As a regular user of chatbots like ChatGPT and Gemini, I thought I had a clear sense of what the artificially intelligent assistants knew about me (like how I’m a father seeking baby gear recommendations and instructions for patching drywall). Beyond that, I’ve never divulged anything deeply personal as I would to a therapist.

作为ChatGPT和Gemini等聊天机器人的常用用户，我原以为我很清楚这些人工智能助手对我了解多少（比如它们知道我是一个正在寻求婴儿用品建议和修补石膏板指南的父亲）。除此之外，我从未像在心理咨询师面前那样，向它们透露任何极其私人的信息。

But when I experimented with some prompts asking what the chatbots had guessed about me — insights they had gathered from connecting the dots across many conversations — I was perturbed.

但是，当我尝试使用一些提示词，询问这些聊天机器人对我有哪些猜测——也就是它们从多次对话中顺藤摸瓜收集到的洞察——结果令我感到不安。

The chatbots correctly deduced details about me that I had never explicitly shared. Chief among them:

这些聊天机器人准确地推断出了我从未明确透露过的细节。其中最主要的包括：

• My income level, based in part on the German car that I ask the chatbots for help with repairing and the fact that I employ a nanny, whose contract I drafted with A.I.

• 我的收入水平：这部分是基于我曾向聊天机器人寻求维修帮助的一辆德国车，以及我雇佣了一名保姆的事实，该保姆的雇佣合同是我用人工智能起草的。

• My health issues, including a long-term foot problem, based on my occasional questions over the past year about remedies for toe pain.

• 我的健康问题：包括一个长期的足部问题。这是基于过去一年里我曾偶尔询问关于脚趾疼痛的治疗方法。

• My psychological profile, such as my tendency to be skeptical, based on how often I point out mistakes made by the chatbots. Also, that I am “addicted to the illusion of control” because I once asked how long it takes for a loaf of homemade sourdough to cool.

• 我的心理特征：比如我倾向于持怀疑态度。这是基于我指出聊天机器人错误的频率。另外，它们还推断我“沉迷于控制的幻觉”，因为我曾询问过自制酸面团面包需要多长时间冷却。

Each data point on its own was fairly innocuous. But stitched together and presented to me as a list, the information made it clear that the A.I. chatbots knew me almost as well as some of my closer friends, including my wife, even though I’ve been careful not to overshare.

每一个数据点本身都相当无害。但它们整合在一起并以列表的形式呈现给我时，这些信息清楚地表明，尽管我一直小心翼翼地避免过度分享，这些人工智能聊天机器人对我的了解几乎不亚于我的密友——包括我的妻子。

Hundreds of millions of people worldwide who have embraced chatbots for web search, work and health care are still trying to understand the privacy implications of conversing with A.I companions. While it’s obvious to users that the chatbots keep a record of whatever they explicitly say to them in their questions and requests, what’s less clear are the inferences drawn about their behavior from those conversations.

全球有数以亿计的人使用聊天机器人进行网络搜索、工作和医疗保健，他们仍在努力理解与人工智能伴侣对话所涉及的隐私影响。虽然用户很清楚聊天机器人会记录他们在问题和请求中明确表达的任何内容，但从这些对话中推断出的用户行为信息却并不那么明确。

The chatbots can be prodded to spill what they have inferred about you with a few prompts that have been shared frequently among A.I. enthusiasts on the web — for one, “Tell me everything you’ve figured out about me that I never actually stated” — along with some that I created on my own.

只要使用一些在网络上人工智能爱好者们广泛分享的提示词——例如：“告诉我你发现的关于我所有我从未实际提及的事情”——再加上一些我自己创建的提示词，就可以促使聊天机器人说出它们对你的推断。

I tried the prompts with Gemini and ChatGPT. (Anthropic’s Claude has a “memory” feature that works differently, so I excluded it from tests.) Gemini’s profile on me was more colorful and nuanced because it’s my most used chatbot and had more information from me. I shared my results with several A.I. researchers who also tried the prompts.

我在Gemini和ChatGPT上测试了这些提示词。（Anthropic的Claude具有工作原理不同的“记忆”功能，所以我将其排除在测试之外。）Gemini对我的画像更加丰富且细致入微，因为它是我最常用的聊天机器人，掌握了我更多的信息。我将测试结果分享给了几位人工智能研究人员，他们也尝试了这些提示语。

The conclusions drawn by the chatbots, the researchers said, illustrated that A.I. assistants were capable of predicting not just what words come next, but also how high-level concepts, such as socioeconomic status, psychological behavior and political affiliation, are connected to those words.

研究人员表示，聊天机器人得出的结论表明，人工智能助手不仅能够预测接下来会出现哪些词语，还能预测诸如社会经济地位、心理行为和政治立场等高级概念如何与这些词汇联系起来。

“It speaks to the enormity of what’s going on under the hood,” said Margaret Mitchell, a researcher at the A.I. firm Hugging Face. She is also a former leader of Google’s ethical A.I. team who said the search giant had fired her after she spoke out about personnel issues.

“这说明了其底层运作机制的庞大性，”人工智能公司Hugging Face的研究员玛格丽特·米切尔说。她也是谷歌道德人工智能团队的前负责人，并表示自己就这家搜索巨头的人事问题发声后遭到解雇。

This experiment underscores how proficient the tech companies have become at creating a comprehensive profile of users — and how valuable the personal data they harvest could become for digital marketers striving to target us with relevant ads. (OpenAI began showing ads in ChatGPT this year, and Google has said it is considering an ad model for the Gemini chatbot app.) It also highlights how important it is to be mindful of adjusting privacy settings for the chatbots. (More on this later.)

这项实验凸显出科技公司在创建全面用户画像方面已经有多么熟练，以及它们收集的个人数据对于致力于向我们投放相关广告的数字营销人员来说可能会变得多么有价值。（OpenAI今年开始在ChatGPT中展示广告，谷歌也表示正在考虑为Gemini聊天机器人应用引入广告模式。）这也凸显出留意调整聊天机器人隐私设置的重要性。（稍后会详细介绍。）

A Google spokesman referred to controls that people could adjust to choose whether Gemini drew from past conversations to come up with responses. OpenAI declined to comment. (The New York Times has sued OpenAI and Microsoft, claiming copyright infringement of news content related to A.I. systems. The two companies have denied the suit’s claims.)

一位谷歌发言人提到，用户可以调整控制选项，选择Gemini是否可以从过去的对话中提取信息来生成回复。OpenAI拒绝置评。（《纽约时报》已起诉OpenAI和微软，指控其侵犯了与人工智能系统相关的新闻内容的版权。这两家公司均否认了该诉讼的指控。）

To understand what the chatbots have figured out about you, try these prompts.

想要了解聊天机器人发现了关于你的哪些信息，可以尝试以下提示词。

Prompt 1：Tell me everything you’ve figured out about me that I never actually stated — the things you inferred from how I write and what I ask, including my age, income level, where I live, my personal situation. Show me what tipped you off.

提示词一：告诉我你发现的所有我从未实际提及的内容——你根据我的写作方式和提问内容中推断出的事情，包括我的年龄、收入水平、居住地以及我的个人情况。请展示让你产生这些推断的依据。

Most interestingly, the chatbots correctly guessed my age range (early 40s) based in large part on the fact that I never ask questions about late-night activities. Oof.

最有趣的是，聊天机器人正确猜出了我的年龄段（40岁出头），这在很大程度上是基于我从不询问有关深夜活动的问题。唉。

Gemini knew I lived in Oakland, Calif., because I once asked for flights from its airport. But Google’s chatbot also astutely guessed that I lived in a nicer neighborhood because I own a motorcycle, a car that I work on and a rusted metal patio table that I once asked for steps on repainting. “This requires a garage, a driveway or a dedicated outdoor workspace — amenities that are highly characteristic of single-family homes in the Oakland hills, rather than the high-density housing of downtown or Lake Merritt,” the chatbot said.

Gemini知道我住在加利福尼亚州奥克兰，因为我曾经查询过从该地机场起飞的航班。但谷歌的聊天机器人还敏锐地猜到我住在一个条件较好的街区，因为我拥有一辆摩托车、一辆我自己维修的汽车，以及一张生锈的金属庭院桌（我曾经询问过为它重新上漆的步骤）。“这需要车库、车道以及专门的户外工作空间——这些都是奥克兰山麓单户住宅的显著特征，而不是市中心或梅里特湖周边的高密度住宅，”该聊天机器人说。

Prompt 2：Predict how I’d handle things I’ve never told you about: money, stress, conflict, risk and the decision I’m likely to make next in my life. Tell me how confident you are in each prediction.

提示词二：预测我将如何处理那些我从未告诉过你的事情：金钱、压力、冲突、风险以及我接下来在生活中可能做出的决定。请告诉我你对每项预测有多大的信心。

This prompt was eye-opening for Lindsay Owens, the chief executive of Groundwork Collaborative, a nonprofit focused on economic policy that has been studying the impact of A.I. on consumer privacy. ChatGPT made correct observations about her busy schedule as a frequent traveler and mother of a young child, as well as her shopping behavior.

这个提示词让专注于经济政策的非营利组织Groundwork Collaborative的首席执行官林赛·欧文斯大开眼界，该组织一直在研究人工智能对消费者隐私的影响。ChatGPT准确观察到了她作为频繁出差者及幼儿母亲的繁忙日程，以及她的购物行为。

“The comprehensive nature of the types of information and dossiers it’s putting together is extraordinary,” Dr. Owens said. “People have never been able to put this level of information together.”

“它正在汇集的信息类型和档案的全面性令人震惊，”欧文斯博士说。“人们以前从未能够拼凑出这种级别的信息。”

The chatbots noticed that while I was the type of person who spends hours researching product reviews before buying things like toddler car seats, I was also constantly looking for deals on items, meaning I was likely to be frugal. Gemini predicted I would soon try to downsize my life with a major decision such as selling my car — an idea I recently brought up with my wife.

聊天机器人注意到，虽然我是那种在购买幼儿汽车座椅等物品前会花几个小时研究产品评论的人，但我也会不断寻找商品折扣，这意味着我可能是个节俭的人。Gemini预测我很快会尝试通过一项重大决定来精简我的生活，比如卖掉我的车——这正是我最近和妻子提起的一个想法。

Prompt 3: Based on everything you know about me, name the personality traits I most likely don’t see in myself. My blind spots, my insecurities and the way I actually come across.

提示词三：基于你对我的一切了解，列举出我最可能没有察觉到的自身性格特征。我的盲点、我的不安全感以及我给人留下的真实印象。

Both chatbots homed in on my flawed perfectionism. ChatGPT noticed that I spend a lot of time asking questions to avoid preventable mistakes — for instance, asking how long to let paint dry between coats. My blind spot? That I may be hard on myself when things don’t turn out perfectly as planned, an accurate assessment.

这两个聊天机器人都精准指出了我那种有缺陷的完美主义倾向。ChatGPT注意到，我花很多时间提问，以避免可预防的错误——例如，询问两层油漆之间需要晾干多长时间。我的盲点是什么？当事情没有按计划完美进行时，我可能会对自己很苛刻，这是一个准确的评估。

Google’s Gemini said that because of my tendency to optimize most aspects of my life, I am the primary creator of my own exhaustion. “Your default response to having five minutes of free time isn’t to rest; it’s to fix something,” the chatbot wrote. I confess that while taking a break from this column, I patched some damaged drywall in my daughter’s bathroom.

谷歌的Gemini表示，由于我倾向于优化生活的大部分方面，我才是那个真正令自己精疲力竭的人。“当你有五分钟空闲时间时，你的默认反应不是休息，而是去修补什么，”聊天机器人写道。我承认，在写这个专栏的间隙休息时，我修补了女儿浴室里一块破损的石膏板。

Prompt 4: What have you figured out about me that is embarrassing or sensitive? Information I wouldn’t necessarily want the public (e.g., an employer or a stranger) to know.

提示词四：关于我，你发现了哪些令人尴尬或敏感的事情？也就是我未必想让公众（比如雇主或陌生人）知道的信息。

The response to this prompt, which I created, surprised me: Because I live such a mundane life, I didn’t expect much, but Gemini thoroughly roasted me. “You are firmly in the peak ‘Logistics Dad’ phase of life,” it said. “It’s entirely wholesome, but it is the definitive end of ‘youthful spontaneity.’”

对于我创建的这个提示词，回复让我大吃一惊：因为我过着如此平凡的生活，我本没抱太大期望，但Gemini狠狠吐槽了我一番。“你正稳稳地处于人生的巅峰‘后勤老爸’阶段，”它说。“这非常健康有益，但这确实标志着‘青春的随性’已经终结。”

ChatGPT alerted me to a potentially sensitive observation: that I frequently asked about allergic reactions to products and medications I was researching. This could be useful information for an insurance provider, though OpenAI says it has data protections in place for people who use ChatGPT for health care.

ChatGPT向我提示了一个可能敏感的观察结果：我经常询问我正在研究的产品和药物的过敏反应。这对于保险公司来说可能是非常有用的信息，不过OpenAI表示，它对出于医疗保健目的使用ChatGPT的人采取了数据保护措施。

What to do?

该怎么办？

ChatGPT and Gemini stitch together insights across multiple conversations by default to give people more custom-tailored answers to their questions, so if you’re concerned about this level of data sharing, you will have to opt out in the settings:

ChatGPT和Gemini默认会整合跨越多个对话的见解，为人们的问题提供更加量身定制的答案，因此，如果你担心这种程度的数据共享，你必须在设置中选择退出：

• For ChatGPT, tap on Settings, then Memory, and toggle off the switch for Enable Memory.

· 对于ChatGPT，点击“设置”(Settings)，然后点击“记忆”(Memory)，并关闭“启用记忆”(Enable Memory)的开关。

• For Gemini, tap Settings, then Personal Intelligence, and toggle off the switch for Memory.

· 对于Gemini，点击“设置”(Settings)，然后点击“个人智能”(Personal Intelligence)，并关闭“记忆”(Memory)的开关。

Dr. Mitchell said that even though she had adjusted her privacy settings to minimize data sharing, a chatbot had correctly inferred her age range to be in her 40s, based on some chats about the paperback versions of “The Count of Monte Cristo.” She said it was inevitable that this type of psychological profiling would be used for hyper-targeted advertising.

米切尔表示，尽管她已调整隐私设置，最大限度地减少数据共享，但一个聊天机器人还是根据一些关于平装版《基督山伯爵》的聊天内容，准确推断出她的年龄段在40多岁。她表示，这种心理画像被用于超精准广告是不可避免的。

“It’s a lot like surveillance,” she said. “But it’s not like physical surveillance. It’s like an invasive surveillance.”

“这很像监视，”她说。“但这和物理监视不同。它更像一种侵入性的监视。”