What Large Language Models Do, and What They Cannot Do

Large language models generate text by predicting the most probable next word based on patterns absorbed from vast datasets. They excel at producing coherent prose, answering questions, translating languages, and summarizing documents. What they do not do is think, plan, or understand meaning the way humans do.

The distinction matters because these systems now influence hiring decisions, medical advice, legal research, and educational assessment. Knowing what they can and cannot accomplish helps users deploy them appropriately and recognize their limitations before errors compound.

What They Actually Do

A large language model processes text as sequences of tokens—fragments of words or punctuation. During training, the system adjusts billions of internal parameters to minimize prediction error across enormous corpora: books, websites, scientific papers, conversations. The result is a statistical map of how words cluster, which phrases follow others, and what structures recur in different contexts.

When you prompt the model, it generates a response by sampling from this learned distribution. Each word emerges from a probability calculation over the entire vocabulary, conditioned on everything that came before. The model has no script, no database of facts, no flowchart of logic. It assembles output token by token, guided entirely by pattern matching at scale.

This approach produces remarkable fluency. The models complete sentences naturally, maintain consistent tone across paragraphs, and switch registers when prompted. They translate between languages by recognizing parallel structures in multilingual training data. They summarize documents by identifying sentences that statistically align with condensed versions seen during training.

Tasks Where They Perform Well

Large language models demonstrate strength in several domains. They draft routine correspondence, generate marketing copy, and rewrite content for different audiences with minimal human direction. Programmers use them to autocomplete code, suggest function names, and generate boilerplate based on comments.

They retrieve information encoded in their training data, functioning as a compressed, albeit imperfect, snapshot of public knowledge up to a cutoff date. Ask about a historical event, a scientific concept, or a common procedure, and the model often produces accurate summaries because those patterns appeared frequently during training.

They assist brainstorming by generating variations on themes, listing alternatives, and recombining ideas in novel arrangements. The models don’t evaluate quality, but they surface possibilities a human can then assess.

In customer service, they handle repetitive queries, route requests, and provide scripted responses faster than human agents. For language learners, they offer conversational practice and grammar correction based on patterns in well-formed text.

What They Cannot Do

Large language models do not reason. They cannot construct a logical argument from first principles, identify unstated assumptions, or follow a chain of inference that wasn’t reflected in their training data. When they appear to reason, they’re pattern-matching against reasoning-like text they encountered, not executing symbolic logic.

They do not possess knowledge in the human sense. They have no beliefs, no model of the world that updates with new evidence, no ability to distinguish what they know confidently from what they’re guessing. Every response is a probabilistic generation, with no internal flag marking certainty or uncertainty.

They cannot verify facts. If misinformation appeared frequently in the training corpus, the model may reproduce it fluently. They hallucinate—generating plausible-sounding but entirely false references, statistics, or quotes—because the objective during training was linguistic coherence, not factual accuracy.

They lack goals, intent, and persistence. They do not plan sequences of actions, monitor progress toward objectives, or adjust strategy when initial approaches fail. Each response is independent; the model doesn’t track what it tried before or learn from mistakes within a session unless explicitly fed its own prior outputs.

They cannot perform precise arithmetic or symbolic manipulation reliably. Calculations that require exact operations on numbers often fail because the model treats digits as tokens, not quantities. It may produce correct results for common calculations it saw during training, but novel problems expose the absence of true computation.

The Boundary Between Capability and Understanding

A model can translate a legal contract from English to Spanish with high accuracy, yet it comprehends neither the legal concepts nor the obligations described. It matches syntactic patterns and word distributions without grasping meaning.

It can write a coherent essay on photosynthesis, describing light-dependent reactions and carbon fixation, while possessing no concept of chemistry, energy, or biology. The text reflects structures absorbed from science writing, not understanding of natural processes.

This gap becomes critical in high-stakes applications. A model may confidently recommend a medical treatment based on symptom descriptions that pattern-match against its training data, unaware that the specific combination in a real patient indicates a contraindication.

Implications for Deployment

Effective use requires treating these systems as sophisticated autocomplete, not as thinking agents. They augment human judgment but cannot replace it in contexts requiring verification, accountability, or reasoning about novel situations.

Organizations deploying large language models implement verification layers: human review of generated content, cross-referencing with authoritative sources, and constraints that prevent the model from operating autonomously in sensitive domains.

Researchers work on architectures that incorporate external knowledge bases, enable multi-step reasoning, and provide uncertainty estimates. These additions address specific limitations but do not fundamentally alter what the underlying prediction mechanism can achieve.

What Remains Unresolved

Whether scaling—adding more parameters, more data, more compute—eventually produces emergent reasoning remains disputed. Some capabilities appear suddenly at larger scales, suggesting qualitative shifts. Others argue that pattern matching, however sophisticated, cannot yield genuine understanding without fundamentally different architectures.

The question of what these models actually encode persists. Researchers probe their internal representations, finding structures that correspond to grammatical rules, factual relationships, and even rudimentary world models. Whether these constitute understanding or simply effective compression strategies remains unresolved.

As these systems become embedded in daily tools, distinguishing what they do well from what they cannot do protects against misuse and misplaced trust. They predict text with extraordinary fluency. They do not think, know, or understand. That distinction defines both their utility and their limits.