Can AI make us better at integrity, or just faster at getting it wrong?

In short: The 10th Week of Integrity opened at Prosus’s AI House in Amsterdam with one question: can AI make us better at integrity, or just faster at getting it wrong? Christian Hauser’s keynote added “prediction” to the compliance cycle and warned about the industrialisation of ethics, a renewed national anti-corruption guide was launched, and a panel argued that no single function owns integrity. The fuller account is below.

The 10th Week of Integrity opened in the AI House in Amsterdam, the space Prosus built to show what AI can do. It would be hard to pick a better-suited location for the morning’s question: can the same technology help us behave better, or just help us get it wrong faster and at bigger scale? This year’s theme is leading with integrity in a digital age, and the three parts of the programme, a keynote, the launch of a national guide and a panel, kept circling that one tension.

The keynote: AI as augmentation, and its new risks

Professor Christian Hauser has researched business integrity for nearly two decades and is currently a visiting scholar at the OECD’s Anti-Corruption Division. He opened by adding a word to the compliance cycle everyone in the room knew by heart. Prevention, detection, response, and now, he argued, prediction. AI can help you see where misconduct is likely before it happens. Then he named the obvious danger in the same breath: prediction is exactly where predictive policing came unstuck, with racial profiling and all the trouble that follows when a model acts on something that has not happened yet.

He walked through where AI actually earns its place, using an OECD list of tasks: recognising things, spotting events, forecasting, personalising, supporting interactions, optimising towards a goal, and making sense of large knowledge structures. All of it maps onto day-to-day compliance work. Training can be personalised and linked to forecasting, so someone heading to a new market gets a reminder about the risks they are about to walk into. Chatbots can take the first line on a helpline. Processes can be streamlined, which matters for a function that still gets treated as a cost centre.

The examples were specific, and that is what made them land. Whistleblowing has been transformed by speech-to-text and translation: a voicemail left in one language can now be processed in another, and European companies, with so many languages to deal with, got good at this early. Voices can be disguised to protect a source. The catch, Hauser said plainly, is that the same trick powers fraud. A snippet of audio no longer than a few seconds is enough to clone someone’s voice, and companies have already wired away millions to a fake CEO. In anti-money-laundering, AI scans transactions in real time and takes some of the strain of the endless false positives thrown up by similar names. In internal investigations, where his own research on AI in ethics and compliance management began, it pulls together data scattered across devices, cleans up duplicate email threads, sets private material aside, and helps find the one document that matters. The AI-generated output presented the 1MDB case as a network, showing who was linked to whom, and automatically placed some of the activity on a map within the Swiss canton of Ticino. Most of us could not have done that by hand.

Then came the part that clearly interests him most: what he calls the industrialisation of ethics and compliance. For most of its life the function has been a craft, with experts handling a problem end to end. AI turns it into something more like an assembly line, each small step done by a person or an AI agent. He did not pretend that was free. Assembly lines breed their own silos, between the people who design the process and the people who run it, and they deskill, because anyone responsible for a single step eventually loses sight of the whole. We have a century of factory history telling us how this goes wrong, he said, so build the new processes with that in mind rather than bolting a tool onto the old ones.

On large language models he was blunt. They guess the next word. Very powerful, not reliable. He had asked one how to spend a day in Amsterdam and been sent confidently to the wrong metro line. More worrying for compliance, he raised intercoder and intracoder reliability: ask two models, or the same model twice, to tag the same text, and the answers can diverge so wildly that no human team would sign off on them. Change the prompt and the answer changes. Tell the model it is talking to an expert and it tries harder than if it is framing the response for a layperson. He pointed to the Epstein files, where AI redaction let victims’ names through, as proof the thing is not right every time. His advice was not to put the tools down but to read their output properly, and to remember that the accountability never moves off a human desk.

He left the room with a simple map. One axis is how much you can see into the algorithm, from fully guided systems where input, output and workings are all explainable, to black boxes even their makers cannot explain. The other is the data, from impersonal, to public personal, to deeply personal. Borrowing Helen Nissenbaum’s idea of contextual integrity, the point that what is fine in one setting is not fine in another, he put the highest risk where a black box meets sensitive personal data, and where the line between private and public gets blurry. So two questions, every time: what kind of model, and what kind of data.

The launch: prevention over cure

The second part of the morning was the launch of a renewed national guide, Verantwoord zakendoen zonder corruptie, doing business responsibly without corruption, with a companion practical guidance document. Annabel Wildschut of the Ministry of Foreign Affairs, who leads on responsible business conduct, presented it as a real joint effort, co-signed by the Ministries of Foreign Affairs, Justice and Security, and Economic Affairs, together with VNO-NCW, MKB-Nederland, ICC Netherlands, evofenedex and Transparency International Netherlands.

The whole thing leans on prevention, voorkomen is beter dan genezen, better to stop it than to cure it. It approaches anti-corruption through responsible business conduct and OECD-style due diligence, with real examples and practical tips for making supply chains more transparent. The brochure gives management the overview; the guidance walks practitioners through the due diligence steps. It is in Dutch for now, with a translation on the way for an international audience.

Michael van Woerden of DCM set it in context, reaching back to earlier editions more than a decade old, and made the point that prevention starts with tone at the top but does not finish there. Responsible business conduct is a useful common frame, he said, but it does not behave the same way across bribery, money laundering, cyber and sustainability. The devil is in the detail. His colleague Obbe Siderius, who introduced himself as a man of practice rather than policy, gave the most down-to-earth advice of the day. Due diligence has to fit your business, your markets and the people you deal with. Smaller firms that try to build one grand, all-covering programme usually fail, because it costs too much and takes too long. Start small. One line of business, one afternoon. And talk to other companies, because firms are oddly secretive about compliance when they could be learning from someone already working in the same country.

The panel: why no single function owns integrity

The panel, Breaking the silos, was moderated by Andrea Cardoso of ICC Netherlands, with Christian Hauser, Annabel Wildschut, Monika Tomczak-Górlikowska, Global Head of Privacy, Digital and AI Governance at Prosus, and Willemijn Gwanmesia of evofenedex.

Hauser started with a story worth more than any theory. He once sent the same research proposal to two people in the same Swiss company. One told him it was worthless. The other said it was exactly what they needed. Same company, opposite answers, because the two sat in different silos and came from different backgrounds. From the outside, that inconsistency is the whole problem, and it is why integrity has to be owned by the organisation, not parked in one team.

Willemijn Gwanmesia pushed the boundary further out. Integrity does not stop at the company gate; it runs the length of the supply chain, and every link counts. Annabel Wildschut separated two kinds of silo, by theme and by function, and described the Dutch habit of pulling business, government and civil society around the table, helped along by international bodies like the OECD and the UN.

Monika Tomczak-Górlikowska was blunt that the pace of technology has killed the silo approach. AI governance, data governance and compliance have to move as one now, and the hard part is turning fast, complicated risk into something an ordinary employee, contractor or supplier can grasp, without stamping on their enthusiasm to use the tools. She wants what she called integrated assurance, risk communicated in one aligned voice, and she declared the old ritual dead, the one where each expert gives the board a tidy five minutes and walks out happy. She drew the contrast between the prescriptive culture, which just blocks AI and keeps control at the price of pushing everything into the shadows, and the open culture, which frees people to innovate but forces a rethink of how you monitor and govern data. Whichever you pick, it is culture change, and culture is the hardest thing of all to move.

When the moderator turned to practice, the answers got concrete. Wildschut explained that the guide deliberately sits in the preventative, due-diligence space, because that is where government and business can actually talk, once you get past the fact that the topic touches the penal code and civil servants have a duty to report. Gwanmesia warned against the lazy reflex of pushing ever more information requirements down the chain, which just adds complexity and fixes nothing, and argued for real conversations with suppliers about the risks that actually matter. Hauser said it starts with shared language, because to a lawyer “integrity” means business ethics and to a data scientist it means system reliability, and that kind of cross-disciplinary talk has to be taught, not assumed. Tomczak-Górlikowska put accountability on the CEOs of the operating companies, not in a neat package posted from head office, which also leaves room for local culture in the many countries of the global South where the group works.

The questions from the floor went for the soft spots. Someone asked whether industrialising compliance contradicts breaking silos. Hauser agreed there is tension: assembly lines split planners from doers, and deskilling is real, so the fix is to design the whole process with proper dialogue between the people organising the work and the people doing it, not to scatter tools around. Another asked about the gap between a good AI policy and any actual value from it. In a lot of companies, he said, the AI policy never even reaches compliance, which turns up late, and that is dangerous. Tomczak-Górlikowska described the quiet turf war over who owns AI governance, and the telling fact that nobody wanted the word in their title. A benchmark policy is not enough, she said. Implementation at this speed and scale will be messy, leaning on automated monitoring and agents to flag risk because no one can review it by hand, and tuned to whatever each company cares about most, whether that is algorithmic transparency, fraud, or creditworthiness.

Hauser got the line of the morning. The gap between policy and practice, he suggested, is partly just human laziness. People flow like water, towards whatever is easiest. Behavioural ethics says to build your systems with the grain, so the easy path and the right path are the same one. Make integrity the harder option and people will find a way around it.

And then the hard part

Mirjam Bakker-Vergouw, closing, would not let anyone off with a neat summary. Leading with integrity in a digital age, she said, is damn hard work. It takes real knowledge of how AI works, not just the language models, so you actually understand the risks. It takes knowing each other well enough to close the gaps between functions. And in the end it stays a human job. The tools are new. The responsibility is the same as it ever was.

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