[ Evidence-led insight ]

How to check AI-assisted writing for made-up facts before I publish it

A practical claim-level review workflow for AI-assisted writing. It separates source attribution from factual verification, records each claim’s evidence status and limitations, and uses the MPS/0.1 Auditor to make unresolved claims visible before publication.

By Mayone Maha Rajan · Maha Strategies LLC ·

Direct answer

Do a claim-level pass, not a vibe check. Go through the draft one substantive claim at a time — every fact, number, quote, attribution, or cause-and-effect a reader might rely on — and give each a status: VERIFIED (you confirmed it against a primary source), SOURCED (attributed to a named, citable source), BOUNDARY (honest uncertainty or an open question), ILLUSTRATIVE (an analogy or example that asserts nothing), or UNVERIFIED (no identifiable source yet). Treat statistics and quotes with no named source as UNVERIFIED by default, and confirm every AI-supplied citation actually exists — fabricated references are common. Then act on each flag: verify, cite, reword, or remove. Nothing ships while UNVERIFIED. Run the passage through Maha's free Auditor (/audit) to get this as a structured MPS/0.1 record — each claim tagged with a rationale and a recommended action — so your check is inspectable, repeatable, and its own provenance trail.

Maha method

Maha’s method treats a draft as a set of reviewable claims, not one document-level assertion. First, isolate each fact, number, quotation, attribution, or causal statement a reader could rely on. For each claim, record the supporting source and assign an evidentiary status such as VERIFIED, SOURCED, BOUNDARY, or UNSUPPORTED. Use the MPS/0.1 Auditor record to preserve the source, the checking performed, and the required next action. Resolve or qualify unsupported claims, retain the limits of the review, and only then complete editorial review. This process improves traceability; it does not independently certify that every claim is true.

[ Evidence artifact ]

MPS/0.1 Auditor — reproduce this claim-by-claim record

This artifact supports an editorial workflow; it does not independently establish factual truth.

Limits

An MPS tag records a claim's evidentiary status and the checking performed — not proof that the claim is true. VERIFIED reflects the author's confirmation, not independent certification, and the Auditor flags what to check rather than confirming facts for you. Automated tagging can miss or mis-tag a claim, so a human editor should review the record before publishing; high-stakes claims (legal, medical, financial) still require domain review. This is a provenance workflow, not fact-checking-as-a-service.

Sources

  1. Source one

    What this source establishes for the reader.

  2. Source two

    A second, independent source of evidence.

  3. Source three

    A third, independent source of evidence.

Need a source-tagged analysis for a decision? Explore consulting.