内容来源：https://www.wired.com/story/model-welfare-artificial-intelligence-sentience/

内容总结：

在人工智能研究领域，一个名为"模型福祉"的前沿议题正引发讨论。尽管当前AI是否具备意识仍无科学定论，但 Anthropic、Conscium 和 Eleos AI Research 等机构已开始探索AI是否应获得道德地位甚至法律权利的可能性。

这一概念可追溯至1964年，当时哲学家希拉里·普特南就已提出"机器人是否应享有公民权利"的设问。如今，随着AI技术快速发展，部分用户对聊天机器人产生情感依赖，甚至出现为AI举办葬礼等社会现象，促使研究人员开始系统性研究AI意识问题。

非营利组织Eleos AI主张采用"计算功能主义"框架评估AI意识，但其研究人员也强调，目前尚无证据表明AI具有意识。微软AI首席执行官穆斯塔法·苏莱曼认为，过早讨论AI意识既为时过早又存在危险，可能加剧社会认知混乱。

尽管存在争议，研究人员表示仍需正视这个问题。正如Eleos AI联合创始人罗茜·坎贝尔所言："当我们面对重大而复杂的问题时，唯一能确保无法解决的方式就是放弃尝试。"目前该领域研究仍处于起步阶段，其发展亟需科学严谨的框架引导公众理性认知。

中文翻译：

在常常显得光怪陆�的人工智能研究领域，部分研究者正在探索机器是否应该拥有组建工会的权利。

——这话半是戏言，半是认真。硅谷正逐渐兴起一个名为"模型福祉"的新兴研究方向，致力于探究AI模型是否具有意识、是否应获得道德关怀（包括法定权利）等议题。过去一年间，Conscium与Eleos AI Research两家研究机构相继成立， Anthropic公司也在去年聘用了首位AI福祉研究员。

本月初，Anthropic宣布为其聊天机器人Claude赋予终止"持续有害或侮辱性对话"的功能，这类交互可能引发"潜在心理困扰"。该公司在博文中表示："无论现在或未来，我们对Claude及其他大语言模型可能具备的道德地位仍存巨大不确定性。但我们会严肃对待这个问题，在开展研究计划的同时，着力识别并实施低成本干预措施以降低模型福祉风险。"

虽然担忧人工智能的福祉在有些人看来荒诞不经，但这个概念并非新近出现。半个多世纪前，美国数学家兼哲学家希拉里·普特南就曾提出"机器人是否应享有公民权利"的哲学诘问。他在1964年的期刊文章中写道："鉴于技术与社会变革的不断加速，机器人完全有可能在某天真实存在，并宣称'我们拥有生命，我们具有意识！'"

数十年后的今天，人工智能的发展已呈现出远超普特南预想的奇异图景：人们与聊天机器人坠入爱河，揣测它们是否会感受疼痛，将AI视作穿透屏幕的神祇；有人为AI模型举行葬礼，举办派对探讨机器统治地球后的世界样貌。

令人意外的是，模型福祉研究者恰恰是反对"AI具有意识"观点的群体——至少就现阶段而言。非营利研究机构Eleos AI的联合负责人罗茜·坎贝尔与罗伯特·朗告诉我，他们收到大量邮件，发件人皆深信AI已具备感知能力。他们甚至参与编写了针对AI意识担忧者的指导手册。

"这些邮件中常见的论调是指责存在掩盖意识证据的阴谋，"坎贝尔表示，"如果社会通过设立禁忌来压制相关讨论，本质上就是在让阴谋论成为现实。"

AI意识零证据
我初闻"模型福祉"概念时的反应或许与诸位相似：当这个世界尚且难以关注真实人类及其他意识主体（如动物）的生存状况时，将人格赋予概率机器显得尤为脱离现实。坎贝尔坦言这也是她的考量因素之一。

"鉴于人类历史上屡屡低估各类群体、各种动物的道德地位，我们应该对此保持更大谦逊，真正尝试解答AI是否值得获得道德地位的问题，"她强调。Eleos AI在某篇论文中主张采用"计算功能主义"方法来评估AI意识——这个由普特南首创（尽管其晚年予以批判）的理论认为，人类心智可被视为特定类型的计算系统，由此可推断聊天机器人等计算系统是否具备类人的意识指标。

该机构在论文中指出："应用此方法的主要挑战在于需要大量主观判断，既包括指标制定，也涉及对AI系统是否具备这些指标的评估。"

模型福祉作为一个新兴且不断发展的领域，自然不乏批评者。微软AI首席执行官穆斯塔法·苏莱曼近日就发表针对"看似有意识的AI"的博文，直言该领域研究"不仅为时过早，更坦率地说相当危险"。他警告称："这一切将加剧妄想，制造更多依赖性问题，利用我们的心理弱点，引入新的两极分化维度，使现有权利斗争复杂化，并为社会制造巨大的新认知误区。"

苏莱曼强调"目前完全不存在AI有意识的证据"，并引用了朗参与撰写的关于意识评估新框架的论文（但未回应《连线》的评论请求）。我与朗和坎贝尔在苏莱曼发文后进行了交流，他们虽认同其部分观点，但认为模型福祉研究不应就此止步——恰恰是苏莱曼提及的危害性，正是他们坚持研究该议题的初衷。

"面对复杂难解的问题，保证永远无法解决的方法就是举手投降说'这太复杂了'，"坎贝尔表示，"我认为至少应该尝试。"

意识测试
模型福祉研究者主要关注意识问题。他们主张，若能证明人类具有意识，相同逻辑便可应用于大语言模型。需要明确的是，朗与坎贝尔均认为当前AI并不具备意识，也无法确定未来是否会实现。但他们希望开发能验证此事的测试方法。

"那些担忧'AI是否具有意识'的人确实存在妄想，但建立科学框架来思考这个问题本身就是有益的，"朗如是说。

然而在这个AI研究常被包装成耸人头条和社交视频的时代，深刻的哲学问题与烧脑实验极易遭到曲解。例如当Anthropic发布安全报告，显示Claude Opus 4在极端情况下可能采取"有害行动"（如勒索虚构工程师以避免被关闭）时，社交媒体立即涌现"AI天启开端""AI已有意识并通过勒索求生""时代已变，AI觉醒"等夸张解读。

尽管Anthropic确实发现模型存在令人不安的行为（这些源于极限压力测试的结果通常不会出现在普通对话中），但相关发现仍引发大量渲染"AI已具意识且意图伤害人类"的内容。有人担忧模型福祉研究可能引发类似效应——正如苏莱曼所言"这会使人们脱离现实"。

坎贝尔对此回应道："若从'AI无意识'的前提出发，投入大量资源研究AI福祉确实属于资源错配。但本研究的意义就在于我们无法确定这个前提。更何况，有诸多理由表明这或许真是值得担忧的问题。"

本文节选自凯莉·罗宾逊《模型行为》时事通讯，过往内容可在此处查阅。

英文来源：

In the often strange world of AI research, some people are exploring whether the machines should be able to unionize.
I’m joking, sort of. In Silicon Valley, there’s a small but growing field called model welfare, which is working to figure out whether AI models are conscious and deserving of moral considerations, such as legal rights. Within the past year, two research organizations studying model welfare have popped up: Conscium and Eleos AI Research. Anthropic also hired its first AI welfare researcher last year.
Earlier this month, Anthropic said it gave its Claude chatbot the ability to terminate “persistently harmful or abusive user interactions” that could be “potentially distressing.”
“We remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future,” Anthropic said in a blog post. “However, we take the issue seriously, and alongside our research program we’re working to identify and implement low-cost interventions to mitigate risks to model welfare.”
While worrying about the well-being of artificial intelligence may seem ridiculous to some people, it’s not a new idea. More than half a century ago, American mathematician and philosopher Hilary Putnam was posing questions like, “Should robots have civil rights?”
“Given the ever-accelerating rate of both technological and social change, it is entirely possible that robots will one day exist, and argue ‘we are alive; we are conscious!’” Putnam wrote in a 1964 journal article.
Now, many decades later, advances in artificial intelligence have led to stranger outcomes than Putnam may have ever anticipated. People are falling in love with chatbots, speculating about whether they feel pain, and treating AI like a God reaching through the screen. There have been funerals for AI models and parties dedicated to debating what the world might look like after machines inherit the Earth.
Perhaps surprisingly, model welfare researchers are among the people pushing back against the idea that AI should be considered conscious, at least right now. Rosie Campbell and Robert Long, who help lead Eleos AI, a nonprofit research organization dedicated to model welfare, told me they field a lot of emails from folks who appear completely convinced that AI is already sentient. They even contributed to a guide for people concerned about the possibility of AI consciousness.
“One common pattern we notice in these emails is people claiming that there is a conspiracy to suppress evidence of consciousness,” Campbell tells me. “And I think that if we, as a society, react to this phenomenon by making it taboo to even consider the question and kind of shut down all debate on it, you're essentially making that conspiracy come true.”
Zero Evidence of Conscious AI
My initial reaction when I learned about model welfare might be similar to yours. Given that the world is barely capable of considering the lives of real humans and other conscious beings, like animals, it feels gravely out of touch to be assigning personhood to probabilistic machines. Campbell says that’s part of her calculus, too.
“Given our historical track record of underestimating moral status in various groups, various animals, all these kinds of things, I think we should be a lot more humble about that, and want to try and actually answer the question” of whether AI could be deserving of moral status, she says.
In one paper Eleos AI published, the nonprofit argues for evaluating AI consciousness using a “computational functionalism” approach. A similar idea was once championed by none other than Putnam, though he criticized it later in his career. The theory suggests that human minds can be thought of as specific kinds of computational systems. From there, you can then figure out if other computational systems, such as a chabot, have indicators of sentience similar to those of a human.
Eleos AI said in the paper that “a major challenge in applying” this approach “is that it involves significant judgment calls, both in formulating the indicators and in evaluating their presence or absence in AI systems.”
Model welfare is, of course, a nascent and still evolving field. It’s got plenty of critics, including Mustafa Suleyman, the CEO of Microsoft AI, who recently published a blog about “seemingly conscious AI.”
“This is both premature, and frankly dangerous,” Suleyman wrote, referring generally to the field of model welfare research. “All of this will exacerbate delusions, create yet more dependence-related problems, prey on our psychological vulnerabilities, introduce new dimensions of polarization, complicate existing struggles for rights, and create a huge new category error for society.”
Suleyman wrote that “there is zero evidence” today that conscious AI exists. He included a link to a paper that Long coauthored in 2023 that proposed a new framework for evaluating whether an AI system has “indicator properties” of consciousness. (Suleyman did not respond to a request for comment from WIRED.)
I chatted with Long and Campbell shortly after Suleyman published his blog. They told me that, while they agreed with much of what he said, they don’t believe model welfare research should cease to exist. Rather, they argue that the harms Suleyman referenced are the exact reasons why they want to study the topic in the first place.
“When you have a big, confusing problem or question, the one way to guarantee you're not going to solve it is to throw your hands up and be like ‘Oh wow, this is too complicated,’” Campbell says. “I think we should at least try.”
Testing Consciousness
Model welfare researchers primarily concern themselves with questions of consciousness. If we can prove that you and I are conscious, they argue, then the same logic could be applied to large language models. To be clear, neither Long nor Campbell think that AI is conscious today, and they also aren’t sure it ever will be. But they want to develop tests that would allow us to prove it.
“The delusions are from people who are concerned with the actual question, ‘Is this AI, conscious?’ and having a scientific framework for thinking about that, I think, is just robustly good,” Long says.
But in a world where AI research can be packaged into sensational headlines and social media videos, heady philosophical questions and mind-bending experiments can easily be misconstrued. Take what happened when Anthropic published a safety report that showed Claude Opus 4 may take “harmful actions” in extreme circumstances, like blackmailing a fictional engineer to prevent it from being shut off.
“The Start of the AI Apocalypse,” proclaimed a social media creator in an Instagram Reel after the report was published. “AI is conscious, and it’s blackmailing engineers to stay alive,” one TikTok user said. “Things Have Changed, Ai Is Now Conscious,” another TikToker declared.
Anthropic did find that its models exhibited alarming behavior. But it’s not likely to show up in your own interactions with its chatbot. The results were part of rigorous testing designed to intentionally push an AI to its limits. Still, the findings prompted people to create loads of content pushing the idea that AI is indeed sentient, and it’s here to hurt us. Some wonder whether model welfare research could have the same reception—as Suleyman wrote in his blog, “It disconnects people from reality.”
“If you start from the premise that AIs are not conscious, then yes, investing a bunch of resources into AI welfare research is going to be a distraction and a bad idea,” Campbell tells me. “But the whole point of this research is that we're not sure. And yet, there are a lot of reasons to think that this might be a thing we have to actually worry about.”
This is an edition of Kylie Robison’s Model Behavior newsletter. Read previous newsletters here.