ChatGPT and false expense reports: an invisible threat on the horizon?

Fraud just a click away

ChatGPT can now generate ultra-realistic fake expense reports in under a minute. Since the integration of the new OpenAI image generator, we have seen a radical change in the quality of productions:

• ✅ The text is perfectly clear and legible
• ✅ The overall result is impressive (folds, tasks, credible details...)
• ✅ Generation takes just 10 seconds, with a simple prompt

The result? The possibility of creating fake cash receipts that are almost undetectable to the untrained eye. Even more worrisome: expense scanner applications save them without any alerts.

A troubling demonstration

Raphael Chenol recently shared a demonstration as brilliant as it is disturbing on LinkedIn. In his post, he presents a real cash receipt side by side and its counterfeit version generated by AI.

Changing a date, an amount, or even the name of a restaurant is child's play. This publication quickly went beyond the professional framework of LinkedIn to be taken up by TechCrunch and even broadcast on the France 2 TV News in a topic dedicated to the excesses of AI.

Flaws that are already visible

On closer inspection, an expert eye can still detect some anomalies in these false documents. In the example shared, VAT is incorrectly calculated and several subtle elements betray the artificial nature of the document.

But this is precisely the danger: most controllers, whether human or automated, don't take the time to perform these thorough checks. And with the constant improvement of AI models, these imperfections will gradually disappear, making fakes completely undetectable in the long run.

Professionals are worried

After relaying this information to my LinkedIn post, numerous experts shared their concerns and thoughts. Here are some of the most relevant perspectives:

A chartered accountant testifies: “We are already starting to see similar fraud attempts. The problem is that our current verification tools are absolutely not designed to detect this type of AI fraud.”

A compliance manager in a large company: “We are going to have to completely rethink our processes for validating expense reports. A simple visual check will no longer suffice.”

A developer of anti-fraud solutions: “We are already working on detection tools that analyze metadata and mathematical inconsistencies. The race between fraudsters and detectors is on.”

A specialized lawyer: “The current legal framework is not at all adapted to these new forms of fraud. Victim businesses will find it difficult to defend themselves legally without formal proof of forgery.”

A shared responsibility

Let's be clear: the objective is not to encourage these fraudulent uses. But it is clear that they are now accessible to all and represent a real challenge for the financial security of companies and administrations.

I am not an AI skeptic, quite the opposite. It's my job, I work there daily and we even train marketing teams in creative AI. But precisely, when you know the possibilities, you also have a responsibility to ask the right questions.

Critical questions for the future

This new reality requires us to question ourselves collectively:

• Are we heading for a new era of invisible fraud? Unlike traditional scams that leave traces, these fake AI-generated documents could become undetectable.
• Are control tools ready for this challenge? Current systems, designed to identify classical anomalies, seem unarmed in the face of this new generation of fakes.
• Who should adapt first? The tax administration? Software publishers? The businesses themselves? The answer probably involves all of these actors.

Emerging solutions

Several approaches are already mentioned by professionals in the sector:

1. Enhanced authentication : require additional proof for expenses over a certain amount.
2. Automated verification of calculations : develop algorithms that systematically check the mathematical consistency of documents (VAT, totals, etc.).
3. Blockchain for traceability : some businesses are considering systems where each transaction would be immediately recorded in an unchangeable chain.
4. Team training : make accounting departments aware of new fraud methods and the subtle clues to look for.

Conclusion: vigilance and adaptation

Generative AI is a great tool that opens up countless creative and productive possibilities. But like any powerful technology, it can be misused for malicious purposes.

The real question is not whether we should hold back innovation, but rather how to adapt our control systems and regulatory framework to this new reality. It is a collective challenge that requires the collaboration of all actors: companies, regulators, developers and users.