A private document may need a local check; a rough email may need a rewrite. For documentation, you may want the same checks to run on every change. Choose the workflow for the job at hand.
This is a guide to using tools, not a ranking of competing products. We build Zero Slop, which appears in several workflows below. For measured comparisons and alternatives, use the four-editor comparison. None of these checks establishes who wrote a passage.
Inspect the writing patterns
For a draft that feels generic, but you cannot yet say why.
Start with Wikipedia's signs of AI writing and our explanation of AI slop. Look for a passage that makes a large claim without evidence, repeats a conclusion, or hides the actor behind abstract language. Name the problem before asking a model to change it.
Try this: Mark one sentence. Ask what fact, instruction, or perspective the reader would lose if you removed it. If the answer is nothing, cut it. If it contains a necessary caveat, keep the caveat and simplify the framing.
Limit: A writing pattern is a clue, not proof of authorship or a rule to apply everywhere. Repetition can help readers; technical language can be precise. Judge the passage in context.
Run a local first check
For private drafts and repeatable before-and-after measurements.
Use the sandbox in local mode or the local CLI to inspect a draft with Zero Slop's scorer. The local scorer runs on your device without sending the draft. It checks weighted writing patterns and reports a heuristic score, not an AI probability.
Try this: Run the same text before and after one deliberate edit. Read the flagged passages alongside the score. A smaller number is useful only if the edit still says what you meant. The scoring guide explains the measurement.
Limit: Local cleanup does not provide the host model's contextual rewrite or understand your full voice. Browser mode also needs its runtime to load; do not assume a first visit works without a connection. Local scoring and hosted rewriting are separate options.
Edit inside your AI assistant
For a rewrite that needs context, audience, and your own phrasing.
Install an editing skill in your existing assistant, or use Try your draft. Give it the source and the intended audience. Ask for a clearer version without added facts, then compare the rewrite with the original before accepting it.
Try this: Name a constraint: an exact quotation must remain unchanged, the email must identify the decision owner, or a research summary must preserve uncertainty. A specific instruction is easier to evaluate than “make this human.”
Limit: The hosted editor sends text for AI processing and has free-service capacity limits. An installed skill uses its host assistant's data handling and inference. Open-source instructions do not make every execution offline. Check privacy and data handling before uploading sensitive work.
Review documentation in CI
For teams that want the same mechanical checks on every change.
Use the documentation-review workflow to flag writing issues in a pull request. Check prose files, pin the tool version, and decide who reviews findings. Start with reports before using a score to block a merge.
Try this: Run the check on a small set of documentation changes. Count useful findings and false alarms. Agree on the scope and exceptions with the people who maintain those documents, then record why a rule became a merge requirement.
Limit: A passing writing check does not establish factual correctness, safety, accessibility, or product accuracy. Keep the existing code tests and a reviewer responsible for the document. Do not auto-merge a rewrite because its score fell.
Verify the final edit
For any draft someone will rely on.
Use a source-by-source review. Check the names, numbers, dates, quotations, and links, then read for meaning: did a possibility become a certainty, did an observation become a cause, or did ownership move to another person?
Try this: Read the original and edit side by side. For each changed claim, point to the source that supports it. If an edit sounds more confident than the source, restore the qualification. Keep a human owner for the final decision.
Limit: An automatic match can preserve a number while changing what it describes. No score removes the need to read consequential claims. If the source itself is wrong or incomplete, an editor cannot supply the missing evidence.
Choose the smallest useful workflow
Do I need all five for every draft?
No. A short internal note may need only a quick read. A public claim or a document with financial, legal, or technical consequences needs source review. Use automation where it saves repeatable effort without removing judgment.
Can these tools identify AI-written text?
These workflows evaluate writing and edits, not authorship. The RAID study documents weaknesses in authorship detectors under changes in generation settings, models, and adversarial edits. Do not use a slop score as evidence against a writer.
What should a team measure?
Track the proportion of findings a reviewer accepts, review time, and edits that need to be reversed. Keep failures and false alarms in the record. A rising number of scans says nothing by itself about better documents.
For context on the cost to readers, see AI workslop at work. For the limits of our measurements, read the benchmark and methodology.