新的“算命AI”：算生算死算你啥时候搬去国外｜科学60秒

2024-01-30 14:01

死亡预言家？

预测的力量是强大的，我们一直在尝试预测未来，在某些领域，我们已经做得很不错了，比如气象预测和气候建模。但从科学角度来讲，我们离水晶球和神谕还很遥远，很多事情我们还没法用可靠的方法进行预测。

但是，如果说人工智能能让我们比以往任何时候都更接近那种超自然的占卜境界呢？比如说，如果人工智能模型能准确预测一个人的死亡呢？

如果你能知道自己什么时候会死，你会想知道吗？

“能预测死亡的人工智能计算器”听起来有点瘆人，但这并没有阻止研究人员成功开发出这种工具，2023 年底发表在《自然-计算科学》（Nature Computational Science）上的一项研究就描述了这么一种名为 life2vec 的深度学习模型。

life2vec 不仅仅是一个“死亡预测器”，它能做的远不止“预测死亡”。但有一点须要注意：目前还很难确切知道它的预测有多准确，研究人员也没有声称它可以明确预测某个人的死亡。life2vec 的训练基于一个包含丹麦居民信息的数据集，因此，它无法对任何非丹麦人进行有意义的预测。

除了预测四年内的死亡外，研发人员还测试了 life2vec 对人们填写性格问卷的预测能力，以及预测某人是否会进行国际搬迁。在所有这三项测试中，life2vec 似乎都表现不俗，准确率高于研究人员比较过的其他预测方法：在死亡测试中，准确率约为 78%，在搬到国外的测试中，准确率约为 73%。

“死亡计算器”为什么还能预测性格特征以及人们是否会跑去国外生活？这些事情有什么关联吗？

作为一个深度学习模型，life2vec 的训练数据为丹麦政府收集的六百万丹麦居民的大量信息，包括就业历史、基本人口数据，以及人们与丹麦医疗系统之间几乎所有互动的信息，因为丹麦实行全民医疗保健。

研究人员将每个人的所有数据整理成时间线，让 life2vec 通过时间序列数据理解到一个人一生中发生的不同事件之间的模式。例如，地点和薪水间存在相关性吗？被诊断患有某种疾病的人是否能存活下来？从事某些职业的人是否更长寿？从事某种职业的人是否具有某种性格特征？

在此基础上，研究人员可以要求深度学习模型根据这些时间线做出预测。life2vec 的有趣之处不仅仅在于它能够做出这些预测，更在于它是如何做到的。这是一种通常用于聊天机器人和语言处理的模式，它的运行方式基本上就像增强型的“自动补全”功能一样。

当你给大模型输入一个生命事件时间轴，然后问它一个特定的问题，它就会根据提示预测时间轴上的下一个相关步骤，有时下一步就是死亡，也就意味着它在预测死亡。这听起来就像《黑镜》（Black Mirror）里的情节，你的手机自动补全功能开始吐出一条信息：“你将在 7 天内死掉。”

一些统计学家和寿命模型专家并不相信这项研究所谓的超过 70% 的准确率，预测死亡是极为复杂而困难的。统计数据具有不稳定性，尤其是在死亡方面。死亡是一种罕见事件，在人的一生中只会发生一次，年轻人不会“经常死亡”。事实上，如果假设人口中 25~50 岁之间的每个人都会在 XX 年死去，那么你就已经是一位准确性非常高的“死神”了。

不过，life2vec 也并非全无用处。首先，它可以帮助我们了解疾病预后和健康结果；从社会学的角度来看，life2vec 还可以用来梳理隐藏的社会偏见，例如年龄或原籍国与职业进步之间的意外关联。

这样的工具肯定也存在道德风险。研究作者仔细地注明了丹麦的隐私法和反歧视法对 life2vec 所有使用方式的限制，学术和政府研究人员必须经过申请才能将其用于特定目的，并有责任保护人们的数据。

研究作者苏恩·莱曼（Sune Lehmann）也表示，他从事这项研究是因为他信任丹麦政府，但在美国或者其他没有全面数据隐私法的国家，他不会那么心安理得地做这项研究。

在美国，预测性警务（predictive policing）和法官在人工智能算法的帮助下做出量刑决定，在所有这些案例中……[查看全文]

Can AI Predict Your Death?

Lauren Leffer: Prediction is powerful. For as long as people have been thinking about the future, we've been trying to predict it. And, in some areas, we've gotten pretty good (think, like, meteorology and climate modeling).

Tulika Bose: But scientifically-speaking — we're still pretty far from the stuff of crystal balls and oracles. There's lots of things people have no reliable way to forecast, despite what your horoscope may suggest.

Leffer: Absolutely. But… what if artificial intelligence could get us closer to that fantasy/supernatural realm of divination than ever before? What if, for instance, an AI model could accurately forecast a person's death?

Bose: Wait, what?!

Leffer: Yeah. It sounds scary, right?

Bose: Uh, yeah.

Leffer: — and it might be, but probably not for the reasons you expect. I'm Lauren Leffer, tech reporting fellow at Scientific American.

Bose: I'm Tulika Bose, Senior multimedia editor.

Leffer: And you're listening to Scientific American's Science, Quickly podcast.

Leffer: So Tulika, death– we're all going to do it one day, but if you could know when you're going to die, would you want to?

Bose: Uh no, no, no no.

Leffer: You didn't even let me finish the question.

Bose: Yeah, I don't wanna know.

Leffer: Yeah, me either– seems like an obvious case of TMI.

Bose: Yeah! For sure.

Leffer: But the inherently unsettling nature of "an AI calculator that spits out death predictions" didn't stop researchers from developing just that. The tool is called life2vec, and was described in a study published late last year.

Bose: And it's really an AI widget that says when you're gonna die?

Leffer: Sort of. Yes and no. Life2vec isn't just a death predictor. It's way more general than just mortality forecasts. It can offer more than that. One BIG CAVEAT right up top: it's hard to know exactly how accurate any of its predictions are at the moment– no one is claiming the tool can definitively predict an individual person's death. It's also trained on a very specific data set, namely one that contains information on residents of Denmark? So it can't offer meaningful predictions about anyone who isn't Danish.

Bose: Ok, all of that's kind of a relief. But can we go back to the "more general thing"? What else can it predict?

Leffer: Yeah, so maybe it's a big deal for Danes. But on top of forecasting mortality over a four year period, the developers also tested their tool's ability to predict peoples' answers to a personality questionnaire —

Bose: Interesting —

Leffer — and to forecast if someone would make an international move. In all three of these tests, life2vec seemed to perform pretty well, with higher accuracy than other prediction methods the researchers compared it to. For the death test, about 78% accurate, for moving abroad, about 73%.

Bose: Whoa.

Leffert: But again, it's hard to know exactly what accuracy means here.

Bose: Hold up, the death calculator could also predict personality traits and if people would live abroad? How? How are any of those things related?

Leffer: Let's take a step back. Life2vec is a machine learning model, which basically just means it's a big computer program made to detect patterns in very large sets of data. The researchers built this model and trained it on tons of government-collected information on millions of people living in Denmark. They had employment history, basic demographic data, and also information on just about every interaction between people and the Danish medical system, because Denmark has universal healthcare.

Bose: Wow, that must be nice.

Leffer: Truly. Anyway, the researchers took all this data for each person and organized it into timelines. The life2vec model was fed all these timelines, and trained to pick up on patterns between the different events that show up in a person's life. You know, how are location and salary related? Do people diagnosed with certain diseases survive? Do people with certain professions tend to live longer?

Bose: Do people with certain career paths have certain personality traits?

Leffer: Exactly. That's how you can go from something like demographic info to personality questionnaire.

Bose: That's so interesting.

Leffer: From there, the researchers could ask versions of the model to make predictions based on those timelines. The extra interesting thing with life2vec isn't just that it makes these predictions–again we've been doing that forever–it's how it does it. It's a type of mode usually used for chatbots and language processing. It basically runs like souped-up autocomplete.

Bose: Hah wait so the death predictions here are "autocompleting" someone's life?

Leffer: That is one way to put it. You give the model the life event timeline, you ask it a specific question, and it predicts the next relevant step in the timeline, based on the prompt. Sometimes that means predicting death.

Bose: Uh, kay yeah. This sounds like a Black Mirror episode. Your phone's autocomplete starts to spit out "you'll die in seven days" type messages.

Leffer: Pitch that to Netflix.

Bose: I'm on it. We can be co-writers. So a little bit ago, you kept hedging on life2vec's accuracy, what was that about?

Leffer: Yeah, so I talked with some statistical and life modeling experts about this and none of them were super convinced by those 70+ percent accurate numbers that the researchers reported in the study. My favorite quote came from Christina Silcox, director for digital health at Duke Margolis Center for Health Policy. I asked her how powerful she thought life2vec's mortality predictions are and she said, quote, "I would not quit my job and go to the Bahamas based on this thing."

Bose: Haha ok, that's kind of a burn?

Leffer: It's more of a statement on how hard it is to predict something like death. And to assess the accuracy of something so new–there's nothing obvious to compare life2vec to.

Bose: You can't just see how much better it is than flipping a coin or something?

Leffer: Nope! Data and statistics are wonky, particularly around death which is a rare event, if you think about it. It only happens once in a person's life and young people don't tend to die very often. In fact, if you were to just assume that everyone between like 25-50 in a population was going to live in a given year, you'd already be a really accurate death oracle. You'd be right most of the time.

Bose: Ah ok, so you've gotta figure out what it means to actually be good at predicting death, and then compare life2vec to that.

Leffer: Totally! And that's kind of a question mark.

Bose: So what could you use this for, if not to tell you when it's time to go permanent vacation mode? If we don't really know how good it is, what's the point?

Leffer: Ooooh ok so there are lots of potential uses–down the road, with more testing. For one, the model could help us understand disease prognosis and health outcomes. And from a sociology angle, you can use life2vec to tease out hidden societal biases, like unexpected links between age or country of origin and professional advancement.

Bose: All of that sounds theoretically cool, but I'm stuck on this idea that–actually– this will be used in a bad and scary way. Like, Minority Report-style punishing crimes before they happen, sort of stuff.

Leffer: It's a valid concern!

Bose: Thank you for validating.

Leffer: Haha of course. There is definitely an ethical risk with a tool like this. The study authors were careful to note all the ways that Danish privacy and anti-discrimination laws restrict how life2vec can be used. It's not just going to be a freely available tool. Academic and government researchers will have to apply to use it for specified purposes and then have a responsibility to protect peoples' data.

Bose: Cool, Ok–that's better than nothing I guess.

Leffer: Yeah, if it makes you feel any better– the study author I spoke with, Sune Lehmann, told me that he worked on this research because he trusts the Danish government, but in the U.S.--or another country without a comprehensive data privacy law– he wouldn't have been so comfortable.

Bose: You know, as someone living in the U.S., that doesn't make me feel better.

Leffer: Ok, fair. Then to double down and make you feel worse: We kind of sort of already live in a version of the terrifying Minority Report world. In the U.S., there's predictive policing and judges making sentencing decisions with the help of AI algorithms. In all of these cases...[full transcript]

“

论文信息

Savcisens, G., Eliassi-Rad, T., Hansen, L.K. et al. Using sequences of life-events to predict human lives. Nat Comput Sci 4, 43–56 (2024).

DOI: 10.1038/s43588-023-00573-5

扫码关注“领研网”微信公众号

订阅最新“科学60秒”英语新闻

不再漏掉任何一次新知 plus 练耳的机会~

微信扫码关注该文公众号作者

戳这里提交新闻线索和高质量文章给我们。

来源: qq

点击查看作者最近其他文章