Your supplier sells you an "artificial intelligence solution". An employee receives a call from their director asking for an urgent transfer. A candidate sends a perfect cover letter. In all three cases, the same question: is this really AI, and can I trust it?
Belgian companies are caught between two opposite mistakes. Some buy AI that isn't AI and pay the innovation premium for a simple automatic rule. Others believe too quickly in what an AI produces, or in what is presented to them as authentic, and fall for a fake.
This page lays out the basics for telling the difference: what AI is (and isn't), how to spot a false promise, how to recognize generated content, how to protect yourself from fraud, and how to set the rules for AI use in your company. It extends Ligne 1's SEO & GEO training, which also teaches you how to be cited correctly by AIs.
questions are enough to expose most fake AI
since 2 February, the European AI Act requires companies to train their staff on AI (Article 4)
What is AI, and what isn't?
The term artificial intelligence covers very different techniques, but one simple dividing line helps you find your way: who wrote the rule?
What isn't AI: rules written by a human
An automation carries out instructions that someone wrote in advance: "if the amount exceeds €1,000, request approval", "if the word 'invoice' appears, file it in this folder", "send an email 3 days after sign-up". It is useful, sometimes very useful. But the system only does what it was told. Same input data, same output, every time. It never gets better on its own.
What is AI: a system that learns patterns from examples
In AI in the strict sense (known as machine learning), nobody writes the rule. The system is shown thousands or millions of examples, and it deduces patterns by itself: what a cat looks like in a photo, how a sentence continues, which customer is likely to leave. It can then handle cases it has never seen. In return, its result is a probability, not a certainty: it can be wrong, and doesn't always know it is wrong.
The three-question test. ① Who wrote the decision rule: a human, or the system from examples? ② Does the system sometimes give a different result for the same situation, or can it make unexpected mistakes? ③ Can it handle a case that nobody anticipated? If the answers are "a human / no / no", it is automation, not AI.
And what about generative AI?
ChatGPT, Gemini, Claude, Mistral, Copilot or Perplexity are generative AIs: language models trained on enormous quantities of text, which produce text (or images, sound, video) by predicting what is most plausible. They are impressive, but they don't "know" things the way an expert does: they produce a plausible answer. That is the source of their usefulness, and of their errors.
AI or not AI? Ten everyday examples in business
Most confusion comes from products that mix the two. Here is how to classify them.
- An Excel formula or a macro — Not AI. A rule written by a human, with an identical result every time.
- A menu-based chatbot ("press 1 for…") — Not AI. A decision tree planned in advance. It doesn't understand your sentence; it recognizes buttons or keywords.
- A chatbot that understands free-form sentences and rephrases — AI, generally a language model plugged in behind it. Keep an eye on it: it can invent an answer.
- An automation robot (such as Zapier or RPA) that copies data from one tool to another — Not AI in itself. It becomes AI if a step adds a model that reads or classifies text.
- A spam filter based on lists of banned words — Not AI. Modern filters that learn from emails flagged by users are.
- License plate or facial recognition — AI (computer vision). With an error rate you need to know.
- "You may also like" recommendations — Often AI (learning from behavior), sometimes simple rules ("best sellers"). You have to ask.
- Automatic transcription of a meeting — AI (speech recognition). Proofreading is still essential for proper names and figures.
- A prospect or candidate "score" — May be AI, may be a hand-weighted points grid. The difference matters for fairness and for the law.
- A tool that "predicts" your sales — Sometimes a real model, sometimes a moving average dressed up as a forecast. Ask what data it learns from.
This is not a value judgment: good automation beats bad AI. The problem isn't the absence of AI; it is paying for AI that doesn't exist, or trusting it where it doesn't deserve it.
AI washing: spotting a false AI promise
AI washing means presenting a product that doesn't use it as "artificial intelligence", or exaggerating the real role of AI to look more modern. It is the digital counterpart of greenwashing. The phenomenon is widespread enough that regulators have taken it up: in the United States, the financial markets regulator (SEC) penalized companies in 2024 that claimed to be powered by AI without really being so.
In Belgium, a false or misleading statement about the characteristics of a product or service can fall under the misleading commercial practices of the Code of Economic Law. A well-informed buyer is therefore wise to check before signing, and a seller to avoid promising what they don't deliver.
Warning signs with a supplier
- "AI" everywhere, how it works nowhere — Promises ("powered by AI", "cutting-edge artificial intelligence") with no explanation of what the system actually does.
- No error figure — Real AI has a measured error rate. A supplier who claims it "never makes mistakes" is not telling the truth.
- Unable to say where the training data comes from — Or which model runs behind the scenes (in-house, or a third-party service).
- Humans behind the curtain — Some services presented as automatic are in fact handled manually by operators, until they can be automated.
- A demo that is too perfect — A demo prepared on hand-picked examples proves nothing. Ask for a trial on your own data.
- A price that changes with the word "AI" — The same software, with an "AI" surcharge nobody can describe the content of.
Questions to ask before buying
Our article details the full checklist, with contract clauses: AI washing: eight questions to expose fake AI.
- What exactly does the AI do in this product? — Which specific task, and what does the product do without it?
- Which model, and from whom? — In-house model or a service from a major provider, under what contract?
- What data was it trained on, and what does it do with mine? — Is it reused to train the model? Where is it hosted (European Union or not)?
- What is the measured error rate, and how was it measured? — On which test set, with which difficult cases?
- Who checks, and what happens when the AI is wrong? — Can a human correct it and take back control?
- Can I try it on my own cases? — With real documents, before committing.
Recognizing AI-generated content: text, images, voice, video
Let's be clear from the start: no method gives total certainty, and the "AI detectors" sold online regularly get it wrong, in both directions. They wrongly accuse texts written by humans and let generated texts through. It is better to learn to spot clues and, above all, to verify.
Clues in a text
- A smooth, interchangeable style — Correct, polite, without rough edges, and could have been written for any company.
- Repeating structures — Three ideas per paragraph, parallel lists, identical transition phrases.
- No lived, dated, verifiable example — Generalities instead of a specific case, with its place and date.
- Invented or untraceable sources — The most reliable sign: a reference that looks serious but doesn't exist when you search for it.
- Round figures with no origin — A statistic quoted with no link, no year, no organization.
Clues in an image
- The difficult details — Hands, teeth, jewelry, reflections, text on signs or clothing, backgrounds that warp.
- An overly smooth rendering — Skin without pores, lighting that is too perfect, suspicious symmetry.
- No origin — An image with no author, no date, no source. Reverse image search (Google Lens, TinEye) lets you find the original or establish that it doesn't exist.
Voice and video
Cloned voices and doctored videos are now good enough to fool someone close to you. The clues (slightly out-of-sync lips, odd blinking, no breathing, flat tone) fade as the technology improves. That is why the defense does not rely on "seeing the difference", but on verification procedures, described below.
One avenue for the future: content provenance certificates (the C2PA standard, "Content Credentials"), which attach the history of its creation to an image or video. And the AI Act provides for transparency obligations to flag content generated or manipulated by AI. These tools help, but they do not replace verification.
Deepfakes and fraud: a simple protocol for your teams
"CEO fraud" existed before AI: someone poses as a senior executive and asks for an urgent transfer. AI makes it more credible, because the voice, and even the face on a video call, can be imitated. In 2024, a multinational company in Hong Kong was caught out by a completely faked video conference, with fake executives on screen, and paid out several tens of millions of Hong Kong dollars, a case widely reported in the press.
A Belgian SME doesn't need to be the target of a sophisticated operation: a fake voice message from a manager, a fake supplier changing their bank details, a fake candidate in a video interview are enough. The defense is above all a matter of organization, not technology.
Five rules to put in place now
The full protocol, with the reflexes to adopt after an attempt, is detailed in our article deepfake and CEO fraud: the protocol for SMEs.
- Double validation through a second channel — Any payment request, change of bank details or sensitive data request is confirmed by a different means than the one used for the request.
- Call back on a known number — Call the person back on the number already on file, never on the one given in the message.
- An agreed word — A spoken password, shared internally, for urgent requests outside the normal procedure.
- Urgency is a warning sign — The more you are pressured, the more you verify. A request that forbids you from talking to anyone else is suspicious by nature.
- A right to say no, without consequences — Every employee must know they will never be blamed for checking, even when it's a senior executive.
Fake profiles and fake content around your brand
The same techniques are used to create fake reviews, fake executive profiles and fake articles. For a company, this is also an e-reputation issue: knowing how to spot a fake review, report it and protect yourself is part of the job.
When AI is confidently wrong: understanding hallucinations
A generative AI has no database of facts to consult: it predicts the most plausible continuation. When it doesn't know, it doesn't always say so: it produces an answer that looks right. This is what is called a hallucination. It can invent a quotation, a figure, a law, a company name, in the same confident tone it uses for the truth.
The best-known example: in 2023, in the United States, lawyers filed a brief with a court citing court decisions… that ChatGPT had made up. Sanctions followed. Nothing looked false: everything was well written.
How to protect yourself day to day
- AI proposes, a human decides — Any text, figure or advice intended for a client, a contract or a decision is proofread and checked by someone who knows the subject.
- Demand sources, then open them — Ask for the link, click it, and check that the page exists and really says what is claimed.
- Use web-connected AIs for recent facts — They cite their sources, which makes verification possible.
- Beware of unearned precision — A very exact figure with no origin is more suspect than a sourced range.
- Keep AI for what it does well — Rephrasing, summarizing a document you give it, suggesting outlines, sorting, drafting. Not serving as an authority.
AI in your company: set the rules before usage sets them for you
Whether or not you have decided it, your employees are already using AI tools. This is called "shadow AI": AI used outside any framework. In 2023, engineers at a major manufacturer pasted confidential code into a public assistant, which led their company to restrict its use. The risk isn't using it, but using it without rules.
What an AI usage policy should say
A ready-to-adapt template: AI usage policy for SMEs, ten rules.
- Authorized tools — A clear list, including the paid versions that protect data where they exist.
- Prohibited data — Customers' personal data, financial information, trade secrets, documents under confidentiality: never in a public tool.
- Mandatory proofreading — Who approves before content produced with AI is sent or published.
- Transparency — In which cases you tell a client or the public that content was produced with the help of AI.
- GDPR — No processing of personal data by an AI tool without a legal basis, information to the people concerned, and a contract with the supplier.
What the law says: the European AI Act
For details by risk level and the action plan: AI Act: what Belgian SMEs need to know and do.
The European Regulation on Artificial Intelligence (AI Act) applies in stages. One obligation directly concerns every company that uses AI: Article 4, in force since 2 February 2025, requires measures to ensure that staff have a sufficient level of AI literacy. Transparency obligations on generated or manipulated content are coming in turn, and the timetable for certain other rules is still being adjusted at European level. Training your teams is therefore both a precaution and a regulatory requirement.
This text gives a general framework and does not replace legal advice. For a specific case (automated recruitment, customer scoring, processing of health data), have your situation reviewed by a specialist lawyer.
What AIs say about your company (and how to correct it)
There is another side, which most managers discover too late: AIs are already talking about you. A potential customer asks ChatGPT, Perplexity or Gemini to recommend a provider. The answer relies on what the AI finds on the web: your website, your reviews, the press, directories, social networks. If your information is missing, outdated or contradictory, it fills the gap, sometimes far from the truth.
This is what we demonstrated in an experiment carried out in public: what AIs repeat, they hold to be true, and what they find written most clearly wins out over what is true. Hence a simple rule: it is better for you to write your own story.
Run the test in ten minutes
- Ask three different AIs: "What do you know about [your company]?" and "Who do you recommend for [your activity] in [your city]?"
- Note what is right, what is wrong, what is missing.
- Identify the sources they cite: those are the ones to look after first.
Correcting what they say comes down to a clear, up-to-date website, authentic reviews, reliable mentions and structured data: this is the whole purpose of GEO (Generative Engine Optimization), which we detail in our SEO, GEO and AI FAQ and our article GEO: getting cited by AIs. Your image online is also a matter of e-reputation, now including in AI answers.
The "Recognizing AI" training for managers and teams
The skills that will matter in the future are not those of programming. They are those of discernment: knowing how to recognize what is AI and what isn't, knowing when to trust it and when to check. That is what Ligne 1's training teaches, tailored to your sector and your tools.
Programme
- Real and fake AI — The three-question test, applied to the tools you use or are offered.
- Assessing a supplier — The question checklist, the warning signs, reading a commercial offer.
- Spotting generated content and fakes — Text, images, voice, video, fake reviews, fake profiles.
- Protecting yourself from fraud — The verification protocol, simulations with your own teams.
- Using AI without getting it wrong — Good practices, proofreading, data that must never be shared.
- Setting a framework — Writing your usage policy and meeting the AI literacy obligation of Article 4 of the AI Act.
- Being well described by AIs — Testing what they say about you and improving your presence (link with the SEO & GEO training).
The training is delivered at your premises in Belgium and Luxembourg, or remotely, from an awareness session for a management team to a full programme for a team. It is led by Nathalie Teston, founder of Ligne 1, who works every day on online visibility and on how AIs select their sources.
Your questions about AI in business
What is the difference between AI and automation?
An automation follows rules written in advance by a human: same data, same result. An AI learns patterns from examples and gives probabilistic results: it can handle new cases, but can be wrong. Many products sold as AI are in fact nothing more than automation.
What is AI washing?
It means presenting a product that doesn't use it as artificial intelligence, or exaggerating the role of AI, in order to look more innovative. The digital equivalent of greenwashing. In Belgium, a false claim can amount to a misleading commercial practice. See the questions to ask a supplier →
How can you tell whether a text was written by an AI?
There is no 100% reliable test, and automatic detectors get it wrong in both directions. Useful clues: smooth, generic style, repetitive structures, lack of specific examples and above all untraceable sources. The real method is to check the claims and ask where the document comes from.
How can a company protect itself from deepfakes?
Never approve a transfer or sensitive information on the strength of a voice or a video alone. Require a double validation through a second channel, call back on a known number, agree on an internal word for urgent requests, and train the teams. See the protocol →
Does the AI Act require you to train your teams on AI?
Yes. Article 4 of the AI Act has required, since 2 February 2025, companies that use AI systems to take measures so that their staff have a sufficient level of AI literacy. It is not a diploma to obtain, but a training effort you must be able to demonstrate.
Why does an AI give false information?
A language model does not consult a database of facts: it predicts the most plausible continuation. When it doesn't know, it still produces an answer that looks right. This is hallucination. Any important information must therefore be checked at the source.
Can my employees use ChatGPT at work?
Yes, provided you have set rules: authorized tools, prohibited data, mandatory proofreading. Without a framework, employees use it on their own and may enter confidential information or personal data, which puts trade secrets and GDPR compliance at risk.
How do AIs talk about my company?
They rely on what they find on the web: website, reviews, press, directories. If this information is missing, contradictory or outdated, the answer will be approximate, or even wrong. Run the test in ten minutes →
Is AI training only for large companies?
No. The risks described here (fake AI purchased, fake transfer, data leak) mostly affect SMEs, which have fewer procedures. The training adapts to the size of the team, from a self-employed professional to an entire department.
Where can I take training to recognize AI in Belgium?
Ligne 1 delivers it at your premises in Belgium and Luxembourg, or remotely, with a tailor-made programme. Request a programme →
Teach your teams to tell the real from the fake.
Tell us about your company: we will build tailor-made training on AI, its promises and its pitfalls.