The main types
AI-generated content spans every medium. Text: articles, reviews, comments, replies from LLMs. Images: photos, illustrations, product shots from diffusion models. Audio: synthetic voices and music. Video: avatars and fully generated clips from tools like Sora.
Each medium has its own tells. AI text tends toward even sentence rhythm, hedged framing, and a fixed bank of transition phrases. AI images often show fine-detail artefacts: garbled text inside the image, anatomical errors in hands and faces, lighting that doesn't fully cohere. Synthetic audio can flatten the micro-pauses and breath patterns that characterise natural speech.
Why it's a problem at scale
The concern is volume and intent rather than AI itself. When synthetic content is mass-produced to game ranking systems, it crowds out original work, degrades the quality of training data (a feedback loop sometimes called model collapse), and leaves readers unable to tell a researched article from a generated one. Researchers feel this pressure first, then platforms and creators.
Content credentials and watermarking are emerging as part of the answer. DALL·E images carry C2PA metadata, for example. But credentials are easily stripped and not yet universal, so until provenance signals are embedded and preserved end-to-end, detection means combining metadata checks with behavioural and visual pattern analysis.
How detection works in practice
No detector is conclusive on its own. Text detection is probabilistic: it reads statistical patterns (burstiness, vocabulary frequency, perplexity) that correlate with model output, but editing or paraphrasing reduces detectability and false positives hit careful human writers. Image detection is more reliable when provenance metadata is intact, since a C2PA credential is strong evidence. Metadata stripped by re-saving leaves you with visual tells alone.
verif·ai reads the page you're already viewing. For images, it checks embedded content credentials on-device and flags anything marked as AI-generated. For text, it scores the visible blocks and highlights passages that read as likely AI, so you can evaluate them in context rather than pasting paragraphs into a separate tool.
frequently asked
Is AI-generated content bad?
Not inherently. The problem is low-effort, mass-produced synthetic content published to manipulate attention or rankings, which crowds out original human work and degrades trust.
How can I tell if something is AI-generated?
Combine signals: visual or linguistic tells, provenance metadata when present, and a detector's confidence score. verif·ai reads image content credentials and scores text on the page you're reading in one click.
Can AI-generated images be reliably detected?
When the original file carries embedded provenance metadata (like C2PA credentials from DALL·E), detection is reliable. Images already on the web are frequently re-saved and stripped of metadata, so visual tells and a detector's trained pattern analysis are your main tools there, and no single check is conclusive.
What is AI slop?
"AI slop" is informal shorthand for low-quality, mass-produced AI content published at scale to game search rankings or social feeds, filling space without adding information. The term captures the worst of the volume problem, not all AI-generated content.
Published June 7, 2026 · Last updated June 16, 2026