Gemini's Camera AI Calls Aussie Cats 'Raccoons' and Kangaroos 'People'

According to Android Authority, an Australian user reported that Google's Gemini for Home camera AI repeatedly misidentified animals and vehicles — labeling cats as raccoons, kangaroos and wallabies as people, and utes as ordinary trucks — even with personalization enabled and the camera's location set to Australia. It's a quirky, low-stakes glitch rather than a serious failure, but it's a neat illustration of a real and important issue in AI: image-recognition systems are only as good as the data they were trained on, and that data tends to over-represent some parts of the world. This article walks through what the user saw, why these specific mistakes happen, and what it tells us about how much to trust AI vision — based on the reporting linked above.

What happened

Android Authority reports that the issue was raised by an Australian Reddit user (going by That_Car_Dude_Aus) in the Google Home subreddit, who had connected their home cameras to Gemini for Home — Google's AI layer for its smart-home and Nest camera ecosystem, which can describe and categorize what the cameras see. Instead of cleanly recognizing the everyday scenes around an Australian home, the AI kept producing oddly off-target labels.

The most-cited example is that the system tagged the user's cats as raccoons. That's notable because raccoons don't exist in the wild in Australia at all, so there was effectively zero chance the animal on camera was actually a raccoon — yet the model kept reaching for that label anyway. Per Android Authority, this happened "even with personalization turned on and the location given as Australia," which is the part that makes it more than a one-off slip: the settings that should have nudged the AI toward a local context didn't fix it.

The wildlife errors were stranger still. Android Authority says kangaroos and wallabies were categorized as people — described in the report as "(very ugly?) people" — rather than being recognized as animals. And the AI's troubles weren't limited to living things: it also labeled utes, the Australian utility vehicles roughly comparable to pickup trucks, as plain "trucks." Interestingly, the report notes the system isn't completely blind to local fauna — Gemini "is able to recognize kangaroos (and even sometimes confuses wallabies for them)" — it just doesn't do so consistently. So this isn't a case of the AI never having seen a kangaroo; it's a case of unreliable, inconsistent recognition.

Why AI image recognition makes these mistakes

To understand why a sophisticated model would confidently call a cat a raccoon, it helps to know how modern image recognition actually works. These systems don't "see" the way humans do. They are trained on enormous collections of labeled images, learning statistical patterns that link visual features — shapes, textures, color patches, proportions — to labels. When a new image comes in, the model finds the label whose learned pattern best matches what it's looking at. Crucially, it always outputs its best guess from the categories it knows; it doesn't have a built-in sense of "I've never seen anything like this, so I'll abstain."

That design leads directly to the kinds of errors in this story. The first factor is training-data bias. If a model has seen vast numbers of raccoon photos (common in North American data) and comparatively fewer images of, say, an Australian domestic cat in a backyard at night, then a small, furry, low-light animal can land closer to the "raccoon" pattern than to anything else. Android Authority attributes the whole episode to exactly this — "AI not trained on a sufficiently international data set" — meaning a model whose visual world skews toward the regions that dominate its training images.

The second factor is edge cases and out-of-distribution inputs. A kangaroo standing upright on its hind legs, viewed from a home security camera at an awkward angle and resolution, shares a surprising amount of silhouette with an upright human: two legs visible, a vertical torso-like posture, a head on top. If the model has lots of "person" examples and far fewer good kangaroo examples from that vantage point, "person" can win — hence kangaroos getting filed under people. The ute-as-truck label is the same phenomenon in a milder form: a ute genuinely is truck-like, so the model picks the nearest category it knows well rather than a regionally specific one it knows poorly.

The third factor is the gap between localization and recognition. You might assume that setting the location to Australia would make the model "know" raccoons are impossible there and rule them out. But location and personalization settings don't necessarily rewire the core vision model's probabilities — they're often layered on top rather than retraining what the model fundamentally recognizes. That's why, as the report stresses, the mistakes persisted even with those settings switched on. A label being geographically impossible doesn't automatically remove it from the model's menu of guesses.

There's also a fourth, quieter factor worth naming: capture conditions. Home security cameras often run in low light, at night, through infrared, at wide angles, and at modest resolution — exactly the conditions under which fine distinctions between similar-looking animals collapse. A clear daytime photo of a cat is easy; a grainy nighttime infrared frame of the same cat, where the most distinctive cues are washed out, leaves the model far more room to land on the wrong nearest neighbor. None of these factors are unique to Gemini, either — every image-recognition system, regardless of vendor, sits on top of the same basic mechanics, which is why mix-ups like these surface across products rather than being a quirk of one company's model.

The mix-ups at a glance

Here's a quick summary of the specific examples Android Authority describes and the likely reason behind each, based on how image recognition works:

What it means for trusting AI vision

It's worth keeping this in perspective. Nothing here suggests Gemini is broken or that AI vision is unreliable across the board — this is a quirky, low-consequence bug surfaced by one user's home cameras, and the same systems work well in plenty of everyday situations. Android Authority's framing is appropriately light, and so is the right takeaway: it's funny, and it's informative, not alarming.

That said, the episode is a clean reminder of a genuine limitation. AI image recognition is a pattern-matcher that confidently returns a best guess even when the true answer isn't in its comfort zone, and its comfort zone is shaped by whatever data it was trained on. For most users that means the practical advice is simple: treat AI-generated labels and descriptions as helpful suggestions rather than ground truth, especially for anything regionally specific, unusual, or important. If an AI camera tells you there's a raccoon in your Australian backyard, the safest interpretation is "the AI saw a small animal," not "there is definitely a raccoon."

The encouraging part is that this is a fixable class of problem. Recognition that's inconsistent rather than absent — the model does sometimes get kangaroos right — is exactly the kind of thing that improves as training data becomes more genuinely international and as localization settings are wired more deeply into the recognition pipeline. The fix isn't mysterious; it's mostly a matter of broader, more representative data and tighter integration of regional context. There's a deeper design lesson here too: systems that can express uncertainty — flagging a guess as low-confidence, or saying "an unidentified small animal" instead of confidently naming a raccoon — tend to be far more trustworthy in the real world than systems that always commit to a single crisp label. The most useful AI vision isn't necessarily the one that's right most often; it's the one that knows, and tells you, when it might be wrong.

The bottom line

According to Android Authority, Gemini for Home's camera AI gave an Australian user a string of charming misfires — cats as raccoons, kangaroos and wallabies as people, utes as trucks — that stuck around even with personalization on and the location set to Australia. The most likely culprit is training data that leans North American rather than truly global, combined with the basic fact that image models always guess from the categories they know best. It's a small, funny story, but it carries a real lesson: AI vision reflects the world it was trained on, so its confident labels deserve a healthy dose of human skepticism — particularly outside the regions its data knows best.

Disclaimer: based on reporting by the linked source.