AI detection
glossary

most AI-detection writing is marketing copy. these entries cite primary sources, hedge where the evidence is thin, and tell you where the technology stops working. plain language, no hype. the same bar amige. holds its own verdicts to. see the machine.

• deepfake →
  synthetic audio, image, or video that depicts a real person doing or saying something they didn’t.
• ensemble detection →
  running multiple AI-content classifiers in parallel and combining their outputs. the foundation amige. builds on. amige. goes further: it routes each scan to the detectors strongest for it, names the likely maker as a best guess, and abstains when the reads conflict.
• model attribution →
  naming the most likely generator behind a piece of content, beyond the binary AI / human call. e.g. ’looks like Midjourney v6.’
• AI watermarking →
  two unrelated technologies sharing a name. invisible neural signals baked into generator output, and C2PA cryptographic provenance metadata.
• GAN →
  generative adversarial network. the architecture that ran image generation until diffusion took over, still common in face-swap deepfakes.
• diffusion model →
  the architecture behind Midjourney, Stable Diffusion, DALL-E 3, Sora, and most modern image and video generators.
• LLM →
  large language model. what’s behind ChatGPT, Claude, Gemini. transformer + autoregressive next-token prediction at trillions of parameters.
• perplexity →
  how surprised a language model is by a piece of text. the workhorse signal of AI-text detection.
• burstiness →
  the variance of perplexity across a document. why flat-rhythm writing (AI or ESL) gets flagged together.
• false positive rate →
  the proportion of human work wrongly flagged as AI. where most 99% accuracy claims fall apart.
• C2PA →
  the cryptographic provenance standard. Content Credentials manifests signed by Adobe, OpenAI, Sony, Leica, and more.
• SynthID →
  Google’s invisible watermark family. embedded by default in Gemini, Imagen, Veo, Lyria. detectable via Google’s verifier.
• prompt injection →
  adversarial attack where untrusted input hijacks an LLM. #1 on the OWASP LLM Top 10. distinct from jailbreaks.
• RAID benchmark →
  the largest public AI-text detection benchmark. 6.2M generations, 11 generators, 11 adversarial attacks. ACL 2024.
• hallucination →
  LLM output that’s fluent, confident, and factually wrong. structural to how next-token prediction works.
• transformer →
  the architecture under most modern LLMs and most image / video generators. self-attention is the key innovation.
• detection routing →
  why amige. doesn’t run every detector on every file. routing picks the experts strongest for a given input, the same gating idea behind mixture-of-experts and LLM routers, applied to AI-content detection.
• selective abstention →
  when a detector returns “uncertain” instead of guessing, that’s selective abstention. the reject option, the risk-coverage trade-off, and why answering less can mean answering righter on the cases you keep.
• model drift →
  why an AI detector quietly gets worse over time, and what monitoring, recalibration, and retraining do about it. concept drift, dataset shift, and the calibration fix, cited.
• AI detector accuracy →
  whether AI detectors are accurate depends on the content and the test, never a single percentage. independent benchmarks show 99% claims collapse on unfamiliar or paraphrased text. RAID 2024, FTC 2025, Stanford 2023.

more entries coming. if there's a term you'd like covered, the easiest signal is to scan something and share the permalink.