Inspiration
F.acT: What do I need to be aware of when using AI and how is it possible that AI tools are still providing unreliable answers to some of our questions?
Jürgen Schmidt: The wrong answers of a language model are called hallucinations. This happens because of the basic architecture of these systems. Ultimately, it is a probability calculation that uses a "look-left-only" principle to calculate the next word in a sentence with the highest possible probability. We can already see that this has relatively little to do with "intelligence" as we humans understand it. But the results are incredibly good - at least in most cases. If a language model has too little basis for a particular question or the context in which the question is asked is not sufficiently narrowed down, it can happen that the probability of a sentence is relatively far removed from the actual reality. This would then be called a hallucination.
There is also the problem of bias. This means that the information can only ever be as good as the data material used to train the model. In this case, people with darker skin can be recognized far more poorly because there are far fewer photos. However, it also happens that, for example, the term man in connection with the term hospital clearly results in doctor and, conversely, in this context the term woman is used to refer to nurse. This is because this connection is made much more frequently in the data material and the AI naturally adopts this. In almost all cases, bias is something that can be found in training and ultimately only represents our society's distortions of the truth.
Back in 2018, Judea Perl wrote "The Book of Why", where she discusses how we could develop a so-called "reasoning" system in which AI models can check their own answers for credibility. However, I would always recommend checking the information provided by a language model and checking the sources given.
F.acT: Sustainability in tourism poses a variety of challenges for the industry. How can AI contribute to this transformation?
Jürgen Schmidt: Artificial intelligence and ecological sustainability are currently topics that do not yet go well together. For 60 requests to an LLM, we currently use around 1 liter of water. Due to the enormous spread of AI models, we have tens of thousands of requests per second. We can all work out for ourselves that this is a massive intervention in our resource budget.
Nevertheless, there are many areas in which AI can have a positive impact. After all, AI does not consist exclusively of language models, even if the latest developments make it look that way. Artificial intelligence can be used excellently, for example, to derive future developments from historical data, i.e. for demand forecasting. Visitor flows can be directed intelligently and with a focus on ecology. This can also result in more intelligent destination management. Dynamic pricing could also be used to influence the behavior of tourists and thus promote more ecological activities.
Another very exciting area would be environmental monitoring that uses IoT, sensors, drones and other data signals to observe and visualize the balance of ecology. Here, too, there are already countless projects that have been successfully implemented.
Language models could also be used to good effect, e.g. in interactive information systems that are designed for ecological behavior and aim to provide guests with more information on the topic of climate change. This can have a direct impact on people's behavior and strengthen the destination at the same time.
