
It wasn’t a single moment, but rather a period of several months around 2012, whilst I was a professor in Aachen. Two events happened almost simultaneously: the Google Car drove autonomously through a small American town, and IBM Watson won Jeopardy. Until then, AI had been confined to the laboratory – Deep Blue played chess, with fixed rules, complete information and no surprises. And suddenly, a vehicle was driving through a perfectly ordinary small town, where a child could run out onto the road at any moment. That was the tipping point: artificial intelligence can navigate the complexity of everyday human life. If you were to summarise the First Industrial Revolution in a single word, it would be ‘steam engine’. If you were to summarise the Fourth in a single word, it would be ‘AI’. And to put it more fully: technical systems are becoming intelligent. This is not just another stage of digitalisation, but an entirely different category.
Many people are still stuck with the image that emerged in 2022. I distinguish between three phases. Phase one, 2022/23: small, specific questions. ‘Translate this text for me.’ ‘Paint me a picture in the style of van Gogh, but without sunflowers.’ Essentially, a better search engine. Phase two, from 2024: AI as a colleague – or at least as part of a colleague. ‘Here are the interview questions; draft some answers for me – you know me, after all.’ That’s exactly what I, as a board member, would have given to my assistant in the past. Phase three, today: AI as an operating system. The analogy for this comes from your world. My computer isn’t secure because I check whether the virus scanner is running, but because processes are constantly running in the background that monitor the system itself. They work using pattern recognition: if there are unusual database accesses at three in the morning, or if sending and receiving processes fall outside the usual parameters, anomaly detectors are triggered – and then the system intervenes before I even notice anything. AI is developing along the same lines: a system that runs alongside me, checks my inbox and says: ‘You’ve got a meeting at eleven; you’re not prepared – here’s a draft already.’ Anyone discussing productivity gains in writing today is still only talking about phase one.
Firstly: What is our overall AI strategy? Everywhere I look, I see ‘lighthouse’ projects – something is being trialled here, something else rolled out there. That’s all well and good, but it doesn’t replace an answer to the question of where we want to be in three to five years’ time. Instead, the discussion centres on whether ChatGPT is currently better than Claude or vice versa, and whether the Chinese models have now caught up. Such discussions fall far short of the mark. Secondly: How is the business model changing around the core product? Not in HR, not in support – at the core. If I’m building cars today: will I still be building cars in five years’ time? Or will I no longer be selling vehicles at all, but instead managing fleets and selling guaranteed mobility, because the value will then lie in scheduling, data and forecasting, and no longer in the vehicles itself? I have opportunities I didn’t have before, and competitors that didn’t exist before. Thirdly: How is expertise developed within the organisation? Not by senior management purchasing 42 training modules, but by employees themselves accepting that they need to develop further. You can’t just sit on the bus and wait for it to take you to expertise. You have to get your bike out, oil the chain and cycle up the hill yourself. And you need someone in charge – not necessarily with the title of Chief AI Officer, but in that role: at board level or immediately below it, with the mandate to bring together the scattered initiatives across the entire organisation. This is the same debate we had a few years ago about the Chief Digital Officer. How else did people think it would work – that things would converge of their own accord when nobody is bringing them together?
There are two reasons, and the second runs deeper. Firstly, early AI projects often emerge from a small, close-knit team – a ‘Gallic village’. They boost the organisation’s self-confidence. But they are islands lacking bridges and ferry connections. They rely on technologies that are obsolete two years later, cannot be integrated into the overall architecture, and eventually die out – even if they were technically sound. Secondly, and this is hardly ever discussed: the generation time for AI is six months. (Just for comparison: for a combustion-engine car, it’s six years.) When we say today that few AI applications make it into production, we’re looking at something that was conceived two years ago and evaluated a year ago – otherwise we wouldn’t be able to assess it at all. So we’re assessing an AI that was in nursery school, whilst today’s AI is sitting its A-levels. Incidentally, compliance and governance are not the problem with the transition; they are the showstopper that comes before it. You almost have to congratulate every pilot project that actually exists: either compliance was sidestepped – or it simply wasn’t even asked about.

It’s not a hard and fast line; it shifts in line with the systems’ capabilities. Ask an LLM the same substantive question today and three years ago: the rate of hallucinations has fallen dramatically. Three years ago, it was more like brainstorming – ‘let’s see what ChatGPT has to say about it’. Today, I delegate the task of writing a cover letter to a business partner entirely to AI. It varies depending on the subject: where to go on holiday in Thailand, and booking the hotel straight away – I can hand that over. Should I quit my job? – I wouldn’t delegate that. My father was a technician and held the position of engineer in nuclear power station construction. When the first calculators became affordable, he sat there with his slide rule and double-checked the calculations because he didn’t trust the machine. For two years, the calculator had to prove that it could do the maths. Then assistance turned into delegation. Today, nobody would commission an engineering firm that calculates the structural integrity of a house using pen and paper instead of sound numerical methods. Essentially, the question is framed incorrectly. But it is a good one, because it reveals the limits of one’s own thinking as soon as one accepts that one is dealing with a non-biological, yet intelligent, entity.
"Trust arises through experimentation, there is no other way."
Trust arises through experimentation; there is no other way. The history of technology offers many examples of this, the railway being one: in the 19th century, there was serious medical debate as to whether the human body could even withstand such speeds and vibrations. There was a full-blown debate about so-called ‘railway sickness’, and in the 1860s the surgeon John Eric Erichsen described ‘railway spine’ – a neurological condition which he attributed to travelling by rail. These concerns were not dispelled by expert reports, but simply by people actually travelling by rail. And if we had regulated the car a hundred years ago in the same way we regulate AI today, we would have had to know in advance that driving licences, road signs, seatbelts and airbags would be necessary. Back then, nobody knew what an airbag was. Clearly, this is how it works: you introduce a technology even though you do not fully trust it, gain confidence because it works when you use it yourself – and put up a traffic sign where it does not. You regulate retrospectively, in those areas where the risks actually arise.
Falsification and manipulation are nothing new, and historically, society’s response to them has almost always been a belated one. The printing press suddenly made it possible to reproduce religious texts, political pamphlets and Bible translations on a massive scale. Princes, reformers and churches used the new medium to spread their respective interpretations; the backlash took the form of printing privileges, censorship and bans. With the internet, the fear was similar: anyone could write nonsense on Wikipedia. In principle, that can indeed happen – only it is then nonsense that is potentially up to date and just as quickly correctable. In contrast, the last printed edition of Brockhaus could not yet include what deep learning means today. These encyclopaedias were virtually unrivalled in terms of care and editorial quality. However, they could only ever reflect the state of knowledge at the time of going to press. Every new media technology thus initially undermines part of the existing framework of trust. When provenance, authority and authenticity are no longer immediately recognisable, new procedures must emerge to establish credibility in different ways: through traceable sources, institutional accountability, editorial oversight or – in academia – through peer review. However, such mechanisms almost always emerge only after the new medium has already shaken the old certainties. I do not for a second consider the argument that we could regulate deepfakes in such a way that they no longer occur. What helps is the ability to exercise sound judgement. Never before has such a large proportion of the population had access, over so many years, to schooling, vocational qualifications, higher education and freely available knowledge. We are the best-educated society that has ever existed. And we allow sixteen-year-olds to vote. So we must also have confidence in people’s ability to distinguish the plausible from the implausible. Schools must provide the foundation; it is up to each individual to do the learning.
A regulatory framework makes sense. The AI Act in its current form does not. Its conceptual foundations were laid in April 2021. ChatGPT was launched at the end of 2022; by early 2023, it had become clear that generative AI was not just another application, but a technological paradigm shift. Nevertheless, in December 2023, a set of regulations was politically adopted whose basic architecture dated from the pre-ChatGPT era. Generative AI could now only be retrofitted into this architecture. This is precisely where the problem lies: a legislative process that takes almost three years cannot adequately address a technology with development cycles of just a few months. The right approach would have been to say: ‘Good idea, important groundwork – but the technological basis has changed, so let’s start afresh.’ Little would have been lost. The stakeholders were knowledgeable and well-versed, the network was in place, and the right questions had been asked. Given the rapid pace of development in AI, a risk I identify today may no longer be a risk tomorrow – and I may not even be aware of the most significant one yet. We should also ask ourselves why the world’s leading technological regions largely refrain from imposing a comprehensive, horizontal regulatory framework on the technology as a whole at the outset. The answer is: those who take this approach are not in the driver’s seat. And those who are not in the driver’s seat do not determine the direction. The correct approach would be the reverse: first regain technological leadership, then develop the regulatory framework step by step – as with the motor car, with a clear goal, ideally zero road deaths. But not to pretend beforehand that we know where the risks lie.
"All the more reason why we must move into the front seat, rather than continuing to ride along like a bobblehead dog on the rear shelf."
The fact that our data flows to the US – or to China; you can now choose – is solely because we did not open up to these technologies in time and have therefore failed to build our own, sovereign solutions that we could fall back on today. This is, of course, a retrospective view of spilt milk that cannot be put back in the bottle. All the more reason why we must move into the front seat, rather than continuing to ride along like a bobblehead dog on the rear shelf. And there’s a new opportunity, albeit a somewhat counterintuitive one: the open Chinese models. It reminds me a bit of the old ‘Linux versus Windows’ debate – and, of course, there’s a business model behind these open models too: anyone who chooses Qwen is expected to use Alibaba Cloud – and then finds themselves with a problem in China rather than in the US. But I don’t have to do that. Just as I used to be able to run my own Linux distribution without buying SUSE or Red Hat, I can actually host these models locally, decoupled from the provider’s infrastructure. That’s a form of insourcing: back to using our own staff, but with complete control. I’m currently discussing this with one of the major German universities for their administrative systems. This isn’t about real-time processes, such as those in a bank, but rather about human resources management, enrolments, examination procedures and accurate records. If the open-source model lags a few weeks or months behind the absolute cutting edge of technology for such applications, that’s completely irrelevant – in return, I have full autonomy. Incidentally, what DeepSeek demonstrated in January 2025 is no voodoo, but simply an outstanding feat of engineering for which one can only congratulate the Chinese – achieved with a fraction of the money, training time and development time spent on the major American models. And that is precisely the point: this achievement does not depend on massive financial investment, nor on the backing of a corporation such as Microsoft or Amazon. A well-resourced research team in Germany or Europe could have achieved the same.
I see three layers. The classic warfare level: intelligent drones, autonomous systems. Below that, hybrid warfare: disinformation, cyber operations, influence operations. And then there is the individual level, which has so far been given far too little attention. In cybersecurity, there is a term for this: ‘harvest now, decrypt later’. Data is being captured and stored today, even though it is encrypted and therefore currently worthless – in the expectation that powerful quantum computers will be able to decrypt it in ten or fifteen years’ time. This fundamentally alters the timeline of vulnerability. Think this through to its logical conclusion. Someone is today completely insignificant politically, economically and socially, and leaves behind traces that nobody is interested in. Two decades later, that same person is at the helm of a company, a government body or a political party – and suddenly, material that was considered secure at the time it was created becomes exploitable. In the best of the worst-case scenarios, the person concerned recognises the danger and steps down from their position. In the worst-case scenario, they remain in place – and become vulnerable to blackmail. Then it is no longer they who are the actual target, but the institution they represent. I consider this to be the most dangerous development in this field, precisely because it only becomes apparent after a long delay and therefore does not alarm anyone today. And that is exactly why no one should now entertain the idea of regulating quantum computing. The result would be that we would not understand the technology – whilst those who use it against us certainly would.
AI brings defence forward in time. In the oldest approach, an attack was only detected once the bomb had struck. The next step in development was to detect the trajectory of a projectile and intercept it whilst still in the air – the Iron Dome demonstrates this principle impressively. In future, the aim will be to intervene even earlier: for example, by detecting unusual procurement patterns for chemicals, components or other dual-use goods before they can be turned into a specific weapon or an attack plan. This requires technological leadership and data fusion – in short, Palantir’s business model. The perpetrator of the 2016 Berlin Christmas market attack was already known to the authorities; the problem was not so much a complete lack of information but rather the evaluation, integration and consistent use of that information. AI might have been able to piece together a complete picture from scattered contacts, movements and warning signs at an earlier stage – before Amri hijacked a lorry and used it as a weapon against the Christmas market. The price is greater surveillance of the individual. I am half-Swedish. In Sweden, you are given a personal identity number at birth which stays with you for your whole life – mobile phone contracts, nursery, the dentist; it is almost more important than your name. Ask Germans, on the other hand, whether Sweden is a country that closely monitors its citizens, and you’ll be met with head-shaking: “Sweden? That’s Bullerby, isn’t it?” That’s exactly how we think of that country here – Astrid Lindgren, red wooden houses, Midsummer. A friendly idealisation that has little to do with the actual data situation in the Swedish state. The perception is therefore completely different, even though the level of surveillance there is considerably greater than here. Every step in human development has led to greater transparency and, on the whole, has done more good than harm. Nevertheless, I realise that such an approach can go wrong. And to put it quite bluntly: people in Germany do not currently place unlimited trust in their state. If you’re building a European alternative to Palantir that keeps pace technologically – excellent, and don’t let that put you off. [Editor's note: secunet, in collaboration with its partner INNOSYSTEC, offers the secure and robust data integration and analysis platform SCOPE x SINA Cloud.] Nevertheless, a significant proportion of the German electorate will ask how much transparency they are willing to grant their own state. And at its core, this is not a technical question, but one of trust: how much trust do I place in an institution that collates data about me, and on what is this trust based? As long as that trust is lacking, even the best European solution will be of no help. This question will have to be renegotiated – and it must be strictly separated from the question of sovereignty. They are two entirely different matters.
We will continue to develop these systems to make them more human-like, because we can only design intelligence from our own, human perspective. It is simply not possible for us to develop a system specifically towards a ‘Martian human intelligence’, because we do not know what such an intelligence would be like. With humanoid robots, there is also a purely pragmatic aspect: I want it to tidy my kitchen, so I give it hands with five fingers and roughly human proportions – otherwise I’d have to screw different handles onto the cupboards. In future, we will work in hybrid teams comprising humans and technical intelligences. We have long been familiar with this, though until now only in a biological context: the drug detective and his dog form a hybrid team that transcends species boundaries, as does the blind person with their guide dog. What is new is that the truffle-hunting pig can now be robotic – in which case it needs no break at night and may even have a better sense of smell. The human-like nature of the intelligence is a key factor in its success: it allows me to communicate more quickly and reliably than with a system whose intelligence is inaccessible to me. When it comes to the human capabilities mentioned in your question, I would also be cautious about giving answers that hold true over time. Any such answer assumes that the difference between human and technical capabilities will remain. A few years ago, the thinking was: ‘Do something with social skills; caring for people by people is what we’re left with.’ Today, a great many people find it extremely comforting to confide in a counsellor or therapist who is an artificial intelligence – because they tell it things they wouldn’t say to a human being.

What is overestimated is the focus on the models themselves. You can have the most ingenious LLM – but if you don’t have high-performance computing (HPC) on which to run it, improving the algorithm won’t help at all. And if you have a problem requiring high real-time performance but no 5G, no 6G and no redundant layer via satellite communication, it’s of no use to you that the algorithm could theoretically have solved it. What is therefore underestimated is, firstly, computing power – HPC and, looking ahead, quantum computing – and, secondly, connectivity, including satellites and redundancies. And thirdly, and most significantly underestimated: virtual worlds and simulation techniques. Since the metaverse hype has died down, the topic has barely been in the public spotlight – even though intensive work on it continues in research and industry. This raises a very practical question: if I ask an AI to draw up a security plan for Berlin Airport based on double the usual passenger volume – how do I actually assess the answer? To do this, I need virtual spaces that I can explore together with the AI. Virtual worlds are the communication interface between technical and biological intelligence.
Of course. The starting point was a structured interview conducted with me and subsequently transcribed. On this basis, I developed the text with the support of several AI systems: researching, structuring, condensing, refining the language, checking for contradictions and editing it several times. The thoughts, arguments and assessments are my own. AI was neither a ghostwriter nor an oracle in this process, but an extraordinarily powerful working tool. I owe readers not the most laborious creative process, but the best possible text. To forgo the best available tools in doing so – I would consider that rather an insult to their time and their intelligence. The fact that politicians have for decades been delivering speeches prepared by speechwriters and assistants has always been accepted. As soon as artificial intelligence is involved, this suddenly becomes a debate about the purity of authorship. This is less a concern for authorship than a defensive reaction to the fourth narcissistic wound: now not even thought and language are exclusively human territory. What matters is not whether a text was created with a pen, a word processor, human assistance or artificial intelligence. What matters is who takes responsibility for it. Therefore, the following applies to this text (or strictly speaking to the German version of this text): I stand by it with my name – for every thought, every statement and every sentence.

Do you have any questions or comments about this article? Then contact us using the contact form on the right.
secuview is the online magazine of secunet, Germany's leading cybersecurity company. Whether cloud, IIoT, home office, eGovernment or autonomous driving - there can be no digitisation without security.
