AI Detection Glossary

A technique where imperceptible perturbations are added to inputs (images, text, audio) to fool AI detection models into producing incorrect classifications. Adversarial attacks are used to bypass deepfake detectors and text classifiers, making robustness testing essential for any detection system.

A phenomenon where a generative AI model produces factually incorrect or entirely fabricated information with apparent confidence. Common examples include invented citations, fictitious people, and fake statistics. Hallucinations are a useful detection signal, since human-written text rarely contains verifiably fabricated references.

The practice of embedding hidden, machine-readable signals into AI-generated content at creation time to enable later verification of its AI origin. Techniques include altering token selection distributions for text (e.g., Google's SynthID) and frequency-domain signatures for images. Increasingly mandated by AI regulation frameworks such as the EU AI Act.

An unintended visual imperfection in AI-generated images or video — such as malformed hands, asymmetric features, incoherent text, or unnatural backgrounds. Artifacts are primary signals for both human inspectors and automated classifiers to identify synthetic imagery, though they grow subtler as models improve.

A synthetic voice recording created using AI to mimic a specific person's vocal characteristics. Modern systems require only seconds of reference audio. Detection relies on acoustic analysis of unnatural prosody, spectral artifacts, and statistical anomalies inaudible to humans.

A measure of variation in sentence length and complexity within a text. Human writing naturally alternates between short and long sentences (high burstiness), while AI-generated text tends toward uniform sentence length (low burstiness). A key signal used alongside perplexity in AI text detection tools like GPTZero.

An open technical standard for embedding cryptographically signed provenance metadata into media files, recording creation history, tools used, and editing chain. Co-founded by Adobe, Microsoft, and others. C2PA credentials are tamper-evident — modifying the file invalidates the signature. Adoption is growing in journalism, social media, and hardware cameras.

A machine learning model trained to categorize content as "human-generated" or "AI-generated." Text classifiers analyze statistical patterns in language; image classifiers examine visual features and frequency signatures. No classifier achieves perfect accuracy — all have false positive and false negative rates — and performance degrades as generative models evolve, requiring continuous retraining.

Synthetic media — typically video, image, or audio — in which a person's likeness or voice has been convincingly replaced or manipulated using deep learning. The term combines "deep learning" and "fake." Types include face swaps, voice cloning, and full synthesis. Detection methods analyze facial boundaries, blinking patterns, lighting, and physiological signals.

A generative AI architecture that creates images by learning to reverse a gradual noise-addition process. Starting from pure noise, the model iteratively denoises to produce coherent images guided by text prompts. Used by Stable Diffusion, DALL-E 3, and Midjourney. Diffusion outputs have different artifacts from GAN images, requiring distinct detection strategies.

False information deliberately created and spread to deceive or manipulate. Unlike misinformation (unintentional), disinformation implies malicious intent. AI-generated deepfakes, fake news articles, and fabricated audio have dramatically increased the scale and sophistication of disinformation campaigns.

Technical data embedded in image files by recording devices — camera model, GPS, timestamp, settings. Genuine photographs contain rich EXIF data; AI-generated images typically have absent or software-generated metadata. Useful as a screening signal, though easily stripped or forged.

A deepfake technique replacing one person's face with another's in video or image while preserving original body movements. Modern methods require only a single reference image. Detection relies on boundary artifacts, lighting mismatches, and unnatural eye movement at the face-neck edge.

In AI detection, a false positive occurs when authentic human content is incorrectly flagged as AI-generated. A false negative occurs when AI-generated content is missed and classified as human. Balancing these error rates is a central challenge — aggressive detectors produce more false positives, while conservative ones miss more AI content.

A deep learning architecture with two competing networks — a Generator creating synthetic data and a Discriminator distinguishing real from fake. GANs were the dominant technology behind early deepfakes (StyleGAN, DeepFaceLab). Their outputs leave characteristic artifacts — asymmetric features, wavy backgrounds, frequency-domain fingerprints — that remain key detection signals.

AI systems capable of creating new content — text, images, audio, video, or code — rather than merely analyzing existing data. Encompasses LLMs (GPT, Claude), image generators (DALL-E, Midjourney, Stable Diffusion), and audio/video synthesis tools. The rapid improvement of generative AI is the core driver behind the need for AI content detection tools.

OpenAI's family of autoregressive language models (GPT-3.5 through GPT-4+). GPT models are the most commonly detected AI text type, forming the core of classifiers like GPTZero. ChatGPT normalized AI text generation for mainstream users, triggering widespread deployment of detection tools.

Content combining human-authored and AI-generated components — such as an LLM draft substantially edited by a person. The most challenging category for detection because human editing disrupts statistical AI signals. Studies show light editing can reduce classifier accuracy from above 90% to below 60%.

An AI image editing technique that fills selected regions with plausible content based on surrounding context and optional text prompts. Only part of the image is synthetic, making detection particularly challenging. Forensic methods analyze noise patterns and statistical inconsistencies at region boundaries.

A measure of vocabulary richness in a text — the ratio of unique words to total words. AI-generated text often exhibits lower lexical diversity than human writing, favouring common, high-probability vocabulary. Combined with perplexity and burstiness, lexical diversity analysis helps detection tools identify AI-authored content.

A neural language model trained on massive text corpora using transformer architecture, capable of generating fluent natural language. Examples: GPT-4, Claude, Gemini, LLaMA. LLMs leave statistical traces in outputs — particularly in token probability distributions — that detection classifiers exploit, though signals diminish as models improve.

The ability to critically evaluate media content — including AI-generated material — for authenticity, bias, and intent. As automated detection tools have inherent limitations, media literacy education is essential for empowering individuals to question suspicious content, verify sources, and understand the capabilities of generative AI.

Structured information embedded in digital files describing properties, origin, and history. Types include EXIF (camera data), XMP, IPTC, and C2PA manifests. Genuine photographs carry rich metadata; AI-generated images often have sparse or software-generated metadata. Easily stripped or modified, limiting standalone reliability.

False information spread without deliberate intent to deceive. AI contributes to misinformation when users unknowingly share hallucinated facts, AI-generated images misrepresenting events, or deepfakes believed authentic. Combating AI-enabled misinformation requires both detection technology and media literacy education.

A computational model of interconnected layers that transform input data through weighted operations to produce predictions. Deep neural networks are the foundation of all modern AI — language models, image generators, and deepfake detectors alike. Key types include CNNs (images), RNNs (sequences), and transformers (parallel sequence processing).

The branch of AI focused on enabling computers to understand, interpret, and generate human language. NLP underpins both AI text generation (LLMs) and AI text detection (classifiers that analyze linguistic patterns). Core NLP tasks relevant to detection include tokenization, sentiment analysis, and statistical language modelling.

A measure of how "surprising" text is to a language model — how uncertain the model is about each next word. Human writing exhibits higher perplexity (more creative word choices) than AI text, which selects high-probability tokens for predictable prose. A primary detection signal in tools like GPTZero, often combined with burstiness analysis.

The practice of designing optimized instructions for AI models to elicit desired outputs. In detection evasion, users prompt LLMs to "write like a human," vary sentence structure, or include deliberate errors to artificially raise burstiness and perplexity. Detection researchers study prompt patterns to improve classifier robustness.

An open-source text-to-image diffusion model by Stability AI. Its open nature made it the most widely deployed image generator, spawning thousands of community fine-tunes. Runs on consumer GPUs without API content policies or watermarking, complicating provenance. Detection requires classifiers trained on diverse community variants.

Any digital content — text, image, audio, video — fully or substantially generated by AI rather than directly captured or authored by a human. Encompasses AI articles, generated artwork, deepfake videos, and cloned audio. Not inherently harmful, but potential for deception makes detection and provenance verification critical.

A parameter controlling the randomness of an LLM's output. Low temperature (e.g., 0.2) produces predictable, repetitive text; high temperature (e.g., 1.0+) produces more varied and creative output. Temperature affects detectability — low-temperature AI text is easier to detect due to its high predictability, while high-temperature text mimics human variability.

The fundamental unit of text processed by LLMs — typically a word, sub-word, or character sequence. LLMs generate text by predicting probability distributions over tokens sequentially. The distribution of token probabilities — which tokens were predictable versus surprising — is the statistical basis for perplexity-based AI text detection.

The neural network architecture introduced in 2017 ("Attention Is All You Need") that replaced recurrent processing with parallelizable self-attention. Foundation of all modern LLMs (GPT, BERT, Claude, LLaMA, Gemini) and many image generation models. Understanding transformers is key to understanding why AI text exhibits its characteristic statistical properties.

Synthetic video depicting a real person performing actions or speaking words that never occurred. Types include face swaps, reenactment, body puppeting, and full text-to-video synthesis. Detection methods include temporal consistency analysis, physiological signal detection, audio-visual alignment verification, and model-specific artifact identification.

AI synthesis of a realistic imitation of a specific person's voice from a short audio sample — as few as 3–10 seconds in modern systems. Used in entertainment and accessibility but also in voice phishing attacks and political disinformation. Detection relies on spectrogram analysis, prosody irregularities, and liveness verification.

The practice of embedding identifying or authenticating information into digital content. AI watermarking embeds provenance signals at content creation time (see AI Watermarking). Robust watermarks survive compression, resizing, and format conversion. Standards are being developed through C2PA and national AI regulation frameworks to enable transparency about AI-generated content at scale.

A detection approach where a classifier identifies AI content from models it was never trained on, by generalizing from universal properties of AI generation. Important because the universe of generative models grows too fast for model-specific training. Current zero-shot detectors achieve 70–85% accuracy on unseen models versus 90–97% on known models.