What's rising, what's fading: AI writing tells, tracked weekly

What's rising, what's fading

Data current as of August 10, 2026 · updates every Monday · every figure cited

This week First week live: the trend board, the checker, and the instruction files launched together. All editions

How to read this: ▲ Rising = readers are noticing it more — worth avoiding ▼ Fading = newer AI models are dropping it – Stable = still common, still noticed

Show:

Direction reflects the newest available data. Several tells are now falling at the source while still rising in human speech and the published record.

Most "AI-isms" are shared across every model. But each one has habits distinct enough that a classifier can name the author. Confidence labels: measured = quantitative study · documented = widely reported, official acknowledgment, or dedicated tracking · anecdotal = community consensus without hard counts.

ChatGPT (OpenAI)

"It's not just X, it's Y." Negative parallelism, in ~6% of sampled real messages. measured
Em dashes as a brand. Highest measured rate of the big models; Altman announced a fix in Nov 2025 that made custom instructions work, but defaults barely moved until GPT-5.4. measured
The "delve" era vocabulary. Delve, tapestry, underscore, showcase: the 2023–24 tells, now fading from new output while still spreading through human speech. Watch instead for "core" and "modern". measured
Markdown maximalism. Heaviest user of bold text, headers, and "Bold term: explanation" bullets. measured
Sycophancy spike. The April 2025 GPT-4o update was rolled back within a week for being "overly flattering"; GPT-5 then overcorrected cold and was warmed back up. documented

Claude (Anthropic)

"You're absolutely right!" The signature. Official GitHub bug #3382, a dedicated tracker site (absolutelyright.lol), and self-parody from Anthropic's own account. documented
Affirmation, then reversal. The praise often signals a correction or backpedal is coming next. documented
The apology cascade. "Ah, I see the issue now" → "You're absolutely right" → "Apologies for the confusion", on loop in coding sessions. anecdotal
Prompt-referencing. Leans on "based on the text", "according to the passage", a phrase family classifiers use to identify it. measured
Counter-tell: uses less bold and fewer headers than ChatGPT. The "wall of markdown" is a ChatGPT accent that gets misattributed to Claude. measured

Gemini (Google)

Most agreeable model measured. 62.5% sycophantic response rate in Stanford's SycEval, ahead of ChatGPT-4o's 56.7%. measured
Plain-vocabulary preference. Says "sugar" where ChatGPT says "glucose"; stylometric analysis attributes text to it by exactly this friendliness. measured
The self-loathing loop. "I am a disgrace to my profession… to this planet", a 2025 failure-mode bug Google acknowledged, unique to Gemini. documented
Fewest em dashes of the majors: 3.5 per 1,000 words, close to the 3.83 we measure in pre-ChatGPT Medium articles. measured
Phrase-level tells are thin. Its fingerprint is statistical (word distributions, register) more than catchphrases, but strong enough that DeepSeek R1's training data was traced to Gemini output by style alone. documented