The Representation Gap

The Evidence Chain, Part 3 of 10. Part 2 dealt with actions a human approved. This part deals with something stranger and more common: actions a human never approved one by one, because they asked the machine to decide.

In August 2026, a United States appeals court looked at an AI agent shopping on Amazon and had to answer a question that sounds like a philosophy seminar but decided a lawsuit: who was actually doing it? Its answer was you. “It is the user who ‘accesses’ Amazon’s computers,” the Ninth Circuit wrote, “with the help of the Assistant,” and however advanced the agent is, “it is a tool, not a person for statutory purposes” (Amazon.com Services, LLC v. Perplexity AI, Inc., 2026, p. 15). On that reasoning, the court vacated a preliminary injunction that would have kept Perplexity’s Comet browser agent off Amazon.com.

Read that as if it were about you, because it is. If you hand an agent your shopping, your travel, your rental search or your job applications, the law’s starting point is that what it does is something you did. That is often exactly what you want: it means a platform cannot treat you as a trespasser for using a tool of your choice. But it also means that every small choice the agent made on the way, which seller, which fare, which wording, which trade-off, arrives on your account as if you had made it yourself. Who actually made those choices, and can anyone later tell? That is the representation gap.

What the court saw inside the tool

The opinion’s description of how the agent works is worth slowing down for. When a Comet user asks the Assistant to find something on Amazon, it takes screenshots of the browser view, sends them from the user’s computer to Perplexity’s servers and receives instructions back from those servers on how to navigate the site; in the court’s words, it “cannot operate wholly independently; it relies on direction from the user and instructions from Perplexity’s servers” (Amazon.com Services v. Perplexity, 2026, p. 7). Amazon had told Perplexity before Comet’s launch that its AI products would not be permitted on the store, and much of the dispute turned on Perplexity’s decision not to send a user-agent string that would identify the Assistant as an AI agent and let Amazon block it; the parties disputed whether Perplexity knowingly altered that identifier after Amazon first succeeded in blocking it (pp. 7–8 & n.1).

The court decided a narrow question, whether Perplexity “accessed” Amazon’s computers under the federal Computer Fraud and Abuse Act and California’s equivalent, and it was careful to say what it was not deciding. It did not establish “a new legal regime governing agentic AI,” and it did not address whether, in other contexts including tort claims, Perplexity could avoid liability for the Assistant’s actions (p. 17). It also noted a consequence of Amazon’s theory that should matter to anyone who uses these tools: treating the agent’s activity as unauthorized access could expose users themselves to criminal liability for facilitating it (p. 16). The court protected users from that. Good. Amazon then asked the full Ninth Circuit to rehear the case, and in September no judge on the court even requested a vote (CourtListener, n.d.), so the ruling stands. But look at the shape of the arrangement it described: one tool, two sources of direction, and a legal rule that assigns the result to only one of them.

The bet the law made, and keeps making

This is not a one-off. Electronic-commerce law made a deliberate bet a quarter of a century ago. The Uniform Electronic Transactions Act, the basis of e-commerce law in almost every US state, says a contract can be formed by the interaction of electronic agents even if no individual reviewed their actions (Uniform Law Commission, 1999, §14), and a United Nations convention said the same internationally in 2005 (United Nations, 2005, art. 12). In 2024 the bet was renewed for the age of AI. The UNCITRAL Model Law on Automated Contracting, which the UN General Assembly recommended to all states in December 2024, provides that, absent an agreed procedure, an action carried out by an automated system “is attributed to the person who uses the system for that purpose,” and that attribution cannot be denied on the sole ground that the outcome was unexpected (United Nations Commission on International Trade Law [UNCITRAL], 2024, art. 7(2)–(3)). Its optional unexpected-actions article lets a party escape an action only when it could not reasonably have expected it and the other side knew or should have known that (art. 8).

That bet made sense for machines that executed rules people had chosen in advance. An AI agent breaks the chain in the middle. You give it an objective, it interprets the objective, it makes a series of discretionary choices and then it acts. The legal rules attach the whole sequence to “the person who uses the system.” But the Comet record shows that the person using the system and the party steering it can be different, and that the steering happens on servers the user never sees.

Who decides, who can intervene, who pays

Put the power questions to the case. Who decides what the agent does at each step? Partly you, through your request, and partly the provider, whose servers send the navigation instructions. Who can intervene? The platform tried to, by blocking; the court noted that Amazon remains free to regulate access through its terms of service (p. 21 n.5); the provider can change its instructions at will; and the user, in practice, mostly watches. Who carries the consequences? Under the attribution rules, the user. Even Amazon’s own evidence gestured at the tension: its declarations claimed the Assistant “may not select the best price, delivery method, or product recommendations for a customer” (p. 19). The court found that harm too abstract to support an injunction, which is fair as a matter of injunction law. As a matter of representation, it is the whole question. When an agent picks a worse option for you, was that your choice?

This is a global question, not an American one. The UNCITRAL model law is designed to be copied into national legislation everywhere, and its attribution default will shape how people in Lagos, Jakarta and São Paulo are bound by agents built in a handful of other countries. It is also a question about the people who will delegate the most, for the longest. If you are young enough that an agent may handle your first lease, your first loan application and your first job search, you are going to live inside these attribution rules for decades. You deserve to know, now, that the law’s default treats the agent’s choices as yours, and you are entitled to ask the companies building these tools to prove which choices were.

The prediction, and what would change my mind

My prediction is that the tool theory will hold for access and come under strain for choice. Courts and legislators will keep saying the user is the one acting when the question is whether a platform can exclude them, and will increasingly have to ask whose decision it was when an agent’s discretionary choice causes a loss, especially where the agent’s provider has commercial interests in the outcome. The research question for this part is not whether agents should act for people; they already do. It is whether anyone can later show which of an agent’s choices were genuinely the principal’s, which were the machine’s and which were the provider’s. If disputes over agents’ discretionary choices turn out to be rare, or easily settled by ordinary attribution, I am wrong about the strain. The Ninth Circuit itself said the legal understanding of agentic AI “will doubtless change” (p. 17). I agree, and I would rather the records exist before it does.

For Mercury Intelligence members: the section below sets out a delegation record that separates instruction, inference, discretion and provider steering, walks the Comet case through it, maps the clauses from UETA to the new UNCITRAL model law and the AI Act, and gives you a representation audit you can run this month.

---

Support independent investigation. The companies building agents, the platforms they act on and the law firms arguing over them all have a stake in how “who decided?” gets answered, and very few people are reading the court records on behalf of the person whose name ends up on the order. That is the work your support pays for. Give once ($5, $10, $25, $50 or any amount), give monthly ($10 or $25), or become a patron; organizations can underwrite the work, and anyone can give without money by sharing this piece or telling us about an agent that decided something for you. All five doors are at fairbydesign.org/patrons-2/.

References

Amazon.com Services, LLC v. Perplexity AI, Inc., No. 26-1444 (9th Cir. Aug. 4, 2026). https://cdn.ca9.uscourts.gov/datastore/opinions/2026/08/04/26-1444.pdf

American Law Institute. (2006). Restatement (Third) of Agency. American Law Institute Publishers. https://www.ali.org/publications/restatement-law-third/agency

Cloud Security Alliance. (2026, August 1). EU AI Act’s high-risk deadline: Deferred, not cancelled [Research note]. https://labs.cloudsecurityalliance.org/research/csa-research-note-eu-ai-act-high-risk-deadline-omnibus-20260/

CourtListener. (n.d.). Amazon.com Services, LLC v. Perplexity AI, Inc., No. 26-1444 (9th Cir.) [Docket: petition for rehearing en banc, Aug. 18, 2026; order denying rehearing en banc, Sept. 10, 2026]. Free Law Project. Retrieved October 7, 2026, from https://www.courtlistener.com/docket/72502129/amazoncom-services-llc-v-perplexity-ai-inc/

Electronic Signatures in Global and National Commerce Act, 15 U.S.C. § 7001 (2000). https://www.law.cornell.edu/uscode/text/15/7001

Kolt, N. (2026). Governing AI agents. Notre Dame Law Review, 101(1), 335–386. https://ndlawreview.org/governing-ai-agents/

Moffatt v. Air Canada, 2024 BCCRT 149. https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html

Parikh, S., & Surapaneni, R. (2025, September 16). Powering AI commerce with the new Agent Payments Protocol (AP2). Google Cloud Blog. https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol

Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data (General Data Protection Regulation). (2016). Official Journal of the European Union, L 119, 1–88. http://data.europa.eu/eli/reg/2016/679/oj

Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). (2024). Official Journal of the European Union, L, 2024/1689. http://data.europa.eu/eli/reg/2024/1689/oj

Turpin, M., Michael, J., Perez, E., & Bowman, S. R. (2023). Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023). https://proceedings.neurips.cc/paper_files/paper/2023/hash/ed3fea9033a80fea1376299fa7863f4a-Abstract-Conference.html

Uniform Law Commission. (1999). Uniform Electronic Transactions Act. https://www.uniformlaws.org/committees/community-home?CommunityKey=2c04b76c-2b7d-4399-977e-d5876ba7e034

United Nations. (2005). United Nations Convention on the Use of Electronic Communications in International Contracts. UNCITRAL. https://uncitral.un.org/en/texts/ecommerce/conventions/electronic_communications

United Nations Commission on International Trade Law. (2024). UNCITRAL Model Law on Automated Contracting (with Guide to Enactment). United Nations. https://uncitral.un.org/en/mlac