Reviewing HCP meeting photos for signs of manipulation

In many life sciences companies, a photo taken at a meeting with healthcare professionals is part of the record that the meeting took place as reported: that it happened on that date, at that venue, with those attendees. It supports the meal or venue expense, the attendee list behind it and, in many markets, the transparency reporting that follows. When the photo is false, all of that is false with it.

Meeting photos are harder to review than receipts. There are no sums to check and no card transaction for the picture itself, and a staged photo of a real room with real people is, pixel for pixel, genuine. This guide covers how meeting photos get faked, a checklist that works regardless of the technique, and where image analysis helps.

Why the stakes are high

Regulators look closely at meetings and speaker programs involving healthcare professionals. The HHS Office of Inspector General's November 2020 Special Fraud Alert on speaker programs listed, among its warning signs, attendees with no legitimate business reason to be there and professionals who attend programs on the same topic again and again. Photos are one of the few records that show who was actually in the room, which is why a fabricated or reused one is a serious finding and not a paperwork error.

Five ways meeting photos are faked

A real photo from another event

The most common and the simplest: a genuine photo from an earlier meeting, a different team's event, or the internet, submitted for a meeting that was smaller, different or did not happen. Every pixel is real, so no image analysis will flag it.

What catches it: comparison with photos already submitted (the same room, the same faces, the same angle), a reverse image search for anything that might be public, and details that date the scene: seasonal clothing, holiday decorations, a slide or screen showing a date, a menu that changed.

Attendees pasted in

A person is cut from another photo and added to the scene, to raise the head count or to put a particular professional at the table.

What catches it: light that falls from a different direction on the added person, a missing or wrong shadow on the chair or table, a halo or overly clean edge around hair and shoulders, a head size that does not fit the person's distance from the camera, and a person who is sharper, softer or grainier than the people next to them. The tamper check is designed to outline regions like these, and the error level view can show the pasted area with a different block pattern.

A generated or AI-extended scene

An AI model generates the whole meeting, or fills empty seats and extends a small gathering into a larger one.

What catches it: the AI check, first of all. By eye, look at the places generators still struggle with: text on slides, name tents and signage that is garbled or nearly right, hands and fingers, repeated or near-identical faces in the background, place settings that do not match from seat to seat, and reflections in glasses and windows that do not show the room.

A face swapped

The body and the scene are real, but the face of one person is replaced with another's. Face swaps are not covered by VeriSight's checks. By eye, look for a boundary around the face where skin tone, grain or sharpness changes, light on the face that does not match the light on the neck and hands, and teeth or earrings that blur. If you suspect a face swap, escalate to specialist analysis rather than relying on a clean result.

A real photo of a staged scene

Colleagues pose as attendees, the photo is taken before the guests arrive, or the same room is used for a meeting that took place on another day. The image is authentic in every technical sense.

What catches it: only corroboration. The sign-in sheet, the venue's bill, the calendar and the people themselves.

A checklist for every meeting photo

The five techniques need different detection, but the same checklist surfaces most of them. Work through it before you upload anything.

• Head count. Count the people in the photo and compare with the sign-in sheet and with the number of covers on the venue's itemized bill. All three should agree, or the difference should have an explanation.
• Identities. Are the listed professionals visible and recognizable? Are name tents legible and do they match the attendee list? Is anyone present who should not be, such as family members or colleagues listed as attendees?
• Place. Can you identify the venue? Does it match the invoice, and does it look like a place where an educational meeting could be held?
• Time. Does daylight in the windows fit the reported time? Do clothing and decorations fit the season?
• More than one photo. Ask for several photos from the event. Real events produce sequences: the same people in the same clothes from different angles, with plates and glasses that change over the evening. Fakes tend to be single images, and inconsistencies between photos of one event are hard to fake.
• The original files. Photos straight from the phone keep their capture time and image detail. Ask for the originals, not a collage, a screenshot or a photo pasted into a report template.

What image analysis can and cannot tell you

Of the four views in the visual review, error level analysis is the most useful on scene photos. The ink channel views and the ink color map are designed for printed text; on a photo of a room they mostly show the color of the light. That can still help when a pasted person was photographed under different light from everyone else, but read them as a supporting view here, not a primary one.

A high AI score on a meeting photo also needs the same caution as on any other image: beautification filters and built-in AI photo features are common on group photos and can raise the score without any intent to deceive. Our guide to reading the AI score covers how to separate the two.

Writing it up

A finding on a meeting photo is strongest when it ties the image to the other records. “The photo shows six people; the sign-in sheet lists eleven attendees and the venue billed eleven covers; the same photo was submitted for a meeting on March 14” needs no detection score to be convincing. Use image analysis to find where to look and to support what the records show, and say which checks you ran on which file.