4 Limitations of GenAI

Non-exclusive list of GenAI’s limitations

Factual inaccuracies in the output

Large language models are probabilistic text predictors, not fact databases. They frequently generate plausible-sounding false statements, fabricate historical events, or invent non-existent citations and links.

• Using GenAI to request academic sources on a research topic, only to receive nicely formatted citations for books and journal articles that do not actually exist.
• Asking for a step-by-step resolution for a calculus problem, where the model makes a confident mathematical error in step two but arrives at an arbitrary answer.

Lack of common sense and real-world understanding:

GenAI models don’t understand the world in the human sense and can produce illogical or nonsensical results through its statistical patterns of probability. It lacks human intuition and knowledge that humans understand naturally like spatial relations, cause and effect, and every day logic.

• Prompting GenAI for a chemistry lab procedure and receiving instructions that sound grammatically correct but suggest unsafe or impossible physical steps (e.g., assuming glass cannot break on concrete).
• Asking a chat tool for public transit routes, and receiving out-of-date or completely fabricated schedule times.
• Requesting an interior design layout and receiving a visual rendering with physically impossible dimensions, such as furniture scaled incorrectly for human use.
• Asking GenAI for advice on a minor dorm room mess and receiving an absurd suggestion like using a hairdryer to dry a flooded carpet, due to a lack of physical context.
• A systematic review on GenAI tools used in mental health education revealed significant deficiencies in contextual reasoning, accuracy, cultural sensitivity, and emotional engagement.

Data bias and representation gaps

Generative AI outputs reflect the patterns, prejudices, and systemic gaps present in their training datasets. Performance and accuracy are inherently limited by the quality and diversity of that data, which can perpetuate harmful stereotypes or present monocultural perspectives as universal facts.

• Asking GenAI to summarize a historical political conflict, and receiving a response that adopts a heavily Western-centric perspective while marginalizing non-Western primary accounts.
• Prompting an image generator to create “a successful CEO” or “a leading surgeon,” resulting in images that exclusively depict light-skinned, middle-aged men unless explicitly instructed otherwise.

• One study on AI teaching assistants indicated that models can interact with students differently based on racial stereotypes embedded within their training data.

Lack of social skills and context

Because GenAI is not human, it cannot think, feel, or process real-world emotional intelligence. It cannot “read the room,” understand unspoken group dynamics, or interpret sarcasm, cultural norms, and human emotional nuance.

• Using GenAI to draft a message during a tense group project disagreement, resulting in a cold, robotic response that unintentionally escalates the conflict.
• Relying on a chatbot for personal guidance during a stressful period, only to receive generic, clinical responses that lack true human empathy.