Category: propTech

Whether text creation, language translation, email management or media production, AI has long been part of everyday life for many people. A study by the University of Zurich shows that more than half of the population already uses tools such as ChatGPT or Gemini. The younger generation in particular is open to the new technology.

High economic potential with risks
In companies, AI is seen as a driver of rationalisation. Studies forecast potential worth billions for the Swiss economy. Especially in the areas of pharmaceuticals, logistics, education and software development, experts are expecting profound efficiency gains. However, the euphoria is clouded by the first warning signs.

Dubious content instead of real information
In the journalistic and media environment, AI is already being used for the mass production of content, from sports reports to financial data. However, in many cases, meaningfulness, source clarity and factual accuracy fall by the wayside. Europol warned as early as 2023 that up to 90% of online content could be synthetic by 2026.

When machines learn from machines
A central problem is that AI models are based on existing data. However, these are increasingly AI-generated themselves. The result is a self-reinforcing feedback loop in which quality and factual accuracy drop rapidly. If models are trained with synthetic data, the results deteriorate drastically, even to the point of complete system collapse.

The Grok case and the ethical dimension
A recent incident shows just how dangerous this development can be. The chatbot “Grok” developed by Elon Musk recently disseminated anti-Semitic content and praised Adolf Hitler. The cause is presumably uncontrolled training on manipulated or synthetic content. The case illustrates how urgently rules for data validation, ethical guidelines and quality standards are needed.

Synthetic data is no substitute for reality
Synthetically generated data only depicts historical patterns and can perpetuate or reinforce existing biases. Without new, high-quality training data, further development will come to a standstill. At the same time, the handling of sensitive or manipulative content raises fundamental questions about responsibility, transparency and regulation.

Modern AI systems such as ChatGPT or Gemini master language with impressive naturalness. This is made possible by so-called transformer models, which are particularly good at recognising relationships between words thanks to their structure. But how does the transition from mere syntax to real semantics work?

From form to meaning
A recent study in the Journal of Statistical Mechanics provides the first experimental evidence that neural networks go through a clear turning point during language learning. Initially, they are guided by the order of words; a sentence such as “Mary eats the apple” is recognised primarily by its structure. However, as soon as a critical mass of training data is reached, the model begins to decode the meaning. This transition is abrupt, comparable to a physical phase change.

First patterns, then understanding
This learning process is similar to human language acquisition. Children also first recognise patterns and sequences before they grasp meanings. In AI systems, this role is played by the self-attention mechanism of the transformers. It allows the model to prioritise each word in context and determine its relevance to the meaning of the sentence.

Statistical physics meets machine learning
The leap in learning described above can be explained using concepts from thermodynamics. Just as water changes from a liquid to a gaseous state at 100 °C, the behaviour of a neural network also changes as the amount of data increases. The many interconnected neurons change their strategy collectively, a statistically describable change.

More data, more meaning, more responsibility
The more data a system receives, the easier it can form semantic concepts. However, as the size of the model increases, so does the challenge of ensuring transparency, security and efficiency. The findings on the learning leap open up new ways of controlling AI in a targeted manner, for example through conscious data management or adaptive architecture designs.

Implications for research and application
The models analysed are simplified, but they reveal fundamental principles. Knowledge of the phase transition provides valuable information on how AI systems can be trained more robustly, quickly and precisely in the future. At the same time, it opens up new perspectives in didactics, in the human-AI interface and in the ethical reflection of machine intelligence.

The moment AI begins to understand
Artificial intelligence does not simply learn language, it undergoes a qualitative transformation. Only when it has sufficient knowledge does it recognise not only where a word is, but also what it means. This moment, when position becomes meaning, marks the beginning of true machine language competence and perhaps the key to deeper understanding between humans and machines.