PenForensic

PenForensic: Academic Integrity Analysis for Educators PenForensic is a Chrome extension that helps university and TAFE educators assess whether student submissions may contain AI-generated content. It is a decision-support tool designed to inform your professional judgement, not replace it. PenForensic does not deliver verdicts. It provides forensic indicators backed by peer-reviewed research so you can make better-informed decisions while following your institution's academic integrity procedures. HOW IT WORKS Open PenForensic from the Chrome side panel, select the student's academic level, paste the submission text, and run the analysis. The tool processes the text through 10 independent forensic markers running concurrently, then presents a structured report with an overall suspicion score, a confidence assessment, and detailed evidence for each marker. You can optionally provide a sample of the student's known previous writing as a baseline. When a baseline is provided, PenForensic generates a Writing DNA comparison that measures how closely the submission matches the student's established writing style across 11 dimensions. 10 FORENSIC MARKERS GROUNDED IN PEER-REVIEWED RESEARCH Every marker in PenForensic is grounded in published, peer-reviewed research. The full citations are displayed inside the extension so you can verify the research yourself. 1. Probability Curvature: Estimates whether the text sits at a probability maximum, a pattern characteristic of machine-generated content where small word changes cause disproportionate drops in fluency. 2. Dual Model Contrast: Evaluates text from two contrasting analytical perspectives, assessing naturalness versus originality. When text reads as highly natural but lacks originality, it may indicate AI generation patterns. 3. Generative Fingerprint: Tests whether prompting a language model with excerpts from the submission produces suspiciously similar continuations, and measures statistical bias toward high-frequency tokens. 4. Information Density: Measures how uniformly information is distributed across sentences. AI-generated text tends to maintain unnaturally consistent information density compared to human writing, which naturally varies in complexity. 5. Robustness and Benchmark: Tests whether AI signals persist after the text is paraphrased, and compares vocabulary and structural metrics against human versus AI reference distributions. 6. Intrinsic Dimensionality: Estimates the intrinsic dimensionality of the text's linguistic feature space. Human writing typically occupies a higher-dimensional manifold than AI-generated text, meaning humans use a more diverse combination of linguistic features. 7. Psycholinguistic Cognitive: Looks for traces of genuine cognitive processes in the writing, such as self-monitoring markers, uncertainty expressions, parenthetical asides, and natural variation in syntactic complexity. 8. Semantic Coherence Drift: Measures how topic similarity between sentences evolves across the text. Human writing naturally drifts in coherence as writers develop arguments and occasionally go on tangents before refocusing. AI text tends to maintain unnaturally steady coherence throughout. 9. Stylometric Fingerprint: Profiles the distribution of function words and part-of-speech patterns, then compares them against known AI writing profiles using statistical divergence measures. 10. Linguistic Error and Fairness: Identifies ESL error patterns such as article omission and agreement mismatches as signals of human authorship, because AI rarely produces these specific errors. This directly addresses the documented fairness problem in AI text detection, where non-native English speakers are disproportionately flagged by other tools. CONFIDENCE AND TRANSPARENCY PenForensic does not give you a single number and leave you to guess what it means. Every analysis includes: - A confidence tier (High Confidence, Moderate Confidence, Uncertain, or Insufficient Evidence) that tells you how much weight to place on the results. - A marker agreement score showing what percentage of the 10 markers agree with each other. When markers disagree, the tool tells you. - A false positive risk assessment (Low, Medium, or High) that warns you when the results are more likely to be inaccurate. - Recommended next steps tailored to the confidence level, such as conducting a follow-up conversation with the student or requesting additional writing samples. - An agreement breakdown showing exactly how many markers flagged concern, how many found no concern, and how many were uncertain. SENTENCE-LEVEL HIGHLIGHTING The evidence tab includes a heat map view that colour-codes every sentence in the submission based on individual suspicion signals. You can see exactly which sentences raised concerns and which appeared consistent with human writing. Each highlighted sentence shows the specific signals that triggered the score. WRITING DNA COMPARISON When you provide a baseline sample of the student's known writing, PenForensic builds a stylometric fingerprint for both the baseline and the submission, then compares them across 11 dimensions. The result is a match percentage with a visual breakdown showing where the two writing samples align and where they diverge. This helps you assess whether the submission sounds like the student who supposedly wrote it. FOLLOW-UP QUESTION GENERATOR PenForensic generates targeted follow-up questions you can use in a conversation with the student about their submission. Questions fall into three categories: conceptual depth questions that test understanding, source verification questions that probe knowledge of references, and counter-argument questions that assess critical thinking. Each question includes a rationale explaining why it was generated and a copy button for easy use. STUDENT DATA PROTECTION PenForensic automatically strips personally identifiable information before any text is sent for a…