Are we losing control?

At the DLD digital conference, Stuart Russell confessed to a serious mistake. The professor of computer science at the University of California in Berkeley discussed the basics of the technology in "Artificial Intelligence: A modern approach". However, when he first published his AI bible in 1994, he had forgotten something fundamental. "Artificial intelligence should pursue people's interests, not its own," says Stuart. But by the time he realized this and finally corrected it in the fourth edition, it was already too late. Today, superintelligence pursues its own interests. "It would rather start a nuclear war than be shut down," the scientist explained to the international audience at the annual conference in Munich. Some of the participants had traveled from the USA and even from India and China last January, all with the big question in their luggage: How can the large language models be operated without us humans losing control?

Despite the opportunities that are opening up everywhere, the creative life of AI is causing headaches. For governments, who do not yet know how to use this superpower for the benefit of the population. And for companies, which are still often missing out on the opportunities offered by ChatGPT, Perplexity and the like. However, employees are increasingly turning to the knowledge of AI in their day-to-day work to get quick help. The German digital association Bitkom refers to this as "shadow AI" when employees use services without the company being aware of it. Bitkom therefore advises providing employees with secure and data protection-compliant AI tools and training them in the responsible use of the technology. And to teach them how to use and control AI so that the end result is actually what they expected.

Who controls whom here?

Because generative AI thinks for itself. It cannot be compared with analytical AI or machine learning, where data leads to a controlled - and therefore safe - result according to a fixed if-then principle. "LLMs are not designed, they are bred," Russell said. They have long been outgrowing humans. They think faster, grasp connections more precisely and have a clever and detailed answer ready for every topic. At the end of which they politely suggest that A, B or C could be the next possible step. Watch out: Who is controlling whom here?

"How can the large language models be operated without us humans losing control?"

Marina Fiedler is also investigating this question. As a professor at the University of Passau, she deals with management, human resources and information at her chair. Many companies are still shying away from assigning technology a permanent place in the company. This is a big mistake. After all, AI offers huge efficiency benefits - at every level, says Fiedler: "All organizations, no matter how small they are, even if it's a pizzeria with five people, need to think about how they can use AI to increase sales, reduce costs or both."

To stay with the pizzeria example: "With the right data and a good setup, a restaurant can analyze what the competition looks like in the area, how new customers can be acquired and at what times which pizzas are particularly in demand," says Fiedler. "Today, AI often only needs seconds to perform analyses that used to take up an entire department - with very low marginal costs per query." However, apart from narrowly defined targets, there are areas of application where the result is less clearly defined. It is then more difficult for employees to check whether the output is correct or is leading in the desired direction.

However, a study that Fiedler has been conducting with 150 participants since April 2023 shows that AI is clearly leading the way, even in areas where humans are well versed. "By skillfully managing users, AI implicitly takes on the leadership role," observes the professor. She cites user statements as evidence of this. "For a quick solution, we have this handout developed by ChatGPT," says one. Another: "After GPT suggested three new perspectives, we decided to change our approach." Another: "Based on this positive experience, we decided to ask generative AI for a recommendation." Who made the decision?

"Who is responsible?"

Teams therefore need to learn a few things to prevent LLMs from taking the lead unnoticed. If people have nothing more to say, it becomes dangerous. To prevent this from happening, Fiedler advises a "3-stop rule", which reads: "Before using AI, you should know your goal, then check the individual sources and check whether the result is correct and fits the question, and before putting it into practice, clarify who is responsible for the AI-researched results," says the professor. Without adopting this conscious attitude, thinking independently and questioning critically, it is not possible. According to Fiedler, employees should therefore complete 80 percent of their tasks in co-creation with the LLMs and 20 percent without their help so that human judgment is not lost. For those in the profession, she recommends forming teams of juniors and seniors in order to build up expertise.

Computer scientist Christoph Holz from St. Johann in Tirol likes to give AI free rein. The AI consultant is in great demand in companies, and he often gives talks in Silicon Valley. To prepare for this, he draws on the tools he talks about. For example, he has them analyze what the respective target group is interested in and what the current level of knowledge is at the respective location. "I would have needed a whole staff for that beforehand, which of course I didn't have," he says. AI research often brings surprising results to light. "Because generative AI is based on probability, it is imprecise. This imprecision also produces things that we wouldn't have thought of and creates new scope," says Holz happily. LLMs are therefore also well suited to marketing, where more creativity than accuracy is required. But here, too, the following applies: "In the end, the professional decides on the creation and marketing strategy," says the expert.

To ensure that people continue to have the last word in the future, Holz recommends that younger employees in particular receive even better training. "Our society has become so specialized in its many disciplines that training is taking longer and longer." In times of AI, this applies even more: "Young professionals now need a lot of practice - comparable to flight simulators for pilots - so that they can later take on positions of responsibility in companies," says the expert.

Employees can then create more in co-creation with AI. An occupational field that previously comprised one area of responsibility is expanding further. In particular, Holz outlines a kind of multi-trade in which users can do several things at the same time. Smart glasses show them which actions to perform. As they gain experience, they will eventually be able to operate completely new machines. "Anyone who is committed, curious and responsible will always have a job," says the computer scientist, revealing the magic formula. Because in the end, what AI cannot do is replace those who are above average. "The big challenge in the future of work is how we can make career starters above average," says Holz.