Artificial Intelligence and Critical Thinking: Beyond the Hype

A Necessary Warning about Generative Artificial Intelligence and the Erosion of Critical Thinking

In our relentless pursuit of efficiency, knowledge, and quick solutions, Generative Artificial Intelligence (AI) (such as Large Language Models or LLMs, exemplified by ChatGPT, Claude, Gemini, etc.) has become an omnipresent and astonishingly powerful tool. It can write code, explain concepts, summarize texts, generate ideas, and much more. However, its use, especially in professional contexts where precision and truthfulness are paramount, requires a level of discernment and critical thinking even greater than what we apply to traditional information sources.

Recently, I had a personal experience that served as a wake-up call. While preparing material on C++ performance optimizations, I based part of my research on case studies of companies that had achieved significant improvements. To enrich this content, I turned to ChatGPT, requesting additional examples of system-level optimizations using C++. One of the cases it suggested with apparent conviction was how Twitter, around 2011, had managed to speed up its search engine by about 3.5 times thanks to a rewrite of critical parts in C++. The narrative was coherent, and the data was specific. I insisted on verifying this information with the AI, and it assured me of its correctness, even offering plausible details. Trusting this AI-validated information, I used this case in a publication.

Days later, I wanted to delve even deeper into the technical details of that supposed Twitter case to expand the analysis... and I was met with a bewildering surprise: the alleged original article or primary source seemed to have disappeared or was unfindable. Searching other sources and using different AI tools (like Grok, which has more direct access to recent web information), I discovered the truth: while there was a significant improvement in Twitter's search performance around that time, and while it may have involved various technologies, the specific narrative of a massive rewrite in C++ as the main cause of a '3.5 times' speedup was, at best, a distortion or, more likely, a convincing 'hallucination' by the AI. The core of the anecdote I had shared was incorrect. I deleted the post immediately.

This experience left me with a crucial reflection:

• When I use AI for software development tasks (generating boilerplate, debugging, exploring APIs), my own knowledge and experience in the domain allow me to quickly detect when the AI makes a mistake or produces suboptimal or incorrect code. I am trained for that scrutiny in that context.
• However, in a different context, such as research for an article or a publication where the topic is slightly outside my deepest specialization, it is easier to let my guard down and assume the truthfulness of a well-written and seemingly authoritative response from the AI.
• And that is a dangerous mistake: Generative Artificial Intelligence can 'lie' or fabricate information in an extraordinarily convincing way, not out of malicious intent (as it lacks intentions), but due to the very nature of its operation (predicting the next most likely word based on patterns from the vast data it was trained on). It reproduces truths, half-truths, common errors, biases, and fictions without intrinsically distinguishing them, unless we, the humans, apply a rigorous critical filter and verify the information against reliable primary sources.

Today, in an era where misinformation is rampant and truth seems to become more relative, this inherent risk in Generative AI is enormous. The same rigor, healthy skepticism, and verification discipline that we apply (or should apply) in our professional development work must be imperatively extended to all areas where we interact with and depend on AI. We must never lose our critical awareness, not even (or especially) when an AI response sounds perfect, eloquent, and exactly like what we wanted to hear. AI is a tool, and like any powerful tool, its responsible use falls entirely on the user.