Is the score a guarantee of AI vs human authorship?

No. Scores are heuristic. Edited AI text, domain-specific human writing, or mixed workflows can confuse any classifier.

What should I do if the score is high?

Review sourcing, authorship, and whether the piece meets your editorial bar. Consider adding human examples, quotes, and primary research.

What does Semantic assist do?

When enabled, it calls your server embeddings endpoint once with your text plus two short reference passages. Similarity is compared between an essay-style anchor and an informal anchor, then mixed (~30%) with the local heuristic. It costs an API round-trip and needs OPENAI_API_KEY (or equivalent) on the server.

Content & Writing

Updated

Live

AI Content Detector

Estimate the probability that a passage was written by an LLM.

Content

Multi-signal heuristic plus optional dual-anchor embeddings (essay-style vs informal human probe in one API batch) — not forensic proof; use with editorial judgment.

Fetch a page

Semantic assist — one /api/embeddings call embeds two fixed anchors plus your text (OpenAI). Blends ~30% of a dual-anchor axis. Needs 40+ words. Disable to stay fully local.

AI score

Reliability: high

Possibly AI

Higher = more signals typical of polished LLM prose (uniformity, repeated phrasing, stock transitions). Semantic assist compares your text to embedded "essay" and "informal" anchors. Human editors and mixed styles can still land high or low.

0.33

Sentence CV

Lower = more uniform length

0.89

MATTR

Lexical diversity (windowed)

Bigram repeat

Structural repetition

33%

Starter dominance

Repeated openers

0.83

Heavy phrases / sent

Strong AI clichés

6 sent / 82 w

Total

Flagged phrases (8)

Heavy-weight patterns weigh more in the score than light connectors.

In conclusion ×1

it's important to note ×1

navigating ×1

cutting-edge ×1

holistic approach ×1

Furthermore ×1

Moreover ×1

stands as a ×1

X / Twitter LinkedIn Facebook

Start here · What is an AI content detector (in plain English)?

Large language models write text with statistical patterns. Detectors look for those patterns and output a rough score or label — not a court verdict.

False positives happen: stiff human writing, templates, and technical docs can look machine-like. False negatives happen when AI text is heavily edited.

Use this tool as triage: flag work that needs a human read — especially guest posts, YMYL topics, and client deliverables.

Semantic assist is optional: leave it off for fully local heuristic scoring in the browser; enable it only when you want a one-shot embedding blend via your server API.

When to use this tool

Editorial queue hygiene
You received ten freelance drafts today. Run a quick pass to see which ones deserve a deeper originality or fact check.
Syndication and partnerships
Partners send republished content. A high score is a prompt to verify authorship and whether terms allow AI-generated copy.
Learning what machine-like text looks like
Paste your own first draft next to a heavily edited version — compare how scores move so you calibrate your gut.

Examples

Walk through these with the form above — they are practice scenarios, not live data.

Guest post triage

Try this

Paste the first 300–500 words of an article that feels oddly smooth — uniform sentence length, no concrete anecdotes.

What to look for

If the score is high, open a second tab: check sources, author history, and whether facts trace to primary references before you publish.

Short tutorial

Follow in order the first time you use the tool; later you can skip to the step you need.

Step 1 — Paste a representative slice
Use at least a few paragraphs. Ultra-short clips are noisy. Avoid only the intro if the body is where the substance lives.
Step 2 — Read the score as a flag, not proof
Low score does not guarantee human authorship. High score does not guarantee AI — it means run your normal quality bar plus sourcing.
Step 3 — Pair with human review
Claim check, quotes, and personal experience still need eyes. Use Claim & Fact Extractor or your editorial checklist next.

More detail

Paste a passage to get a rough probability-style signal that it may have been produced by a large language model.

The on-page checker mixes lexical and structural signals. Optional Semantic assist (checkbox) sends one batch to your embeddings API: your text is compared against two fixed anchors (polished “essay” style vs informal first-person prose), then blended with the heuristic score — still an approximation, not proof.

This is not a legal or academic forensic test — it is an operator aid for triage and internal QA.

FAQ

Is the score a guarantee of AI vs human authorship?: No. Scores are heuristic. Edited AI text, domain-specific human writing, or mixed workflows can confuse any classifier.
What should I do if the score is high?: Review sourcing, authorship, and whether the piece meets your editorial bar. Consider adding human examples, quotes, and primary research.
What does Semantic assist do?: When enabled, it calls your server embeddings endpoint once with your text plus two short reference passages. Similarity is compared between an essay-style anchor and an informal anchor, then mixed (~30%) with the local heuristic. It costs an API round-trip and needs OPENAI_API_KEY (or equivalent) on the server.