How committed are you to critical thinking about the technologies you use?

Another post here on LinkedIn about the importance of not just using LLM tools (like ChatGPT) but also thinking critically about how we use them. Especially because there are some things users can’t ask LLMs—those questions rely on our human intelligence, specifically, Critical Thinking.

The focus of this new post is on the conclusions we can draw from prompt suggestions, which have become the most productive way to get good results when using LLMs like ChatGPT. These prompts are linguistic constructions that closely resemble the technical expressions we find in the logical structure of algorithms.

The first inference we should make is that technical and logical articulation is highly effective in a tool whose foundation is built on mathematical logic. While training allows certain nuances of natural language to be understood, a more technical, unambiguous approach remains the best method for reasoning. Since LLMs’ reasoning models are based on human logic, what I’ve written here applies to ChatGPT in the same way it does to human cognition.

As we continue our exploration, having clarified that logical writing should not be abandoned when using LLMs, the next step is to ask: if we create or use a well-constructed prompt, does that mean we will fully learn from the response? Does a good prompt equal complete comprehension of the answer (everything there is to understand within the context)? My answer is no.

Readers of my previous post already knew the answer. Here are some scenarios that make this clear:

A) Someone uses a prompt but doesn’t fully understand what that prompt can return. If I’m not designing the prompt based on my own knowledge and weaknesses, it’s difficult to know which specific objectives are or are not being achieved by the LLM. The more generic the usage, the less meaningful the analysis of the result will be. This is the first case of acceptance, either in full or partially, in the dark.

B) Someone uses the prompt but has no knowledge of the topic they’re getting an answer for. If the response doesn’t contain anything that feels counterintuitive to the user, it’s just accepted. If there’s an error, it’s hard to spot because, after all, they didn’t know enough to actively look for mistakes. This is the line of thinking that draws a clear distinction between a teacher using ChatGPT (as an aid and never as a source of unvetted new information) and a student, who often lacks the depth of analysis needed.

Note, readers, that scenarios A and B complement each other.

C) For years, there have been countless books aimed at providing the most accessible paths for learning a given topic. Yet, a significant number of readers still struggle with learning, often due to either a deficiency in reading comprehension or an inability to achieve the necessary level of abstraction. They tend to rely on concrete examples that only work when applied to identical situations. This is a common scenario I encounter in the classroom and in discussions with others. Since LLMs primarily function as text generators, only extensive research involving a large number of users across various segments and uses can determine whether complex text is properly understood. And research like this is not in the interest of those who are already convinced (without evidence or coherent arguments) that LLMs are a mature and problem-free solution.

These three scenarios, if ignored, will jeopardize the future of LLM usage, much like how today we realize that the unrestricted use of smartphones in the classroom can hinder the learning of a significant portion of society. No technology is free from challenges or incapable of creating problems, as they are applied by humans who have many vulnerabilities, often concealed from critical self-reflection and from the importance of influencing major achievements.

How committed are you to critical thinking about the technologies you use?