In recent years, artificial intelligence (AI) platforms such as ChatGPT, Gemini (by Google), and DeepSeek have entered the public and professional consciousness as tools that can generate text, assist in problem solving, and automate a wide variety of cognitive tasks. These tools are underpinned by technologies called large language models (LLMs), which have stirred both enthusiasm and concern. In this essay, we will explore how these platforms work, their strengths and weaknesses, and their limitations — particularly in terms of reasoning, logic, and truthfulness.
At the core of ChatGPT, Gemini, and DeepSeek are LLMs, massive neural networks trained on vast corpora of text data. Their primary function is to predict the next word in a sequence, given a prompt or preceding text. This might sound trivial, but when scaled up to billions (or trillions) of parameters and trained on internet-scale datasets, the resulting systems can generate convincingly human-like text, mimic tone and style, and answer questions in a coherent manner.
The "language model" part comes from their origin in natural language processing (NLP), a subfield of AI focused on making machines understand and produce human language. The "large" part, meanwhile, refers to the number of internal parameters — essentially, the knobs and dials inside the model that are tuned during training.
Training a model like ChatGPT involves exposing it to vast swaths of text — books, articles, websites, code, and more. This data is tokenised (split into word-like units), and the model learns to assign probabilities to different possible next tokens. Over time, it develops statistical correlations between language patterns. For example, if it sees the phrase “peanut butter and”, it learns that “jelly” is a likely follow-up in English-speaking contexts.
Importantly, the model doesn't understand these concepts the way a human does. It doesn’t “know” what peanut butter is, or why jelly goes with it. It only knows that these words often occur together.
Training is typically done using high-performance computing clusters over weeks or months, and the process consumes significant energy and financial resources. Once trained, the model is "frozen" — it doesn’t keep learning unless retrained or fine-tuned on new data.
Once trained, these models are wrapped in user-friendly interfaces and sometimes additional layers of processing to make them useful for end users. When you type a question or prompt into ChatGPT, for instance, the backend converts your input into tokens, feeds it through the model, and produces a set of predicted next tokens — essentially guessing, word by word, what the most statistically appropriate response is.
ChatGPT (developed by OpenAI) is designed with conversational flow and general knowledge in mind, often used for brainstorming, writing help, coding support, and Q&A.
Gemini (formerly Bard, developed by Google DeepMind) is tightly integrated with Google’s search infrastructure and increasingly focused on multimodal capabilities — i.e., the ability to process text, images, and other input types together.
DeepSeek is a Chinese-developed LLM project that focuses on bilingual and scientific applications. It’s been making strides in maths and technical understanding, offering models optimised for reasoning and code generation.
These abilities stem from their training on huge and diverse datasets. If something has been written about online in accessible form, there's a good chance an LLM will be able to generate content related to it.
Despite their impressive fluency, these models are not guaranteed to be factually accurate. A key concern with LLMs is their tendency to “hallucinate” — that is, to generate confident-sounding statements that are completely false, unsupported, or even nonsensical.
This happens because the models do not verify facts. They don't consult databases, or check logic like a formal reasoning engine. Instead, they base their output purely on pattern-matching. So if you ask for a list of books written by a fictional author, the model might invent some. If asked a question about an obscure legal statute, it might produce an answer that seems plausible but is entirely fabricated.
In fact, some platforms like Gemini have integrated live search capabilities to mitigate this problem — but even then, factual reliability is far from perfect.
Because of this, all output from platforms like ChatGPT, Gemini and DeepSeek should be regarded as provisional — a first draft rather than a final answer. Critical use of these tools requires human judgment. Where facts matter (e.g., legal, medical, historical, or scientific contexts), it's essential to cross-reference with reliable sources. The responsibility for accuracy ultimately rests with the user.
Some newer models, such as ChatGPT with the “browsing” or “search” option enabled, try to pull up current information from the web to inform answers. However, even then the information is filtered through a statistical language generator, not quoted verbatim unless explicitly instructed to do so.
This brings us to a fundamental limitation: these AI platforms do not think. They do not possess beliefs, consciousness, or intentions. While they may sometimes appear logical or reflective, these are illusions of language — emergent properties of the statistical training process.
They cannot:
At best, they can simulate logical processes by mimicking the structure of arguments found in their training data. This is sometimes effective — especially in technical fields like mathematics or software development — but it lacks reliability under scrutiny.
Many people assume that a fluent AI model must be "intelligent" in a human sense. In truth, the fluency is surface-level. The model doesn’t “know” what it’s saying; it doesn’t understand cause and effect or physical reality. It cannot “know” it is wrong, unless you tell it. It cannot introspect or learn from mistakes on the fly.
Similarly, the idea that AI can “take over” roles requiring creativity or moral judgment is premature. While AI can certainly assist in creative endeavours (writing lyrics, drafting code), it lacks genuine intention, taste, and lived experience. These are qualities still uniquely human — for now.
Despite the impressive façade, LLMs like ChatGPT, Gemini and DeepSeek cannot:
In short, they are tools — sophisticated, yes, but not agents. They are best thought of as predictive word engines with extremely good training, not synthetic minds.
LLMs are incredibly helpful when used as idea generators, summarisation tools, and language assistants. They are less suitable for tasks requiring high-stakes accuracy (such as legal advice, scientific discovery, or medical diagnosis) without human review.
To use AI responsibly, one should:
Research continues to address these shortcomings. Future models may incorporate better reasoning modules, integration with real-time databases, and more rigorous alignment with human values. OpenAI, Google DeepMind, and others are experimenting with methods to improve factuality, reduce bias, and allow limited memory between sessions (with user control).
Nonetheless, the gap between “appearing intelligent” and “being intelligent” remains vast. True reasoning, grounded thinking, and ethical judgment are, for now, still uniquely human capabilities.
AI platforms like ChatGPT, Gemini and DeepSeek represent a remarkable convergence of language, computation, and pattern recognition. They are game-changing tools — but they are not oracles, nor should they be mistaken for independent intelligences. Their outputs are often useful, sometimes entertaining, and occasionally wildly wrong.
Like all powerful tools, they must be used thoughtfully, critically, and with full awareness of their strengths and blind spots. Used well, they can amplify human creativity and efficiency. But left unchecked, they can mislead, misinform, and create illusions of competence where none exists.
In the end, the responsibility for understanding — and truth — still lies with us.