Even in cases decided thirty years apart, there remains a necessary volitional nexus between human conduct and machine action for there to be liability.

The Ninth Circuit recently confronted a distinctly modern question in Amazon.com Services, LLC v. Perplexity AI, Inc., No. 26-1444 (9th Cir. Aug. 4, 2026). Amazon objected to the use of Perplexity’s AI-powered browser to interact with Amazon’s website on behalf of users. In addressing Amazon’s claims under the Computer Fraud and Abuse Act (18 U.S.C. § 1030) and a similar California statute, the court treated the human user as the person accessing Amazon, with Perplexity’s technology functioning merely as the tool through which that access occurred. As a result, Perplexity avoided liability. The decision is narrow, but the action-attribution question behind it is not.

Perplexity’s technology does more than simply display webpages. Its agent can interpret what appears on a page, determine what steps are necessary to accomplish a user’s request, and then take a sequence of actions on the user’s behalf. The user supplies the objective and sends the agent off to accomplish it, while the software determines much of what happens in between. That makes the relationship between human conduct and machine action more complicated than it is with a conventional browser.

But it may also make one part of the attribution analysis relatively clear. There is a meaningful connection between the user’s volitional act of directing the agent and the resulting activity on Amazon’s computers. By “volitional nexus,” I mean the connection between a person’s purposeful conduct and the particular act carried out by the machine that gives rise to legal consequences. The agent may choose the means, but it is pursuing an objective that the user supplied.

That way of looking at the problem has an older analogue. In Religious Technology Center v. Netcom On-Line Communication Services, Inc., 907 F. Supp. 1361 (N.D. Cal. 1995), a user posted allegedly infringing material to a Usenet newsgroup. The operation of the network caused copies of the material to be made automatically on computer systems operated by intermediaries, including Netcom. The copyright owners argued that those automatic copies made the intermediary operators direct infringers.

Following the volition

Netcom became important for its focus on volitional conduct. Netcom operated the equipment that made the copies, but there was no sufficiently meaningful connection between Netcom’s own volitional conduct and the particular copying that gave rise to the claim of direct infringement. Its computers were responding automatically to conduct initiated by the user. The court therefore looked beyond the immediate operation of the machine and asked whose conduct had actually set the infringing event in motion.

That helps bring Perplexity into greater relief. In Netcom, the connection between the intermediary’s volitional conduct and the liability-producing copies was weak, while the connection to the user who posted the material was much stronger. In Perplexity, the connection between the user’s volitional conduct and the resulting computer access is more direct still because the user expressly dispatches the agent to accomplish a particular objective. The technology separating the human instruction from the resulting act is far more sophisticated, but the underlying attributional instinct looks remarkably similar.

The same problem became harder as internet services assumed a larger role in producing automated results. In Viacom International, Inc. v. YouTube, Inc., 676 F.3d 19 (2d Cir. 2012), the parties disputed whether YouTube’s automated processing and display of user-uploaded videos involved sufficient volitional conduct to support direct infringement. The Second Circuit ultimately focused on the DMCA safe harbor and did not resolve the direct-infringement dispute on volition grounds. Still, questions about volitional conduct and automated processing became harder as online platforms did more than merely transmit what users supplied.

The Supreme Court’s decision in American Broadcasting Companies, Inc. v. Aereo, Inc., 573 U.S. 431 (2014), complicated the picture further. Aereo’s system operated automatically after a subscriber selected a television program, yet the Court treated Aereo itself as performing the copyrighted works. The decision showed that automation alone does not determine where legally significant conduct will be located. Increasingly elaborate technological systems can make the connection between human volition and machine action harder to trace.

Whose act is it?

Agentic AI makes that tracing exercise more important. An AI agent can examine changing circumstances, select among different courses of action, and undertake intermediate steps that the user never specifically directed. At the same time, its activity may remain tightly connected to the objective supplied by the person who sent it out to act. The relevant question may therefore be less whether the machine acted and more how closely the machine-produced event can be connected to the volitional conduct of the person whose legal responsibility is at issue.

That question needs more than a label; it needs factors. Read together, Netcom, Aereo, and now Perplexity suggest at least three things courts are implicitly weighing: (1) how specifically the human defined the objective versus left the means open-ended; (2) whether the intermediary’s system was general-purpose or built to produce the specific outcome that triggers liability; and (3) how much of the resulting event tracks ordinary, foreseeable use rather than a path the intermediary itself engineered.

That third factor is what actually distinguishes Aereo from Netcom, and it matters because Aereo is the case that seems to cut against a pure “who set it in motion” test. A subscriber’s button-press was just as proximate a trigger as the user’s post in Netcom, yet the Court still found Aereo, not the subscriber, to be performing the copyrighted work. The distinction is that Aereo’s entire system was architected around a specific liability-triggering act, the individualized retransmission of copyrighted broadcasts, dressed up in a way designed to look like private, user-initiated copying rather than public performance. Netcom’s copying, by contrast, was a generic, protocol-level side effect of running a newsgroup server, not a function built toward the infringing outcome.

So the deeper principle Aereo adds to Netcom isn’t that automation defeats the volitional-conduct inquiry; it’s that a system purpose-built to produce the liability-triggering result can pull the operator back into the frame even when a human nominally pressed the button first.

Run the Ninth Circuit’s actual Perplexity opinion through that framework and it lands closer to Netcom than to Aereo. The court’s own description of the mechanics matters: the Assistant “cannot operate wholly independently,” it depends on both user direction and server-side instruction, and, critically, Perplexity’s servers “never directly access Amazon’s servers” at all; only the user’s browser does, with the Assistant relaying screenshots and receiving guidance. That is a system general enough to be pointed at any website, not one architected specifically to defeat Amazon’s access controls.

The one fact that could have flipped the analysis, Perplexity allegedly altering the Assistant’s user-agent string after Amazon began blocking it, is exactly the kind of purpose-built, evasion-specific design choice that would edge the case toward Aereo‘s territory. Tellingly, the court flagged that dispute in a footnote and declined to resolve it, leaving open on remand the very question the framework highlights, namely, not whether the Assistant acted, but whether Perplexity built it, at least in part, to defeat the specific authorization boundary Amazon had drawn.

So the variable to watch in the next case isn’t “how autonomous is the agent,” which is a one-way ratchet that will only ever increase. It’s whether the agent’s design targets the specific restriction, protection, or outcome that gives rise to liability, or whether it remains a general tool that happens to be capable of reaching that outcome incidentally. (It is too late in this piece to meaningfully compare this to the doctrine articulated in Sony Corp. of America v. Universal City Studios, Inc., 464 U.S. 417 (1984) that a purveyor is not liable for copyright infringement if its product is sold for legitimate purposes and has substantial non-infringing uses, but the connection is there if you would like to follow it on your own.)

An agent that negotiates around a paywall a user couldn’t otherwise access, spoofs credentials, or is tuned to specifically defeat a known blocking mechanism looks like Aereo. An agent that a user directs toward an ordinary task, using generic capabilities that happen to touch a restricted resource along the way, looks like Netcom, and, on the facts as the Ninth Circuit found them, like Perplexity.