AI Grouped 13 Children by Face in Their Church Livestreams

The media coordinator found the problem in a spreadsheet.

She had exported it from the congregation’s video dashboard in September 2024 while checking an archive of 63 weekly streams. Each row represented a face the system believed belonged to one recurring person. The spreadsheet recorded a label, the number of appearances and a link to a public page containing matching clips.

There were 47 rows. Thirteen belonged to children.

One child’s full name sat beside 31 appearances. The earliest came from a service more than a year before, when the child had crossed behind a speaker for a few seconds. Later matches included a holiday program, a reading and several moments in the front rows. Anyone using the archive’s search bar could enter the name and move through those appearances.

The congregation had never planned to build a directory of children. It had subscribed to the video service for $74 a month because a small volunteer team was struggling to prepare each recording. The software generated captions, divided long services into chapters and made transcripts searchable. Homebound members used those features.

So did people with hearing loss and parents trying to find the part of a service in which their child had participated.

Face grouping came with the same package. At first it looked like another organizing shortcut.

A useful archive becomes a people index

For 14 months, volunteers uploaded a weekly livestream and reviewed the automated captions. The system detected faces in selected video frames, compared them with faces found in other recordings and gathered likely matches into clusters. A volunteer could then give a cluster a name to locate future clips more easily.

The software did not learn a child’s name merely by seeing the child’s face. In this case, people supplied the names. One volunteer labeled a cluster while preparing a clip from a youth program, and the system applied that label to earlier and later appearances it considered a match.

That distinction mattered, but it did not reduce the exposure. A person had identified one appearance. The machine connected it to 30 more.

Without face grouping, finding those moments would have required watching hours of video, recognizing the child each time and recording every link. The AI performed that linking in the background, including during services where the child was not speaking and had no reason to expect that crossing the camera’s view would create another entry in a personal index.

A parent noticed after typing a child’s name into the archive search. The resulting page showed thumbnails from multiple months, including ordinary attendance rather than public participation. One result had also appeared through a general web search, which meant the people pages were not confined to members who already knew how the archive worked.

The parent had previously asked schools and community groups not to publish identifiable images of the family. The reason involved a difficult custody arrangement. Most volunteers did not know that history, and the parent had not expected to disclose it in order to attend worship.

The spreadsheet made the scale visible. Of the 13 minors, five had full names attached to face clusters. The rest had partial labels or descriptions entered by volunteers who wanted to organize program footage. Two vulnerable adults also appeared in the index, including one person who had asked not to be shown online.

The congregation knew the public pages had existed and that at least one had reached a general search result. It could not establish whether anyone had downloaded the clips, saved the thumbnails or copied the associations between names and attendance. Removing a label would change the live archive. It would not retrieve copies already made.

What the machine recognized

Face grouping systems usually turn detected faces into numerical representations called embeddings. The numbers encode visual relationships that help software compare one face image with another. They are not a photograph, and they do not contain a name, but they can support a judgment that two images are similar enough to place in the same cluster.

The threshold matters. If it is too strict, one person may be split into several clusters because lighting, camera angle or changes in appearance make the images look less alike. If it is too loose, different people may be merged.

Both failures appeared in the 47-row spreadsheet.

One child had two separate rows, one from the main camera and another from lower-resolution program footage. Two siblings had been merged under a single name in several clips. The software displayed its groupings as settled results, although the cluster boundary was only a probability-based decision made across imperfect frames.

The errors created a second kind of risk. A wrong match could place one child’s name on another child’s image, or suggest that a person had attended a service where they had not been present. Volunteers had treated the feature as an editing aid, so they had corrected obvious mistakes only when they needed a clip. Nobody had reviewed the people index as a public record.

The tool also could not reliably decide who was a minor. Volunteers discussed keeping adult labels while suppressing children, but estimating age from a face would have introduced another uncertain inference. A teenager might be classified as an adult. A young-looking adult could be treated as a child.

The congregation would still need knowledge from outside the video to make the distinction.

This was why the problem could not be reduced to a careless upload. A conventional folder of weekly recordings would still carry privacy risks, but it would not automatically connect a brief appearance in one service with dozens of moments across the next year. The system converted scattered footage into a record of recurrence.

That record could reveal more than a face. Repeated appearances showed who regularly attended, which family members sat together and when someone stopped coming. Spoken announcements and searchable transcripts supplied more context. The software did not infer a family dispute or a person’s vulnerability, yet it arranged ordinary signals so another person could make those inferences quickly.

The argument was about belonging

The volunteers agreed to hide the people pages while they reviewed the archive. Agreement ended there.

Several members wanted all recordings involving children removed. Others worried that a broad retreat would cut off people who depended on captions or could no longer attend in person. One member had used transcript search every week to revisit readings without scrubbing through an entire service. A family that had first encountered the congregation online said the archive had made an unfamiliar community easier to approach.

The accessibility gains were real. They did not require the public face index.

The media coordinator demonstrated this by processing a new recording with automated captions and chapters while leaving face grouping unpublished. The result still needed human review because names and religious terms were sometimes transcribed incorrectly, but the useful parts remained. The congregation could separate features that had arrived bundled together.

The harder dispute concerned the camera itself. A public livestream inevitably captured people who entered its view, even if no face labels were shown. Some members believed signs near the sanctuary entrance provided enough notice. Parents answered that notice of filming was not the same as notice that software would track recurring faces across 14 months.

That difference changed relationships inside the congregation. Media volunteers had thought of themselves as helping people participate. Parents now saw that the same work had created records they could not inspect without knowing where to look, while volunteers felt blamed for activating a feature presented as part of routine video organization.

No one had intended surveillance. Intent did not alter the spreadsheet.

The congregation kept livestreaming but changed the frame so fewer attendees appeared. It retained automated captions, added a delay for review and stopped publishing named face clusters. Older people pages were removed from the public archive over six weeks. The volunteers did not attempt to replace them with automated blurring because the available tool sometimes missed profiles and partially covered faces.

The original spreadsheet remained the review document. The coordinator added a column recording whether each public people page had been removed, then checked the links again after general search results began disappearing. The rows for children stayed visible to the small review group until each associated label had been cleared.

Some members still wanted the entire archive taken down. Others believed the revised system preserved enough of what had made the streams valuable. The congregation had not settled whether old services should remain online indefinitely, and parents did not agree on one consent standard.

Questions people ask

Can face grouping identify a child by name?

Face grouping can connect similar faces across recordings without knowing a name. In this congregation, a volunteer named one cluster, and the system carried that label across every clip it considered a match. The machine supplied the scale and linkage; the human-supplied label made the cluster searchable by identity.

Why was this different from posting an ordinary livestream?

An ordinary stream can expose anyone visible on camera, but finding one person across months of footage takes substantial effort. The automated index reduced that effort to a name search and assembled recurring appearances on one page, including moments when the person was only passing through the frame.

Did the congregation have to give up automated captions?

No. The volunteers found that captions and chapter suggestions could remain active while the public people index was disabled. Caption errors still needed review, especially around names, but members who relied on searchable transcripts continued using them after the face labels were removed.

What happened to the people index?

The public pages were removed over six weeks, while the congregation kept a restricted record of the cleanup because it did not know whether copies existed elsewhere. At the next archive review, the coordinator opened the same 47-row spreadsheet and found the public-link column blank.