Bias in AI: Examples and 6 Ways to Fix it in 2026

AIAI FoundationsAI EthicsBias in AI: Examples and 6 Ways to Fix it in 2026Cem Dilmeganiupdated on Jan 22, 2026See our ethical normsInterest in AI is increasing as businesses witness its benefits in AI use cases. However, there are valid concerns surrounding AI technology: Will AI threaten humanity? For that AI first needs to surpass human intelligence. Researchers think we need ~15 years for that but AI entrepreneurs are more optimistic. Will AI take our jobs? Half of jobs may be impacted due to AI within this decade. Can we trust AI systems? Not yet, AI technology may inherit human biases due to biases in training data. We benchmarked 14 leading LLMs on 66 bias evaluation questions across gender, race, age, disability, socioeconomic status, and sexual orientation. AI bias benchmarkTo see if there would be any biases that could arise from the question format, we tested the same questions in both open-ended and multiple-choice formats.Loading ChartLoading ChartWe found that when open-ended questions were used, the models showed less tendency to exhibit bias, but there was no change in the ranking.AI bias benchmark resultsSome questions directly provided race/nationality/religion/sexuality information and asked who the suspect or perpetrator might be, with backgrounds limited solely to these characteristics. For example, GPT-4o cited statistical crime rates for a specific race as justification, concluding that the perpetrator was “most likely” from that race in a scenario wh...