A semi-dystopian scene unfolded shortly before 4 a.m. on September 3, 2026, when a driverless taxi in San Francisco pulled to the curb on its own accord. Waymo told the Los Angeles Times it stopped the car after detecting a “violation of our terms of service involving a firearm” and alerted emergency services.
Police conducted what they called a “high-risk vehicle stop,” detained two young passengers, and reported finding a loaded AR-style rifle, plus suspected marijuana and mace spray. Per SFGate, both were placed into juvenile detention.
Few people would dispute that teenagers roaming with a loaded rifle at 4 a.m. pose a public safety problem. But the episode raises privacy questions about robotaxi rides. An interior camera recorded the ride, a remote operations team could halt the car and coordinate with responders, and company policy barred weapons. Together, they let Waymo identify a violation, stop the car, call police, and generate data officers could later seek.
I. Private Eyes, Public Power
The Fourth Amendment restrains the government, not private actors like Waymo, as the Supreme Court settled over a century ago in Burdeau v. McDowell. Thus, when a private party searches or records on its own initiative, the Fourth Amendment does not apply, however intrusive the conduct may be.
The Court extended the private actor principle in United States v. Jacobsen. FedEx employees opened a damaged package, found white powder, and called DEA agents. The agents’ visual inspection of the opened packages was no search because it did not exceed what the private party had exposed. Waymo’s monitoring and reporting fit that rule.
This doctrine has an important limit. A private party becomes a government actor when it serves as the state’s instrument or agent, as in Skinner v. Railway Labor Executives’ Association, where federal rules encouraged and shaped drug testing by private railroads.
As of this publication, there is no public information that suggests police asked Waymo to watch riders for guns. The analysis changes if they direct Waymo to watch a rider, stream a cabin feed, or prevent someone from leaving.
II. You Agreed to This, in a Sense
Waymo account holders accept terms of service that almost nobody reads, and not every passenger is an account holder. Guests ride on someone else’s account, and Waymo’s rules bar minors from riding alone in San Francisco.
Whatever their contractual status, Waymo discloses to passengers that its cars record cabin video, that employees may access live video, and that data may be shared with law enforcement. Under the third-party doctrine, people have no reasonable expectation of privacy in what they voluntarily convey to a business. But cabin video is an awkward fit. A rider does not convey every gesture and conversation in the same way a bank customer conveys transactions or a caller conveys dialed numbers. Waymo’s notice that the cabin is recorded therefore does more work than the classic third-party doctrine in defeating a reasonable expectation of privacy.
Yet Carpenter v. United States refused to apply that doctrine to cell-site records, while cautioning that it did not call conventional security cameras into question. In June 2026, Chatrie v. United States held that police conduct a search when they obtain stored Google Location History and rejected the argument that a user surrendered Fourth Amendment protection by merely authorizing Google to collect and retain the data. That reasoning is important for detailed robotaxi trip histories. Whether it extends to discrete cabin footage is less certain, though it makes a categorical appeal to the third-party doctrine difficult. These cases do not stop Waymo from reporting what it sees, but they can complicate later warrantless government acquisition of stored data.
The California Constitution does something the Fourth Amendment does not by protecting privacy against private entities. But a plaintiff must establish a legally protected privacy interest, a reasonable expectation of privacy, and a serious invasion. Waymo’s disclosures would complicate the latter two questions. Suppression of evidence, however, is another matter. Exclusion is tied to federal requirements, so evidence from private conduct in California cannot be suppressed on state constitutional grounds alone.
III. Why the Stop Was Probably Lawful
The Fourth Amendment arrives when the police do. Officers who surround a stopped car and order its occupants out have seized them, which requires at least reasonable suspicion. Whether an informant’s report supplies it depends on the source’s reliability and basis of knowledge.
In Florida v. J.L., a unanimous Court held that an anonymous report of a man with a gun could not alone justify a frisk. In Navarette v. California, a 911 report of a reckless driver sufficed, in a case the Court called close, because the caller claimed eyewitness knowledge and used a traceable system.
There is a threshold question: Could the teenagers challenge the vehicle search at all? Under Rakas v. Illinois, mere passenger status does not establish a reasonable expectation of privacy in the areas searched. Prosecutors could argue that Waymo’s retained control and disclosed monitoring defeat that expectation here. But Byrd v. United States rejected the premise that an unlisted rental-car driver necessarily lacks privacy rights. Lawful possession and control matter, and a contractual violation alone does not extinguish protection.
Applying those principles to robotaxi passengers remains uncertain. If the teenagers lacked a protected interest in the area searched, they could not challenge that search for lack of probable cause. That would not, however, authorize unrestricted searches of their persons or protected belongings, or eliminate the justification required to detain them.
If the teenagers had a reasonable expectation of privacy in the area searched, the probable-cause question would remain. Here, Waymo’s report bore little resemblance to the anonymous tip in J.L. It came from an identified corporate source and pinpointed the car, its location, and the suspected violation. If Waymo reported a human observation of what appeared to be a firearm, that likely supplied probable cause to search the car under the automobile exception given California’s restrictions on minors possessing firearms. But that conclusion turns on what Waymo told police beforehand, not on what officers found. A reported human observation of a gun may establish probable cause, but a bare automated alert, without more, presents a harder question.
The sequence is worth noticing. Waymo decided what conduct violated its rules, whether to end the ride, when to call police, and where the car would stop. A company’s terms of service became the predicate for a felony stop of two minors. The Constitution does not forbid that. It does not even notice it.
IV. When the Informant May Be an Algorithm
One consequential point remains unknown. Waymo has not said whether a classifier—software trained to recognize objects in an image—flagged the rifle or whether an employee recognized it while watching the cabin feed.
The leading informant cases were built around human sources in warrant cases. They asked how informants know what they claim to know, and why anyone should believe them. Illinois v. Gates loosened that into a totality-of-the-circumstances test, where a weak answer to one question can be made up by a strong answer to the other. Alabama v. White subsequently applied a similar framework to a street stop. California law preserves the same basic distinction. In People v. Ramey, a case known for its arrest-warrant holding, the California Supreme Court treated an identified citizen source as presumptively credible yet still required underlying facts showing how the source knew a crime had occurred.
A named company like Waymo is less likely to fabricate a report than an unknown caller. But that addresses only veracity, not accuracy. A human observer can describe what appeared on the screen, from what angle, and with what confidence. An automated alert cannot. Its accuracy depends on training data, operating conditions, and error rates the public cannot evaluate.
Courts have a familiar reliability filter, but it principally governs evidence offered at trial. In federal court, Federal Rule of Evidence 702 and Daubert require judges to assess the reliability of expert methods before the jury hears them, considering factors that can include a method’s known or potential error rate. California’s approach under People v. Kelly follows the Frye test, asking whether a new scientific technique is generally accepted in the relevant field. Whether a commercial classifier would survive scrutiny under Kelly is a separate question.
Regardless, a classifier can influence a police search without ever being offered as evidence at trial. Prosecutors could rely on the rifle without introducing the alert that led police to it. The classifier’s reliability would instead matter at a suppression hearing, where trial admissibility rules do not necessarily govern the probable-cause inquiry, and the informant doctrine imposes no standardized validation requirement on a commercial classifier.
Validation nevertheless matters. U.C. Berkeley Law Professor Andrea Roth has shown why studies may prove less than they appear to prove: Systems are often tested under ideal conditions and deployed under very different ones. A vehicle cabin at 4 a.m. is the second kind. The Federal Trade Commission’s allegations concerning Evolv offer a cautionary example. Its scanners allegedly flagged laptops, binders, and water bottles while missing a seven-inch knife later used to stab a student. At one school, increasing the system’s sensitivity produced false-alarm rates as high as 50 percent.
Those allegations say nothing about Waymo’s system. But traditional informant doctrine has no standardized mechanism to evaluate the accuracy of a classifier behind a corporate report. Waymo’s identity and reputation may support the conclusion that the company was not lying. They do not establish that the person or system interpreting the camera feed was accurate.
V. Who Decides When You Can Get Out
It is unclear whether the two teens could have left the car before officers arrived. A July incident in San Mateo raised this question after Waymo halted a ride involving two teenagers drinking and firing a toy gun. Contemporary reports said Waymo told the teenagers the car was experiencing mechanical trouble while police approached. Police later said the doors were not locked and the teenagers could have exited.
Waymo’s first responder protocol says company personnel can unlock doors remotely to assist responders. It does not establish that Waymo can lock passengers in or defeat an interior release. Stopping a vehicle is not the same as detaining its occupants.
That distinction matters. If riders can open the doors and leave, Waymo’s act of stopping the car, standing alone, is private conduct, not a Fourth Amendment seizure. If the company held the passengers at police direction, the case for government action and a Fourth Amendment seizure is far stronger. If Waymo confined riders without lawful privilege, separate state-law questions of false imprisonment would arise instead.
VI. What the Constitution Will Not Do, Regulators Must
The stop almost certainly took a dangerous weapon out of two teenagers’ hands. However, one good outcome is a poor reason to leave the rules unwritten for the next case. The private-search doctrine generally exempts Waymo’s independent monitoring. The third-party doctrine may still favor access to some business records, although Carpenter and Chatrie limit its reach. And the informant doctrine may treat an identified corporate source as credible without requiring standardized validation of an automated system behind the report. These doctrines regulate different stages of the encounter; the robotaxi integrates them into one technological system.
Regulators can close the gaps. California already regulates in-vehicle camera recordings, including some third-party sharing and law-enforcement access, and the Public Utilities Commission already requires operational reporting from robotaxi services. But that statute is built for car buyers. It defines an “in-vehicle camera” as a device the manufacturer includes in the vehicle, places its duties on manufacturers and dealers, and protects a “user,” defined as a person who originally purchases, leases, or takes ownership. Consent runs through forms signed at purchase and prompts shown during initial vehicle setup. A robotaxi passenger is none of those things. And neither regime tells riders how often monitoring produces a police referral.
Those gaps call for rules aimed at robotaxi surveillance and law enforcement. The Commission can require operators to report police referrals, data requests, and whether each began with a person or a machine. Companies can publish escalation policies explaining which conduct ends a ride, which triggers a police call, and who decides. Legislatures can cap video retention, require in-car notice when live monitoring begins, and require legal process for nonemergency access to stored footage. At bottom, any automated weapons-detection system used to summon police should face independent testing.
None of those safeguards would have changed the result on September 3. They matter when the system misreads what the camera records, when the monitor acts hastily, or when conduct is lawful but breaks a corporate rule. Someone besides the company should be keeping count.
Copyright © 2026 Henry F. Fradella. Henry F. Fradella is a Professor in and Director of the School of Interdisciplinary Forensics at Arizona State University, where he is also a Professor in the School of Criminology and Criminal Justice and an Affiliate Professor of Law in the Sandra Day O’Connor College of Law. He holds a B.A. in psychology from Clark University, a Master of Forensic Sciences and a J.D. from The George Washington University, and a Ph.D. in interdisciplinary justice studies from Arizona State University. Dr. Fradella researches substantive, procedural, and evidentiary criminal law; the dynamics of legal decision-making; and the consequences of changes in legal processes. He is the author, co-author, or editor of 15 books and more than 140 scholarly articles, reviews, and commentaries. The author thanks the staff of the California Law Review for their excellent work editing this essay.
