How Griefy detects AI content
Griefy is an academic quality-control system. This page gives full transparency on the detection engine behind one of its review angles — AI authorship — including the engines we use and what results actually mean.
The AI Detection Engine
Griefy uses an enterprise-grade, industry-tested AI-detection engine (Copyleaks) as the authoritative source for AI-authorship probability. The same engine is trusted by thousands of academic institutions and publishers. It returns a single, calibrated probability score from 0 to 100, along with a confidence level (low, medium, or high) based on how decisive the signals are.
Authoritative AI score
The Copyleaks engine analyzes your full text and returns the probability that it was AI-generated. This score is the one we report — it is not a guess or an estimate produced by a general-purpose chatbot.
Example output:
AI probability: 42% with medium confidence — meaning the engine sees moderate AI signals but is not certain.
How the LLM Complements the Engine
The detection engine returns an overall probability, but it does not tell you whichsentences triggered it. For that, Griefy runs a second analysis using Claude, which reads your paper and flags the specific short phrases and sentences that exhibit the strongest AI-authorship signals. Claude also performs a separate originality pass to surface passages that may match external sources without attribution. Claude's probability estimate is always overridden by the authoritative Copyleaks score.
Two roles, one report: Copyleaks decides the headline AI probability; Claude points you to the exact passages to review and rewrite in your own voice.
The Six Signals Claude Evaluates
When flagging specific passages, Claude looks for six statistical and stylistic signatures of AI authorship.
Perplexity
Measures how predictable each word choice is. AI language models produce highly probable word sequences; human writing varies more and uses less predictable vocabulary combinations.
Burstiness
Humans write in rhythmic bursts — short sentences then long ones, informal then formal. AI produces uniformly medium-length sentences. We measure variance in sentence length across the full paper.
Lexical diversity
Type-token ratio over long passages. AI tends to recycle a narrower vocabulary, while human writing draws on a wider, more varied word set.
Structural predictability
Detects rigid, formulaic paragraph structures (e.g. a topic sentence, three supporting points, a summary) that are characteristic of machine-generated prose.
Hedging-language authenticity
Flags over-use of hedge phrases ("it is important to note", "it is worth considering") and assesses whether hedging reads as authentically human or machine-templated.
Transition uniformity
Identifies overuse of identical transitional phrases ("furthermore", "moreover", "in addition") — a common AI fingerprint — versus the more varied connectives of human writing.
Sensitivity Levels
Sensitivity controls how aggressively Claude flags individual passages for your review. It does not change the authoritative Copyleaks score — only how many passages are surfaced for you to look at.
Conservative
Only flags passages with strong, consistent AI signals. Lowest risk of false positives. Best when you want to avoid over-flagging your own writing.
Balanced (default)
Flags passages where multiple signals agree. Recommended for most academic submissions.
Strict
Flags any passage showing a significant AI signal. Use before submitting to institutions with aggressive AI-detection policies.
Originality Checking
Griefy extracts distinctive sentences from your paper and searches the live web (via Google Search) for matching content. Passages with significant word overlap to an external source are flagged with a link to that source, so you can add a citation or rewrite the passage in your own words. Correctly quoted and cited material is not flagged.
Citation Checking
Citation checking runs in two layers. First, Claude evaluates in-text citation format, reference-list formatting, orphaned references, and missing citations against your selected style (APA 7, MLA 9, or Chicago). Second, Griefy verifies references against the CrossRef database — checking that DOIs resolve and that cited works actually exist — so fabricated or incorrect references are caught, not just misformatted ones.
Limitations
No AI detector can guarantee 100% accuracy. Heavily rewritten or humanized text reduces detection rates across all tools. Griefy always shows the AI probability and a confidence level alongside every result, plus a plain-English explanation of the flagged passages. A high AI probability is a signal to review your writing — not proof of intent.
Griefy's detection engine is designed to be upgraded as better models become available. We will publish methodology updates whenever the detection approach changes.