Would an AI have written this anyway?

The only score that measures what a piece adds beyond what AI already knows and what is already published. Claim by claim, with the reasons, in any language.

Yours An AI would say it Known elsewhere Known

Free to start: ten quick reads with your email. Full verdicts from a dollar.

See an example result

Paste or link
Any language, 120 to 6,000 words.
Weighed claim by claim
Each statement is checked for whether it brings anything new, and how well it holds up.
See what's yours
Your own contributions, credibility and effort, with the reasons.

What no detector, checker or grader measures.

AI detectors ask who typed the words. Plagiarism checkers match strings. SEO graders count the terms of pages that already rank. Verdict asks the question editors and readers actually ask: does this piece add anything?

Beyond AI and beyond the web, at once

Each claim is checked against what a capable language model would say unprompted and against the published web. Only what passes both is yours.

A verdict with reasons, not a probability

Every claim is marked yours, something any AI would say, known elsewhere or common knowledge, with its credibility and a note. You can argue with it, line by line.

Cannot be gamed with confident nonsense

Invented statistics and unsourced figures earn almost nothing; in our test they cost four points. New only counts when it holds up.

Reads older pieces as of their date

A piece that was first is credited as the source of the ideas that spread from it, with today's reading shown beside the original.

Side by side: Originality.ai, GPTZero, Copyscape, Turnitin, Grammarly, Surfer SEO.

The number to know before you publish.

Verdict is built on what Google's quality raters are told to reward, what editors look for and what readers notice: originality, effort, accuracy and first-hand experience. A low verdict means the piece will be read as an echo of what is already out there, however well it is written. A high one means it is a source.

Every verdict comes with the claim view and, when you ask, concrete suggestions to raise it: what to source, what to cut, where to add what only you know.

Writers score drafts before publishing and watch the verdict climb with each revision.
Editors gate submissions and audit a back catalogue in one batch, with reasons they can send back to the author.
Publishers show the badge, score on publish with the WordPress plugin, and rank by what adds something.
Agents call it over MCP to check their own drafts before they ship.

Community Edition, Apache 2.0

Open source. Host it yourself.

The quick read, the benchmark that produced every number on this page, the API clients, the WordPress plugin and the MCP integration are open. Run the quick read on your own machine, reproduce our results, or build on the API. The hosted verdict pays for the model work; the community keeps it honest.

Star expulp/verdict Read the benchmark

# Run the quick read locally, no key, no model
git clone https://github.com/expulp/verdict && cd verdict
pip install -e .
verdict-quick read https://example.com/essay   # one piece, in your terminal
verdict-quick serve --port 8080                # or the same API on your own host

# Reproduce the benchmark against the hosted verdict
python -m benchmark.assemble && python -m benchmark.report scores.csv

What is open

The quick-read engine (sentence echo map, predictability, provenance checks), the benchmark harness and every result table, the Python and JavaScript clients, the WordPress plugin, the MCP setup.

What stays hosted

The claim judge, its prompts and weights, and the corpus of earlier submissions that protects first publishers. That is what a verdict pays for, and what keeps the score hard to game.

How to take part

Open an issue with a piece the verdict got wrong; it goes into the calibration set. Add test pairs in your language. Build a client or an editor plugin. Contributors get verdicts on the house.

Quick reads are free. Full verdicts start at a dollar.

Quick read

Free

What's already online, sentence by sentence. How distinctive the wording is. Hidden characters. A provisional score. Ten per account, unlimited once you've bought anything.

Read a piece

Full verdict

$1 each, or 10 for $5, 50 for $20

Every claim marked as yours, known, or already online, with its credibility, the effort the piece shows, and a verdict you can share.

API and bulk

Same prices

Score drafts before publishing, audit an archive, gate submissions. Verdicts bought here work on the API too. Batch mode and publisher plans are next.

API reference

Verdicts are valid for 90 days. Payments by Stripe: cards, Apple Pay and Google Pay.

Writing or publishing every week? Monthly plans.

Writer

$9/month

150 verdicts a month. Emailed reports and share cards.

Pro

$29/month

600 verdicts a month. Full API access and batch mode.

Publisher

$99/month

3,000 verdicts a month. Badge and plugin, team keys.

Monthly verdicts are added on each payment and stay valid for 90 days. Cancel any time from your account.

Three real pieces, read the same way.

92 A security researcher's own investigation: first-hand findings, most claims the author's.

90 then, 62 today A 2019 essay whose ideas have since been repeated everywhere. Credited as the source.

13 A competent AI-written article on a real incident: nothing a reader couldn't get elsewhere.

What it gets right, and where it is still weak.

Verdict is in early access. It is a reading with reasons, not a certificate. Everything below was measured on 638 pieces the tool was never tuned on; the full tables are in the open repository.

638: pieces tested, 7 kinds, 11 languages
4 in 5: real essays beat their AI twin
−4: points for adding five invented statistics
44 / 45: paraphrased copies caught

Holds up

80%
Real essay vs an AI article on the same topic. The essay scores higher in 80% of pairs, by 22 points on average.
−4
Invented statistics don't pay. Five fake figures added to an AI article lowered its score by 4 points on average. A scorer that rewards "new facts" without checking them shows the opposite.
44/45
Paraphrased copies are caught. A rewrite of a piece already analysed is flagged as a restatement, with the original's date.
5/5
Personal essays and how-tos. After the genre work, the author's piece beat its AI twin in every pair, by 29 and 46 points.
then
Older pieces are read as of their date. Ideas that spread afterwards are credited to the piece, with today's reading shown beside it.
any
Any language. The free quick read uses no model at all.

Still weak

0
No human ratings yet. "Excellent" in our test means essays that reached the Hacker News front page, which measures interest, not originality. Some of the 1-in-5 misses may be right.
±7
Run-to-run noise. Near a tier boundary, two runs can land on different labels. Read the claims, not only the number.
?
Accurate but thinly indexed facts read as new. A machine rewrite of real local news, with facts our search does not find, can score well.
en
Search recall is weaker outside English and on small blogs, so "known elsewhere" is a floor, not a census.
beta
Fiction is experimental. Weighing observations claim by claim does not yet separate a human story from a competent AI one; fiction results carry a notice.
new
Pasted text is judged as new. Submit the link if the piece was published earlier, so it can be dated.

Verdicts that fail are not charged. If a result looks wrong, reply to any email from us with the link: every disputed case goes into the calibration set.

Detectable isn't the question. Worth reading is.

AI detectors and watermarks ask who typed the words. Readers, editors and search engines ask whether the piece adds anything. Verdict measures the second.

Watermarks prove a tool was used, not that the work is empty

Anthropic began watermarking Claude's text in August 2026, with a statistical signal in word choice rather than hidden characters. Google's SynthID detector opened to the public in October 2026, for images, video and audio.

Neither tells you whether the ideas are the author's. A watermark survives in a piece a writer merely proofread with AI, and vanishes from an AI draft that was reworded by hand. Anthropic's own explainer lists the cases where the mark breaks down and says it cannot prove authorship. For text, no vendor offers a public detector yet: Anthropic has announced one, Google's covers other media, and OpenAI marks images only.

Sources: BleepingComputer on Anthropic's text watermark, The Decoder on the announced detection API, eWeek on the SynthID Detector launch, Google's SynthID documentation.

AI detectors measure style, and style is the wrong target

Detectors are classifiers trained on how machines tend to phrase things. Vendors report near-zero error rates on their own benchmarks; independent tests find accuracy falling on edited or paraphrased text.

A Stanford study found seven detectors flagged the majority of essays by non-native English speakers as machine-written while classifying native speakers' essays almost perfectly. As more people write with AI assistance, "was a model involved" stops separating good work from filler. A piece can be written entirely by hand and say nothing new, or drafted with help and carry a real finding.

Sources: Liang et al., GPT detectors are biased against non-native English writers (2023), Pangram Labs technical report on its classifier (2024).

Google rates on originality and effort, and so does Verdict

Google's Search Quality Rater Guidelines tell raters to give the lowest rating when the main content is copied, paraphrased or AI-generated with little effort, originality or added value. Verdict is anchored on the same criteria.

The guidelines define originality as content "not available on other websites" and effort as the extent to which "a human being actively worked to create satisfying content". They name the tells of paraphrased content: only commonly known information, heavy overlap with reference sites, or a summary of one page with nothing added. Google also holds a patent on scoring a document by the information it adds beyond what the reader has already seen.

Sources: Google Search Quality Rater Guidelines (PDF, sections 3.2 and 4.6), Search Engine Land on the January 2025 update, Google patent US20200349181A1, contextual estimation of link information gain.

The research behind measuring what a text adds

Recent work shows that language models converge on the same points and that human writing adds more new ideas per piece, and gives ways to measure novelty and factual support claim by claim.

Wenger and Kenett found answers from different models far more alike than answers from different people. Lu et al. introduced a Creativity Index that measures how much of a text can be rebuilt from existing web text, and found human writing scored far higher. Document-level novelty metrics such as NovAScore score a piece by its atomic facts, and Google DeepMind's long-form factuality work checks each fact against search. Verdict draws on this line of research without being a copy of any one method.

Sources: Wenger and Kenett, creative homogeneity across LLMs (2025), Lu et al., AI as Humanity's Salieri, ICLR 2025, NovAScore, COLING 2025, Wei et al., long-form factuality in LLMs (2024), Measuring information potential in text corpora (2025).

Where Verdict fits

	Expulp Verdict	AI detectors	Plagiarism checkers	SEO content graders
The question answered	Does this piece add anything beyond what's known and published?	Was this written by a machine?	Is this text copied?	Does it cover what already ranks?
Still meaningful when people write with AI help	Yes	No, by design	Yes	Yes
Reasons, claim by claim	Yes	Sentence-level probability	Matched passages	Term coverage
Catches paraphrased ideas, not only copied words	Yes	Not its job	Rarely	No
Rewards invented facts as "new"	No, unverified figures earn little	Not scored	Not scored	Often, difference reads as gain
Languages	Any	Mostly English	Many	Several

Examples of each category: GPTZero, Pangram and Originality.ai; Copyscape and Turnitin; Surfer and Clearscope. Based on each product's public documentation, October 2026. Detail: vs AI detectors, vs plagiarism checkers, vs SEO graders, vs writing assistants.

	Expulp Verdict	Grammarly	Jasper	Notion AI, ChatGPT
What it does for a draft	Tells you whether it was worth writing, and which parts	Fixes grammar, tone and clarity	Generates marketing copy	Drafts, rewrites and summarises
Knows what an AI would have said anyway	Yes, that's the point	No	No	No
Checks the published web for the same ideas	Yes	Plagiarism check on words only	No	No
Weighs the credibility of each claim	Yes	No	No	No
Best used	Before you publish, and on what you've already published	While editing	While drafting	While drafting

Verdict complements writing tools rather than replacing them: write with whatever you like, then find out what the result adds.

	Expulp	Medium	Substack	LinkedIn	X
What gets a piece seen	Originality: what it adds beyond AI and the web	Engagement and reading time, plus curator Boosts	Subscriptions and recommendations	Engagement	Engagement
Checks whether a piece adds anything new	Yes, claim by claim	No	No	No	No
Shows writers what an AI would have said anyway	Yes, on every piece	No	No	No	No
Popularity counters drive ranking	No	Yes	Yes	Yes	Yes

Based on each platform's public documentation as of October 2026.

Write what an AI couldn't? Expulp is being built for you.

A publishing network where reach follows originality, not volume. Founding editors get early access, free verdicts and a say in how ideas are ranked.

What Expulp is

You publish on Expulp, or connect what you already publish elsewhere.
Every piece gets a verdict: what it adds beyond what an AI knows and what is already online, with the reasons.
Reach follows the verdict, not follower counts or posting frequency. Readers find what is actually new.

Founding editors shape the ranking rules, get free verdicts while the network is built, and are first in when it opens.

	Newsletter platforms	Social networks	Expulp
Reach comes from	Your subscriber list and the platform's recommendations	Engagement, follower count, posting rhythm	What the piece adds, measured claim by claim
What gets measured	Opens, reads, subscriptions	Likes, reposts, impressions	Originality, credibility, effort
Rewrites and recaps	Compete on equal terms	Often win on timing	Read as echoes and ranked below the source
Writers who do well	Those with an audience already	Those who post most	Those who bring something new

The same verdict, as an API

Score drafts before you publish, audit a back catalogue, gate submissions, or run hundreds of pieces as one batch. Your verdicts work on the API with the key from your account.

Full reference · Clients and open source

# Free quick read (no key)
curl -X POST https://verdict.expulp.com/v1/quick \
  -H "Content-Type: application/json" -d '{"url": "https://example.com/essay"}'

# Full verdict (one credit), then poll until status is "done"
curl -X POST https://verdict.expulp.com/v1/verdicts \
  -H "X-API-Key: oi_your_key" -H "Content-Type: application/json" \
  -d '{"url": "https://example.com/essay", "notify": true}'
curl https://verdict.expulp.com/v1/verdicts/Q2x… -H "X-API-Key: oi_your_key"
# → {"status":"done","result":{"score":74,"tier":{"label":"Fresh Voice"},"verdict":"…","claims":[…]}}

# pip install nothing: clients/python/verdict_client.py is standard library only
from verdict_client import Verdict
v = Verdict("oi_your_key")
quick = v.quick(url="https://example.com/essay")        # free
full = v.verdict(url="https://example.com/essay")      # one credit, waits for the result
print(full["score"], full["tier"]["label"], full["verdict"])
for c in full["claims"]:
    print(c["status"], c["text"])         # original / known / echo-ai / echo-web

import { Verdict } from "./verdict.js";   // clients/js/verdict.js, no dependencies
const v = new Verdict("oi_your_key");
const full = await v.verdict({ url: "https://example.com/essay" });
console.log(full.score, full.tier.label, full.verdict);
const left = await v.credits();

# Up to 500 pieces, processed asynchronously; poll or give a webhook
curl -X POST https://verdict.expulp.com/v1/batches \
  -H "X-API-Key: oi_your_key" -H "Content-Type: application/json" \
  -d '{"items": [{"id": "a", "url": "https://example.com/one"}, {"id": "b", "text": "…"}],
       "webhook_url": "https://your.site/hooks/verdict"}'
curl https://verdict.expulp.com/v1/batches/b_… -H "X-API-Key: oi_your_key"
curl https://verdict.expulp.com/v1/batches/b_…/results.csv -H "X-API-Key: oi_your_key"
# webhook body is signed: X-Verdict-Signature = HMAC-SHA256(your key, body)

# Verdict is an MCP server (streamable HTTP). Add it to Claude, Cursor, or any MCP client:
{
  "mcpServers": {
    "verdict": {
      "url": https://verdict.expulp.com/mcp,
      "headers": { "Authorization": "Bearer oi_your_key" }
    }
  }
}
# Tools: quick_read (free), verdict (1 credit; improve=true adds editing suggestions),
#        get_verdict, improve_verdict, start_batch, batch_status, credits
# Claude Code:  claude mcp add --transport http verdict https://verdict.expulp.com/mcp --header "Authorization: Bearer oi_your_key"