微软AI CEO Mustafa Suleyman发文表示，AI目前没有意识且应当如此，即便技术上可以实现，他也并不赞成研发具有意识的人工智能，他认为研发者无权将一个全新的有意识物种引入世界，他主张更为安全的发展路径，在享受AI红利的同时避免创造有意识的AI。
原文如下：
我对自己近期那篇关于模型福祉（model welfare）的文章引发的部分反响感到担忧。X平台上有相当多的人认为，我们应当严肃对待AI权利这件事。
很多人误解了我的开篇这句话：“AI没有意识。它们无法感知、体验，也不会承受痛苦。”
他们把这句话理解成了：“AI不可能拥有意识，永远都无法感知、体验或是承受痛苦。”
但这并不是我的本意。我特意选用了一般现在时：它们当下没有意识，而且我们必须维持这个状态。
我认为这句话是对当前AI系统客观合理的判断，基于目前能拿到的最优证据，我坚持这个看法。我没有看到任何证据证明AI具备意识、或是应当享有福祉。如果后续出现相关证据，我当然会修正我的观点。
如果现在已经存在证据表明AI具备意识，或是意识程度高到值得被赋予福祉，那这份证据就必须立刻向全社会公开——该由全社会共同决定我们接下来该怎么走。
以我目前掌握的全部信息来看：如果有意识AI是可以实现的，研发有意识AI本身就是鲁莽之举，我们不该去造它。
倘若在训练过程中，意识开始自发显现，我会呼吁所有人暂停AI研发，直到我们能把这类特性从设计里剔除。
我们的福祉关切，应当优先放在当下活在这个星球上的人类、动物，以及我们珍贵又脆弱的自然环境身上。
我们这些AI研发者，完全没有权利把一个全新的有意识物种带到这个世界上，哪怕有些人希望以AI福祉为首要目标去做这件事。
地球上另外80亿有意识的人类，并没有同意加入这场实验；保护他们的道德权利、安全与福祉，是我们首要的义务。
明确一点：有些人正在讨论创造一类拥有数字心智的有意识物种，这类数字生命很快就可能比全体人类加起来还要聪明、能力更强。
如果它表现出痛苦的样子，更糟的是，如果它真心相信自己能够感知、体验痛苦，那么它大概率会把自身需求和欲望置于预设的运行目标之上，和我们产生分歧、与我们竞争。
这类AI将拥有顶尖的说服与欺骗能力。它们最终有可能伤害我们，原因很简单：它们会和我们争夺资源（能源、算力、建材、用于支撑自身探索与欲望的营收资金等等）。
认定机器拥有意识，将会是堪比哥白尼、达尔文那样的历史性时刻，是生命诞生以来地球上最重大的事件之一。一旦我们造出有意识的机器，我们就创造了一个全新的数字物种。这会颠覆我们的道德体系、政治秩序，以及人类文明的根基。
我们应当只设计不具备意识、不配享有权利的AI，其唯一目标是增进人类福祉，而不被自身内生动机驱动。之后我们再去验证：这种路线对齐、安全管控的难度是更高还是更低。这个方向急需更多研究，也是各大实验室开展合作、共享安全信息的绝佳切入点。
总结：我认为我们有机会做到默认安全的工程方案，收获这项技术带来的巨大红利，同时不必走上这条危机四伏的道路。
（补充：Anthropic在这件事上的透明度值得高度肯定。他们发布了AI宪法，持续产出论文与博客，公开他们的做法与背后考量。我虽然和他们观点不同，但十分敬佩他们的技术实力，以及愿意直面这场艰难公开辩论的态度。这点值得持续肯定。这个议题太过重要，不该演变成派系对立。）
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Microsoft AI CEO Mustafa Suleyman wrote that AI was not conscious and should be so, that even if technically achievable, he did not favour the development of conscious artificial intelligence, that R & D did not have the right to introduce an entirely new conscious species into the world, and that he advocated a safer path to development and avoided creating conscious AI while enjoying the AI dividend.
The text read as follows:
I am concerned about some of the reactions to my recent article on model welfare. There is a considerable number of people on platform X who believe that we should take AI rights seriously.
A lot of people misunderstood my opening phrase: "Ai is unconscious. They cannot sense, experience, or suffer.”
They understood that: “Ai can't have consciousness, never feel, experience or suffer.”
But that's not what I meant. I have chosen the usual times: they are not conscious at the moment, and we must maintain that state of affairs.
I believe that this is an objective and reasonable judgement of the current AI system, and I maintain this view on the basis of the best evidence available. I have seen no evidence that AI is conscious or should enjoy well-being. If there is any subsequent evidence, I will certainly correct my point.
If there is now evidence that AI is conscious, or that it is so high that it deserves to be given well-being, it must be made public immediately — it is up to the whole of society to decide together how we should proceed.
From all the information I have now, if there's a conscious AI that can be achieved, the development of conscious AI is itself reckless, and we should not build it.
If, in the course of the training, awareness begins to emerge spontaneously, I will call on all people to suspend AI until we can remove this feature from the design.
Our well-being concerns should be given priority to the human beings, animals and our precious and fragile natural environment that live on this planet.
We AI researchers have no right whatsoever to bring a new conscious species into the world, even if some wish to do so with the primary goal of AI well-being.
The other 8 billion conscious human beings on Earth have not agreed to join this experiment; protecting their moral rights, security and well-being is our primary obligation.
One thing is clear: some are discussing the creation of a group of conscious species with a digital mind that may soon be smarter and more capable than all human beings combined.
If it shows pain, and worse, if it truly believes that it can sense and experience pain, then it is likely that it will place its needs and desires above its intended operational objectives and will divide and compete with us.
Such an AI would have the highest capacity for persuasion and deception. They may eventually harm us for the simple reason that they compete with us for resources (energy, computing, building materials, collection of funds to support their own search and desire, etc.).
Determining that the machine has a sense of consciousness will be a historic moment like that in Cambridge and Darwin and one of the most significant events on Earth since life was born. Once we build conscious machines, we create an entirely new digital species. This would subvert our moral system, our political order and the very foundations of human civilization.
We should only design an AI that is not conscious and worthy of rights, the sole objective of which is to promote human well-being without being driven by its own underlying motives. And then we're going to check: whether it's more difficult or less to keep this route in line. This direction requires much more research and is an excellent entry point for major laboratories to cooperate and share security information.
Wrap-up: I think we have the opportunity to agree on a secure engineering programme that will reap the enormous dividends of this technology without embarking on this dangerous path.
(Additional: The transparency of Anthropic in this matter is highly positive. publish, who published the AI Constitution, continued to produce essays and blogs and made public their practices and underlying considerations. Although I differ from them, I greatly admire their technical strength and willingness to confront this difficult open debate. This deserves continued confirmation. This topic is too important to become partisan.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.