Technological developments in artificial intelligence

Hey Gemini, what is intelligence?

“Intelligence is a complex cognitive ability that encompasses the understanding, processing and application of information. It enables people to solve problems, recognise connections, learn from experience and adapt to new situations.”

[Answer from Gemini 2.5 Flash]

The concept of intelligence has long been directly associated with humans or other living beings. Even the intelligent voice assistant Gemini understands intelligence as a human characteristic. To understand exactly what artificial intelligence (AI) is and what processes take place in the background, it can be helpful to compare it with human intelligence.

The basis for human intelligence is the brain. Billions of nerve cells – known as neurons – work together there. They are connected to each other via contact points called synapses. Information is transmitted from one nerve cell to the next via these synapses with the help of chemical messengers such as dopamine or serotonin. This is how thoughts, memories and feelings arise. Learning manifests itself through the strengthening or weakening of specific synapses. Our experiences, attitudes and character traits are thus reflected in the micro structure of the brain.

Artificial intelligence is based on an artificial network that is roughly modelled on the human brain. Here, too, there are “neurons”, but these are not real cells, rather elementary calculation rules. These record data, process it and pass on the results. Instead of chemical messengers, AI uses arithmetic operations and numerical values to learn and recognise patterns. An artificial neuron also has a “strength” like a synapse. A neural network is an interconnection of a large number of such neurons, whose strengths are geared towards a specific task on the basis of training data. A network trained in this way can seemingly make its “own decisions” – even if it ultimately only reacts according to certain rules set by humans.

With the advent of increasingly powerful computers capable of processing ever larger amounts of data, the development of groundbreaking algorithms and high-performance hardware in providers' server farms – especially GPUs (graphics processing units) from the field of computer games – as well as the widespread use of powerful end devices in the form of smartphones and tablets, AI has now found its way into a large part of our society and has found real-world applications in almost all areas of life. This field of research has been in existence for 70 years. The first theoretical considerations as to whether machines can “think” actually took place as early as the 18th century. However, it took many years before the first breakthrough was achieved.

British scientist Alan Turing is considered a pioneer in the field of artificial intelligence. His concept of the “imitation game” (1950) is now known as the Turing test and was the first to demonstrate how cognitive processes can be performed by machines.

This approach is considered to be the precursor to artificial intelligence – a term that was first used in 1956 at the Dartmouth Conference, when a programme called Logic Theorist was written that could prove mathematical theorems.

Based on the research of American psychologist and computer scientist Frank Rosenblatt, who invented the first artificial neural network model in 1957, American scientists David Rumelhart and Geoffrey Hinton, who later won the Nobel Prize in Physics, demonstrated in 1986 how such artificial neural networks can be effectively trained using error feedback (backpropagation), thus laying the foundation for modern deep learning systems. 

In 1997, the Deep Blue supercomputer developed by the American technology company IBM wins a chess match against the reigning world chess champion Garry Kasparov. This victory is considered a turning point in the history of AI and shows that machines can surpass human decision-making abilities in clearly defined areas.

AI received its ultimately decisive boost in 2012 with the development and application of artificial neural networks. In particular, the victory in the ImageNet competition by a deep neural network called AlexNet, developed by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton, accelerated this surge. The model outperformed all competitors in image classification and demonstrated that deep learning has practical applications and is significantly more powerful than traditional methods. 

During the 2010s, various AI models such as GPT-3 and AlphaFold began to rise rapidly, leading to AI now penetrating many areas of society and becoming part of people's everyday lives. 

With the technological breakthroughs of recent years, public debate on the subject of AI has also intensified. This repeatedly involves a mixing of different terms that are often related to each other but mean different things. In particular, the conflation of the term AI with buzzwords such as automation, algorithms, machine learning and deep learning means that many people do not have a clear idea of how AI actually works.

In addition to technological innovations, particularly in the field of deep learning, the “triumphant advance of AI” is based above all on the widespread availability of the internet and user-friendly interfaces, i.e. simple, intuitive interfaces such as chat windows or voice assistants, which make AI immediately accessible to most people today without any prior knowledge.

A particularly dynamic development is currently taking place in what is known as generative AI. It enables the creation of entirely new content – such as text, images, speech, or even videos – that is almost indistinguishable from human-made products. This fundamentally changes not only how we interact with technology, but also how content is created – from everyday communication to professional media production. Technically speaking, such generative AI models are based on deep learning techniques, i.e. artificial neural networks with many layers that analyse large amounts of data and learn from it. Large language models (LLMs) in particular have received a new boost since the release of the ChatGPT tool in 2022, and the market has since been expanded to include other powerful tools such as Perplexity, Gemini and DeepSeek.

Users cannot understand how generative AI tools generate their output, what training data they use, and to what extent the information they provide is actually correct and therefore reliable. In the still young field of explainable AI research, methods are being developed to make AI-generated suggestions or decisions traceable after the fact. However, it faces considerable challenges, particularly with respect to large language models (LLMs). LLMs generate output based on context, with the relevant context derived from the prompt – the textual user input – and the text already generated. There is currently no technological approach for the external reconstruction of these contextual references. Current research therefore aims to enable LLMs to engage in self-explanation processes. However, due to their purely statistical prediction principle, they have neither a substantive understanding nor actual insight into their internal decision-making processes. Accordingly, self-declarations can be both accurate and freely constructed. This problem shows that it is always important to critically examine in which areas of application a potential risk of misinformation appears acceptable and in which areas a high level of transparency is absolutely essential.

The computer scientist Zeynep Akata conducts research in the area of explainable AI and develops AI that combines visual, linguistic, and conceptual elements, and thus makes its decisions comprehensible to humans. For this, she will receive the “ZukunftsWissen” (Future Knowledge) Award from Leopoldina and the Commerzbank Foundation in 2025. In the interview, she introduces herself and her research and talks about what the award means to her.

The example of combining visual, linguistic and conceptual information in modern AI models shows that explainable AI can offer real added value, especially when combined with innovative learning approaches. The AI learns not only from image data, but also from linguistic descriptions, e.g. of typical characteristics of a particular animal or clinical picture. Especially in the case of so-called low-shot or zero-shot learning – i.e. learning with few or no specific training examples – AI models must generalise, link concepts and draw on existing knowledge. If these models also use language or visual cues, they can potentially make their decisions easier to understand. This creates an important prerequisite for explainable AI: In applications such as medical image analysis, AI can not only make a diagnosis, but also visually or verbally explain which characteristics led to the decision. Such approaches strengthen transparency and trust among users.

However, the widespread use of AI easily obscures the fact that many systems are actually capable of much less than the term “intelligence” suggests. AI supports humans in certain areas, such as navigating in a car or creating texts, but in these cases it relies on human input in the form of specific commands or the activation of functions (e.g. traffic jam reports). Contrary to popular fears and horror scenarios from the entertainment industry, AI models that base their actions on their own consciousness and human abilities such as empathy, emotionality or creativity, or that act on their own initiative, still only exist in theory.

Published: September 2025