He Spent 17 Months in Jail for a Crime He Didn't Commit. Now Amazon Is Being Sued Over the Face Match.

The probe image was grainy, taken from a distance and from above the suspect. A hood covered the forehead. A medical mask covered the rest of the face. St. Louis area investigators fed the photo into a regional facial recognition system anyway, and Christopher Gatlin spent the next 17 months in jail for a MetroLink assault he did not commit (Biometric Update).

The St. Charles Rock Road station attack on December 7, 2020 left a security guard with a traumatic brain injury. He told police he could not remember the attackers. The submitted image was, in the words of the federal court record, inadequate for reliable identification (federal complaint, E.D. Mo.). The regional system, Mugshot Recognition Technology, searches a database of more than 250,000 mug shots. Its own user manual warned that a high-quality image was required for likely candidates.

Now Amazon Web Services has been added as a defendant in Gatlin's federal civil rights case. An amended complaint filed last Thursday names Amazon Rekognition, AWS's cloud-based facial recognition service, as the identification technology used in the probe (Biometric Update; Bloomberg Law). Gatlin v. Welle et al, Case No. 4:2025cv00043, has expanded from a suit against the police officers and the region to one that also holds a vendor to account for what its tool was asked to do.

Gatlin's lawyers are not arguing that facial recognition as a category is defective. They are arguing that Rekognition was used to identify Gatlin from a probe image the regional system had already flagged as inadequate, and that naming a vendor in that chain creates a duty the vendor cannot disclaim. The complaint's defect theories treat the manual's image-quality warning as a floor: if a facial recognition tool is sold for identification, the vendor is on notice about what its outputs can and cannot support.

The theory lands on an identification chain the suit alleges was broken at multiple points. Investigators had no prior information linking Gatlin before the algorithmic candidate return. When the victim was shown a six-person lineup in August 2021, he initially leaned toward another man. An officer then prompted him to focus on the suspect's complexion and clothing, and the identification shifted to Gatlin (Biometric Update). A Washington Post investigation found that police often ignore identification standards after AI facial recognition matches, a pattern the lineup sequence illustrates.

The ACLU has tracked more than a dozen wrongful arrests tied to police use of facial recognition tools, and frames Gatlin as part of that documented pattern rather than an isolated case (ACLU). The civil rights group's catalog has become the reference set plaintiff lawyers cite for what they call a structural failure of input quality, not a one-off mistake by a single department.

The amended complaint also matters because it tests where the responsibility line falls in biometric procurement. Police departments buy facial recognition as a service. The vendor ships an API, a manual, and a contractual disclaimer. When the input is a hooded, masked, distant frame that the system says is too poor to use, the suit asks whether the vendor has any duty to refuse, warn, or document. The answer will shape how future contracts are written, and how procurement officers price the risk of deploying these tools in criminal-justice contexts.

The amended complaint is recent, the Rekognition-specific defect theories remain allegations rather than adjudicated findings, and AWS has not yet filed its response. The 17-month detention number comes from Gatlin's account and the federal court record, as reported by Biometric Update. The image-quality warning is documented in the user manual for the regional system, not specifically in Rekognition's own documentation as quoted in the complaint.

Watch next: AWS's response to the amended complaint, any motion testing whether vendor duty extends to misuse outside the tool's stated envelope, and whether the court treats the manual's quality warning as a binding specification or an advisory note.