SlopTotal: AI text detector

How it works

23 engines read the text

Neural classifiers trained on AI text, statistical tests that measure how predictable each word is, and checks for the phrasing AI models overuse.

One calibrated score

The engines' votes are combined by a model fitted on SlopBench, so a score of 55 means the same thing in every report: fewer than 2 in 100 human texts score that high.

Every vote shown

You see each engine's score next to the total. When the engines disagree, the report says so.

Reading the score

The bands are placed by how often human-written text reaches them, so a high score is rare for a person's writing.

0–30
Clean, likely human-written
Few AI signals across the engines. The text reads like human writing.
30–45
Low risk
Some engines see minor signals. Formal or heavily polished human writing often lands here.
45–55
Suspicious
The engines disagree. It could be AI-assisted, edited AI output or formal human writing. A reason to look closer, not a conclusion.
55–80
Likely AI-generated
Strong signals from several engine families. Substantial AI generation with limited editing is likely.
80–100
Slop detected
Broad agreement across the engines: clear statistical and linguistic markers of unedited AI output.

The engines

Each engine uses a different method. Some carry weight in the score, all are shown in every report.

Engine	Method
TMR Detector RoBERTa-base RAID-trained detector — 97.3% accuracy, 2.3% FPR	neural
ReMoDetect Reward-model detector targeting RLHF-aligned LLMs (GPT-4, Claude, Llama-chat)	neural
Binoculars Cross-entropy ratio between two LMs — agreement signals AI-generated text	statistical
Fast-DetectGPT Conditional probability curvature — AI tokens sit at peaks of the model's distribution	statistical
Perplexity GPT-2 perplexity scoring — low perplexity suggests AI-generated text	statistical
Cross-Perplexity Compares perplexity across GPT-2 Large and DistilGPT-2 — agreement amplifies signal	statistical
Fakespot RoBERTa-base AI text detector from the APOLLO system	neural
E5-Small Ultra-lightweight E5 LoRA detector — 33M params, #2 on RAID benchmark	embedding
BERT-tiny RAID BERT-tiny (4.4M params) trained on RAID dataset — ultra-fast inference	neural
OpenAI Detector RoBERTa fine-tuned by OpenAI to detect GPT-generated text	classifier
ChatGPT Detector RoBERTa trained specifically on ChatGPT-generated text	classifier
Desklib DeBERTa DeBERTa-v3-large fine-tuned on GPT-4, Claude, Llama — RAID benchmark leader	neural
SuperAnnotate RoBERTa-large optimized for low false-positive rate — RAID benchmark	neural
GLTR Token rank distribution — AI text uses disproportionately top-ranked tokens	statistical
Log-Rank Average log-rank of tokens under GPT-2 — lower rank means more predictable (AI-like)	statistical
DivEye Surprisal diversity — AI text has unnaturally uniform per-token surprisal	statistical
Burstiness Measures per-sentence perplexity variance — flat rhythm suggests AI	linguistic
Linguistic Markers Detects AI-preferred words and phrases (delve, tapestry, landscape...)	linguistic
Structural Analysis Analyzes em-dash usage, sentence uniformity, list density, paragraph regularity, colon patterns	linguistic
Vocabulary Richness Measures type-token ratio, hapax legomena, and Yule's K diversity	linguistic
Formulaic Patterns Detects cliché AI openings, closings, filler patterns, and structural repetition	linguistic
Readability Uniformity Checks if readability stays unnaturally consistent across paragraphs	linguistic
Sentiment & Hedging Detects excessive hedging, forced balance, vague positivity, and explanatory patterns	linguistic

Is this text written by AI?