AI-Generated Fake Invoices: Detection & Prevention Guide

Key takeaways

• Yes, an invoice or receipt generated by AI often passes visual inspection: on the AppZen audit platform, the share of flagged fake receipts that were AI generated rose from 0% in March 2025 to 70.8% by mid May 2026, in fourteen months.
• The amounts are deliberately small: the average AI generated receipt was $101, with a median around $32, just under the usual vigilance thresholds.
• On the penalty side, Article 1737-I of the CGI (General Tax Code, France's main tax statute) sets a fine of 50% of the invoice amount for issuing an invoice that does not correspond to an actual delivery or service, on top of the disallowance of deducted VAT.
• Reliable detection no longer relies on the naked eye: it combines file metadata, arithmetic consistency, IBAN (International Bank Account Number) checks, and cross referencing with purchase orders and delivery notes.

Can a fake generated by ChatGPT really fool an accountant?

Yes, and the data confirms it. An accountant reviewing 300 supplier invoices a month has only a few seconds per document: logo, amount, VAT, date. That is exactly what image generators now reproduce best.

The shift has been measured. AppZen detected 1,471 AI generated receipts submitted by 745 employees across 174 companies over the twelve months ending May 15, 2026, and roughly a third of those employees submitted more than one. Before this, receipt fraud relied on templates from specialized websites: these "template" fakes accounted for 95% to 100% of caught receipts a little over a year ago, compared with 29% in May 2026.

What has changed is not the trick, it is the arithmetic. Language models produce internally consistent documents: VAT correctly calculated from the pre-tax amount, payment terms matching the stated conditions, supplier details in the right format. The arithmetic that used to give away hand-made fake invoices is now handled automatically.

The check that used to involve verifying that 20% VAT adds up correctly no longer catches anything: that is precisely the part the machine gets right without fail.

How big is the phenomenon really in 2026?

Two independent sources point in the same direction, but the figures need careful reading. The 70.8% figure is a share of flagged fraudulent receipts, not a share of all expense claims, nor a confirmed fraud rate across the economy.

The self-reported data is starker. An Emburse survey conducted by Atomik Research among 2,000 US and UK employees between May 5 and 8, 2026 found that 40% of US respondents and 29% of UK respondents admit to having generated a fake receipt using AI; among them, 40% used AI tools paid for by their employer and 9% built their own tool. In detail, 19% fabricated a purchase that never happened, 15% inflated the value of a real expense, and 6% recreated proof for a receipt that was genuinely lost.

For supplier invoices, the signal exists too: Ramp states that its software detected more than one million dollars in fake invoices over three months. And AI driven document fraud mainly hits banking, insurance, and finance departments that process invoices and expense claims at volume.

What signs still give away an AI generated document?

No single signal is enough: there is no one feature whose absence reliably exposes a fake. Useful detection is a stack of independent checks.

Here are the ones that still work, in the order a finance department can apply them.

Metadata. The creation date, the producing software, or provenance information can reveal an image generator, unless a screenshot has wiped them out. A PDF "invoice" whose producer is a drawing tool, or a photo with no device EXIF data, deserves a question.

Formatting comparison. Compare the document against genuine templates from the same issuer to spot an unusual font or header. This is the most effective check in practice: you already have a dozen invoices from the same supplier on file.

Bank details. An IBAN whose country code, check digits, or bank identifier do not match the stated institution is a strong fraud indicator. Requests to change bank details remain the typical vector for supplier payment diversion.

External consistency. A fake is consistent with itself, rarely with your own records: purchase order, delivery note, contract, SIREN number (French business registration number), actual existence of the supplier.

What does a company risk by recording a fake invoice?

The tax risk is often underestimated, because it is not calculated on the tax evaded. The fine under Article 1737 of the CGI is based on the transaction amount, not the tax evaded: a discrepancy involving 400,000 euros in flows automatically generates a 200,000 euro fine. It is also a standalone penalty, applied regardless of any adjustment to the tax base.

Three levels to keep in mind:

1. Tax fine: 50% of the amount in question (Article 1737 CGI), with disallowance of wrongly deducted VAT.
2. Invoicing failure: 50% of the transaction amount, reduced to 5% if the supplier proves, within thirty days of a formal notice, that the transaction was properly recorded.
3. Criminal penalty: forgery and use of forged documents is punishable by 3 years in prison and a 45,000 euro fine, rising to 5 years and 500,000 euros for tax fraud.

The French Constitutional Council (Conseil constitutionnel) upheld this severity: the 50% rate was not found to be manifestly disproportionate, since the breaches in question concern a transaction carried out by professionals in the course of their business and necessarily involve intent.

Does mandatory e-invoicing solve the problem?

Only partially, and over a specific scope: domestic B2B flows. The reform took effect on September 1, 2026: all companies, regardless of size, must be able to receive electronic invoices, and large companies and mid-sized enterprises (ETI) must issue all their invoices in electronic format and transmit their e-reporting data. The obligation to issue invoices and perform e-reporting extends to small businesses, SMEs, micro-enterprises, and freelancers on September 1, 2027.

The anti-fraud benefit is real: an invoice that passes through a Plateforme de Dématérialisation Partenaire (PDP, a government-registered e-invoicing platform) arrives as structured data, linked to an issuer identified in the official directory, rather than as an image emailed as an attachment. The French government's stated goal is to fight VAT fraud and simplify reporting obligations.

The gap persists elsewhere. Expense claims and documents outside the scope (foreign suppliers, file attachments, proof of address, or payslips) still arrive as PDFs or photos. And a fake can perfectly well be issued from a legitimate account. For the regulatory framework, Sygnet details the rules and penalties in its guide to e-invoicing in France and the blind spots of the directory in its analysis of guarantees and fraud risks linked to the directory.

How should you actually adapt your validation workflow?

Take an accounting firm of 40 people processing 300 supplier invoices a month per client. Exhaustive human review is impossible; the goal is therefore to make fakes costly to produce, not to inspect every one by eye.

Four useful decisions:

• Require the original file, not a screenshot. A cropped photo or a screenshot wipes out metadata. Reject them as a rule above a certain threshold (for example 50 euros), since AI generated fakes are smaller, around $100 versus a little over $180 for a template based fake.
• Lower the thresholds for random checks. AI fraud hides below thresholds, not above them. A random 5% sample of small amounts is worth more than a 100% check above 500 euros.
• Lock down changes to bank details with a phone call back to a number already on file, never the one shown on the invoice.
• Automate the matching of orders, deliveries, and invoices. Software with duplicate detection, metadata inspection, and image manipulation analysis provides proof, not just suspicion.

The goal is not to recognize a fake, it is to make it impossible to pay a document that leaves no trace in your own systems.

Sygnet publishes a document extraction platform used in particular by accounting firms; to measure what an automated check actually detects on your own documents, the extraction accuracy evaluation checklist offers a testing method before deployment.

FAQ

Is an accountant liable if they record a fake invoice?

Their job is to present the accounts, not to investigate. But the fine under Article 1737 hits the company, and the 50% or 5% fine cannot be applied unless the tax authority has notified the offender in writing. In practice, the firm has every interest in documenting its checks (matching, missing documents, alerts flagged to the client) to demonstrate due diligence in the event of a tax audit.

Can an AI generated image be detected automatically?

Yes, tools exist and arrived quickly: Dext rolled out an AI generated image detection feature in early April 2025. But a check based on the image alone no longer reliably detects AI generated fakes. Several layers need to be cross checked: file, image, data, device, person.

Which documents are most commonly falsified in France today?

Beyond invoices, identity documents and proof of solvency. Passports and ID cards are used to open accounts and for KYC (know your customer) onboarding; proof of address, payslips, and tax notices feed into credit or rental applications. A letting agent verifying a payslip should therefore cross check the salary against the tax notice and, where possible, the employer itself.

At what volume should these checks be automated?

There is no official threshold, but the logic is one of cost per anomaly avoided. Generative tools make it possible to produce visually authentic and arithmetically consistent documents in under ten minutes, which makes manual review structurally a losing proposition from just a few hundred documents a month onward. An ROI calculator is more telling than an arbitrary threshold.