An overfit model performs extremely well on training data but fails to make accurate predictions on real world data. This generally happens when models are too complex or niche and fail to understand the noise that will most likely come from real world data. Thus, the model has a hard time making generalizations across new data.

In many cases, there isn’t enough “new data” for AI models to train on. Some models, like LLMs, require massive amounts of training data, and data collection efforts can’t always keep up with the rapid societal, political, and economic changes we see month to month.

This occurs when AI systems are influenced by pre-existing beliefs. If a model builder unconsciously disregards data that doesn’t align with their perceptions, the model will likely be unable to find new patterns and instead reinforce the pre-existing beliefs.

This occurs when computer systems produce biased results that are discriminatory or unjust.

This occurs when the data used to train an AI system is not representative of the application's true intention. An example is building an AI model for car loan approvals but only collecting data for the top 1% of earners. The data should be just as diverse as the population, not a subset.

This is a form of bias where annotation groups unconsciously inject their own biases into the labeled data, resulting in various groups being discriminated against.

This occurs when certain groups of features are systematically excluded from the data used to train the AI model (e.g., an LLM that excludes certain languages).

This occurs when the context in which data was collected differs from the context in which the model is applied. An example of this would be a UK AI company developing and training a model for the US market with British English and terminology as the core of its training dataset. This might not translate very well to US lingo, terminology, and regional differences.

This is a common type of bias where individuals are grouped based on a few characteristics that lead to stereotyping (e.g., generalizing data based on age, race, income, location, etc.).