What are the main limitations of LLMs?

Large language models have hard boundaries: they cannot perform real-time tasks, reason about novel visual data, access the internet, perform complex arithmetic perfectly, or make truly independent decisions. They work best within narrow domains with clear objectives.

Can LLMs access the internet?

No. Large language models cannot browse the web, check current events, or retrieve real-time information. They can only work with information included in their training data (with a knowledge cutoff) or provided in your prompt.

Do LLMs understand images?

Some models like GPT-4o and Claude 4.6 Sonnet can analyze images you provide, but they struggle with handwriting, small text, highly abstract graphics, and images requiring real-time context.

Can LLMs do math reliably?

LLMs are weak at complex arithmetic because they tokenize numbers as text. They can handle simple calculations and can follow step-by-step instructions for math, but for high-precision calculations, use calculators or code execution instead.

Can LLMs make real decisions?

LLMs generate plausible-sounding responses based on patterns, not true reasoning or independent judgment. They cannot take moral responsibility for decisions. Always treat LLM outputs as suggestions requiring human review, especially in high-stakes contexts.

What is hallucination in LLMs?

Hallucination is when an LLM confidently invents false information, citations, or facts that were not in its training data or your prompt. It happens because models optimize for generating coherent text, not for truth.

How can I work around LLM limitations?

Combine LLMs with other tools: use APIs for real-time data, integrate code execution for math, add human review layers for high-stakes decisions, and use retrieval-augmented generation (RAG) to reduce hallucination.

Are newer LLMs better at reasoning?

Newer models like Claude 4.6 Sonnet and GPT-4o reason better than older versions, but fundamental limitations remain: they cannot truly access external information, perform novel visual reasoning, or guarantee correctness without external validation.

AI Limitations — What LLMs Can't Do

Coming soon — understanding the hard boundaries of large language models, where they fail, and how to design systems that account for these limitations.

Key Takeaways

LLMs cannot access the internet, perform real-time reasoning, do complex math reliably, or make true independent decisions
Hallucination: LLMs confidently invent facts, especially on niche or recent topics. Combine with RAG or fact-checking for high-stakes use.
LLMs struggle with handwriting, abstract images, small text, and visual reasoning (except GPT-4o and Claude 4.6 Sonnet vision)
Math weakness: LLMs tokenize numbers as text. They can follow step-by-step instructions but fail on novel calculations. Use code execution instead.
No novel reasoning: if a task requires reasoning never seen in training data, LLMs often fail or produce incoherent output
No true judgment: LLMs generate plausible responses based on patterns, not reasoning or responsibility. Require human review for high-stakes decisions.
Cost/latency at scale: running LLMs on large datasets or in real-time systems becomes expensive and slow
Workarounds: combine LLMs with APIs (real-time data), code execution (math), RAG (knowledge grounding), human review (decisions), and vision models (images)