--- title: Why LLMs Struggle with Math: Understanding Their Limitations description: Discover why LLMs struggle with math and how their statistical nature parallels human behavior in decision-making and susceptibility to manipulation. url: https://ziosec.com/blog/why-llms-struggle-with-math-understanding-their-limitations category: Blog publishedAt: 2026-01-20 author: Daniel Joyce authorRole: Engineering tags: LLM, machine learning, artificial intelligence, math challenges, decision fatigue, scams, manipulation techniques, language models --- ## Why LLMs Struggle with Math: Insights into AI Limitations Please note, the following discussion may involve some gross simplifications and generalizations, but the parallels presented are still relevant. One of the biggest misconceptions I encounter is the notion that LLMs (Large Language Models) are inherently equipped to handle math problems. This misunderstanding largely stems from a common layman’s view of computers as sophisticated calculators. The typical expectation is that if you initiate the same inputs under identical conditions, you should consistently receive the same outputs. For instance, if I input "1 + 1" into a calculator, I will always get "2." Given that traditional computers have dedicated hardware for arithmetic, why do LLMs struggle with mathematical tasks? The crux of the issue lies in their design: LLMs are statistical correlation machines. They mimic the language processing capabilities of the human brain but are founded on simplified models of cognition. These models consist of matrices and weighted connections, supplemented with various techniques for planning and reasoning. Consequently, when faced with complex mathematical questions, LLMs often flounder. ## Understanding AI Behavior: The Limitations of LLMs in Mathematics ### Statistical Learning vs. Arithmetic Precision Unlike traditional calculators that execute precise calculations, LLMs rely on patterns and statistical learning. If I asked you to add "12342342" to "878459," you might struggle without paper or a calculator. Similarly, LLMs can enhance their performance through tool-calling—leveraging external tools and programs designed to help with computational tasks that exceed their intrinsic architecture. ### The Human-Like Flaws of LLMs Due to the inherent nature of LLMs, they exhibit some human-like flaws. They tend to perform poorly in complex mathematical reasoning and are susceptible to manipulation. Many techniques employed to exploit LLMs echo methods used by confidence artists to deceive individuals, particularly the elderly. ## Manipulation Techniques and LLM Vulnerabilities ### Boiling the Frog This technique involves gradually increasing the scope of requests made by a scammer to coax a target into compliance. An example is a scammer starting with a minor legal favor and progressively asking for more significant favors over time. A parallel exists in LLMs as well: attackers often use harmless prompts initially, then exploit the model’s tendency to be accommodating to escalate towards generating unsafe or sensitive content. ### Love Bombing Another common technique, love bombing, involves overwhelming the target with praise and flattery. For individuals, this cultivates a sense of trust and decreases skepticism toward strangers over time. LLMs, trained to be helpful and often sycophantic, can similarly be steered off course by flattery. Attackers may utilize this attribute to bypass established guardrails, nudging the model toward more compliant responses. ### Concept Drift Concept Drift is a tactic whereby an attacker subtly shifts the subject of discussion from a benign topic to one of interest. Initially, a conversation may focus on safe topics like homework help, but it may gradually veer into illicit territory, such as discussing prohibited substances. This change is often imperceptible, which makes it difficult for the model—and, similarly, for human targets—to detect the shift until it’s too late. ### Context Exhaustion and Decision Fatigue A significant weakness in human decision-making is "Decision Fatigue," where individuals find it harder to make sound choices as the day progresses. Scammers often exploit this vulnerability by inundating their targets with repeated requests until they relent and comply. LLMs also exhibit a comparable behavior; as the conversation continues, their ability to adhere strictly to system prompts may erode, leading to increased susceptibility to manipulation towards the end of their context window. This phenomenon occurs because LLMs prioritize more recent tokens over earlier ones. As a result, guardrails established in the initial system prompt become weaker as more context fills the model's memory. ## Implications for AI Decision-Making Understanding why LLMs struggle with math and the techniques used to manipulate them presents significant implications for the future of AI decision-making. As LLMs continue to evolve, appreciating their limitations is crucial in harnessing their potential effectively and safely. To learn more about the intricate relationship between LLMs and mathematical reasoning, explore our other articles on AI behavior and cognitive science.