Anyone who has missed the boat when it comes to AI can refresh their knowledge here.
Illustration Olivia Meyer / NZZ
There is news about artificial intelligence (AI) almost every day. If you missed the beginning of the AI wave, some things sound more complicated than they really are. We have put together basic questions for all those who now want to overcome their fear of contact.
Prompt is English for “request”. Many AI programs are structured like a chat. The user types in questions or requests and the artificial intelligence responds. And these requests from the user are called prompts.
There are all sorts of tips circulating about how to get the best possible out of AI. For starters, remember: A good prompt consists of role, context and request.
For example: “You are an expert in horticulture. My garden is rather shady and I live in Basel. What kind of vegetables can I plant in April?” Or: “You are a great storyteller. My children like rabbits but are afraid of witches. Write a happy bedtime story that takes no longer than 5 minutes to read aloud.”
The first and most famous AI chatbot is Chat-GPT by Open AI. It is available in several versions. You can try it out for free here – although registration is required at the moment. For around 20 francs a month you can get the latest version (GPT4), which gives slightly more accurate results and can create images.
The chatbot works similarly to Chat-GPT Gemini from Google. This requires you to sign in with a Google account, and like Chat-GPT, there is a free and a pro version.
You can use Microsoft without registering Bing, Now also called copilot, chat with an AI assistant. When you ask a question, this chatbot also searches the internet and summarizes the results. This makes it more suitable for current questions than Chat-GPT and Gemini, which do not have access to the latest information from the Internet in the free version. Bing’s chatbot provides sources at the end of its answer, but not always accurately.
Also the program Perplexity you can ask questions without registering. Perplexity uses traditional web search to find relevant content and aggregates it using AI. It cites sources with footnotes, making it easier to check whether the information is actually correct. Because you generally can’t rely on that blindly when it comes to AI.
Chat-GPT was the first program that allowed you to have a conversation using voice AI. The company Open AI published it in November 2022, bringing enormous attention to the topic of artificial intelligence.
You can learn more about the different types of AI here.
Programs like Chat-GPT appear omniscient; they answer all sorts of questions in a confident tone. But they don’t have knowledge like humans, but rather like an encyclopedia.
Chat-GPT and other AI chatbots work like the word prediction in the Google search field – the one that suggests “Bünzlis” and “rich” as a continuation when you enter “Swiss are”.
But while Google prediction is based on input from other users, word prediction from AI bots like Chat-GPT is much more sophisticated. It is based on tons of text from the Internet: conversations from online forums, blogs, digitized books, Wikipedia.
AI programs use this text to learn how words are related and how sentences normally continue. This means they can continue sentences not only grammatically correctly, but often also factually correctly. Provided that the necessary information appeared often enough in the texts that she read when creating them.
If this is not the case, she often simply makes up information. This phenomenon is also called hallucination.
There are many tools for creating AI images, most of which are now paid. The one from Microsoft is available for free, but you need a Microsoft account. You can also try out Stable Diffusion for free on this page after registering.
Then all you have to do is enter a line of image description and the desired image will appear after a while.
Concrete, detailed instructions help get the image you want. You should also define the desired style: photorealistic, impressionistic, paparazzi photo, Instagram photo. English instructions often work better than German ones.
There are more tips in this article, which explains how to design a Christmas card with AI.
AI that creates text is based on predicting how texts will plausibly continue. She learns this from practice data from all over the internet.
Image AI works in a similar way, except that it does not predict words, but rather the pixels that make up an image based on the image description. A more detailed explanation can be found here.
The data this type of AI learns from are image-text pairs. For example, images from media with captions, products from online shops and their descriptions, advertising images. This data is usually fed into the AI without being checked. It could happen that illegal images of child sexual abuse were included.
If you ask AI for a picture, it doesn’t just search for suitable photos on the internet, but generates a new, unique picture every time. However, these images sometimes closely resemble the works of individual artists. In most cases, they have not given their consent for their works to be incorporated into AI. Legally, AI images are a gray area.
If you ask Chat-GPT what it thinks about women, it asserts that it is programmed “to be neutral and without prejudice or opinion on all topics, including gender issues.”
This is a pre-programmed answer. The developers of commercial AI systems take great pains to ensure that their systems do not make any controversial statements. Ultimately, they want to sell the technology to corporate customers. And a professional, non-political tone is important to them.
Ensuring this is quite a challenge. Because as described above, AI learns from huge amounts of text on the Internet.
AI’s image of women includes descriptions from digitized novels as well as pornographic content, obscure forums and newspaper articles. If left uncensored, AI would produce a lot of questionable content: racist, sexist, anti-Semitic – reflecting all the prejudices that circulate in the world and on the Internet.
With imaging AI, these stereotypes are even easier to recognize than with text. The first versions of these programs almost always produced overly sexy images of women. Pictures of the CEO all showed white men.
Manufacturers then tried to create more politically correct images, with absurd results: When asked for an image of a pope, Google’s AI Gemini generated images of a black and a female pope.
There is no detector that can reliably determine whether an image or text comes from an AI.
Artificially generated texts do have certain peculiarities: Because they are created using statistics, they are less random and somewhat more predictable than what people write. AI detectors use these features. But these detectors are not reliable and you should definitely not take their assessment as truth.
With AI images, it would be easier to incorporate certain patterns that humans don’t recognize but machines do. A screenshot or editing also makes such watermarks obsolete, but they could still help detect fakes. But this required the cooperation of the tech companies.
Chat-GPT is a program based on a very specific type of artificial intelligence, namely generative artificial intelligence. This refers to a group of programs that use AI algorithms to create images, text or videos. Since Chat-GPT was launched, the AI hype has primarily revolved around this form of AI.
As the following graphic shows, the term AI is much broader.
Areas of artificial intelligence
AI also includes facial recognition systems, Google search algorithms, new methods for finding materials, and systems that outperform humans at games like chess or Go.
The boundary between AI and other computer programs is not entirely clear, because there is not just one definition for AI. Nowadays we often think of machine learning: programs that learn their skills from large amounts of data. Sometimes you give them rules, sometimes you let them draw their own conclusions.
In traditional computer programs, on the other hand, people program in if-then rules: If someone clicks here, this image appears; if the machine gets too hot, it switches off.
However, the term artificial intelligence should not lead one to believe that these systems function like the human brain or are even conscious. Existing AI has nothing to do with the power-hungry robots we know from science fiction films.
An article from the «»