ChatGPT’s new image tool can fake IDs, prescriptions and bank alerts, researchers warn.

Image-generating artificial intelligence has always been able to make convincing scenery, portraits and product mockups. What it has struggled with, consistently and for years, is text. Letters inside generated images tended to warp, misspell themselves or dissolve into gibberish, and that weakness quietly limited how useful the technology could be for anyone trying to fake an official-looking document. That limitation appears to have largely disappeared.

Researchers and journalists testing OpenAI’s newest image-generation system say it can now render clean, legible text inside realistic-looking documents, a capability that closes one of the last practical gaps standing between casual AI image generation and convincing document fraud.

What the testing found

The findings came out of hands-on experiments described in reporting picked up by Yahoo Finance, which detailed how a journalist working with the tool produced more than 100 fraudulent images over the course of testing, including fake identification cards, fabricated prescription labels, invented receipts, spoofed bank fraud alerts and doctored news screenshots. The images were not crude approximations; testers described them as readable and detailed enough that a person glancing at one quickly, whether a bank teller, a pharmacist or a hurried customer service representative, could plausibly mistake it for the real thing.

That combination of speed and legibility is what separates this generation of tools from earlier ones. Producing a passable fake ID or fabricated bank notification used to require either genuine graphic-design skill or specialized fraud kits circulated in criminal forums. Generating one now takes a written prompt and a few seconds of processing time, collapsing a skill barrier that previously kept a meaningful share of would-be scammers out of the document-forgery business entirely.

Why the improvement matters for fraud

The core technical shift is in how the underlying model handles typography. Earlier image generators treated text as just another visual pattern to approximate, which is why generated signs, labels and forms so often came out with warped or nonsensical lettering. Reporting on the new system describes a marked improvement in rendering coherent, readable text directly inside generated images, extending to the kind of small-print detail found on government IDs, pharmacy labels and bank statements.

For fraud specifically, that matters because so much of consumer-facing document verification still relies on a human glance rather than forensic examination. A cashier checking an ID, a support agent reviewing a screenshot of an alleged bank alert, or a landlord glancing at a submitted pay stub is typically looking for overall plausibility, not scrutinizing individual pixels. Tools capable of producing clean, coherent text inside a realistic template make that kind of surface-level check far less reliable than it used to be, which is precisely the concern researchers raised in their assessment of the technology.

The scam categories most at risk

Fabricated bank fraud alerts sit near the top of the list of concerns because they plug directly into one of the most effective scam scripts already in wide circulation: a message warning a person their account has been compromised, paired with urgent instructions to move money or share login credentials to “secure” it. A convincingly rendered fake alert, sent by text or email alongside a doctored screenshot, adds a layer of visual credibility that has historically been missing from that scam format.

Fake prescriptions and IDs raise a different set of risks, touching on everything from underage purchases to pharmacy fraud to identity verification for financial accounts. Because so many verification processes, particularly at the point of sale or in a pharmacy line, depend on a quick visual check rather than a database lookup, a realistic-looking fake document can pass scrutiny that was never designed to catch AI-generated forgeries in the first place.

What safeguards exist, and their limits

Developers of these tools have not ignored the risk. Reporting on the technology noted that both OpenAI and Google have introduced measures intended to slow misuse, including embedding metadata in AI-generated images and, in Google’s case, a hidden watermarking system paired with a detection tool designed to flag synthetic images after the fact. Those measures can help platforms, banks and law enforcement identify AI-generated content when they know to look for it.

The limitation is that these protections are not foolproof and were never designed to be. Metadata can be stripped from an image file with widely available editing software, and watermarking systems generally only work if the image passes through a detection tool that is actually checking for them, which rarely happens in the split-second interactions where fraud actually occurs. The proliferation of open-source image models outside the control of any single company adds another layer of difficulty, since those tools carry no built-in safeguards at all.

What it means for consumers and institutions

The practical upshot is that the burden of catching AI-generated fakes is shifting away from visual inspection and toward verification processes that do not rely on a document simply looking legitimate. Financial institutions are increasingly leaning on backend fraud-detection systems, transaction pattern analysis and multi-factor confirmation rather than trusting a submitted image at face value. Consumers, meanwhile, are being encouraged by fraud-prevention resources to treat unexpected bank alerts, prescription requests or identity-verification prompts with skepticism regardless of how convincing any accompanying image looks, and to confirm account activity directly through a bank’s official app or verified phone number rather than reacting to a message in isolation. The Federal Trade Commission’s identity theft recovery resource outlines steps for anyone who suspects a fraudulent document or account activity has already affected them.

As image-generation tools continue to close the remaining gaps in realism, researchers say the broader lesson extends well beyond this one product. The technical hurdle that once made convincing document forgery a specialized skill has largely fallen away, and the institutions and individuals who once relied on a quick look to catch a fake are being pushed toward verification methods that assume any single image, no matter how convincing, could be synthetic.

Morning Overview produced this article with AI assistance and reviewed it against the cited sources.

More from Morning Overview