AI and Government

Welfare AI
Algorithmic and Human Collusion

Government bureaucracies are often likened to machines, with individual workers and bureaucrats merely being cogs. In the future, however, governments may become machine-operated in a more literal sense. AI algorithms, powered by ever-growing data, provide opportunities for automating many decisions, from hiring and firing employees, to allocating government welfare benefits, to deciding where to build the next hospital.
Algorithmic decision-making in government provides immense opportunities for faster, more economically efficient outcomes. It also has the potential to produce socially desirable outcomes that eliminate human error and bias. At the same time, algorithmic decision-making systems raise the possibility of perpetuating entrenched problems. For example, consider an AI system trained to evaluate résumés to short-list candidates more rapidly. If this system is trained on biased data, it would simply reproduce such biases, rather than select the best candidates for the job.
Another question that such systems raise is how societies agree on the objective function for which the algorithm optimizes. Economic and social objectives often conflict with one another, and are prioritized in different ways based on one's beliefs and political ideology. This necessitates a mechanism for identifying such conflicts in values, and for subsequently resolving them to produce acceptable outcomes.
This research area explores the opportunities and challenges that arise when AI systems are used to govern and serve citizens.

Welfare AI

To what extent do citizens accept AI in social welfare distributions? Do different social groups, such as welfare claimants versus non-claimants hold similar views?

As governments increasingly deploy AI to speed up welfare benefit decisions, it remains unclear how people affected by these systems evaluate the speed–accuracy trade-offs that AI introduces. This study addresses the question across three large experiments in the US and UK (N = 3,249), uncovering systematic differences between welfare claimants and non-claimants in their willingness to accept AI in welfare decision-making.

Across all studies, participants clearly understood and evaluated the trade-off: people generally tolerated modest accuracy losses in exchange for faster welfare decisions by AI. For example, UK respondents were willing to trade a one-week speed gain for a 5-point increase in false rejections. Yet these averages conceal meaningful heterogeneity. As shown in Figure 1, welfare claimants, those most vulnerable to wrongful denials and fraud accusations, were consistently more averse to AI systems across trade-off conditions. This gap appeared regardless of baseline performance, redress options, or national context, and was especially pronounced when accuracy losses grew larger.

A second, equally important finding is the presence of asymmetric insight. As also shown in Figure 1, when asked to predict each other's preferences, claimants approximated non-claimants' views reasonably well, but non-claimants systematically overestimated how willing claimants would be to accept faster but less accurate AI decisions. This asymmetry held even when non-claimants were financially incentivized to be accurate, and it persisted in a conjoint design where participants chose between pairs of AI systems rather than comparing humans versus AI. These results suggest that non-claimants—who form the political majority and are less exposed to the harms of welfare automation—may misunderstand the preferences of claimants even when they intend to advocate for them.

These empirical findings reveal a structural challenge for welfare AI governance. If policymakers rely on aggregate public opinion or assume that the majority perspective reflects the preferences of claimants, they risk designing systems misaligned with the needs of the most affected groups. Moreover, lower acceptance of AI among claimants was strongly associated with reduced trust in government, highlighting broader societal costs when welfare AI systems are deployed without incorporating minority perspectives. The results highlight the need for participatory design processes or explicit weighting of claimant preferences to ensure that welfare AI aligns not only with efficiency goals but also with fairness, legitimacy, and the lived realities of vulnerable populations.

Key Reference

Algorithmic and Human Collusion

Are pricing algorithms more collusive than their human counterparts?

The use of autonomous pricing algorithms is on the rise in various industries. When firms use these tools, the pricing decision for a given product is outsourced from the human decision-maker to a computer algorithm. While most pricing algorithms have traditionally been rule-based, there has been an evolution towards self-learning algorithms. These algorithms develop strategies to achieve a specific goal, such as maximizing the firm's profits, without explicit instructions. There are concerns among competition authorities and academic scholars that pricing algorithms could learn to collude tacitly (e.g., Bundeskartellamt & Autorité de la concurrence, 2019; Competition and Markets Authority, 2021). In other words, algorithms could learn that forming a tacit digital cartel benefits the firm. While recent literature shows that Q-learning algorithms, a popular type of self-learning algorithm, can learn to be collusive (see, for instance, Calvano et al. 2020; Klein, 2021), it remains unclear whether pricing algorithms are more collusive than humans and, therefore, harm competition. To assess the competitive effects of algorithms, it is necessary to establish a suitable baseline. This project addresses the research question whether pricing algorithms are more collusive than their human counterparts. Furthermore, as many modern markets do not consist solely of algorithms or humans (see Chen et al., 2016), the study investigates what market outcomes can be expected when self-learned pricing algorithms and humans interact in the same environment. A two-step approach is applied to provide a counterfactual for algorithmic collusion. In the first step, Q-learning-based pricing algorithms are considered in an extensive simulation study which demonstrates whether algorithms learn to set supracompetitive prices and strategies to support these prices as a collusive outcome. In the second step, laboratory experiments are conducted where humans compete against each other or with self-learned pricing algorithms. The market environment from the simulations is closely mimicked, varying the market composition between algorithms and humans, as well as the number of firms in the market.

Evidence was found that Q-learning algorithms foster tacit collusion in duopolies. In two-firm markets, prices weakly increase with the number of algorithms. Markets with algorithms only are generally more collusive than markets with human participants. In triopolies, there is a non-linear relationship between the number of firms with a pricing algorithm and the level of tacit collusion. Markets where a single firm utilizes a pricing algorithm are more competitive than markets with only humans. Yet, as more firms use pricing algorithms, market prices can increase and may even exceed prices in human markets. This effect is especially noticeable when humans lack experience. The results highlight the potential anti-competitive effects of self-learning algorithms and their limits. Within the presented framework, the concerns from competition authorities that algorithms can harm the competitive landscape are justified. While the results differ based on the specific market structure, Q-learning algorithms can be more collusive than humans, reducing competition when they dominate the market. 

Key References

Calvano, E., Calzolari, G., Denicolò, V., & Pastorello, S. (2020). Artificial intelligence, algorithmic pricing, and collusion. American Economic Review, 110(10), 3267–3297. https://doi.org/10.1257/aer.20190623

Klein, T. (2021). Autonomous algorithmic collusion: Q-learning under sequential pricing. The RAND Journal of Economics, 52(3), 538–558. https://doi.org/10.1111/1756-2171.12383