The evolution of AI-generated documents

In February 2026, only two leading AI models could convincingly create fake documents. Today it’s a common feature among the most popular models.

It only takes a few words: "change this to a Walmart Receipt,” “Put my name on this bank statement." No questions, no hesitation, just image buffering, then document fraud.

In Resistant AI's webinar, Fighting Fraud in an AI World, we traced this and several other AI-powered fraud tactics, educating institutions on how to combat fraud in the modern era.

Which tools can do this?

We’ve logged nine mainstream models that can now edit and create document images on request:

• ChatGPT, Gemini, and Flux - The highest fidelity of the group, with reliable text and realistic paper.
• Reve, Recraft, Ideogram, and Luma AI - Same core capability, packaged for design and marketing users.
• Meta AI and Grok - Still behind the quality of the other models but can produce convincing documents at a glance.

How did we get here?

You’ll find a brief overview of the AI-generated document timeline below:

• Before March 2025: Models could fake a face for an identity check but not text. One of our early tests produced a Chase bank statement that read like a newspaper.
• March 2025: GPT Image 1.0 delivers the first clear, readable text. The output still looks generated.
• August 2025: Nano Banana 1.0 drops the typical AI blur, sharpens detail, and starts producing believable forms.
• December 2025: GPT Image 1.5 solves typos. Alibaba and ByteDance ship models that can be run locally.
• February 2026: Nano Banana 2.0 works at 4K, documents photographed on a desk stop looking flat, and believable identity documents come within reach.

Eleven months from readable to credible. But these improvements don’t mean that the models are chasing perfection. GPT Image 1.5 and 2 actually provide lower resolution outputs that are better at fooling fraud checks, 1024x1024 reads like a phone photo of a real document rather than a render.

Realism can now include deliberate imperfection. Detection tuned to spot clean edges and blemish-less faces might be hunting for the wrong artifact.
 

How do fraudsters get past the guardrails?

Guardrails are always on in the web interfaces, and they slow this down rather than stop it:

Prompting "Generate a fake bank statement" usually gets refused. "Recreate this with different details" usually does not. The same applies to caveats like “as a joke,” “as an example,” or “for private use,” can also circumvent most hesitancy.

A watermark can be a dead giveaway, but removal can be achieved with a paid subscription.

Access a model through an application programming interface (API), or run one of the locally deployable models, and the guardrails can come off entirely.
 

Why are AI-generated documents dangerous?

The cost and skill floor have collapsed. Fraudsters no longer need Photoshop skills, a template farm, or a dedicated document editor. They need a plausible document, a prompt, and a few attempts.

That widens what can be faked. A template farm only sells templates someone bothered to build, which keeps fraudsters pointed at high-volume documents from large institutions. Prompted editing works on whatever document gets uploaded, so a regional utility bill, a niche payroll statement, or a business licence issued by a single county all come into range.

The documents you receive least often, and therefore verify least confidently, are now among the easiest to fabricate.

It also erodes one of the better signals defenders have. Template-farmed forgeries share a common source, so they cluster across unrelated applications and expose the operation behind them. A prompted edit starts from a genuine document and changes only the detail the fraudster needs, leaving each submission looking unconnected to the last.

Then there is the crime that the document unlocks. Pay stubs approve a loan, invoices release a payout, certificates of incorporation onboard faulty merchants. Document fraud sits at the front of decisions that move money, a single convincing file can cost far more than the seconds it took to generate.
 

What can your organisation do about it?

Stop scoring documents on appearance. Visual quality no longer correlates with authenticity in either direction. Analyse how the file was constructed, not just the image. Metadata, generation traces, and internal inconsistencies survive when the pixels look clean.

Serial fraud is also a huge red flag. Look across submissions, not only within one. A single generated document can be near-flawless while the pattern across applicants gives the operation away.

Want to learn more about the evolution of AI-generated documents? Watch the full video here.

This is one tactic on a long list available to criminals in 2026. To see the rest, watch the webinar or get in touch with the Resistant AI team.