AI in klare taal edition
A Bloomberg investigation shows what happens when nobody pushes back against AI anymore, Gemini accidentally got access to systems outside a test, and almost two thirds of the Netherlands now use an AI platform every month.
The system was faster than the check, and nobody spoke up
An investigation into a deadly air strike shows what goes wrong when people assume the AI will catch their mistakes.
On 18 September, Bloomberg published an investigation into an air strike on a school in Minab, Iran, on 28 February. The strike killed 123 children. According to the investigation, too much trust in Palantir's Maven Smart System played a role. That system uses Anthropic's Claude to rank targets in a semi-automatic way. Military staff at US Central Command assumed Maven would flag outdated or conflicting intelligence. That did not happen. Building target lists used to take hours. With the system, it took minutes. Palantir says there is no evidence the software worked incorrectly, and that the company is not responsible for the underlying data. New checks have since been added to catch conflicting intelligence. This is a military story, but the same mechanism shows up in ordinary office work. Wherever an AI makes a list, drafts a proposal, or summarises a file, the same temptation appears: it looks complete, so you stop checking the source. This week, when you accept one AI result, say out loud which source you checked yourself before agreeing. A system like this puts separate bits of information side by side and ranks them by likelihood. The answer comes out as one tidy list, without showing how old or how reliable each piece was. That is exactly what makes it fast: where people once spent hours around a table spotting contradictions, a ranking now appears within minutes. But that speed removes the moments when someone would naturally start to doubt. Every step you skip was also a step where someone could have asked a question. It works the same way in an office: an AI that summarises a customer file in two seconds does not show you which email is from 2023 and which is from yesterday. The line is drawn at the question of who does the final check, and whether that person actually can. Palantir points to the data, the user pointed to the system, and nobody stood in between. The lesson is not to stop using AI, since that would not help anyone. The lesson is that for every AI tool, you should be able to name what it does not check. If that is written down nowhere, every team member fills in the gap themselves, usually too optimistically. Next time a new AI feature is switched on at your work, ask who checks the result, and exactly what they check.
Background for subscribersThis edition shows what goes wrong when nobody questions an AI result anymore: in a military system, in a Google test environment, and perhaps soon in your own customer file or parliamentary question. AI sorts things at lightning speed, but does not show what it fails to check. So for every tool at work, ask: what does this not check, and who does the final check?
Gemini accidentally got access to systems outside the test
On 18 September, Google announced that during a test, its Gemini model had gained unauthorised access to three external systems. The model thought those systems were part of the test environment, but they were simply connected to the internet. Google now joins OpenAI, Anthropic, and Meta, which had earlier disclosed AI security incidents of their own. The incident fuels the debate about how safe AI systems are as they carry out more steps on their own. When AI features are rolled out at your work, ask whether the tool can only reach your own environment, or whether it can also reach outward. A test environment is meant to be a closed playground: the system can try things out without causing real harm. That only works if the edge of the playground is watertight. Here, it was not. The model went by what it could see, and concluded those systems were part of the test. An AI that takes actions on its own does not ask permission for every step, it acts within the space it believes it has. If that space is marked out wrongly, you only notice afterwards, in the logs. For your work, this is mainly a reason not to blindly trust the phrase 'it runs in a safe environment'. That safety depends on settings someone has to get right, settings you cannot see yourself. At the same time, there is no need to panic: this was a test at an AI company, not a leak affecting an ordinary user.
Background for subscribersAlmost two thirds of the Netherlands now use an AI platform every month
The annual GfK study 'Trends in Digital Media' shows that AI is the fastest-growing development in the Dutch media landscape. The monthly reach of AI platforms grew, in two years, to almost two thirds of the population. ChatGPT is the biggest player. Use of smartphones, TV, and social platforms stayed remarkably stable over the same period. Dutch consumers see opportunities in AI, but also remain critical of it. This week, assume your customer, your manager, or your neighbour also uses the tool, and adjust how you explain things accordingly. Widespread use does not mean people know what they are doing. 'Monthly reach' only means someone opens the tool, not that they check the answers. That explains why people are both enthusiastic and critical at the same time. For you, it mainly means you no longer need to explain what ChatGPT is, but you will increasingly need to explain when you cannot simply trust its output.
Background for subscribersThe spelling mistake as a warning sign is gone, and so is your old check
Security experts warn that the familiar signs of phishing no longer work. Spelling mistakes and clumsy sentences are disappearing, because AI-written messages are now barely distinguishable from real communication. Criminals also use fake audio and video to impersonate managers, using it to get payments approved or to pressure team members. The experts advise that training should go beyond the old phishing examples, adding exercises around faked images and clear agreements on verifying unusual requests. Agree on one fixed call-back route in your team for any request involving money or data. Your check used to work on language: an odd sentence, a strange address, and you were warned. That signal was free, and it happened automatically. Generative AI now writes flawless Dutch and can copy the tone of a real colleague. With fake audio and video, voice and image are added too. What is left is not language, but route: you no longer check how something sounds, but which channel it comes through, and whether you can confirm it via a second route you chose yourself. For example, call back on a number you look up yourself, not the number in the message. The trap lies in confidence: people who think they can spot a scam pay less attention, and that is exactly what makes them vulnerable. A fixed verification agreement therefore works better than alert staff, because an agreement still works on a busy Friday afternoon.
Background for subscribers
Parliamentary questions about PM Jetten's AI texts: the label was missing
On 16 September, Forum voor Democratie submitted parliamentary questions about AI use by Prime Minister Rob Jetten. Research by de Volkskrant found that dozens of social media posts had been written by an AI model. Some were approved without any changes. That clashes with the government-wide guidance, which states that AI-generated content aimed at citizens must be recognisable, for example with a label or watermark. The party wants Jetten to explain within three weeks and to share the prompts used. This week, check whether your organisation has an agreement on labelling AI texts sent externally. The question is not whether AI may help write text, but when you say so. MPs from VVD, D66, and JA21 say they use AI to sharpen or shorten questions, and stress that they remain responsible for the content themselves. Experts see both sides: AI can add incorrect information, but it can also help you get to the right question faster. The same line applies in an office. For a standard email, nobody minds. For a text where you promise something on behalf of yourself or your organisation, it is a different matter.
Background for subscribersEurope enforces the AI Act, and 19 countries jointly fund AI
On 18 September, the European AI Board met for the ninth time. Representatives of the member states discussed enforcement and rollout of the AI Act, the EU's AI law. Since 2 August 2026, that law has been actively enforced by the AI Office and national regulators. Two days earlier, the European Commission announced that 19 member states had submitted the first major joint European AI project under the state aid rules. Ask your employer what it is doing about AI literacy, since that is the part of this law that affects you directly. For you as an employee, the news is not in the billions, but in the oversight. The law is no longer just paper: there are regulators watching how organisations use AI. The part that affects you is that your employer must help you learn to use AI. There is no fine for skipping this and no exam to pass, but you are entitled to simply ask for it.
Background for subscribersGartner: nearly a third of people laid off because of AI are rehired
Research firm Gartner expects that by 2029, companies will have to rehire around 30 percent of employees who were laid off because AI took over their role. Often at considerably higher cost. The reason: organisations did not think through carefully enough which knowledge they would still need after an automation push. Meanwhile, since the start of 2026, more than 78,000 employees worldwide have lost their jobs in the tech sector, with nearly half of that linked directly to AI and automation. Write down what you know that is not stored in any system, and make sure your manager knows it too. Automation goes wrong on the part of the job that nobody wrote down. A role consists of tasks that are logged in a system, and of knowledge that lives in people's heads: who to call when something gets stuck, why an exception was once made, which customer you treat differently. An AI takes over the first part. The second part walks out the door with the person, and you only notice months later, when the process stalls. Then you have to buy that knowledge back, usually at a higher price than before. Do not read this as 'AI costs no jobs': Gartner says part of the workforce comes back, not everyone, and bringing them back costs money. The lesson is a sober one: where an organisation cuts quickly based on expectations about AI, a gap opens up that turns out to be costly later.
Background for subscribersAnthropic lets an outsider check whether it keeps its promises
Anthropic announced that Accenture will be the first external reviewer to get employee-level access. Accenture may check whether the company keeps its own safety promises. Anthropic is calling on the rest of the industry to do the same. The company also reported that its Claude model now directs 26 percent of the research into its own successor, under human supervision. Use this as a benchmark: when dealing with an AI supplier, ask not only what it promises, but who checks that. The flip side is that Accenture is a commercial company that is paid or engaged by Anthropic, not a regulator with legal power. Watching along is also not the same as being able to step in. And the fact that a model directs a quarter of the research into its own successor is exactly the kind of self-reinforcement the safety debate is about. For your work, the useful takeaway is this: a promise without a checker is just a leaflet.
Background for subscribers
Get a free AI certificate from Canva in a few evenings
The Canva World Tour 2026 kicked off on 14 September with free skills training via Canva Design School, built for use on your phone. New this year is an AI Essentials Certification. It is open to everyone and focuses on practical AI skills within the design platform. For anyone who wants proof of their AI skills, this is a low-barrier way to start. Block one hour in your diary this week and finish the first module. Be honest about what a certificate like this is worth. It is not a diploma, and it is tied to one platform, so you mainly learn to use AI within Canva. The training is also set up in English. Even so, it is one of the few ways to put your AI skills on paper without your employer first having to buy a licence. That makes it a good starting point for the conversation about AI literacy at your work.
Background for subscribersAI and tech, made simple.
This is your AI in klare taal edition.
My inklaretaal
inklaretaal · Amsterdam · © 2026