3. Ethical concerns

Limitations of Large Language Models

 
Based on their technical characteristics, large language models, despite their spectacular performance, exhibit a number of problematic properties, including:

  1. Hallucinations: the model generates information that is false or fabricated without basis, even though it may appear convincing and authentic. This can occur especially when answering questions about specific facts or when creating details for which the model has no clear data points from its training.
  2. Inaccurate generalization: Machine learning iteratively approximates a statistical estimation function. The system operates stochastically by principle. Therefore, perfect accuracy is not possible. Patterns may be detected where none exist, or false connections may be made between concepts that are actually independent.
  3. Bias and stereotyping: Language models learn from large datasets that come from the internet or other sources. These data may be biased or contain stereotypes.
  4. Lack of contextual understanding: The language model may not correctly capture the context of a situation or question and therefore respond irrelevantly or incorrectly. Sometimes, the language model falls back on general superficial knowledge.
  5. Violation of copyright and privacy: Language models may occasionally reveal sensitive or private information derived from the data used for training.
  6. Security risks: Risks arise if the model, when generating answers, exposes sensitive, dangerous, or age-inappropriate content from its training material, such as instructions for building bombs, producing drugs, etc.
  7. Difficulty with logical reasoning and arithmetic: Since outputs are based on correlations and statistical predictions, language models (without appropriate plugins) struggle with complex logical reasoning. However, given their functionality, it is remarkable that LLMs are still able to acquire (limited) reasoning abilities from general textual knowledge and develop a certain "understanding," e.g., of causality.
Therefore, it is important to critically evaluate the outputs of language models, even if they sound linguistically convincing.

-     Optional Task: Test the limits of language models – find examples that demonstrate the limitations of AI chatbots, such as ChatGPT, and post your example in the appropriate category on the digital pinboard. Do you see connections to the described technical concepts of GPTs?