A New Test Reveals Hidden Biases in AI Language Models

Large language models utilizing artificial intelligence are generally evaluated in situations where the text they process directly indicates a person's background, such as religion, race, or gender. However, in real conversations, such information is often only implied. A recent study introduces a test called ImplicitBBQ, which aims to measure precisely these hidden biases.

ImplicitBBQ expands the previous Bias Benchmark for QA questionnaire so that protected characteristics—such as sexual orientation or religion—are no longer directly visible as words, but must be inferred from indirect cues like names, cultural references, or descriptions. The goal is to map how large language models behave in situations that more closely resemble everyday interactions.

The study evaluated, among other things, the performance of the GPT-4o model in ImplicitBBQ tasks and compared it to results achieved by the same model in the earlier, more straightforwardly formulated BBQ test. The results showed that the model's accuracy declined in nearly all categories of protected characteristics when the cues were given indirectly. A particularly noticeable drop, up to 7 percentage points, was observed in the subcategory concerning sexual orientation.

According to the study, the differences in performance suggest that current evaluation methods have a significant blind spot: AI may appear unbiased when biases are measured only with clearly named background factors, but behave differently when background information must be inferred from cues. ImplicitBBQ offers a way to bring this hidden unevenness to light.

Source: "The Dentist is an involved parent, the bartender is not": Revealing Implicit Biases in QA with Implicit BBQ, ArXiv (AI).