everything
we've written.

a panel of independent detectors, fused into one honest verdict, with the disagreement on display. limitations included.

one scan, start to finish: a trained model routes it to the detectors built to read it, a panel weighs in, the reads are fused and capped, and “uncertain” is a real outcome. the visual walkthrough.

an honest, use-case ranked list of the best AI detectors for image, video and text. what each is best at, where it falls short, and who it’s for.

what a classifier actually reads in an image, why every honest result is a probability, why a panel fails less badly than one model, and what “uncertain” really means.

don’t take “99% accurate” on faith: a 30-sample blind protocol, a pre-2022 false-positive baseline, a paraphrase stress test, and the transparency checklist. applies to us too.

reading an image’s C2PA history and Google’s SynthID watermark: what a valid credential proves, what a missing one doesn’t, and why platforms strip them.

verify a suspect news photo the way an editor would: corroboration first, signed provenance second, a panel read third. every result a probability, not proof.

a fully synthetic face leaves nothing for reverse image search to find. the visual tells, why “no match” isn’t an all-clear, and how to check properly.

before you send a deposit: a reverse image search to catch stolen photos, a generation check to catch invented ones, and the behavioral red flags around both.

synthetic audio, image, or video that depicts a real person. three subtypes (face-swap, lip-sync, full synthesis). 2024 election context.

running multiple classifiers in parallel. why it’s less brittle than singletons and when it isn’t.

naming the specific generator behind a piece of content. fingerprints, the closed-set caveat, where attribution falls over.

two unrelated technologies sharing a name. SynthID / Stable Signature on one side, C2PA Content Credentials on the other.

the architecture that ran image generation until diffusion took over. still common in face-swap deepfakes.

the architecture behind Midjourney, Stable Diffusion, DALL-E 3, Sora. forward noising plus reverse denoising.

large language model. what’s behind ChatGPT, Claude, Gemini. transformer plus autoregressive next-token prediction.

how surprised a language model is by a piece of text. the workhorse signal of AI-text detection.

the variance of perplexity across a document. why flat-rhythm writing (AI or ESL) gets flagged together.

the proportion of human work wrongly flagged as AI. where most 99% accuracy claims fall apart.

the cryptographic provenance standard. Content Credentials manifests signed by Adobe, OpenAI, Sony, Leica, and more.

Google’s invisible watermark family. embedded by default in Gemini, Imagen, Veo, Lyria. detectable via Google’s verifier.

adversarial attack where untrusted input hijacks an LLM. #1 on the OWASP LLM Top 10. distinct from jailbreaks.

the largest public AI-text detection benchmark. 6.2M generations, 11 generators, 11 adversarial attacks. ACL 2024.

LLM output that’s fluent, confident, and factually wrong. structural to next-token prediction, not a bug.

the architecture under every modern LLM and most image / video generators. self-attention is the key innovation.

detection routing sends each scan to the detectors best suited to it instead of running everything equally, borrowing the mixture-of-experts gating idea from Jacobs et al. (1991) through Switch Transformer and system-level LLM routing. how a scan finds the right expert.

the academic name for a detector that declines to answer: the reject option, selective classification, and the risk-coverage trade-off, from Chow’s 1970 result to today’s AI-image rejection mechanisms. why returning “uncertain” is a design feature, not a failure.

model drift is when a model’s training data falls out of step with the real world, so accuracy decays, which is what happens to an AI-content detector every time a new generator ships. why no detector is set-and-forget.

are AI detectors accurate? not as a single number. peer-reviewed benchmarks show tools advertising 99% accuracy drop sharply on out-of-domain text and collapse under paraphrasing, and the FTC has already moved on inflated claims.

the closest multi-modal rival: image, video, audio, text, API-first. amige. reads the same modalities against a panel of independent detectors, fuses one verdict, and names the maker too.

the free academic text checker, QuillBot-powered, plagiarism bundled. text only. amige. adds image, video and model attribution.

the institutional academic-integrity standard for text, sold to schools, no individual access. amige. is the consumer tool for image, video and text.

the education incumbent. owns text in classrooms. doesn’t do image or video. different jobs, different tools.

the content-marketing detector. bootstrapped. strong RAID numbers. plagiarism + fact-check bundled. text only.

ten years in market. the broadest LMS network, source code detection, enterprise + education distribution.

the GPTZero-alternative for schools. plagiarism bundled, tight LMS integration, ESL false-positive caveats.

free-tier text detector with 15K-char no-signup ceiling. independent tests show 15-33% FPR on academic writing.

the actually-accurate text detector. tied best on RAID with 97.7% adversarial accuracy. text-only.

B2B writing-assistant suite with AI detection bundled. SSO, SCIM, self-hosted. very different shape of product.

the category leader for stylized image generation. V7 and V8 are meaningfully harder to detect than V5 was.

the highest-traffic AI product in history. characteristic structural tells, post-launch vocabulary spikes, and documented false-positive risk on non-native English.

OpenAI’s text-to-video. Sora 2 added audio and better physics. C2PA watermarks and the celebrity-likeness scandal.

Anthropic’s frontier text family. Opus 4.8 is the current flagship. characteristic hedges, em-dashes, structured bullets.

Google DeepMind’s frontier model. embedded across Search AI Mode, Workspace, Pixel. SynthID by default.

OpenAI’s pre-GPT-Image image generator. deprecated May 2026. still in the wild on social media.

the open-weights family plus the Civitai ecosystem. the hardest detection case because of the fine-tune long tail.

Black Forest Labs’ flow-matching model. the strongest open-weights challenger to Midjourney. very hard to spot by eye.

Google’s image generator inside Gemini. SynthID by default. when the watermark survives, attribution is near-certain.

ByteDance + Tsinghua face-swap model. not a general image generator. a DreamID flag means deepfake.

Google DeepMind’s diffusion image family, distinct from Nano Banana. clean high-key aesthetic. SynthID by default.

Google’s text-to-video. Veo 3 added synchronized native audio. SynthID on video AND audio tracks.

the indie / creator-favorite video gen. Gen-1 through Gen-4.5. most embedded in professional film and TV.

the free-tier consumer video gen. 25M+ users reported during 2025. smooth motion, soft textures. the ’I tried it for fun’ tool.

the TikTok-native social video gen. Pikaffects (Cake-ify, Crumble, Explode). stylized over photoreal.

owned the in-image text rendering niche from 2023. design-y compositions, crisp typography, marketing workflows.

Adobe’s enterprise ’commercially safe’ image (and video) gen. C2PA signed on every output. ~24B assets generated.

OpenAI’s GPT-5, the routed “unified system” that became the ChatGPT default in August 2025 and now runs through a 5.x line up to GPT-5.5. how its writing reads, why flatter output dodges detectors, and how amige. estimates it.

how to spot Grok Imagine, xAI’s image-and-video generator built into the Grok app: the visible tells across stills and clips, the 1.5 image-to-video update, the Spicy-mode deepfake scandal, and how amige. reads it.