AI Detection Glossary 2026 — 60 Essential Terms

A measurement of how surprised a language model is by a piece of text. AI-generated text typically has lower perplexity (more predictable) than human writing (more variation). Detectors compare a candidate text’s perplexity to a reference distribution.

Source: GPT-2 paper, Radford et al. 2019

Variation in sentence length and structure across a passage. Human writing is bursty (short sentences mixed with long); AI tends toward uniform length. Burstiness + perplexity together form the foundation of detectors like GPTZero.

Information-theoretic distance between two probability distributions. Used by detectors to compare token probabilities under candidate text vs typical AI distributions.

The likelihood a language model assigns to each individual token (word fragment) in a passage. Detectors visualize per-token probability heat maps to surface AI patterns.

Raw, unnormalized scores a language model produces before they are converted to probabilities. Some detectors use logit distributions for white-box detection of specific model families.

Measure of randomness/unpredictability. Low-entropy passages (highly predictable) raise the AI-likelihood score; high-entropy passages (creative, unpredictable) lower it.

A detector that has no access to the model weights that may have produced the text. Uses surface features (perplexity, burstiness, n-gram patterns). Examples: GPTZero, Originality.ai.

A detector with access to the originating model’s logits or probabilities. Higher accuracy but only works against specific known models. Rare in commercial tools.

A detector that looks for invisible statistical signatures intentionally embedded by the generating model (e.g. Google SynthID for text). Highest accuracy but only works on watermark-aware models.

A detector that analyzes writing style features (function-word frequency, punctuation patterns, sentence parsers) rather than token statistics. More robust against paraphrasing.

Human-written text incorrectly flagged as AI-generated. The most damaging detector error in academic and hiring contexts. Rates vary 0.5%-12% across major detectors and content types.

AI-generated text incorrectly classified as human. The detection-evasion outcome targeted by paraphrase tools and humanizers.

AI-generated text correctly identified as AI.

Human-written text correctly classified as human.

A 0-100% value indicating the detector’s certainty. Most detectors are NOT calibrated probabilities — a "75% AI" score does not mean 75% chance the passage is AI.

Software that paraphrases AI-generated text to evade detection. Effective against perplexity-based detectors but defeated by stylometric and watermark methods.

Detection-evasion technique using a second language model to rewrite a passage. Reduces detection accuracy by 30-60% across most commercial tools.

Input crafted specifically to fool a detector while preserving meaning. Often involves character substitutions (Cyrillic ‘а’ for Latin ‘a’) or whitespace insertion.

Process of removing or weakening an invisible watermark via paraphrase, translation round-trip, or word-substitution attacks.

Translating AI text into another language and back. Disrupts detection signals but degrades fluency, leaving secondary detectable artifacts.

Google DeepMind’s text watermarking system released 2024. Modifies token sampling probabilities in a way detectable with high accuracy by paired detector.

Industry standard for cryptographic content credentials. Embeds signed metadata about how an asset was created and modified. Adoption: Adobe, Microsoft, OpenAI, Meta as of 2025.

Consumer-friendly name for C2PA metadata. Visible as a small "CR" icon on supporting platforms; click to see provenance chain.

International Press Telecommunications Council standard for embedding rights, source, and provenance data in news media. Updated 2024 to align with C2PA.

ISO 24138:2024 standard for content identification using perceptual hashing. Identifies content even after format conversion.

Watermarking method that biases token sampling toward a pseudorandomly-determined "green list" without changing meaning. Detectable via statistical test on token sequence.

A digital signature attached to a content asset that proves authorship and detects tampering. Used in C2PA Content Credentials.

A neural network trained on text to predict the next token. Modern LLMs (GPT-5, Claude 4.7, Gemini 2.5, Llama 4) range from 7B to 1T+ parameters.

A large model trained on broad data, intended to be adapted to many downstream tasks. The pre-fine-tune state of an LLM.

Additional training of a foundation model on task-specific data. A fine-tuned model produces outputs that may be harder to detect than base-model outputs.

Training method where humans rate model outputs and the model is reinforced to produce preferred responses. Source of much "AI assistant style" detected by tools.

A sampling parameter (typically 0.0-2.0) controlling randomness. Low temperature produces predictable, easily-detected text; high temperature produces creative but error-prone text.

Decoding strategies that restrict token selection to the most likely candidates. Strongly affects output detectability; nucleus sampling (top-p) is the most common in 2026.

AI detection feature integrated into Turnitin academic plagiarism platform. Used by 16,000+ institutions; precision/recall published only at aggregated level.

Institutional rules governing how AI tools may be used in submissions, hiring, or publication. Common 2026 categories: prohibited, disclosed-use, allowed-with-citation, fully-allowed.

Requirement that authors declare AI assistance in a piece of work. Practiced by many academic journals and several major newspapers as of 2025.

Use of AI detectors on resumes, cover letters, and writing samples. Legal exposure rising in 2025-2026 as false-positive impact disproportionate on non-native English speakers.

European Union regulation effective 2025-2027 that requires AI-generated content to be labeled and watermarked under certain conditions.

TP / (TP + FP). Of items the detector labeled AI, what fraction actually were AI. High precision = few false alarms.

TP / (TP + FN). Of all the actual AI content, what fraction the detector caught. High recall = few missed AI items.

Harmonic mean of precision and recall: 2 * (P * R) / (P + R). Single-number summary of detector quality.

Area Under the Receiver Operating Characteristic curve. Measures detector quality across all confidence thresholds.

Whether a detector’s confidence scores reflect actual probabilities. Most commercial AI detectors are poorly calibrated.

A held-out collection of human-written and AI-written passages used to evaluate detectors. Common 2026 benchmarks: GPABench, RAID, AI Text Detect Challenge.

The minimum burstiness expected from genuine human writing. Some technical genres (legal contracts, code documentation) naturally have low burstiness — leading to high false-positive rates.

AI detection on programming source code. Far less reliable than text detection because programming languages are syntactically constrained.

Content with both human and AI sections. Hardest case for detectors; often requires sentence-level rather than document-level analysis.

Detection of AI-generated images (Midjourney, DALL-E, Stable Diffusion). Different methodology from text — relies on artifact detection, prompt forensics, and metadata.

Detection of synthetic speech and deepfake audio. Spectral analysis, formant patterns, ad-hoc breath signatures.

Documented pattern where AI detectors flag writing by non-native English speakers as AI 5x-10x more often than equivalent native-speaker writing. Stanford 2023 paper.

Detectors over-flag certain content types (formal academic writing, technical documentation) and under-flag others (informal social media, dialogue). Source of disparate-impact concerns.

A detector’s ability to maintain accuracy under deliberate evasion attempts.

Consumer-focused AI detector launched 2023 by Edward Tian; combines perplexity, burstiness, and stylometric features.

Commercial AI detector integrated with content marketing workflows. Higher claimed accuracy than free tools but mixed independent benchmark results.

Plagiarism + AI detection platform widely used by enterprises. AI module released 2023, frequently updated.

AI detector with strong performance on academic content per benchmarks; integrates with several university systems.

Software that processes AI text to evade detection. Examples: Undetectable.ai, Quillbot, Stealth Writer. Effectiveness varies by detector.

A directed graph representing the chain of edits and authorship for a piece of content, intended to be queryable in 2026 platforms.

Identifying which specific model (GPT-5, Claude 4.7, Gemini 2.5) produced a piece of content. Active research area; commercial tools approach 60-75% accuracy in 2026.

Detecting whether a model was trained on outputs of another model (e.g. Llama distilled from GPT-4). Important for IP and licensing compliance.

Real-time embedding of provenance metadata as content is generated, vs post-hoc detection. The direction industry standards are moving in 2026.