Document Detection

AI document detection is the practice of using machine learning, computer vision, and deterministic forensics to decide whether a document is authentic, tampered, or fully AI-generated.

Traditional review relies on a human comparing fonts, layouts, and totals. AI document detection automates that work at scale and adds checks no human can do reliably — pixel-level error level analysis, PDF object inspection, MRZ checksum validation, font fingerprinting, and detection of generative-AI signatures.

TrueDoc is built specifically for this category: a universal fake document checker that scores authenticity, lists forgery signals, and explains every finding.

TrueDoc runs a layered detection engine. The first layer extracts text and structure via OCR and PDF parsing. The second layer runs deterministic forensics — Error Level Analysis (ELA), metadata inspection, font and kerning checks, MRZ checksums, and math reconciliation on financial documents.

The third layer evaluates AI-generation signals: telltale artifacts from current-generation text-to-image models, layout regularities common in generated PDFs, and inconsistencies between rendered content and document structure.

Findings from all layers feed a final scoring layer that produces a Document Trust Score plus a categorized list of Critical Issues, Warnings, and Positive Indicators.

Inconsistent fonts, kerning, or text alignment across a single document. Edited PDF metadata (Producer, Creation Date, ModDate) that disagrees with the visible content. Missing or invalid MRZ checksums on IDs and passports. Math that does not reconcile on paystubs or bank statements.

Pixel-level splicing artifacts visible under ELA. Compression boundaries that suggest cut-and-paste edits. Layout regularities and rendering artifacts typical of generative-AI image and PDF models.

Recycled real documents reused across multiple submissions, identifiable by document hash and content fingerprint.

Government-issued IDs, passports, driver's licenses, and residence permits. Bank statements, paystubs, tax forms, and proof-of-income documents. Invoices, receipts, purchase orders, and proof-of-payment screenshots. Contracts, NDAs, employment offers, and reference letters. Diplomas, certificates, and academic transcripts. Utility bills, lease agreements, and proof-of-address documents.

Screenshots of payments, transfers, dashboards, and chat conversations — frequently faked to support scams and disputes.

AI-generated fake documents are now produced in seconds at near-zero cost. Lenders approve loans against forged income docs, landlords sign leases against fake paystubs, marketplaces refund disputed payments against doctored screenshots, and HR teams hire candidates with fabricated credentials.

A single missed forgery typically costs more than a year of automated review. Modern teams need a document detection layer that catches generative forgeries and produces audit-ready evidence — the operational counterpart to the free document detection tool on this page.

Every TrueDoc analysis returns a Document Trust Score — a single number that summarizes overall authenticity risk based on the underlying forensic and AI-generation signals.

The score is paired with structured evidence: Critical Issues that strongly indicate fraud, Warnings that warrant human review, and Positive Indicators that support authenticity. Teams use the score to auto-approve clean documents and route flagged documents to manual review.

For programmatic workflows, the same score is exposed via the document fraud detection API and the document trust score API.

Document detection is not a single check — it is a family of techniques applied together. Authenticity detection asks whether a document is genuine or forged. Tampering detection looks for edits to a real document: replaced names, altered totals, spliced regions, or modified PDF objects. AI-generated document detection targets synthetic files produced by current-generation image and PDF models, using layout regularities, generator fingerprints, and rendering artifacts as signals.

Template and layout detection compares a document against known issuer templates (a bank's statement layout, a government ID format) and flags deviations. Metadata detection inspects PDF Producer, Creation Date, ModDate, XMP fields, and EXIF tags for edits that disagree with the visible content. OCR-based content detection cross-checks the extracted text and math against the document's rendered structure, catching swapped numbers and inconsistent totals.

TrueDoc runs all of these in one pass on every upload — the reason a single document detection call surfaces the full risk picture rather than one narrow signal.

Error Level Analysis (ELA) recompresses the image and highlights regions whose compression signature differs from the surrounding pixels — the classic signature of splicing and photoshopped edits. PDF object inspection walks the file's incremental update history, embedded fonts, and object streams to catch content that was edited after the document was first saved.

Font and kerning fingerprinting compares glyph shapes and letter spacing against known issuer templates; forged documents almost always slip on this because attackers rarely have the original font. MRZ checksum validation applies the ICAO 9303 check-digit formula to the machine-readable zone on passports and IDs — a broken checksum is deterministic proof of tampering.

Metadata cross-checks compare PDF and EXIF timestamps against the visible content. Generative-AI artifact detection looks for the telltale patterns current text-to-image and PDF models leave behind. Hash and content-fingerprint checks catch recycled real documents being reused across multiple scams.

These three terms are used interchangeably online, but they solve different problems. Document detection asks 'is this file authentic?' — it analyzes the document itself for forgery, tampering, and AI-generation signals, and returns a trust score with evidence. It works on any document type and does not require a source of truth.

Document verification asks 'does this document match a source of truth?' — it cross-references the document against issuer databases, bank APIs, or government registries. It is stronger where those integrations exist (bank statements via open banking, passports via ICAO PKD) and unavailable where they do not.

Identity verification asks 'is the person presenting this document who they claim to be?' — it combines a document check with a live selfie or biometric match. Identity providers (Persona, Onfido, Jumio) focus here. TrueDoc is a document detection tool: fast, universal, and file-only. Enterprise plans integrate with verification and identity layers when a workflow requires them.

Lending and BNPL: fake paystubs and edited bank statements are the single largest source of first-party fraud loss in unsecured lending. A document detection pass before underwriting catches AI-generated income proofs and doctored balances that manual review misses.

KYC and onboarding: AI-generated IDs and passports now pass casual review at scale. Document detection flags synthetic IDs, invalid MRZ checksums, template mismatches, and recycled real IDs reused across applicants — the fraud patterns identity-only verification is not designed to catch.

HR and hiring: fabricated diplomas, reference letters, and employment verification letters are increasingly generated by LLMs. A document detection step in background checks flags credential fraud before offers go out.

Marketplaces and P2P: proof-of-payment screenshots, shipping labels, and invoices are routinely faked to drive refunds and chargebacks. Document detection on submitted evidence dramatically reduces dispute loss without blocking legitimate users.

A useful document detection result is more than a red/green label. It answers three questions in one pass: is the file authentic, has it been tampered with, and was any part of it AI-generated? Each answer should ship with the underlying evidence — the exact PDF object that was edited, the metadata field that disagrees with the visible content, the region of the image where the compression signature breaks, the MRZ line whose checksum fails.

TrueDoc's document detection engine returns a Document Trust Score, a categorized list of Critical Issues, Warnings, and Positive Indicators, and — for every finding — a short human-readable explanation. Reviewers do not need to trust a black box: they can see why a document was flagged and defend the decision downstream.

That structure makes document detection usable in the real workflows where forgery hurts most. Underwriters can auto-approve clean income docs and route only flagged files to manual review. KYC teams can catch AI-generated IDs and template mismatches identity-only checks miss. HR and background-check teams can screen diplomas and offer letters for LLM fabrication. Marketplaces can detect doctored payment screenshots before refunds are issued.

If you need to check a document for fraud right now, upload it above — the free document detection tool returns a full report in seconds, no signup required.

To tell if a document is AI-generated, look for six signal families that current-generation image and PDF models still leak. First, layout regularity that is too clean — spacing, margins, and grid alignment that no human-produced or scanned document ever achieves. Second, font and glyph inconsistencies: generative models often mix subtly different glyph shapes for the same character across a page, or render kerning that no real typesetter would ship.

Third, rendering artifacts around edges of stamps, signatures, logos, and table borders — soft halos, warped curves, or ghosting that betray a diffusion or text-to-image pipeline. Fourth, structural mismatch between the rendered pixels and the underlying document: an AI-generated PDF frequently has no real text layer, or a text layer that disagrees with what the eye reads. Fifth, metadata that either is missing entirely, or names a Producer / Creator that does not match the document type (a bank statement 'produced' by a screenshot tool, a passport image 'created' in a chat app).

Sixth — and most reliable — generator fingerprints: statistical patterns in pixel noise, compression signatures, and object structure that map back to specific model families. A modern AI document detector runs all six checks in one pass and returns a Document Trust Score with per-signal evidence, so a reviewer can see exactly which signals fired and why the document was flagged as AI-generated.