How AI Can Inherit Human Bias

Many people assume that artificial intelligence makes better decisions because it is objective. Algorithms do not have emotions, preferences, or personal agendas. Yet AI can still reproduce many of the same biases that influence human judgment. The bias usually does not exist in the algorithm itself. Instead, bias enters through the training data, the assumptions built into the model, the objectives selected for optimization, and the thresholds that trigger decisions. As a result, AI models can reproduce the same cognitive biases as project managers.

One common example is optimism bias, described extensively by Bent Flyvbjerg in project research. Project teams often underestimate costs and schedules while overestimating the likelihood of success. AI models can make the same mistake if they are trained on unusually successful projects or configured with overly optimistic assumptions. Similarly, anchoring bias can persist when early budgets or schedules remain embedded in forecasting models long after new information suggests they should be revised.

AI can also reinforce uniqueness bias. Organizations sometimes believe their projects are too different to benefit from historical comparisons. When models rely only on internal data and ignore external benchmarks, they reinforce that same belief. Overconfidence bias presents another risk. Project leaders may trust AI recommendations simply because they appear mathematical and objective. Similarly, the model itself may project a level of statistical confidence that the underlying data cannot justify.

Perhaps the greatest risk is that these biases become less visible once they are embedded in an analytical model. Decisions begin to appear objective because they are supported by algorithms, even when the underlying assumptions remain flawed. Effective project governance requires questioning not only the decisions produced by AI, but also the data, assumptions, objectives, and design choices that shape those decisions. AI should strengthen human judgment, not simply automate its biases.