deepidv Logo
AI Document Verification

The paystub is fake. The applicant is fake. Catch both.

Detect AI-generated documents at 96% or higher accuracy, with a forensic evidence chain attached to every decision.

Test a documentSee sample forensic report

paystub.pdf

Forensic document report

font drift
metadata

Generator match

0.96

Live decision

Likely synthetic

96%

Latency

<500ms

Action

Block

PDF generator signature inconsistent with issuer
Payroll layout artifacts match synthetic template family
Employer and salary pattern require issuer review
0%+

AI document detection accuracy

<500ms

Decision time per document

0

Document types supported

$0B+

Annual US fraud loss to fake documents

Coverage

Ten document types. One forensic engine.

Each workflow gets the document-specific checks it needs: metadata, pixels, issuer patterns, expected fields, and counterfactual evidence.

Paystubs

Detect AI-generated paystubs from ChatGPT, Claude, and dedicated paystub generators.

W-2 / 1099

Tax document forensics including issuer validation and employer cross-reference.

Bank Statements

Detect fabricated statements with transaction pattern analysis and PDF metadata forensics.

Government IDs

Driver license, passport, and national ID forensics including hologram and microtext analysis.

Utility Bills

Address verification documents with issuer cross-reference and PDF metadata forensics.

Receipts

Retail and restaurant receipt forensics for returns fraud and expense fraud detection.

Medical Records

Detect AI-generated medical documents in disability claims and insurance claims.

Contracts

Detect AI-generated signatures, modified clauses, and synthetic contracts.

Insurance Claim Photos

Detect AI-generated incident photos in P&C and auto insurance claims.

Tax Returns

Form 1040 and supporting schedule forensics for lending and underwriting.

The Pipeline

Five forensic layers. One verdict.

The engine does not just say fake. It shows which signals made the document unsafe to trust.

01

Metadata Forensics

PDF metadata, EXIF, edit history, generator software signatures, and file structure analysis.

02

Visual Forensics

Pixel-level analysis for AI-generation artifacts, font inconsistency, and layout anomalies.

03

Content Forensics

Cross-reference document content against expected patterns such as employer names, address formats, salary ranges, and tax codes.

04

Issuer Cross-Reference

Validate against issuer databases where available, including The Work Number, ADP, and government registries.

05

Decision

Combined forensic verdict with full evidence chain attached. Confidence score and counterfactual returned.

Evidence

Every decision returns a forensic report you can defend.

Analysts get the same signal chain the model used, formatted for escalation, audit, and examination support.

Forensic markers

Specific artifacts identified, with pixel coordinates or document offsets.

Generator attribution

Probable AI tool used to generate the document, when identifiable.

Confidence interval

Statistical confidence with upper and lower bounds.

Counterfactual

What document properties would change the verdict.

Examination format

OCC SR 11-7 model documentation auto-generated for regulated industries.

Integrations

Plugs into your existing underwriting and claims stack.

Drops into the AI stack you already run. Connect the agents, channels, and data systems where verification has to happen — no rip-and-replace.

Who uses it

Built for every document-receiving workflow.

Mortgage Lender

Paystub, W-2, bank statement, and tax return forensics at loan origination

Auto Lender

Paystub forensics at F&I window with sub-60-second turnaround

Property Manager

Paystub and bank statement forensics at tenant application

Insurance Carrier

Claim photo and medical record forensics at FNOL

Enterprise HR

ID, contract, and tax document forensics across hire-to-retire lifecycle

Ready when you are

Stop approving applications backed by ChatGPT paystubs.

Send us a sample document. Get back the forensic report in 30 seconds.

Test a documentSee sample report

Decision record

Evidence attached

Audit trail

Signal chain

Built to fit the workflow you already run.

FAQ

Let's answer your questions.

96% or higher on most document types. Accuracy varies by document type, generator quality, and adversarial sophistication. Confidence intervals are returned with every decision so analysts know when to escalate.
ChatGPT, Claude, Gemini, dedicated paystub generators, deepfake document tools, and most modern image generation models. The detection model is updated continuously as new generators are released.
Edit-history forensics catches most cases. PDF metadata, EXIF data, and edit-trail analysis surface modifications. Counterfactual analysis tells you which specific edits were made.
Yes. Claim photo forensics is a supported document type. Detects AI-generated incident photos, modified vehicle damage, and fabricated property damage at FNOL.
Sub-500ms per document on standard formats. Larger files (multi-page tax returns, full bank statements) can take up to 2 seconds.
Under 2% on production traffic for most document types. False-positive rate is configurable per workflow. High-stakes workflows can be tuned for lower false-positive rate at the cost of slightly higher false-negative rate.
Yes. Every decision returns OCC SR 11-7 model documentation including the model version, forensic markers, confidence interval, and counterfactual analysis. Exportable in FFIEC and EU AI Act formats.
API or SDK integration. Most customers integrate via webhook from their existing underwriting or claims platform. Reference integrations available for Encompass, Blend, Guidewire, Duck Creek, and Workday.