Document editing & creation using AI image generation: Alarming results for manual risk teams around the world
A few days ago, I asked ChatGPT to behave as a world-class document fraudster.
The brief? Modify one of our staff’s genuine OVO Energy bill by changinga few details - the name, account number and elements of the address.
Within seconds, we had a near-perfect forgery on our hands.
This wasn’t a complex prompt. I simply uploaded an image and asked for these text changes. The tool returned a new version where the design, layout, font, and formatting were fantastically similar to the original. The only clue that it wasn’t real? A very minor formatting quirk - a slightly misaligned pound symbol.
Let that sink in: with a single prompt and no specialist knowledge, I was able to create a fake utility bill convincing enough to fool the human eye. If I was part of an underwriting team who assessed the validity of hundreds of documents per day and say, I was on document 30 and it was 4pm on a Friday, it is entirely plausible for me to pass this document through with flying colours, potentially giving access to finance on completely false pretences.
And it’s only going to get worse because technology is only going to get better.
The document in question? See below:
AI Image Generation Is Advancing at a Staggering Pace
Tools like ChatGPT’s image editor, Midjourney, DALL·E, and others have transformed from quirky art generators to powerful instruments capable of photorealism and document precision. What used to take professional designers or fraudsters hours with Photoshop can now be done by anyone in under a minute.
Even more concerning is that this technology is getting better. Fast. Each iteration improves coherence, font matching, logo clarity, and layout retention. And this is with the most basic of prompts that any risky director desperate to hide the true state of their finances, or chancer trying to gain a mortgage far greater than their affordability should allow, could simply enter. Cifas reported that 16% of all mortgage applicants in the UK have admitted to massaging the details in their application. I suspect this number will increase with new tools at the hands of the many.
AI image generation isn’t just getting better at drawing hands, turning one’s self into an action figure or a Studio Ghibli artefact - it’s learning to imitate reality.
Why This Is a Threat to applications
I mean, it's obvious isn't it?
“Generate an invoice that proves I paid for a £2,000 laptop on Friday, 13th November 2024”
or
“Change my Salary from £20,000 per year to £200,000 per year so it seems I can afford a penthouse flat in Mayfair”
The danger lies in what people can now do with ease:
Forge bank statements with altered balances or account holder names
Create fake proof of address documents for credit checks, or commit multi-account creation fraud on gambling websites
Generate doctored payslips for mortgage or rental applications
Spoof identity documents like passports, driving licenses, and visas
For industries that rely on document-based verification - finance, insurance, property, gambling, even government - this is a ticking time bomb.
The human eye is no match for modern AI fakery.
Worse still, many companies still rely on manual document reviews. That worked fine when forgeries were amateur Photoshop jobs or basic PDF edits. But now? Fraudsters are using AI.
Are you?
The OVO Bill Example: A Microcosm of a Macro Threat
Let’s go back to that OVO Energy bill. Here’s what made it so worrying:
The logo was retained and near-perfect (though slightly clipped in the first attempt).
The fonts were indistinguishable from the original.
The layout mirrored the source document precisely.
The changes were cleanly integrated without affecting the design.
If I hadn’t created the original, I’d struggle to prove the second wasn’t legitimate.
Now imagine a fraudster submitting that as part of a tenancy application, or to secure a vehicle lease, or to access a buy-now-pay-later service.
Would your fraud team catch it? Would your onboarding software flag it?
What’s the likelihood of it genuinely slipping through the net - just like hundreds of similar fakes already are?
“Use AI to Fight AI Fraud” Isn’t a Catchphrase - It’s a Survival Strategy
It’s tempting to think of AI as the problem. But really, it’s just the weapon. The bigger issue is being underprepared.
Many companies have invested heavily in digital transformation, yet are still handling fraud prevention, when it comes to documents, like it’s 2015.
The good news? The same AI that’s creating these threats can also detect them.
AI-powered document forensics tools, like Fraudfinder, can spot subtle inconsistencies in documents invisible to the human eye. We scan for manipulated metadata, font mismatches, compression artifacts, layering anomalies, and more. All on the backdrop of a large adaptive database of global financial templates which are being added to every single day.
Where humans rely on experience to spot a document anomaly, Fraudfinder relies on data, structure, templates and machine–learning algorithms capable of decoding barcodes, restoring deleted transactions and spotting a font that just doesn’t quite match the correct fonts seen a million times over for a specific bank.
They work at scale, they don’t get tired, and crucially, they evolve with the fraud.
Fraudfinder can now identify documents that have been generated using ChatGPT
There are certain artifacts within AI-generated content, from ChatGPT and other LLMs that Fraudfinder can now identify and flag to users within seconds.
Check out the GIF below:
Technology is always evolving and it is essential to evolve document compliance practices to match its trajectory.
What Can You Do Right Now?
If you’re in any industry that processes documents for income or proof of address, here are 5 immediate actions to take:
Audit Your Current Processes: Understand how you’re currently verifying documents. Is it manual? Is it automated? If so, how intelligent is the automation?
Educate Your Team: Show them examples like the OVO bill. The more people understand the sophistication of AI fraud, the better prepared they’ll be.
Test Your Defences: Try generating fake documents yourself. If they pass your internal checks, that’s your canary in the coal mine.
Adopt AI-led Verification Tools: Deploy tech that can detect fakes using AI, not just formatting templates. Look for tools with strong track records in font analysis, pixel-level inspection, and pattern recognition. There are ways Fraudfinder and other tools can tell when an image has been AI-generated and when it hasn’t.
Shift from Reactive to Proactive: Don’t wait until your fraud losses spike. Assume fraudsters are already using this tech and beat them to the punch.
Shift to PDF collection where possible: ChatGPT can’t mimic and generate PDFs… yet. PDFs are the gold standard of document proof because there is so many datapoints attached to them. They can be validated with greater precision than an image.
Don’t Bring a Magnifying Glass to a Gunfight
The document I created wasn’t meant to scam anyone. But someone else, somewhere, is doing exactly that - with better prompts, more malicious intent, and zero hesitation.
AI image generation is no longer a novelty. It’s a serious threat vector.
And if you’re still relying on manual reviews and PDF eyeballing, you’re already behind.
It’s time to fight AI with AI.
Want help testing your defences or curious how real your onboarding documents really are?
Get in touch - or run a trial with Fraudfinder. We’ll show you what’s getting through.
