AI-assisted writing signals

What "AI-assisted writing signals" means

Autotend Forensics surfaces patterns commonly observed in AI-assisted writing. We use that wording deliberately:

• Signals, not verdicts. No detector here returns "this was written by ChatGPT." The pattern observed is consistent with AI-assisted writing; many of these patterns also appear in non-AI writing, especially formal-register prose by non-native English speakers.
• Patterns, not fingerprints. We do not claim model identification. The same signal can correspond to GPT-4o, Claude, Gemini, or simply careful editing.

This is the most controversial signal channel because the public AI-detection industry has a long history of false positives that ruined real students. Autotend Forensics keeps the bias caveat visible everywhere this channel surfaces, and the structural signals (metadata, edit history, paste detection) are always the primary review path.

What the signals are

• Stock openers. "In this essay," "It is widely acknowledged that," "In today's rapidly evolving landscape" — formulaic opening patterns that LLM training data over-weights.
• Synonym churn. A pattern of substituting unusual synonyms for common words ("utilize" for "use," "endeavor" for "try," "myriad" for "many") at higher density than typical student prose.
• Em-dash density. Some LLMs over-produce em-dashes relative to most human writing styles.
• Hedging vocabulary. "It is important to note," "however, it is worth considering," "ultimately, the question becomes" — consistent hedging across an argument.
• Vague specifics. Confident-sounding sentences whose factual content evaporates on inspection ("Numerous studies have shown..." with no citation, "Experts agree...").
• Markdown leakage. Direct copy-paste from a chat interface sometimes leaves markdown syntax in the document (**bold** text that didn't render to bold, stray # headers).
• Self-disclosure. Text from the assistant itself that shouldn't be there ("I'm an AI language model and I cannot...," "As of my last update," "Sure! Here's a draft of...").

What these signals cannot tell you

The known false-positive paths are substantial:

1. Non-native English writers trained in formal academic English score very high on stock-opener, synonym-churn, and hedging detectors. Multiple public studies have demonstrated this bias.
2. Formal-register native writers (students who read a lot, debate-team kids) also score high.
3. Genre conventions matter. A scientific lab report should have hedging and stock phrasing; flagging them is inappropriate.
4. AI-edited but human-written text carries some AI-style word choices without being AI-authored. Almost everyone uses Grammarly / Word's editor / similar tools now.

For these reasons, Autotend Forensics:

• Surfaces this channel as auxiliary, never primary.
• Carries an explicit non-native-English bias warning in the scorecard copy.
• Pairs every flag with a "what this can and cannot tell you" note inline.

What we surface

For each AI-assisted-writing signal Autotend Forensics finds, the inspector shows:

• The phrase or pattern observed.
• Its location in the document (character offset, with highlight).
• The reference corpus the threshold was calibrated against.
• The known false-positive paths for that specific signal.

Use this channel as one input among many. The structural signals (metadata anomalies, edit-history residue, paste-detection flags) are far more reliable, and the conversation with the student should start with those.