Artificial intelligence promises to transform many aspects of human life by automating complex tasks, processing vast amounts of data and personalising services to individual needs. But alongside the benefits that may accrue to everything from medicine to education, AI also causes anxieties. One particular area of concern is around how algorithms function and what unintended consequences they may bring.
In many everyday markets, algorithms already set prices: ride-hailing platforms, online retail and airline tickets are just a few examples. A common worry is that pricing algorithms might push prices above competitive levels, even without explicit communication between providers, thereby harming consumers in the process. Could something similar to this ‘supra-competitive pricing’ happen in financial markets, where mispricing can distort economic decisions?
Supra-competitive pricing is not an intrinsic feature of AI
Our research shows that in financial markets, supra-competitive pricing is not an intrinsic feature of AI. Instead, outcomes depend on how algorithms are designed and on the competitive pressures faced by the financial institutions that are deploying them (Guarino et al, 2025).
We study algorithmic market-makers (AMMs), programmes that set bid and ask prices in asset markets (the prices at which an institution is willing to buy and sell a given asset, and where the difference between them is known as the spread). We focus on a widely used reinforcement-learning method for AI – Q-learning – in which algorithms experiment with different quotes, learn from trading outcomes and update their strategies over time.
Using both simulations and analytical results, we find that poorly specified learning rules can generate undesirable outcomes: in some configurations, AMMs converge to ‘loss-free’ prices at which trade disappears; in others, they produce wide spreads and supra-competitive prices.
But crucially, these outcomes arise from limited learning and limited use of market information, not from anything inherent in algorithms or AI. When AMMs are endowed with even minimal market understanding – such as recognising that a trader willing to buy at one price would also have bought at any lower price – the same learning framework instead produces active undercutting, tighter spreads and convergence towards the competitive benchmark.
Moreover, competitive pressures across financial institutions push market-makers to employ richer data and better-designed learning rules, further reinforcing competitive pricing.
Algorithm design is decisive
Our results show that algorithm design is decisive:
- Naive Q-learning leads to loss-free pricing. Ask quotes are set too high and bid quotes too low. This protects algorithms from losses, but drives informed traders out of the market, leaving only uninformed trading. Prices then fail to incorporate information about the asset’s fundamental value, undermining market efficiency.
- Tuning learning and exploration parameters reduces extremes, but it can still sustain wide spreads and supra-competitive prices. Liquidity deteriorates and traders face worse terms than under competitive market-making.
- Adding counterfactual updating changes the outcome. It uses the logic that accepting a price implies acceptance of any better price. With this minimal market insight, AMMs learn to undercut rivals and prices converge to the competitive levels predicted by theory.
The key message of our research is therefore not that ‘anything can happen’, but that non-competitive outcomes are neither natural nor inevitable. They emerge when institutions deploy algorithms that are poorly aligned with basic market logic or informational structure. In competitive environments, market-makers have strong incentives to implement designs and data usage that drive pricing towards competitive benchmarks.
Concerns about AI pricing making markets less competitive may be overstated
There is growing concern that AI pricing can make markets less competitive. In goods markets, several studies find that algorithms can set and maintain supra-competitive prices (for example, Calvano et al, 2020).
A similar concern has been raised for financial markets (for example, Colliard et al, 2022). The underlying fear is of what’s known as tacit collusion between providers: prices remain above competitive levels not through explicit communication, but because algorithms independently ‘learn’ that higher prices can be sustained.
Our findings suggest that in financial markets, these concerns may be overstated. It is true that reinforcement-learning algorithms can fail to learn competitive pricing if they are naively designed or deprived of relevant market information. But competition changes the incentives: market-makers that price poorly lose order flow, and those that deploy algorithms with a better grasp of market structure, and that use richer information, gain order flow. In our setting, once algorithms incorporate even minimal market logic, competitive forces lead them to undercut and converge to efficient prices.
For regulators and policy-makers, the implication is clear: the main risk does not come from AI in the abstract, but from how it is implemented. Oversight should focus on whether trading algorithms embed sufficient market structure, how they are trained and evaluated, and whether data usage could sustain non-competitive outcomes.
The appropriate regulatory response is therefore not to presume that AI will inflate prices, but to develop standards that encourage designs supporting liquidity, transparency and robust competition.
Risks of AI stem from how it is coded, trained and embedded in real trading environments
Ultimately, whether AI makes prices more or less competitive in practice is an empirical question. Our research demonstrates that different algorithmic designs can generate very different outcomes.
But it also delivers a sharper conclusion: there is nothing inherent in AI that produces supra-competitive prices in financial market-making. When firms face competition, they have incentives to deploy learning rules and information sets that deliver competitive pricing, because doing so is essential to winning order flow and sustaining profitable participation.
For regulators, the challenge is to move beyond broad concerns about AI and instead scrutinise the specific design choices, incentives and feedback processes that shape algorithmic behaviour. The risks posed by AI pricing stem less from the technology itself than from how it is coded, trained and embedded in real trading environments.
