AI for citizen-led prevention: useful, but under control

When a fraud campaign begins, the first signals are often tiny: three citizens forward the same text message, an association reports a fake invoice, a bank sees a domain appear that is almost identical to its own. Taken separately, these elements look like noise. Brought together quickly, they can reveal a campaign before it gains momentum.

It is in this work of joining the dots that artificial intelligence can genuinely help citizen-led prevention. Not by replacing the analyst or deciding on the citizen's behalf, but by shortening the time between the first report and a useful alert.

Start from a concrete problem

The word "AI" easily attracts vague projects. Prevention, on the contrary, requires starting from a measurable question: can we spot more quickly that several messages belong to the same fraud? Can we explain why a link is dangerous? Can we produce a consistent alert in the national languages without waiting several days?

In 2025, nearly ten million suspicious messages were forwarded to Safeonweb. That volume exceeds what a human team can review one by one. Belgian mechanisms such as BAPS use these reports to identify malicious links and redirect users to a warning page.

AI can improve the initial triage, but the value comes from the whole chain: simple reporting, fast analysis, validation, proportionate blocking and clear feedback to the public.

What the machine does well

A system can group messages whose wording differs but whose scenario is identical. It can detect that several domains imitate the same bank, extract the numbers or addresses used and rank reports by urgency.

It can also help to:

• recognise a campaign circulating simultaneously by SMS, email and messaging apps;
• spot variations across several languages;
• identify the most exposed sectors or regions;
• summarise a hundred reports for the analyst;
• propose a first draft of an alert tailored to the audience concerned.

These tasks are repetitive, high-volume and structured enough to benefit from automation. They free up human time for investigation, contact with victims and decisions that have consequences.

What the machine cannot decide on its own

A recently registered domain is not necessarily fraudulent. An urgent request can be legitimate. A message that is unusual in one municipality may be normal in another.

The system does not always know the social and institutional context. It may confuse an official campaign with an imitation or give too much weight to a popular signal. If the training data mainly contains French-language reports, the results may be weaker in German or Dutch.

The decision to block, publish an accusation or pass information to an authority must therefore remain supervised. The greater the impact on people, the more human validation, traceability and the possibility of appeal are needed.

Explaining the alert

An effective warning does more than display "dangerous website". It gives a reason the person can understand:

• the address imitates an official service but uses a different domain;
• the site asks for a bank code that should never be shared;
• several independent reports describe the same scenario;
• the domain was created recently and redirects to a page already linked to fraud.

This explanation serves two purposes. It helps the person decide right now and builds a reflex for the next attempt. It also allows a legitimate site to challenge a mistaken block.

A good prevention system therefore does not only aim to be right. It aims to show enough evidence for its decision to be checked.

A realistic day in a citizen-led system

At 8:15 a.m., several people forward a text message about a so-called energy refund. The wording varies slightly and the links use five domains.

The system groups the messages, notes the similarities and finds that the pages copy the same visual identity. At 8:30, an analyst reviews the whole set, checks with the organisation being imitated and confirms the fraud. The domains are added to the protection mechanisms. An alert written in French is adapted into Dutch and German, then proofread.

By 9 a.m., municipalities, banks and associations have a consistent message to pass on. New reports enrich the existing campaign instead of creating twelve separate files.

This scenario does not require "general AI". It requires information to flow well, models limited to a single task and a team capable of deciding quickly.

Attackers use the same tools

Fraudsters produce cleaner texts, adapt their scenarios to several languages and automatically collect public information. Synthetic voices make certain requests more convincing. Above all, AI allows them to increase volume and speed.

The answer cannot be to automate defence blindly. It must combine:

• strong, phishing-resistant authentication;
• independent procedures for verifying payments;
• rapid blocking of known infrastructure;
• simple advice on how to break off a suspicious interaction;
• support for victims without blame.

Technical prevention and the human ability to say "I'm stopping to check" reinforce each other.

Protecting the data of those who report

A forwarded message may contain a name, a case number, an address or a private conversation. Not all of these elements are needed to recognise a campaign.

The platform must mask or delete personal data as soon as possible, restrict access, log every consultation and set a retention period. Data used to train or test a model must be kept separate from operational files and be the subject of an explicit decision.

A vicious circle must also be avoided: collecting ever more messages because "it might improve the model". The question to ask is the opposite: what is the minimum amount needed to achieve the objective?

Indicators the public can understand

A system funded to protect citizens should publish something other than an abstract accuracy rate.

Useful indicators are:

• the time between the first report and the validated alert;
• the proportion of blocks that were challenged and then lifted;
• the quality of results in each national language;
• the number of visits diverted away from a fraudulent domain;
• the volume of personal data deleted automatically;
• the response time when someone asks for a correction.

These figures force a look at real usefulness, not just the model's technical performance.

Opening up the capability to people on the ground

Municipalities, schools, associations and small businesses receive valuable signals, but rarely have a structured channel to share them. Simple, documented formats would allow a campaign to be reported without passing on the entire contents of a file.

In the other direction, local players need ready-to-use alerts: a short description, the actions to avoid, the official contact details and an update date. An open interface does not mean making sensitive data public; it means making prevention information reusable.

Official sources

• Safeonweb — phishing continues to affect Belgian internet users (FR)
• CCB — Belgium's cyber reality in 2025 (FR)
• BOSA — Charter for the responsible use of AI in public services (NL)