The right questions to ask AI and evaluating the answers
When discussing artificial intelligence, we must begin with a fundamental question: what are prompts? These are the instructions that the user provides to AI systems, specifying what they wish to obtain from the algorithm. The design of specific text-based commands is now the subject of a growing number of dedicated courses. However, the issue goes beyond simply teaching prompt engineering: it is equally important to teach people to recognise when, even with well-crafted instructions, AI systems are unable to provide an adequate response, why this happens, and how to identify such situations.
The key issue is the development of critical and lateral thinking, which is more important than ever – not only for formulating and using prompts appropriately. It is precisely this critical thinking ability that students need to develop in order to use artificial intelligence systems responsibly: that is, the ability to understand when the results obtained through these tools are actually reliable and when, on the other hand, they are not. One aspect to bear in mind is that not all types of problems and cognitive analyses are handled by artificial intelligence systems with the same degree of reliability.
It should be noted, however, that the research and analysis currently underway in this field show that the limits of reliability are constantly evolving, given the speed at which these systems develop and improve their performance. Nevertheless, the available evidence is already highlighting certain categories of problems in which artificial intelligence systems prove to be particularly reliable. These include text summarisation, particularly of classical texts, which have long been available in various forms online and have, in all likelihood, become part of the data used to train the algorithms. These are often very lengthy and complex essays, which would take a considerable amount of time to read in full. Consider areas such as the analysis of regulatory or case-law corpora, or the review of key contributions from entire schools of thought, or even the analysis of large quantities of accounting documents or statistical data online to arrive at overall assessments.
The ability to summarise content of this kind effectively is, as a rule, one of the most well-established strengths of these systems. Asking artificial intelligence systems to summarise the content of a long and complex text, identifying its key points, or to answer questions about the content itself, the relationships between its various parts, and their meaning, or even to present it in greater or lesser detail and with varying levels of summarisation, appears to be one of the most established and reliable capabilities of these tools.
This means that certain types of analysis can now be easily delegated to artificial intelligence systems, often with surprising results. And what, on the other hand, are AI systems still unable to do? What is now clear is that these systems encounter greater difficulties when called upon to generate new knowledge, particularly in cases where lateral thinking and the ability to establish indirect associations between pieces of information not explicitly present in the texts used to train the algorithms are required. Similarly, tasks requiring truly innovative and creative output, or the development of a multi-step logical process in which the link between the initial premises and the final conclusions becomes more complex and less straightforward, prove more challenging to tackle.
