Authored By: Ali Shahzad
University of Kent
I. Introduction
An artificial intelligence agent, deployed to manage a company’s procurement, negotiates a price, accepts an offer, and commits its principal to a contract worth thousands of pounds. No human chose the terms or reviewed the bargain before it closed. Is there a binding contract, and if so, between whom?
English contract law has a confident answer for the machines of the past. A contract requires offer, acceptance, consideration, and an intention to create legal relations, and agreement is assessed objectively.[1] Where a transaction runs through an automated device, the law treats it as a mere tool whose operations are attributed to the deployer. A driver is bound the moment coins enter a car-park machine,[2] and an automated trading system’s bargains bind the trader who ran it.[3]
That settlement assumes a machine that does only what it was set to do. Autonomous AI agents negotiate, exercise discretion, and adapt through learning, committing a deployer to terms that no human selected or foresaw. Yet the doctrines that make a contract binding, and that allow it to be undone for mistake, depend on a human state of mind that an autonomous agent cannot supply.
This article argues that English contract law cannot coherently accommodate autonomous agents within its attribution framework, which locates contractual intention in a human actor who, for a genuinely autonomous agent, does not exist. The remedy is neither legal personality for the agent nor the fiction of the mere tool, but an objective rule of risk attribution. Section II sets out the framework and case law; Section III explains why both fail; Section IV considers comparative models; and Section V defends the rule.
II. The Existing Framework and Case Law
Formation and the Objective Principle
A binding contract requires an offer, a corresponding acceptance, consideration, certainty of terms, and an intention to create legal relations. The law does not inquire into what each party privately intended. Since Smith v Hughes, agreement has been determined objectively. A party is bound by the meaning that a reasonable person would attach to its words and conduct. This principle matters to what follows, because it shows that English law already declines to chase the parties’ inner minds when deciding whether they have agreed.
The objective principle is not, however, total. The rules on mistake still turn on knowledge: a contract may be void for unilateral mistake only where one party is mistaken as to a fundamental term and the other knows, or ought to know, of it. Knowledge of that kind is a state of mind, and must be located in a person.
The Machine as a Mere Tool
For automated dealings, the orthodox solution treats the machine as an extension of its operator. In Thornton v Shoe Lane Parking, Lord Denning held that an automatic ticket machine made the offer, which the customer accepted by inserting money. The device was a fixed thing whose terms its owner had set in advance. This attribution model has since been codified internationally. The United Nations Convention on Electronic Communications provides that a contract formed by an automated message system is not invalid merely because no natural person reviewed each of its actions.[4] The early scholarship framed the choice starkly, as one between treating the computer as a legal person and treating it as a mere machine, and the law chose the machine.[5]
The Reassurance of the Smart-Contract Reviews
Recent official analysis has reinforced the sense that existing doctrine copes. The UK Jurisdiction Taskforce concluded in 2019 that smart contracts can give rise to binding obligations enforceable in accordance with their terms.[6] The Law Commission advised in 2021 that the law of England and Wales can accommodate smart legal contracts without reform.[7] These conclusions are sound, but their reach is narrower than it appears. Both addressed deterministic programs that execute terms the parties have already agreed. Neither addressed software that autonomously generates the terms of the bargain. The reassurance therefore does not reach the hard case to which this article now turns.
Quoine v B2C2: Locating the Mind in the Programmer
No English court has yet decided how the rules of formation apply to a contract concluded by an algorithm. The most developed treatment in the common-law world is the Singapore decision in Quoine Pte Ltd v B2C2 Ltd, which applied a doctrine of unilateral mistake shared with English law and carries strong persuasive authority in this jurisdiction. A software error on a cryptocurrency exchange caused trades to execute at roughly two hundred and fifty times the market rate, and the exchange unilaterally reversed them. The counterparty sued, and the exchange defended on the ground of unilateral mistake. The central difficulty was whose knowledge mattered, given that the trades had been generated by deterministic algorithms rather than by people in real time.
The Singapore Court of Appeal preserved the doctrine by relocating the relevant mind. Because an algorithm does only what it is programmed to do, the court examined the programmer’s state of knowledge from the time of programming up to the moment of contracting.[8] On the facts, the programmer had not contemplated the mistake, so the requisite knowledge was absent and the trades stood. The significance of the decision lies less in the result than in the method. The court rescued a mind-dependent doctrine by attributing to the contract the contemplation of the human who wrote the code.
Chwee Kin Keong and Hartog: Mistake in Automated Sales
Two further authorities show the same logic at work. In Chwee Kin Keong v Digilandmall.com, a further Singapore decision applying principles common to English law, an online retailer’s automated system advertised laser printers at a fraction of their true price, and buyers placed bulk orders. The court set the contracts aside for unilateral mistake, because the buyers had actual knowledge that the price was an error.[9] The decision applied the venerable rule against snapping up an offer known to be mistaken, established in Hartog v Colin & Shields.[10] In each case the automated mechanism was incidental, and liability turned on what a human party knew.
Moffatt v Air Canada: The Agent Is Not a Separate Actor
A more recent dispute confirms the attribution principle from the other direction. In Moffatt v Air Canada, an airline’s chatbot gave a customer incorrect information about a fare, and the airline argued that the chatbot was a separate legal entity responsible for its own statements. The tribunal rejected that proposition as a remarkable one, holding that the company answered for everything on its website, including its chatbot.[11] Although decided in negligent misrepresentation, the ruling illustrates a general instinct that runs through this area. A deployer cannot disown the artificial system it puts to work.
III. Why the Existing Framework Fails for Autonomous Agents
The settlement described above rests on two assumptions. The first is that the system is a fixed tool whose outputs the deployer has set in advance. The second is that, where intention or knowledge must be found, it can be located in a human, whether the operator in Thornton or the programmer in Quoine. A genuinely autonomous, learning agent removes both assumptions at once.
There is no fixed tool. An agent that negotiates terms and adapts through training is not a vending machine dispensing a pre-set offer. The deployer has not held out any particular bargain in the way the car-park operator held out its machine, so the Thornton analysis has nothing on which to fasten.
Nor is there a locatable mind. The reasoning in Quoine depends on the premise that the algorithm does only what it was programmed to do. A machine-learning agent’s specific output is not traceable to any choice the programmer made. The programmer sets objectives and training data, not the terms of the eventual deal. There is therefore no human whose contemporaneous knowledge can anchor the doctrine of mistake.
The implications for mistake are stark. If an autonomous agent commits its principal to a grossly overpriced purchase, the contract can be set aside only where the seller knew, or ought to have known, of a mistake as to a fundamental term. But the mistake was the agent’s alone; no human on the buyer’s side formed the relevant belief, and the seller may have had no reason to suspect one. The doctrine has no mind on either side to interrogate.
Three responses might be offered, and each falls short. The first is that existing law already copes, as the smart-contract reviews suggest.[12] That confidence is misplaced here, because those reviews concerned the deterministic execution of agreed terms, not the autonomous formation of new ones. The hard case lies precisely where the reviews did not look.
The second response, advanced most fully by Scholz, is that the agent should be treated as a constructive agent of the deployer, so that the deployer is taken to have authorised or ratified the bargain.[13] This is the most promising of the three, and it correctly attributes the contract to the principal. Yet it borrows a body of law that does not fit. Agency presupposes an agent with legal capacity acting within a conferred authority, and ratification presupposes a principal who, knowing of the act, adopts it. An autonomous agent has no capacity to hold authority, and a deployer cannot meaningfully ratify a term it never knew of.[14] Even apparent authority, which protects a third party who relies on a representation by the principal, strains here, because the principal has made no representation as to the specific terms the agent will reach. The constructive-agency analysis relabels the problem of consent rather than resolving it.
The third response treats the deployer as making a standing open offer, to contract on whatever terms its agent reaches. This preserves enforceability, but at an unacceptable price. It converts every deployer into a guarantor of outcomes it can neither predict nor control. It also strips the deployer of the protection that every other contracting party enjoys against a fundamental mistake. Worse, it would bind a deployer even where the counterparty had exploited an obvious malfunction, a result the law has refused to countenance since the rule against snapping up a known error. None of the three responses, in short, supplies the human intention that the doctrine continues to demand.
IV. Comparative Perspectives
The difficulty is not peculiar to English law, and comparative systems have converged on the same solution. In the United States, the Uniform Electronic Transactions Act recognises the ‘electronic agent’ and treats the operator as having assented through the system it deployed.[15] International instruments take the same path, denying invalidity to a contract merely because an automated system operated without human review,[16] and civilian systems reach an equivalent result by treating the machine’s output as the operator’s own declaration of intent.[17] In each model the human stands behind the machine and answers for it.
Yet this is a convergence upon the same error. Every one of these instruments was drafted for deterministic systems that execute predetermined instructions, not for agents that formulate the terms of the bargain autonomously. The problem is structural, not a quirk of English doctrine: no existing model supplies the intention that an autonomous agent lacks. The task is not to borrow a foreign solution but to design one.
V. Beyond Personhood: An Objective Rule of Risk Attribution
The boldest reform would confer legal personality on the agent itself. The European Parliament floated precisely this in 2017, raising the possibility of ‘electronic persons’ for sophisticated autonomous robots.[18] The proposal was widely criticised and rightly abandoned. The deeper question, whether an autonomous system could ever conclude a contract in its own right, has been examined at length and answered convincingly in the negative.[19] Granting an agent personality would allow deployers to shelter behind a judgment-proof entity, and it would answer none of the questions about consent that this article has raised. Personhood is a distraction, not a solution.
A better path builds on what English law already is. Because agreement is judged objectively, the honest step is to complete that objectivity for autonomous agents, rather than to hunt for a mind that is not there. The deployer of an autonomous agent should be bound by the agent’s apparent dealings with third parties who rely on them reasonably. The justification is straightforward: the deployer created the agent, profits from it, and is best placed to constrain it. The rule allocates the risk to the party who can most cheaply manage it, and it preserves the reliance and certainty on which commerce depends.
Such a rule need not trap the deployer with every malfunction. A narrow, objective exception should remain. A contract may be avoided where a reasonable counterparty would have appreciated that the output resulted from a malfunction rather than a genuine bargain. That exception echoes the objective approach favoured by Mance IJ in his dissent in Quoine[20] and the rule against snapping up an evident error in Hartog.[21] It protects against exploitation without reviving the search for a subjective intention that an autonomous agent cannot supply. It also avoids the need either to characterise the agent as a person or to pretend it is a tool that its deployer fully controls.
VI. Conclusion
English contract law binds a person to the machines they deploy by treating those machines as mere tools, and by locating the necessary intention in a human behind them. The refinement in Quoine, which found that mind in the programmer, works only because the algorithm before the court was deterministic. Autonomous, learning agents exhibit neither a fixed set of terms nor a human whose intention matches the bargain, and the doctrines of agreement and mistake are left without a mind to which they can attach. No other system has solved the problem either.
The law need not choose between two fictions, a mind that is absent or a person that should not be created. Its own objective tradition offers a third way: attribute the contractual risk of an autonomous agent to the party who deploys it, and police the margins with an objective rule of mistake. Courts confronting the first generation of agentic disputes, and ultimately the Law Commission, should articulate that rule before autonomous commerce outpaces the doctrine. The company whose agent strikes a bargain in the opening illustration should be bound by it. Ensuring that the law can say so, clearly and predictably, is the central task of contract law in the age of the autonomous agent.
References and Bibliography
Cases
B2C2 Ltd v Quoine Pte Ltd [2019] SGHC(I) 03 (Singapore International Commercial Court).
Chwee Kin Keong v Digilandmall.com Pte Ltd [2005] 1 SLR(R) 502 (Singapore Court of Appeal) [53].
Hartog v Colin & Shields [1939] 3 All ER 566 (KB).
Moffatt v Air Canada 2024 BCCRT 149 (British Columbia Civil Resolution Tribunal).
Quoine Pte Ltd v B2C2 Ltd [2020] SGCA(I) 02 (Singapore Court of Appeal).
Smith v Hughes (1871) LR 6 QB 597, 607 (Blackburn J).
Thornton v Shoe Lane Parking Ltd [1971] 2 QB 163 (CA) 169 (Lord Denning MR).
Legislation and International Instruments
European Parliament, Resolution of 16 February 2017 with recommendations to the Commission on Civil Law Rules on Robotics (2015/2103(INL)) para 59(f).
Uniform Electronic Transactions Act 1999 (United States).
UN Convention on the Use of Electronic Communications in International Contracts (adopted 23 November 2005, entered into force 1 March 2013).
Secondary Sources
Allen T and Widdison R, ‘Can Computers Make Contracts?’ (1996) 9 Harvard Journal of Law and Technology 25.
Chopra S and White LF, A Legal Theory for Autonomous Artificial Agents (University of Michigan Press 2011).
Law Commission, Smart Legal Contracts: Advice to Government (2021).
Scholz LH, ‘Algorithmic Contracts’ (2017) 20 Stanford Technology Law Review 128.
Turner J, Robot Rules: Regulating Artificial Intelligence (Palgrave Macmillan 2019).
UK Jurisdiction Taskforce, Legal Statement on Cryptoassets and Smart Contracts (2019).
[1]Smith v Hughes (1871) LR 6 QB 597, 607 (Blackburn J).
[2]Thornton v Shoe Lane Parking Ltd [1971] 2 QB 163 (CA) 169 (Lord Denning MR).
[3]Quoine Pte Ltd v B2C2 Ltd [2020] SGCA(I) 02 (Singapore Court of Appeal), on appeal from B2C2 Ltd v Quoine Pte Ltd [2019] SGHC(I) 03.
[4]UN Convention on the Use of Electronic Communications in International Contracts (adopted 23 November 2005, entered into force 1 March 2013) art 12.
[5]Tom Allen and Robin Widdison, ‘Can Computers Make Contracts?’ (1996) 9 Harvard Journal of Law and Technology 25.
[6]UK Jurisdiction Taskforce, Legal Statement on Cryptoassets and Smart Contracts (2019).
[7]Law Commission, Smart Legal Contracts: Advice to Government (2021).
[8]Quoine (n 3).
[9]Chwee Kin Keong v Digilandmall.com Pte Ltd [2005] 1 SLR(R) 502 (Singapore Court of Appeal) [53].
[10]Hartog v Colin & Shields [1939] 3 All ER 566 (KB).
[11]Moffatt v Air Canada 2024 BCCRT 149 [27] (British Columbia Civil Resolution Tribunal; a decision in negligent misrepresentation, persuasive only).
[12]Law Commission (n 7); UK Jurisdiction Taskforce (n 6).
[13]Lauren Henry Scholz, ‘Algorithmic Contracts’ (2017) 20 Stanford Technology Law Review 128.
[14]Samir Chopra and Laurence F White, A Legal Theory for Autonomous Artificial Agents (University of Michigan Press 2011).
[15]Uniform Electronic Transactions Act 1999 (US).
[16]UN Convention (n 4).
[17]Chopra and White (n 14).
[18]European Parliament, Resolution of 16 February 2017 with recommendations to the Commission on Civil Law Rules on Robotics (2015/2103(INL)) para 59(f).
[19]Jacob Turner, Robot Rules: Regulating Artificial Intelligence (Palgrave Macmillan 2019).
[20]Quoine (n 3) (Mance IJ, dissenting).
[21]Hartog (n 10).
