What Stipple Checks Before You Trust a Document

A document can look convincing and still fail a basic check. A payslip may contain arithmetic that does not reconcile. A research report may cite sources that do not exist or do not support its claims. An application may be written by a language model, while a scanned file may have been altered after it was issued.

Stipple is built around that point of uncertainty. It brings several checks for submitted text and documents into one web service, with API and CLI access listed as supported platforms. Its central promise is not simply to produce a score. It aims to show the evidence behind a result so that a person can decide what to do next.

Three kinds of checking in one place

The most visible part of Stipple is its AI text detector. A user pastes prose and receives a probability that it was written by a language model, along with the phrases and stylistic signals that contributed to the result. The product is careful to describe this as a judgement about style, not proof of authorship or authenticity.

A second tool checks citations. It is intended for reports and other research-heavy material, where a reference can fail in several ways. The source may be unavailable, the citation may point to something different from what the text claims, or the source may not support the surrounding statement. Stipple also says it can recompute internal maths, which is useful when a report presents totals, percentages, or other figures that appear authoritative.

The third group of checks concerns documents. Stipple says its verifier looks for tampering, AI-generation signals, arithmetic inconsistencies, and provenance traces. Other tools classify financial documents, extract fields from tables and checkboxes, find personal information for redaction, and compare a document pack with a required checklist.

That makes Stipple broader than an AI-writing detector. The common problem is deciding whether something supplied by another person is safe to rely on. The checks differ, but the workflow is similar: submit material, inspect the returned signals, and keep the final decision with a human.

Where the product fits into real work

The use cases are practical rather than consumer-facing. A lender might check payslips and bank statements before making a decision. An insurer could inspect claim invoices. A property manager could review a rental application, while a hiring team could screen submitted documents and written applications.

There are also checks for identity documents, document packs, and adverse media. The identity workflow is described as an Australian check across documents such as passports, licences, and Medicare cards. The adverse-media tool looks for negative news and sanctions or politically exposed person exposure, with same-name matches filtered as far as the system can manage. Those results are signals to review, not automatic findings of wrongdoing.

For researchers, editors, and publishers, the citation and AI-text checks are the more obvious fit. Developers can call the checks through a REST API, and the landing page also describes an MCP server for AI agents. That could put verification inside an automated workflow rather than making every reviewer move files between unrelated tools.

Stipple is free, and it says users can begin without signing up. That lowers the barrier to trying a check on a document or a passage of text. The more important question for a team is whether the returned evidence fits its review process, not whether an automated result can replace that process.

Evidence is useful, but it is not a verdict

The product's strongest idea is its treatment of uncertainty. Instead of presenting an unexplained percentage, it says it returns the relevant phrases, failed checks, reasoning, and a shareable report. It also says the analysis tools are stateless, with submitted text and files assessed without being retained, and that results do not include the raw input. Those are meaningful privacy claims for sensitive material, but organisations should still assess them against their own requirements before sending regulated documents.

The landing page presents benchmark results for its AI detector and says the method, sample, and misses are published. That is better evidence than an unsupported accuracy claim, but a benchmark made from selected samples is not a guarantee for every kind of writing. Language, editing, document type, and context can all affect an AI detector. Stipple itself frames a high score as a reason to look closer.

There is a clear limitation here. The AI detector only assesses running prose. Forms, tables, spreadsheets, and scanned layouts are declined rather than assigned a potentially misleading score. That is a sensible boundary, but it means a user must choose the document-verification route for many files. It also means Stipple cannot settle the authorship of every document that arrives in an inbox.

The same caution applies to forensic signals. A tampering flag or an arithmetic mismatch identifies something worth investigating. It does not, by itself, establish fraud. A clean result does not prove that a document is genuine either. The tool supports review; it does not remove the need for source checks and human judgement.

Who should use Stipple

Stipple is aimed at teams that routinely receive documents, reports, applications, or other material from people they do not fully know. Lenders, insurers, property managers, hiring teams, editors, publishers, onboarding staff, and developers building agent workflows are the clearest matches. Its combination of a no-signup web interface and programmable access also makes it suitable for testing the idea before committing to an integration.

It is not a good fit for anyone looking for an automatic fraud verdict, a definitive way to identify AI authorship, or a detector that handles every visual document equally. It is better understood as an inspection layer that surfaces reasons for suspicion, missing evidence, and inconsistencies. For decisions where those details matter, that narrower role is more useful than a confident score with nothing behind it.