Data leaks, poor content, fragile decisions: recent incidents do not betray a technical failure, but a renunciation of human judgment. How to regain control.
Imagine the director of a national cybersecurity agency copying internal documents from his administration into a consumer chatbot. It’s not fiction. In early 2026, the American press revealed that the acting director of CISA, the federal agency responsible for defending critical networks in the United States, had placed several documents marked “for official use only” in the public version of ChatGPT. The affair was detected by the administration’s own surveillance systems. The agency indicated that access had been authorized and use limited, but the fact remains: the man supposed to embody the requirement for the protection of sensitive data entrusted it to a system beyond any control.
If it can happen there, it can happen to you. And the problem is almost never the tool. He sticks to one gesture: suspending judgment, delegating without checking, delivering quickly. In short, let go of your hand.
1. When data escapes
Shadow AI – the use of artificial intelligence tools without agreement or supervision from IT management – is no longer marginal. According to data published by IBM, the proportion of employees using generative AI increased from 74% to 96% between 2023 and 2024. And 38% of them admit to having entered sensitive professional information without authorization.
The Samsung case remains the most documented illustration. In 2023, in just twenty days, the company recorded three leaks via ChatGPT: one engineer pasted buggy source code from a semiconductor database, a second confidential test models used to spot defective chips, and a third the recording of an internal meeting to summarize. The data entered could feed the training of the model, out of reach of the company. Samsung subsequently banned generative AI tools in the workplace.
The leak does not only concern the technical teams. In February 2026, American federal judge Jed Rakoff deemed communicable the analyzes that a former manager of GWG Holdings, prosecuted for stock fraud, had produced with an AI assistant to prepare his defense before transmitting them to his lawyers. Reason: an exchange with an AI may be assimilated to an exchange with a third party, and thus lose the protection of attorney-client privilege.
In your company, do you know precisely which AI tools your teams are using today, and with what data?
2. When AI invents and no one checks
The second risk is more insidious, because it does not trigger any alerts. AI produces text that is fluid, credible… and false. If no one checks, the error passes.
In October 2025, a consultancy agreed to partially reimburse the amount of fees collected from the Australian government for a report riddled with AI-generated errors: non-existent academic references and even a fabricated quote attributed to a federal court judge.
The world of law collects the same mishaps. As early as 2023, in the Mata v. Avianca case in New York, a lawyer was fined $5,000 for submitting to the court six court decisions purely invented by ChatGPT. In Canada, in 2024, a lawyer had to pay the opposing party’s costs for two fictitious case laws cited in a child custody dispute. The missing reflex is always the same: reread and check.
Alongside these spectacular cases, a quieter risk eats away at daily productivity: work slop, this AI-generated content that appears professional but lacks substance, and forces the recipient to start over again. According to a BetterUp Labs / Stanford study published in the Harvard Business Review in September 2025, each occurrence costs on average nearly two hours of correction, or approximately $186 per month per affected employee.
Three mechanisms fuel these excesses: excess confidence (we suspend our critical thinking in the face of a well-written text), the race for adoption (managers deploy uses faster than controls follow) and excessive delegation (we entrust the machine with the synthesis, the advice, sometimes the decision, without checking the result).
Of your latest high-stakes deliverables, how many have been reread by a human capable of spotting a substantive error?
3. Take back control
Banning does not work: pure and simple blocking pushes uses into the shadows. An isolated charter or training holds no weight compared to the pressure to deliver. To reduce risk structurally, an organization must build three capabilities that directly address the three identified risks.
Visibility of real uses. As long as the company does not know which tools are used, by whom and with what data, it remains blind. This requires detecting not only officially declared applications, but also browser extensions, AI functionalities integrated into business software and uses via personal accounts.
Control at the time of interaction and execution. It is no longer enough to block sites. You must be able to analyze the content of a message before it goes to an external model, and above all monitor the actions that the AI proposes to execute. This requirement becomes critical with autonomous agents, capable of acting directly on information systems.
Local execution for sensitive data. For cases involving professional secrecy, strategic information or high-risk personal data, systematically passing information to an external supplier creates permanent exposure. Running the model on internal infrastructure – or via a secure gateway to a more powerful model when necessary – keeps the data within the company’s perimeter.
These three capabilities do not replace the culture of use, they make it operational. Without visibility, we cannot train or provide accountability. Without real-time control, good reflexes remain insufficient. Without local execution, employees are sometimes forced to choose between productivity and compliance.
This culture is broken down by profession. A lawyer never sends the facts of a sensitive case to an external tool. A manager requires human review of any deliverable with customer or strategic impact, as well as traceability of the use of AI. A developer tests and secures the generated code instead of believing it to be ready for use.
There is evidence that this approach works. A construction sector platform obtained the first global ISO 42001 certification in six months by structuring both its control tools and its validation processes. According to publisher feedback, a large global health technology company has, for its part, reduced unauthorized use of AI to almost zero on more than 60,000 users by combining automated discovery and interaction monitoring.
The cost of inaction
The question is no longer whether AI will be deployed in your business. It is already deployed. The real question is at what cost. Organizations that still treat governance as a secondary topic accumulate hidden costs: time spent on corrections, data incidents, decisions based on fragile foundations and gradual erosion of the quality of deliverables.
Those that build visibility, real-time control and local execution capacity – and that anchor the right reflexes in each business – reduce these costs while making adoption faster and more stable.
The next incidents will not come from an isolated technical fault, but from a series of daily actions repeated without framework. In your organization, is AI governance still a regulation to enforce, or a system of capabilities to build?
