Armin Ronacher tests Pangram, an AI detection tool, by having an LLM generate text based on a detailed prompt, then manually rewriting it from scratch without AI assistance. Both versions get flagged as 100% AI-generated, suggesting that detection tools struggle with text that originated from LLM-structured prompts even after substantial human revision.

• Pangram claims extremely low false positive rates (0.0041% false AI accusations), yet flags clearly human-authored text as entirely AI when the author used an LLM to develop the initial structure
• The rewritten text shares only 50% similarity to the LLM output with no identical sentences, yet still scores as 100% AI, indicating detection models may penalize writing that follows LLM-suggested structural patterns
• Authors who rely on LLMs for organizing ideas face a credibility problem: their work gets flagged as AI regardless of how much manual editing and rewriting happens afterward

Anthropic built a tool that checks whether Claude created image files by reading cryptographically signed metadata attached to downloads, using the C2PA industry standard that camera makers and photo software already use. Text detection works differently and requires a separate API currently available only to certain EU organizations.

• Claude embeds content credentials (signed metadata) in supported file types like PNG, JPG, and SVG to prove involvement in creation or processing
• The checker only reads the embedded credential, not the file itself, and never stores or accesses your uploaded file
• Text watermark detection uses a separate API in private preview, while similar file-checking tools exist from OpenAI and Google DeepMind

Pangram Labs released Pangram 4, an AI text classifier that detects whether text was written by humans or AI systems. The model achieves 99.16% accuracy and can now identify mixed human-AI writing and fine-grained edits better than its predecessor.

• AUROC of 0.9916 with 0.0041% false positive rate and 0.3396% false negative rate
• Can distinguish fine-grained edits and detect interleaved AI assistance in co-authored text
• Shows improved robustness to adversarial attacks and better generalization to out-of-distribution data

Superhuman Inc., the company behind Grammarly, has bought GPTZero to add AI-detection and authenticity checks to its Superhuman Go assistant. GPTZero’s tools flag AI-generated text, fake citations and plagiarism, and will now work alongside Superhuman’s own detector across a million apps and sites.

• Superhuman (Grammarly's parent) acquired GPTZero, adding it to Superhuman Go's existing detector to cross-check AI-generated text, fake citations, and plagiarism across a million+ apps and sites.
• GPTZero brings 19 million users and ~$30 million ARR, having raised $13.5 million and been valued above $88 million by PitchBook before the deal.
• The combined two-model approach aims to cut false positives/negatives as AI-generated content now makes up roughly half of all new online articles.

This article argues that overusing the “X isn’t just about Y; it’s about Z” structure is the most common giveaway of AI-generated text, not em dashes. It shows examples of this negation pattern and notes two runner-up structures—“from…to…” and “whether…or…”—that also signal machine writing.

• The "X isn't just about A—it's about B" negation-contrast pattern is the most common AI writing tell, more than em dashes.
• "From...to..." parallel phrasing and "Whether...or..." constructions are secondary AI tells that show up in sales copy and captions.
• Sometimes unedited AI output leaks through directly, like leftover prompts asking "Would you like a concluding paragraph or bulleted summary?"