How do ChatGPT and Gemini think?

Behind the sleek interfaces of today’s chatbots lies a labyrinth of code, data and unanswered questions. Artificial Intelligence (AI) has become humanity’s most impressive mirror, reflecting our collective knowledge, creativity and confusion.
From Microsoft’s Copilot to Google’s Gemini and OpenAI’s ChatGPT, these systems appear almost magical. Ask them to summarise a report, write poetry, or translate a medical abstract and they reply in seconds. Yet few, even among the engineers who build them, truly understand what happens behind the screen, or how much of it remains a mystery.
At the core of every modern chatbot lies a Large Language Model (LLM), a digital brain with hundreds of billions, sometimes trillions, of parameters. These parameters are not programmed rules, but emergent correlations derived from analysing massive libraries of human textbooks, websites, code and research papers. These internal representations and pathways are highly complex, intricate and nonlinear, making it difficult to trace how a given output is generated from a specific input.
No engineer can point to a single neuron and say, “This one understands irony”. Instead, intelligence emerges. Layers of mathematical connections detect subtle linguistic rhythms and semantic patterns. Why some clusters “light up” when asked moral or creative questions is still unknown. Even at OpenAI and Google DeepMind, teams describe their models as “black boxes” whose inner logic defies full explanation.
The reason an AI writes what it writes is often as mysterious to its creators as it is to its users. Before these systems speak, they undergo an immense and largely unseen process of data cleaning, human feedback and reinforcement. Thousands of annotators across the globe review text, rate AI responses; and guide tone and ethics through Reinforcement Learning from Human Feedback (RLHF).
This phase doesn’t teach reasoning; it teaches acceptable behaviour. Over millions of interactions, the AI learns which answers sound “helpful” or “harmless”, a definition shaped by human judgment rather than complex logic. While companies emphasise ethical data sourcing, the reality is opaque. Much of the internet contains bias and misinformation; and these, inevitably, seep into AI memory.
Proprietary and licensed datasets, including snippets from publishers, academic databases and coding sites, supplement public data. Engineers often work with pre-processed corpora and never see every source. This phenomenon, called data opacity, means even developers cannot always trace the full origin of what their model “knows”.
The synthetic knowledge and the illusion of understanding when ChatGPT or Gemini replies are not recalling facts but performing an ultra-fast probability game, predicting the most likely word sequence. The result feels intelligent because language itself is patterned.
Researchers call this synthetic cognition: intelligence without comprehension. When an AI fabricates a citation or misquotes a figure, it is not deceiving; it is completing a linguistic pattern that statistically fits.
Peering Inside the Black Box reveals that new research in mechanistic interpretability seeks to map how these internal circuits represent ideas such as empathy or arithmetic. Early findings show “concept neurons” that consistently activate for abstract categories such as cities or melodies. Still, most scientists admit they are steering a ship whose navigation system they only partly understand. The gap between creation and comprehension remains wide.
AI today is both a marvel and a mystery, a tool advancing faster than our theories can keep up with. Engineers can train and align it, yet they cannot fully explain its mind. Perhaps that is what makes artificial intelligence profoundly unique: it works beautifully, even when its creators don’t quite know how.