The Algorithm Found a Face. The Detective Still Needed Probable Cause.
A facial-recognition result can look authoritative because a machine produced it. A November 9 federal ruling in Kentucky explains why that appearance cannot carry a warrant. Alain Vizcay Osorio alleges that police arrested him for attempted check fraud after artificial intelligence connected a bank image to his booking photo and a Facebook profile. He says the detective did not corroborate the result, did not disclose the technology's decisive role, and described the process to the warrant court as the use of investigative tools followed by a human comparison. Osorio was arrested in Florida, extradited to Kentucky, and prosecuted before the indictment was dismissed with prejudice after he showed that he was working in Nebraska when the incident occurred. The court has not found those allegations true. It held that, if they are true, the Fourth Amendment claim may proceed. That procedural distinction matters. So does the larger lesson: an algorithm may generate a lead, but probable cause remains a human and constitutional judgment.
The case begins with a lead and ends with an alibi
According to the complaint, a man using the name Ernesto Ramirez tried to cash a suspected fraudulent check at a bank in Florence, Kentucky, in August 2024. The teller declined the transaction, and a security camera captured the man's face. Detective Ryan Kilmer later submitted an image from the footage to facial-recognition technology. The system connected the image to a booking photograph of Osorio and to a Facebook profile under a different name.
The court's account says the detective then signed a criminal complaint stating that he used investigative tools to locate a booking photograph, compared that photograph and a Facebook profile with the bank footage, and observed that the subjects matched. A warrant issued. Osorio, who lived in Nebraska, was arrested in Florida, extradited to Kentucky, and held for more than two weeks before release on bond. The state later dismissed the indictment with prejudice after he showed that he was working in Nebraska at the time of the attempted transaction.
Those are allegations, not established trial facts. Judge S. Chad Meredith was deciding a motion for judgment on the pleadings, so he had to accept well-pleaded factual allegations as true and draw reasonable inferences for Osorio. The detective denies wrongdoing and argued that the complaint failed to allege a constitutional violation and that qualified immunity barred the claims.
The November 9 order permits an amended complaint and allows the federal false-arrest and malicious-prosecution claims to continue. It expressly leaves room for discovery to show that the investigation contained more than Osorio alleges. That is the right characterization: a significant pleading-stage ruling, not a liability judgment and not a final declaration that every facial-recognition-assisted arrest is unconstitutional.
The constitutional problem was the missing middle
The opinion does not treat facial recognition as forbidden technology. Its concern is the inferential gap between a software result and probable cause to seize a person. A search system can return a candidate because two images share measurable features. That does not establish that the candidate is the person in the probe image, much less that the person committed the offense under investigation.
Police policy already recognized that distinction. The Florence policy attached to the complaint described facial recognition as advisory and said it did not establish probable cause without further investigation. Kentucky's statewide model policy is equally clear: a result is not a positive identification, does not establish probable cause to arrest or obtain a search warrant, and must be corroborated through other evidence.
That is not bureaucratic caution for its own sake. Image quality, angle, lighting, database composition, candidate ranking, human confirmation, and the identity information attached to a photograph all affect what the output means. Even a technically accurate similarity calculation cannot answer whether a booking photograph and a Facebook account describe the same person, whether the account name is reliable, or whether independent evidence places that person at the scene.
The alleged investigation skipped that middle work. The complaint says the AI output supplied the only link from the bank footage to Osorio. The rest of the warrant materials described the suspected crime, but not an independent reason to believe Osorio was the man in the bank. That difference is the case. Evidence of an offense and evidence identifying the offender are separate parts of probable cause.
A warrant does not cure a misleading account of the evidence
Osorio was arrested under a warrant, which ordinarily provides a powerful defense to a false-arrest claim. But a warrant does not end the inquiry when it allegedly rests on material falsehoods or omissions made knowingly or recklessly. The court asked a practical question: after correcting the description of the investigation and disclosing the omitted facts, would the remaining affidavit still establish probable cause?
The wording mattered. Saying that a detective used investigative tools and then observed a match suggests a broader process and an exercise of independent judgment. Saying that facial-recognition software produced the only connection, contrary to a policy requiring corroboration, presents a materially different application. Judge Meredith concluded there was good reason to think the warrant would not have issued if the reviewing court had received that fuller account.
This is a disclosure problem as much as a reliability problem. Judges cannot evaluate the weight of technology-assisted evidence if the application collapses the software, the database search, and the officer's review into a vague phrase. The government does not have to write a technical manual into every affidavit. It does have to reveal the facts that make an identification more or less trustworthy when that identification supplies the bridge to probable cause.
Lawyers challenging or defending technology-assisted warrants should therefore resist the temptation to argue only about a vendor's general accuracy. The first question is narrower: what exactly did the tool return, what did a human do next, what independent facts corroborated the candidate, and how was that chain described to the judicial officer who issued the warrant?
New technology does not always require new constitutional precedent
The detective argued that Osorio could not identify an earlier case specifically holding that exclusive reliance on facial recognition violates the Fourth Amendment. That argument goes to qualified immunity, which generally protects an officer unless existing law made the alleged violation clear.
The court rejected the need for a facial-recognition twin. Existing Sixth Circuit law already said an officer may not obtain a warrant through reckless material falsehoods or omissions and that probable cause cannot rest on evidence too unreliable or too untested. Judge Meredith also treated the allegations as the rare obvious case. His point was not that facial recognition is literally random. It was that a new tool does not create a constitutional blank slate when the old rule already requires corroboration and honest disclosure.
The opinion used a Magic 8 Ball hypothetical to make that point. The analogy is intentionally sharp, perhaps sharper than the technology deserves. Facial recognition can be useful and often performs far better than chance. But usefulness does not answer the legal question. The Constitution assigns the probable-cause judgment to accountable officials and ultimately to a court. A machine's novelty cannot supply immunity for conduct that would be obviously unlawful if the same unsupported identification came from any other untested source.
That reasoning will matter beyond facial recognition. Courts will repeatedly confront claims involving AI-generated leads without precedent naming the precise model, dataset, or workflow. The sensible doctrinal move is to identify the established legal function the system entered. If the tool screens, predicts, identifies, or scores, the underlying duty to investigate, disclose, authenticate, or exercise judgment does not disappear merely because software now occupies one step.
The evidence file must preserve the chain, not just the answer
For criminal defense and civil-rights lawyers, discovery should reach the probe image, every returned candidate, confidence or similarity information, database sources, examiner notes, comparison procedures, audit logs, communications with the vendor, training records, policy versions, and the warrant-drafting history. A screenshot of the final candidate is not the system's full evidentiary record. Neither is the officer's conclusion that two faces matched.
For prosecutors and agencies, the operational lesson is equally concrete. Treat the result as a lead. Record the tool and version used. Preserve the original media. Require a documented human comparison. Seek independent corroboration that actually connects the person to the event. Describe the technology's role accurately in warrant materials. And ensure that an officer can explain why the evidence establishes probable cause without asking the court to defer to a proprietary output.
For judges, the reliability inquiry may begin before any expert battle. Was the system's output a ranked candidate list or a claimed identification? Did a trained examiner conduct a blind or structured review? What facts beyond facial similarity connect the candidate to the offense? Was exculpatory information checked? The answers determine whether the technology is one clue among many or the whole case disguised as one clue.
Osorio v. Kilmer is early. The amended allegations may fail after discovery, the parties may settle, or later courts may narrow the reasoning. But the order is worth reading now because it identifies the correct unit of analysis. The legal system should not ask whether AI is accurate in the abstract. It should ask how a particular output became evidence, how that evidence became a sworn statement, and how the statement became state power over a person. The algorithm can begin that chain. It cannot be allowed to hide it.
Sources and further reading
Primary and industry sources used to support this page. External guidance should be reviewed in context and for your jurisdiction.
- Osorio v. Kilmer, memorandum opinion and orderThe November 9, 2026 ruling granting leave to amend, denying judgment on the pleadings, and rejecting qualified immunity based on the allegations at this stage.
- Osorio v. Kilmer federal docketDocket information for No. 2:26-cv-00004 in the Eastern District of Kentucky, including the complaint, policy exhibit, and challenged motion.
- Kentucky Facial Recognition Model PolicyOfficial model policy stating that a facial-recognition result is an investigative lead, not a positive identification or independent probable cause.
- Kentucky Revised Statutes, Chapter 61The state's facial-recognition governance provisions, including KRS 61.9305.
- Reuters, man wrongly jailed after AI mismatch can pursue lawsuit for nowNovember 10 reporting on the ruling, the parties' positions, and the growing Fourth Amendment litigation over AI-generated evidence.