You have seen something wrong inside an AI company. Now what?

In short: If you work inside an AI company and see something that scares you, where do you go, and what happens to you when you get there? Geert Vermeulen, Jennifer Gibson and Simon Gerdemann mapped the reporting routes, the gap between “protection” and a mere remedy after the fact, and the twist that AI now sits on both sides of the report. The fuller account is below.

Most people only think about whistleblowing when they are already in deep trouble, or about to cause some. That was one of the first honest observations in this Week of Integrity session, hosted by Geert Vermeulen, founder and CEO of The Integrity Coordinator, who has set up whistleblowing procedures since 2007 and handled hundreds of reports. His two guests had spent a lot of time in that uncomfortable space: Jennifer Gibson, co-founder and executive director of Psst.org, which helps tech and AI insiders work out what to do when they see something wrong, and Simon Gerdemann, a professor of law at the BSP Business and Law School in Berlin who specialises in whistleblowing and AI law. The question they circled for an hour was deceptively simple. If you work inside an AI company and you see something that scares you, where do you go, and what happens to you when you get there?

Where do you even report an AI concern?

The EU AI Act sets up the first part of the answer, at least on paper. Some obligations are already live, such as labelling AI-generated or modified images and video with a watermark. High-risk systems, the kind used in courts, medical software or heavy machinery, come with heavier rules that phase in over the next year. And there are specific rules for general-purpose and frontier models, the very large systems with outsized impact.

Where any of that is breached, Simon explained, you can report to the EU AI Office in Brussels, which has opened a channel for confidential and anonymous reports. National authorities handle most cases; the AI Office focuses on general-purpose and frontier models, the big engines behind everything else. Usefully, the AI Office treats itself as a one-stop shop: report anything you believe might be a violation and, even if they are not the right body, they will tell you which authority in which country is responsible and, with your consent, forward it.

The honest caveat came quickly. Nobody is fully ready. Not the Netherlands, where the Data Protection Authority is expected to take the national role and already runs a hotline for GDPR breaches, and where the House for Whistleblowers can advise people. Not Germany, where the Federal Office of Justice in Berlin takes reports across many subjects and routes them on. The authorities will take AI reports and process them diligently, Simon said, but he would not put his hand in the fire and promise they yet grasp every technical intricacy of how these models are built and run. We are getting there. It will take time.

A patchwork that stops at the border

Then Jennifer complicated the picture, which is her daily reality. The UK sits outside the EU and does not follow the AI Act. Its Public Interest Disclosure Act lists regulators you can take a concern to, but AI does not really sit in any of their portfolios. So a worker at a large frontier company who wants to raise an alarm often cannot tell where to go, or what they are even allowed to say. The deeper problem is that whistleblowing law on both sides of the Atlantic is built around what is illegal or against an established standard, and in AI there is often neither. What people are raising are safety concerns, with no safety benchmark to measure them against. That gap breeds confusion.

A small but important point of language: Jennifer deliberately says frontier companies, not frontier labs, because there is a live debate about whether these are research labs or simply companies. Either way she meant the leading few, Anthropic, OpenAI, Google DeepMind, Meta and xAI, with some in Europe adding Mistral, and nobody forgetting the Chinese players.

The cross-border trap is the part worth sitting with. Imagine a US worker at a frontier company who reports to the EU AI Office. The office can offer confidentiality, which matters enormously. What it cannot offer is protection from retaliation, because those rules protect European workers, and most of the people inside the big five are not in Europe. So the very people with the deepest inside knowledge are often the least protected when they try to speak across a border.

Three doors, and the one marked “press”

Under the EU Whistleblower Directive there are three routes, Jennifer explained. Internal first. Then regulators. And then, in certain cases, where internal channels fail or regulators are too slow, and there is a genuine, urgent public-safety issue, you can go to the media. That third door is not theoretical in AI. With the heads of some of these companies themselves saying the technology could pose catastrophic risks, insiders have gone straight to journalists, sometimes alongside regulators and sometimes bypassing them entirely, to force a debate the public cannot otherwise have, given how little is disclosed about what happens inside the companies day to day.

The United States is its own maze. There is no federal AI regulation, so individual states are stepping into the void, with California, Illinois and New York active and their attorneys general keen to hear from people. The preferred word there is not whistleblower, which is loaded, but insider: a worker with a concern. And across these systems, part of what an adviser actually does is triage urgency, which of the growing number of pathways is most likely to be protected and to have the biggest impact.

Going to a member of parliament or congress is another option, and a powerful one, because what the politician then says in the chamber is itself protected speech that cannot be pinned back on the person who told them. Whether the act of disclosing to a legislator is fully protected is, as Simon put it, a jury still out, or not yet convened. The spirit of the directive suggests it should be, the case law of the European Court of Human Rights is sympathetic, but you might have to go all the way to the European Court of Justice to be sure.

Protection, or just a remedy?

Jennifer made the sharpest point of the hour, and it is the one compliance people should repeat to themselves. We talk about protections, but mostly what the law gives you is a remedy. Nothing actually stops a company from firing you. The law just gives you something to claim after they do. On that logic, one of the best real protections is the plain confidentiality the AI Office offers, keeping your identity out of it so the retaliation never starts.

Where the EU Whistleblower Directive is notably strong is the reversal of the burden of proof. Once you are a whistleblower, harm done to you is assumed to be because you blew the whistle, unless the employer proves otherwise, which matters because a smart employer will always have a tidy reason for the transfer to the Siberian office. Protection under the Human Rights Convention, by contrast, is thinner, often amounting to a court saying after the fact that an action was wrong.

The Dutch example made it concrete. A current, well-known case has a government whistleblower being dismissed by the very ministry that is also responsible for whistleblower legislation. The directive says member states should provide interim relief, the power to tell an employer they simply may not fire someone while a case runs, but that part has not been fully written into Dutch law, so the Dutch authority cannot yet do it, much as it would like to.

AI on both sides of the report

The twist in this session was that AI now sits on both sides of the table.

On the reporting side, people increasingly use AI to write their reports, which tends to make them longer and more tangled than they need to be. The risks are serious and easy to miss. Conversations with a chatbot are not covered by legal privilege, so they can be discoverable if a case ends up in court. They are not necessarily private, and public chatbots may train on what you type. And if you are a worker inside a frontier company using your own employer’s model to work through your case, there is a real chance you are quietly telling your employer exactly what you are doing. On top of which, a lot of what the model hands back is simply wrong, which eats time.

On the handling side, the advice was use it, carefully, and never as the face of the process. AI earns its keep sifting large volumes, the three thousand emails a junior associate used to read line by line, and synthesising, spotting patterns across many cases. But both experts drew a hard line: not in a public frontier model. For something this sensitive, Simon said, you want a sealed environment, an agent running offline on a machine with no internet, or not at all. And Jennifer was firm that AI should never be the interface with the whistleblower. These are people in a heightened state of stress who have decided to take an extreme step, and nothing upsets them faster than feeling they are talking to a machine that is not hearing them. Keep a human on the line. One nice detail from practice: an investigator takes a particularly complex case, has a safe in-house AI turn it into a podcast, and listens in the car to finally get the cast of characters straight.

Why people stay quiet, and whether to pay them

When Geert’s courses ask people what holds them back, two answers lead every time: fear of retaliation, and the belief that nothing will come of it, so why take the risk. In AI, Jennifer added a third calculation, one specific to this field: can I do more to prevent the harm from the inside than from the outside? That question keeps a lot of people at their desks.

Which raised the uncomfortable topic of rewards. The United States pays whistleblowers a share of the fine, and it has surfaced a great deal of fraud. The UK long held that rewards made no sense, since most people only wanted protection, though that position is softening. Europe is culturally wary of paying people to report others, for reasons rooted in its history, and treats what little it offers as compensation rather than reward. Both guests landed in a similar place for AI: these are very highly paid people who stand to lose enormous packages if they are pushed out, so some compensation to make the balance sheet look even is not unreasonable. It is just not near the top of any European legislator’s agenda yet.

The session ran out of time with a question still arriving in the chat, which felt about right for the subject. The law is unsettled, the technology is moving faster than the rules, and the people with the clearest view of the risk are often the least protected. What came through most clearly was not a tidy answer but a working principle: if you see something inside one of these companies, get good advice before you act, protect your confidentiality above all, and keep a human in the loop on both sides of the report.

The Week of Integrity continues all week, with partner events across the country and online, the seminar on 29 October, the closing event, and a spotlight on International Anti-Corruption Day on 9 December.