1956
At the Dartmouth Conference the idea gets a name: artificial intelligence. At the time it is mainly a research promise, not yet a tool for everyday office work.
What is artificial intelligence?
What is artificial intelligence?
A clear explanation – without tech jargon and without trying to turn maths into magic.
Artificial intelligence is software that isn’t given every rule, but learns patterns from many examples and applies them to new situations. That is how it understands language, recognises images and creates texts. It doesn’t think like a human, but it can take on tasks across several steps.
MIT / Theory of intelligence
Marvin Minsky
„Marvin Minsky at OLPC“ by Sethwoodworth at English Wikipedia, taken by Bcjordan, 2008
Source: Wikimedia Commons · Licence: CC BY 3.0 · Image credits
Editing: Image edited (AI-assisted) from a photo by Bcjordan, CC BY 3.0. Changes: cut out, upscaled, hand position newly generated (not present in the original photo). Edited by Today is Life.
Marvin Minsky wanted to understand what we actually mean when we talk about “intelligence”. His answer was remarkably modern: intelligence is not a single ability but arises from many different processes. That is exactly what helps in understanding today’s AI. A system can write texts, recognise images or solve problems and still be completely helpless in other areas. AI doesn’t have to think like a human to be astonishingly useful.
What is artificial intelligence?
The introduction “tasks that normally require human intelligence” is not wrong, but it has become a bit textbook-like. It explains about as much as saying a car is an object you can use to get around. In the past you could make things comfortable for yourself that way: artificial intelligence is software that understands language, recognises images, makes decisions or solves problems. That is not entirely wrong. It just doesn’t explain very much any more.
The OECD now defines an AI system somewhat more precisely, as a machine-based system that infers from its inputs how to generate outputs such as predictions, content, recommendations or decisions. These systems can work with varying degrees of autonomy and may adapt after deployment. Sounds correct. It also doesn’t sound like something you would read voluntarily over breakfast. So let’s try it another way.
Normal software does what you tell it
For decades we have programmed computers on basically the same principle. If A happens, do B. If the invoice amount is more than €1,000, get an approval. If the user clicks “Save”, save. If the password is wrong, don’t let them in. Of course such programs can become gigantically complicated. But somewhere a person described what should happen.
With many of today’s AI systems it works differently. We show the machine a great many examples and let it learn structures and patterns from them. In machine learning, these patterns are therefore not fully specified as rules by programmers. That is exactly one of the essential differences from classic programming.
Take image recognition for pests. We could try to explain to a computer what a bed bug looks like. Six legs. A certain body shape. Colour. Size. Antennae. That gets funny at the latest when the photo is blurry, the creature is half hidden under a mattress seam and obviously hasn’t read the manual. Instead, you can train a model with a great many images. In doing so it learns statistical features and relationships that let it classify a new image. So it doesn’t “know” what a bed bug is the way a pest controller does. But it can get damn good at recognising one.
And then came generative AI. For a long time, AI systems were mostly specialists. One AI recognised faces. Another forecast sales volumes. A third decided which advertising you are shown. That was AI long ago, even if nobody was yet announcing in the office in the morning that we now urgently had to become “AI-ready”. With large language models and so-called foundation models, something decisive changed. Such models are trained on very large and diverse amounts of data and can then be used for very different tasks. Generative AI creates new content: texts, images, program code, audio or video.
That is why ChatGPT can do more than correct spelling mistakes. A modern language model can summarise a contract, explain program code, draft a letter, structure data, analyse a job advertisement or discuss why your new product idea may not be quite as brilliant as it seemed yesterday evening after the second glass of wine. There is, however, no little electronic being sitting in the computer thinking about your question. A language model works with probabilities and learned relationships. Very simply put, it calculates again and again which continuation is likely or fitting in the given context. It just does so on the basis of structures learned during training, on a scale at which this seemingly banal principle gives rise to abilities that look anything but banal.
And this is exactly where a misunderstanding arises. Because the result looks intelligent, we quickly assume the system has the same kind of understanding that we humans have. We had better not. An AI model can give a brilliant answer and, three sentences later, convincingly talk nonsense. It can recognise connections without experiencing or understanding the world the way we do. That is why control, context and responsibility remain important, especially where a pretty answer turns into a real decision.
Why AI is different today
1997
Deep Blue beats Garry Kasparov at chess. This shows that machines can outperform humans in clearly defined rule worlds. It still has only partly to do with today’s AI.
2011
IBM Watson wins at Jeopardy. Suddenly it becomes visible that machines process language better than many had expected.
2016
AlphaGo defeats the Go world champion. It becomes clear that AI can also get strong in areas that can no longer be neatly broken down into fixed rules.
2022
With ChatGPT, AI suddenly becomes directly usable for millions of people. No longer just researchers and corporations: almost anyone can try out what such models can do.
Since 2024
The real leap is not just in better answers. Models become multimodal, get tools, access systems and work on tasks across several steps. That makes them relevant to companies in day-to-day business.
Common questions
What is artificial intelligence?
Artificial intelligence is software that learns patterns from many examples instead of being given every rule. That lets it understand language, recognise images and create texts. Connected to tools and company data, it can take over whole work steps.
Can AI really think?
Not the way humans think. An AI system has no world understanding of its own, no consciousness and no intentions. It works with patterns, probabilities and the structures it learned in training. That is often enough for astonishingly good results. But it is something different from human understanding.
What is the difference between classic software and AI?
Classic software follows rules that people set beforehand: if A happens, do B. AI models, by contrast, learn patterns from many examples and can apply them to new situations. That is exactly why AI is more flexible, but also harder to fully predict than classic software.
Is ChatGPT already the whole AI revolution?
No. ChatGPT is just the best-known interface to a large language model. It gets really interesting where such models are connected to company data, tools, databases and clear action steps. Then it is no longer about a chatbot but about systems that support or partly take over work across several steps.
What exactly does generative AI mean?
Generative AI creates new content instead of merely sorting or evaluating existing data. That can be text, images, audio, video or program code. The difference from older AI systems matters: many earlier models were specialists for exactly one task. Generative models can be used for very different tasks.
Why does AI make such a difference for companies in particular?
Because it increasingly changes whole work steps, not just individual answers. When a system can search for information, bring it together, draft, document and kick off the next step, that changes the entire process behind it. That is exactly why AI is more than a new tool for companies.
Will AI replace jobs?
Some activities, yes; whole professions not automatically. More often AI changes work instead of making it disappear completely. Routine, preparatory work, summaries, first drafts or simple decisions can often be automated well. That is exactly why what humans do better becomes more important: making sense of things, taking responsibility, setting priorities and dealing with uncertainty.
Is AI dangerous?
It can be, if you trust it blindly or put it into critical processes without control. A model can be convincingly wrong, pick up biases from training data or handle sensitive information in problematic ways. The naive idea that AI automatically delivers objective or reliable results is dangerous, too.
Does a company have to switch everything to AI right now?
No. That would often be the wrong reflex. It makes more sense to look honestly at your own processes and check where AI actually improves something. Sometimes the right answer is a bigger overhaul, sometimes just a single process. And sometimes it turns out that a problem is better solved with clean software or clearer data.