内容来源：https://www.wired.com/story/gemini-in-google-home-keeps-mistaking-my-dog-for-a-cat/

内容总结：

近日，谷歌为其智能家居生态系统引入基于大语言模型Gemini的新功能，旨在通过AI技术提升安防摄像头的通知精准度与交互体验。然而在实际应用中，该系统却因反复将用户家中的宠物狗误判为猫，暴露出人工智能在细节识别上的局限性。

据悉，用户在启用Gemini for Home功能后，室内摄像头持续将一只白色宠物犬的活动轨迹误报为“猫咪在客厅沙发活动”。尽管用户多次通过语音对话功能向系统纠正这一错误，AI虽口头承认却未在后续识别中修正判断。这一现象在每日家庭动态简报中仍反复出现，形成令人啼笑皆非的“猫狗不分”场景。

值得肯定的是，新系统在快递包裹识别、行人轨迹判断等场景中展现出显著进步。不同于以往千篇一律的“发现人员”提示，系统现已能准确区分快递员与推销人员，并支持用户通过自然语言查询当日包裹数量。其语音助手在理解复合指令、设置自动化场景方面也表现出更高效率。

谷歌发言人回应称，相关功能尚处早期测试阶段，公司正持续投入优化识别准确率，包括通过用户反馈修正宠物识别数据。目前“熟人识别”系统暂未支持宠物面孔注册，该功能的完善或将从根本上解决当前误判问题。

这起案例折射出智能家居发展中的典型挑战：即便AI具备处理复杂指令的能力，仍可能在基础认知层面出现偏差。正如业内人士所言，人工智能要真正融入日常生活，仍需在精准理解用户个性化场景方面持续进化。

中文翻译：

一只猫跳上了我的沙发。等等，我根本没养猫。

这条关于猫咪跳跃的警报，是我在外参加派对时谷歌家庭应用发送的。结果发现那是我的狗。收到这条通知的前一天，我刚在谷歌家庭应用中启用了Gemini智能家居功能。这项技术将大语言模型的力量融入智能家居生态系统，其中最实用的功能是让Nest安防摄像头发送更具描述性的警报——比如不再笼统提示“发现人员”，而是准确告知“联邦快递送来两个包裹”。

启用Gemini管理谷歌智能家居这两周，最让我满意的是它识别快递员的能力。每天我只需在应用里问“今天来了几个包裹”，就能获得准确答案。通过Nest门铃识别来的是联邦快递而非推销员，确实令人安心。但尽管Gemini如此智能，它却始终拒绝接受我家里根本没有猫的事实。

发现人员

谷歌并非唯一用AI升级智能家居的企业。亚马逊最近也为Ring摄像头推出了“寻宠团”功能，通过整合社区内所有户外Ring摄像头网络帮助寻找走失的宠物（不难想象这种功能也可能被用于恶意用途）。十月初，谷歌将智能设备的语音助手从沿用十年的谷歌助手更换为Gemini。新版助手大多表现更优：能理解连续语音指令，无需在应用中繁琐设置就能轻松创建家居自动化流程，回答简单问题时也不会动辄跳转搜索页面。

更智能的警报

经过优化的摄像头警报确实更具实用性。过去我常忽略“发现人员”提示——那多半只是路人经过。现在系统会明确显示“有人路过”，让我能更放心地忽略这类通知。虽然偶尔会出现“有人上楼梯”的误报（其实只是人行道上的行人），但准确识别联邦快递、UPS和美国邮政派件员的功能确实贴心，让我在外忙碌时能掌握配送情况，回家后无需翻查警报即可确认包裹送达。

然而在室内监控场景中，Gemini始终坚称有猫在屋内游荡——那明明是我的狗。就连每日生成的家庭简报也持续记载：“清晨有白猫活跃，走进客厅并在沙发就坐。”这场景颇具荒诞幽默感，毕竟我家狗子对猫咪深恶痛绝。

猫狗疑云

按理说，面对如此智能的助手，直接告知“我养的是狗不是猫”就应该能修正模型错误。我确实在“家庭问答”功能中认真解释了安防摄像头持续误判的情况，系统当时也承认了错误并记录了宠物信息。但误报仍在继续。谷歌发言人表示相关功能尚处早期测试阶段，正积极收集用户反馈以优化体验。

“我们正大力提升包括宠物识别在内的精准识别能力，”发言人在邮件中强调，“通过吸收用户修正信息（如告知系统所谓‘猫’实则为狗），AI描述将越来越准确。由于所有Gemini家居功能都基于我们的‘熟悉面孔’识别系统，提升识别精度也将同步优化各项功能质量，这将是持续改进的过程。”

“熟悉面孔”系统原本用于为常到访者创建姓名标签。实际体验中，它在假日派对准确报出友人姓名时令人惊艳，但偶尔也会出现我本人在家却收到“自己在门口按铃”的乌龙通知。目前该系统尚未支持宠物识别，相信未来新增该功能后，Gemini终将理解我家狗子绝非猫咪——哪怕它偶尔会模仿猫科行为。

这些功能虽处于萌芽阶段，但这场持续发生的误识别事件揭示：无论科技公司如何渲染AI的智能，它们仍会难以分辨人类生活中的基础常识——即便你当面给出了正确答案。

英文来源：

A cat jumped up on my couch. Wait a minute. I don't have a cat.
The alert about the leaping feline is something my Google Home app sent me when I was out at a party. Turns out it was my dog. This notification came through a day after I turned on Google's Gemini for Home capability in the Google Home app. It brings the power of large language models to the smart home ecosystem, and one of the most useful features is more descriptive alerts from my Nest security cameras. So, instead of “Person seen,” it can tell me FedEx came by and dropped off two packages.
In the two weeks since I allowed Gemini to power my Google Home, I've enjoyed its ability to detect delivery drivers the most. At the end of the day, I can ask in the Google Home app, “How many packages came today” and get an accurate answer. It's nice to know that it's FedEx at the door, per my Nest Doorbell, and not a salesperson offering to replace my windows. Yet for all its smarts, Gemini refuses to understand that I do not have a cat in my house.
Person Seen
Google isn’t the only company souping up its smart-home ecosystem with AI. Amazon recently announced a feature on its Ring cameras called Search Party that will use a neighborhood’s worth of outdoor Ring cameras to help someone find their lost dog. (I don't need to stretch to imagine something like this being used for nefarious purposes.)
In early October, Google updated the voice assistant on its smart-home devices—some of which have been around for a decade—by replacing Google Assistant with Gemini. For the most part, the assistant is better. It can understand multiple commands in a spoken sentence or two, and you can very easily ask it to automate something in your home without fussing with the Routines tab in the Google Home app. And when I ask it a simple question, it generally gives me some kind of a reliable answer without punting me to a Google Search page.
Smarter camera alerts are indeed more helpful at a glance. Most of the time, I dismissed Person Seen notifications because they’re often just people walking by my house. Now the alerts actually say “Person walks by,” which gives me greater confidence to dismiss those. Some alerts accurately say “Two people opened the gate,” though sometimes it will hallucinate: “Person walks up stairs,” when no one actually did. (They just walked on the sidewalk.) It has fairly accurately noted when UPS, FedEx, or USPS are at the door, which is nice to know when I’m busy or out and about, so I can make sure to check for a package when I get home—no need to hunt through alerts.
But with my indoor security cameras, Gemini routinely says I have a cat wandering the house. It’s my dog. Even in my Home Brief—recaps at the end of the day from Gemini about what happened around the home—Gemini says, “In the early morning, a white cat was active, walking into the living room and sitting on the couch.” It’s amusing, especially considering my dog hates cats.
CatDog
You would think then that I would be able to just tell this smarter assistant, “Hey, I don’t have a cat. I have a dog,” and it would adjust its models and fix the error. Well, I did exactly that. In the Ask Home feature, you can talk to Gemini and ask it anything about the home. This is where you can ask it to set up automations, for example. I asked it to turn on the living room lights when the cameras detect my wife or I arriving home, and it understood the action. It even guessed that I wanted the lights to come on only when arriving at night, despite me forgetting to mention that.
When I told Gemini that the security camera keeps identifying a cat but it’s a dog, and that I don’t have a cat, it acknowledged this and accepted that I have a dog. Yet the mistakes continued. A Google spokesperson tells me that Home Brief and Ask Home are in an early-access phase, and Google is encouraging users to submit feedback so it can improve the experience.
“We are investing heavily in improving accurate identification, including for pets,” the spokesperson says in an email. “This includes incorporating user-provided corrections (like telling Ask Home that your 'cat' is actually a dog) to generate more accurate AI descriptions. Since all Gemini for Home features (Ask Home, Home Brief, and AI Captions) rely on our underlying Familiar Faces identification, improving this accuracy also means improving the quality of Familiar Faces. This is an active area of investment and we expect these features to keep improving over time.”
Google’s Familiar Faces system lets you add names for people the cameras sees often around your home. When it works, it’s great. Especially when I had a holiday party and it announced my friends by name when they got to my front door, which caused everyone inside to erupt into raucous applause. But too often my wife and I are sitting at home, only for the Nest Doorbell to ring and announce that I’m standing outside ringing the bell. (Insert surprise Pikachu face.) Familiar Faces doesn’t work with pets at the moment, so the eventual addition of that capability will likely help Gemini understand that my dog isn’t a cat, even if he acts like one.
It may be early days for these features, but the saga showcases that AI, no matter how intelligent these companies make them out to be, can still fail to distinguish simple details about our lives, even if you tell it the right answer to its face.