Musil studied them in silence a minute. Better days. Sailors having a good time during a port call. Nothing out of the ordinary. Except for the one striking and unavoidable discrepancy. In the shot on the right, the man standing up was in fact Stephen Gelfland. Musil had recognized him at once from Clara Hughes’s smartphone photograph. But he was equally certain the man standing in the left-hand version of the shot was not Gelfland. His clothes were the same in both. His body and posture. Everything the same until Musil looked at his face. Then, he stopped being the same. His face became altogether different, its shape and features clearly matching those of the man shown in the portrait photos.

“One of these posts has been manipulated,” Musil said, pointing to the right-hand photo. “That’s Gelfland over there.” He brought out his own phone and showed him Clara’s picture of her cousin. “You see?”

Ki studied the phone image. “Check,” he said. “How about the guy in the other screens? Who’s he?”

Musil shook his head. “I don’t know.”

Ki looked at him steadily. “What if I told you those photos in the middle are supposed to be Gelfland? The ID shots, the school pic. Every one of them. Plus, the guy standing up in the group shot on the left.”

Musil’s eyes widened. “Did you pull them from the same source?”

“No,” Ki said. “That’s the catch. Different organizations, different databases. Varying levels of network security that go from low to maximal protection. But it looks like somebody hacked all of them to change Gelfland’s real face to a fake. And there’s more.”

“Tell me.”

“The fake faces were created by a generative adversarial network. A GAN. That’s an artificial intelligence that tests itself to improve its results. Like blue-teaming and red-teaming in cybersecurity. They work together. The blue team sets up a computer system. The red team tries to penetrate and expose its flaws. Blue team patches them, red team attacks again. Over and over until the system’s as secure as it can be.” Ki paused. “The GAN does the same thing with its fake faces. It essentially red team-blue teams itself. Puts together a face, then analyzes it with facial-recognition algos. If it identifies any characteristics that might distinguish it from a real face, it reworks them until the fake is perfect. But what’s incredible is that these are the same fake face at a different age. There’s a Fake Gelfland face in every public database and online source that has information about him. One that’s been aged properly in every instance. That takes some kind of special image-simulation technology.”

“But you spotted them as fakes.”