Algorithmic bias Algorithmic bias can both be influenced by Misinformation and disinformation and can be a cause of it.55 The risks of algorithmic bias are heightened when the data used for training an AI model is itself a biased sample. Sometimes, the bias can be obvious. For example, in a hiring process, a set of bios used as examples of good candidates might be drawn from a pool of previous candidates, all of whom might have the same gender, race or nationality. Other times, a bias can be less obvious: for example, a model could be trained on citizens’ previous spending on education, without accounting for certain minority groups typically spending less on education. Synthetic data may be used, aiming to remove bias, but that can itself introduce new biases.56 Examples of biases against citizens include waiting times for a government appointment being assigned on the basis of a questionable set of input data and criteria, or automated responses failing to respond adequately to citizens’ needs. When algorithms are applied to sensitive decisions, biases in training data or assumptions made during model design can perpetuate or exacerbate inequities, further disenfranchising marginalized groups. Predictive policing is one area where algorithmic bias based on race can be a concern.57 Such risks are heightened further when there is no human participation in decision-making. Unless there are clear accountability frameworks in place, the use of automated algorithms makes it challenging to assign responsibility when harmful or erroneous decisions are made, especially when AI is involved. Automated algorithms often operate as “black boxes”, making it difficult for individuals to understand how decisions are made. This lack of transparency and accountability can foster mistrust and skepticism about the fairness and accuracy of decisions taken.In many cases, algorithmic bias can be the result of lack of knowledge, testing or sufficient oversight. How a model is developed, applied and governed is key to mitigating these risks. Independently of the input dataset used, the personal biases of individuals designing the assumptions of the model can also play a role in leading to unjust outcomes. These personal biases may be accidental (for example, the result of those inputting the data having insufficient technical expertise) or intentional, for example, to pursue political aims. One risk that could come into focus more over the next two years is algorithmic bias against people’s political identity.58 Algorithmic political bias might be used intentionally to, for example, affect recruitment into public-sector jobs or access to certain public services or financial services. What makes this risk especially dangerous is that individuals’ political biases are widely known, and those biases can easily find their way into algorithms or data sets. Furthermore, individuals’ political views can increasingly be determined, even against their will, from their online activities.59 Similarly to individual biases, societal biases can also play a role.60 These are likely to become more prevalent as societal divisions deepen. In the GRPS, Societal polarization is ranked #4 over a two-year time horizon. Regionally, Latin America and the Caribbean, Eastern Asia and Europe manifest the most pressing concerns over Societal polarization in the next two years, according to the EOS. Citizen surveillance risks Government technology (GovTech) is entering a new era, as AI, data analytics and digital platforms become the backbone of public administration.61 Technology companies have long worked closely with governments, for example, in the sensitive Mitchell Luo, Unsplash Global Risks Report 2025 36