know, investigate further.’ This prevents the false positives that destroy student trust and protects institutions from wrongful accusations.
When evasion techniques are detected, the text might still be human-written (some people genuinely use thesaurus words). Instead of affecting the AI/Human score, adversarial detection lowers the confidence, telling faculty ‘something unusual is happening, but we’re less sure about the overall verdict.’
A 95% accurate black box is less useful than an 85% accurate system that explains its reasoning. Faculty need to understand why text was flagged so they can make informed decisions and have defensible conversations with students.
Tell us about your academic integrity challenges and let’s explore what a multi-method, explainable detection system could look like for your institution.
No pitch. No pressure. Just a conversation about what might work.