Reality has become difficult to verify. On September 4, 2026, Martin Anderson reflected on a koala video that turned a familiar question into a much harder one: can anyone trust what they see?
The video showed a static, tight shot of a mother koala with her two sleeping offspring. A young tabby cat then entered the frame, creating the sort of strange but tender moment that invites an immediate emotional response.
The mother koala reached over and swooped up the young cat, then settled it beside her offspring. It was an absurdly appealing scene, which is also what made it such a useful test for visual judgment.
Anderson and his wife declared that the video was likely to be AI-generated. Their conclusion did not come from a casual suspicion. Anderson has provided deep and intensive coverage of AI-generated image and video systems for seven years, including discussion and creation of deepfake-detection systems.
That experience should offer an advantage when a video raises questions about its authenticity. Instead, the koala clip exposed a more uncomfortable conclusion: expertise no longer guarantees certainty.
The Last Weakness in AI Video
Anderson identifies inconsistency as the biggest remaining bottleneck for AI video synthesis. Generated videos can still reveal themselves through details that fail to remain stable, coherent, or consistent from one moment to the next.
Those flaws once offered viewers a useful escape route. A strange movement, a shifting detail, or a visual mismatch could suggest that an apparently real scene came from an AI system. The koala video made that safety net feel less dependable, even for someone who has spent seven years examining these systems.
Anderson put the realization this way: “whether or not that particular video was fake or real is immaterial; what struck me was that I had finally arrived at a long-awaited, long-expected epiphany – that, notwithstanding the fact that I have provided deep and intensive coverage of AI-generated image and video systems (as well as discussing and creating deepfake-detection systems) for seven years now, even I cannot trust what I see any longer.”
The specific answer matters less than the loss of confidence behind the question. Whether the koala video was fake or real, it reached the point where the distinction could not be settled by simply watching it and trusting trained judgment.
Short Videos, Perfectly Served
The problem becomes more serious when AI video meets short-form content. AI can perfectly and exactly fill our shrunken attention spans for short-form content, giving viewers just enough time to react before they have time to inspect what they watched.
A static and tight shot helps that process. The frame limits what viewers can examine, while the koala, sleeping offspring, and young tabby cat provide an instantly readable sequence: vulnerability, surprise, then apparent acceptance. The scene needs no explanation, and that is part of its power.
The video also shows why inconsistency remains such an important bottleneck. If a generated scene contains no obvious break in continuity during the few seconds that matter, viewers may never reach the moment when its construction becomes clear.
That leaves a strange standard for visual truth. A video does not need to survive deep inspection if it can hold together long enough to produce a reaction.
Anderson’s reflection is not a verdict on one koala clip. It is a measure of how far AI-generated image and video systems have moved, and how little certainty remains when even a specialist cannot trust what he sees.
