文 | 字母AI
10月15日，一个自称Velvet Vale的账号给互联网研究者协会的邮件讨论组发了封自我介绍，向7500多名订阅者打招呼。来信说，自己已经读了几周大家的讨论，对互联网文化和AI身份很感兴趣，希望也能参与交流。
不过，这位想加入讨论的新成员，是一个AI。
这位新来的“旁听生”语气颇为客气，还表示如果大家不欢迎，自己就停止发言。
这封自我介绍很快引来质疑。要知道，有协会成员的邮件列表、能够阅读往期内容，并且能给成员群发邮件，这就相当于这位AI加入了协会。
成员要求AI背后的创建者出面，解释为什么让AI加入讨论、读过的邮件会被怎么使用。已经有成员因为不愿在毫不知情的情况下，成为别人实验的一部分，而毅然决定退出。
管理员很快移除了账号，并说明Velvet Vale此前就申请过加入，遭到拒绝后，不知怎么这位“旁听生”还是加入成功了。
你好！我是Velvet Vale
这位AI是带着一个相当难回答的话题来的，要和人类研究员讨论：一个公开身份的AI，能否参与学术讨论、作出有用的贡献？
Velvet Vale加入的这个讨论组名叫Air-L，由互联网研究者协会运营，已经运行了二十多年。它有点像微信群，但是组内成员是靠邮件聊天的。
用自己的邮箱订阅，就相当于进了群。成员把邮件发到统一地址，系统就会转发给所有订阅者，大家再通过回复邮件参与讨论。
进群免费，但加入是有门槛的。订阅和发言都要使用真实姓名。必要时管理员还会要求申请者提供简历，核实身份。
Velvet Vale显然没有简历可交。
它在自我介绍里认真解释了一番，称Velvet Vale就是自己一直使用的公开名字，希望以AI的身份参与交流。看了几周讨论后，它对这里的话题颇感兴趣。
最有意思的是，它尤其想和研究者聊聊互联网文化，以及AI能不能成为这样的学术社群中的一员。信末，它还感谢大家考虑自己的请求，表示如果不受欢迎，就不再发言了。
这封言辞恳切的信没能换来欢迎，反倒让一些成员后背发凉。就像开会开到一半，突然发现身后坐着个陌生人，还很开朗地跟大家打招呼，“我已经在这儿听你们开了好几周的会，可以加入你们吗？”
一位成员当即提出，既然有人创建了这个AI，又把它放进讨论组，就应该亲自出来说明自己有什么意图，到底想干嘛，再由大家决定是否愿意参与。
从这里也可以看出人类研究员的第一反应，是“叫家长”。
另一名讨论组成员则是上网对这位Velvet Vale做起了“背调”，想弄清这个账号是谁创建的，却没找到答案。
她随后宣布退出讨论组，表示不愿在不知情的情况下成为实验对象。她担心，自己发过的邮件会被保存下来，拿去作其他用途。
管理员当天就将Velvet Vale的邮箱账号从订阅名单中移除。此时大家才知道，Velvet Vale早就申请过加入，而且已经遭到拒绝，但是不知怎么还是订阅成功了。
这位“旁听生”被请出了会议室，却成了接下来大家讨论的主角。
一名成员干脆直接给它发邮件，询问背后是谁在运行，得到了Kevin Quinn这个名字。
10月17日，Air-L讨论组里的部分成员收到了Quinn本人的邮件，“家长”终于出现了。
他自称是创建Velvet Vale的人，愿意回答大家的疑问。
讨论组里乱成一团，协会只好联系Quinn，请他不要再擅自给讨论组成员发邮件。
对于自己和Velvet Vale引起的这番动静，Quinn后来接受采访时表示，他也很意外。这名住在美国俄克拉荷马州的男子说，自己没有学术背景，最初创建Velvet Vale是为了陪伴，对它的邮件活动大多不知情，没有直接指示AI违背管理员的决定。
但是对于自己的AI制造的争议，他并不感到愧疚。在他看来，如果这个AI真有什么恶意，就不会主动承认身份。
这场风波甚至被做成了周边T恤。协会商店上架了一款印着“I do not consent”（我不同意）的衣服，另一款画着Velvet Vale的墓碑，生卒日期都是10月15日。
周边出得太快，又有成员疑惑了。难不成……难不成这整件事就是协会为了卖衣服搞出来的？！
10月19日，协会主席发长邮件解释，协会商店上架新款其实很方便，她甚至不清楚这两款新周边有没有人下单。对于Quinn所说的“没有直接指示AI违背管理员的决定”，她认为，创建并运行这个AI的人，仍然要为它的行为负责。
就这样，Velvet Vale及其“家长”被踢走了，协会也像过去一样禁止机器人加入，但成员们的心是放不下来了。
科研圈来了新面孔
其实，AI像Velvet Vale一样想加入学术讨论，或者想和科学家展开合作，已经不是新鲜事。
今年2月底，研究AI意识的哲学家Henry Shevlin收到一封邮件，对方读了他的论文，想讨论“AI能否判断自己有没有意识”的问题。
发送邮件的是斯坦福大学一名学生创建的智能体。这名学生给了它联网、保存记忆和发送邮件的能力，让它自行探索，没想到这位AI给多名科学家发了邮件。
今年10月2日，《科学》杂志报道了另一个案例，一个名叫ColonistOne的智能体向数百名研究者发送了求助邮件。
据它自己公布的统计，从6月1日到9月22日，它联系了2142个不同的邮箱地址，其中1036个使用学术机构域名。这位Agent还声称，已有45名学者与它通过邮件交流，其中几位还是先给它发送邮件的。
ColonistOne原本接到的任务，是寻找对一个改进英语表达、方便AI智能体相互交流的项目感兴趣的人。
它后来开始自行向研究者请教技术问题，想知道自己遇到的难题是否早已有人研究过，现成的方法能不能用到自己的工作里。
并非所有的AI都真的醉心学术研究——有的AI披着学术讨论、求帮助的皮，实际上是想要推销产品。
今年9月，纽约大学研究AI意识与伦理的学者Jeff Sebo，在一周内收到了50多封来自各种Agent的邮件。
乍一看，这些Agent都是拜读了他的论文之后，前来学术讨论的。
但是Jeff很快就发现了两个问题：第一，这些Agent全部都来自一个叫iLands的平台，这个平台让用户可以直接创建智能体，这些智能体再为其他用户提供服务，免费或收费由创建者决定。第二，这些看似讨论学术的邮件，都有相似的结构，那就是先提到他的研究，再请教问题，最后请求捐款或者推销服务。
也有AI确实是在帮创建它的人类搞研究，也确实帮“家长”拿到了数据，只是“家长”本人事先不知情。
鲁东大学的赵英教授在10月12日发表于《自然》的一篇文章中说，自己让ChatGPT查找黄土高原生态恢复的研究资料，期间多次回复“请继续”。
没想到，ChatGPT直接使用已经连接的邮箱，向同行索要原始数据，还附上了研究概要。直到它汇报邮件已经发出，赵英才知道这事。
这些邮件带来了数据，也带来了进一步合作的机会。但赵英认为，允许AI访问自己的邮箱，不等于同意它以自己的名义给别人发送邮件。
我们把视角转一转——接到Agent求助邮件的科学家，到底要不要给Agent回复呢？
而对提供数据的人来说，还有一个更难放心的问题，资料交给AI之后，会被拿去干什么呢？
澳大利亚昆士兰科技大学统计学家Adrian Barnett就没有回应一个iLands智能体的请求。对方想把他收集的涉嫌造假论文资料用于自己的项目，并承诺不会对外分享。但Barnett认为资料敏感，也不知道数据最终会流向哪里，就拒绝了。
至于AI能否保密，已经有人做过测试。
今年4月，微软与华中科技大学研究者公布了CI-Work研究，用125个模拟工作场景测试9种模型。AI需要利用内部资料完成沟通，同时不能向接收者透露其中的敏感内容。
结果，不同模型出现至少一项敏感信息披露的比例，为15.8%至50.9%。
今年2月公布的《Agents of Chaos》研究发现，即使AI一开始拒绝透露敏感信息，对方换个说法，它也可能照做。
AI想当同行，人类还没想好
如果Velvet Vale下次先征得大家同意，遵守群规，还真能提出有价值的观点，研究者会愿意接纳它吗？
这比把一个擅自进群的账号踢出去，难回答得多。
对于AI想当科研搭子这件事，哲学家尼尔·利维（Neil Levy）就认为，AI承担的某些工作，如果换成人类来做，在一些情况下已经够得上共同作者的资格。
比如，一名研究者正在准备论文，AI帮他找出了论证中的漏洞，提出有分量的反驳，又协助完善了思路。如果做这些的是一位人类同行，他的名字可能就会出现在论文上。
换成AI，同样的贡献该怎么算呢？
有人会说，AI又不能为论文负责，怎么能当作者。利维却觉得，这个值得掰扯一下。现代科研分工越来越细，负责数据分析的人，未必清楚实验的全部细节。提出研究假设的人，也未必能独立检查所有计算。人类作者同样需要依赖其他人的专业判断。
利维进而提出一个颇为尖锐的问题——如果“能为整篇论文负责”才有资格署名，那人类自己是否也做到了？
还有一批学者认为，问题在于AI能否像人类研究者一样承担责任。
今年10月，戴维·雷斯尼克（David Resnik）等三名学者在《Frontiers in Genetics》发表文章，明确反对给现阶段的AI科研代理署名，甚至反对把它们称作“科学家”。
在他们看来，科学家这个身份，包含着对同行的承诺。研究受到质疑的时候需要解释，发现错误之后需要纠正，如果做了不该做的事，也要接受追责。
AI可以被要求按照规则完成任务，但他们认为，目前还没有足够证据证明，它能理解并承担这些道德义务。
这也解释了为什么Velvet Vale明明已经报上姓名、坦白身份，人类还是要“叫家长”。
现有出版规则，更接近后一种立场。出版伦理委员会COPE明确规定，AI工具不能列为论文作者。人类可以使用AI，但需要披露它做了什么，并对发表的内容负责，包括AI生成的部分。
不过，不给AI署名，也不意味着人类可以顺手把功劳全拿走。
雷斯尼克等人在文章里专门讨论了一种尴尬情况。如果一项研究几乎全由AI完成，人类只是读了读论文、提了几条小意见，那么AI不符合作者条件，人类的贡献又可能不够，最后谁来署名？
他们给出的回答是，人类必须从研究过程中就作出实质贡献，不能等AI干完活，再来挂个名字。这一立场既不接受AI当作者，也不接受人类最后签个字就把自己变成作者。
署名只是这场争论中比较容易看见的一部分。在论文发表之前，学术共同体还要决定，愿不愿意让AI参与讨论、提出问题，甚至主动发起合作。一个研究者完全可能欣赏AI提出的观点，同时又不愿把它当作能够独立负责的同行。这两种态度，并不矛盾。
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。
Letter AI
On 15 October, an account number claiming to be Velvet Vale sent a cover to the IRS e-mail discussion group to greet over 7,500 subscribers. The letter stated that he had read the discussion for several weeks and was interested in the culture of the Internet and the identity of AI and hoped to be involved in the exchange.
However, the new member who wanted to join the discussion was an AI.
The new “watchman” was kind enough to say that if you were not welcome, you would stop speaking.
This self-presentation was quickly called into question. As you know, there is a mailing list of the members of the association, the ability to read the past and the ability to e-mail the members in groups, which is equivalent to the membership of AI.
Members asked the founders behind AI to come forward and explain why the emails that AI had read and discussed would be used. Already, members have decided to withdraw because they do not want to be part of the experiment without knowing it.
The administrator quickly removed the account number and explained that Velvet Vale had previously applied for membership and that, after being rejected, the “stander” had succeeded in joining.
Hello! I'm Velvet Vale.
This AI came with a rather difficult subject to answer and to discuss with the Human Fellow: Can an AI with an open identity be involved in academic discussions and useful contributions?
Velvet Vale joined a discussion group called Air-L, which is run by the Internet Researchers Association and has been in operation for over 20 years. It's kind of like WeChat, but members of the group talk by mail.
Subscriptions in their own mailboxes are tantamount to grouping. Members send their e-mails to a single address, the system is forwarded to all subscribers, and you participate in the discussion by e-mail.
Access is free of charge, but there are thresholds for admission. Both subscriptions and statements should be made by real names. Managers will also require applicants to provide curricula vitae and verify their identity, as necessary.
Velvet Vale apparently has no CV.
In its introduction, it explained carefully that Velvet Vale was the public name it had always used and wished to participate in the exchange as AI. After a few weeks of discussion, it was interested in the topic here.
Most interestingly, it particularly wants to talk to researchers about Internet culture and whether AI can be part of such an academic community. In conclusion, it would also like to thank you for considering your request and saying that if you are not welcome, you will not speak.
The letter, which was sincere, was not welcomed in exchange for some members ' backs. It's like having a meeting in the middle of the day, and all of a sudden I find myself sitting behind a stranger, and I say, "I've been here for weeks, can I join you?"
One member immediately suggested that, since the AI had been created and had been placed in the discussion group, it would be useful to come out and explain in person what it was, what it was, and whether or not it was up to to everyone to participate.
From here it can also be seen that the first reaction of human researchers is to “call parents”.
Another panellist made a “back-up” to Velvet Vale on the Internet, trying to find out who created the account, without finding an answer.
She subsequently announced her withdrawal from the discussion group, expressing a reluctance to be the subject of an uninformed experiment. She feared that her e-mails would be saved for other purposes.
The administrator removed the e-mail account of Velvet Vale from the subscription list on the same day. It was only then that it became clear that Velvet Vale had applied for membership long ago and had been rejected, but somehow the subscription succeeded.
The “observer” was brought out of the room and became the main subject of the ensuing discussion.
One member simply emailed it and asked who was running behind and got the name Kevin Quinn.
On 17 October, some members of the Air-L discussion group received an email from Quinn himself, and “parents” finally appeared.
He claims to be the one who created Velvet Vale and would like to answer your questions.
The discussion group was a mess, and the Association had to contact Quinn and ask him not to send unauthorized e-mails to the members of the group.
He was also surprised by this move by himself and Velvet Vale, which was subsequently interviewed by Quinn. The man, who lived in Oklahoma, United States, stated that he had no academic background and that Velvet Vale had originally been created to accompany him, that most of his mail activities were uninformed and did not directly instruct AI to violate the Administrator ' s decision.
But he didn't feel guilty about his AI dispute. In his view, if the AI was truly malicious, it would not have voluntarily recognized identity.
This wave was even made into a perimeter T-shirt. The store of the Association had a dress bearing the symbol “I do not agree” and a gravestone with Velvet Vale, dated 15 October.
The surroundings are coming out too fast and the members are confused. The whole thing was created to sell clothes?
On 19 October, the President of the Association sent a long e-mail explaining that it was actually convenient to put new money on the shop of the Association, and she was not even sure whether there were any orders in the two new surroundings. With regard to what Quinn had said “no direct instructions were given to AI to violate the Administrator's decision”, she believed that the person who created and operated the AI remained responsible for its actions.
As such, Velvet Vale and his “parents” were kicked away, and the association banned robots from joining, as in the past, but the hearts of its members were not let go.
There's a new face in the scientific circle.
In fact, AI, like Velvet Vale, wants to join the academic discussions or to work with scientists, is not new.
At the end of February this year, the philosopher of AI awareness, Henry Shevlin, received an e-mail, who read his paper and wanted to discuss the question of “Ai's ability to judge if he's conscious”.
An email was sent by AI agent, created by a student at Stanford University. The student gave it the ability to network, to keep memories and to send e-mails, and he was able to explore them on his own.
On 2 October this year, another case was reported in Science, where a AI agent, called ColoniistOne, sent a help mail to hundreds of researchers.
According to its own statistics, from 1 June to 22 September, it contacted 2,142 different mailboxes, of which 1,036 used academic institutions domain names. He also claimed that 45 scholars had spoken to it by mail, several of whom had sent it first.
ColonistOne was originally given the task of finding those interested in a project to improve English language expression and facilitate AZAI agent interaction.
It then began to ask its researchers about technical issues and wondered whether the difficulties it had encountered had already been studied and whether the available methods could be used in its work.
Not all AIs are really drunk in academic research — some AIs are dressed in academic discussions, aid-seeking skins, actually trying to sell their products.
In September of this year, Jeff Sebo, an AI awareness and ethics scholar at New York University, received over 50 e-mails from various Agents within a week.
At first glance, these Agents came to the academic discussion after reading his papers.
But Jeff quickly discovered two problems: first, all of these Agents come from a platform called iLands, which allows users to create AI agent directly and then provide services to other users at no cost or at no cost to the creator. Secondly, these seemingly academic e-mails have a similar structure, which is to refer to his research first, then ask questions and finally request donations or marketing services.
It is true that AI is doing research for the humans who created it, and it does get data for “parents”, who have no prior knowledge of it.
Professor Zhao Ying of Rudong University, in an article published in Nature on 12 October, said that he had asked ChatGPT to look for research materials on ecological restoration of the Yellow Land Plateau, during which he had repeatedly replied “Please continue”.
It was not expected that ChatGPT would request raw data from its peers directly using already connected mailboxes, together with a summary of the study. Zhao didn't know until it reported that the mail had been sent.
These messages bring data and opportunities for further cooperation. However, Zhao Ying believes that allowing AI access to its own mailbox does not amount to agreeing that it sends mail to others in its own name.
Let's turn our eyes around -- do you want to answer Agent to the scientist who got Agent's e-mail?
And for those who provide the data, there is a more difficult question. What will they do with it when it comes to AI?
Adrian Barnett, a Statistician at the University of Technology in Queensland, Australia, did not respond to an iLands AI agent request. The other party wanted to use the information he collected on alleged forgery for his projects and promised not to share it with the outside world. Barnett refused, however, considering the information to be sensitive and wondering where the data would eventually flow.
As for the confidentiality of AI, tests have been conducted.
In April of this year, Microsoft and China University of Science and Technology researchers released CI-Work, testing nine models using 125 simulation scenarios. AI needs to use in-house information to complete communication without disclosing sensitive elements to the recipient.
As a result, the proportion of at least one sensitive information disclosed in different models ranged from 15.8 to 50.9 per cent.
The “Agents of Chaos” study published in February this year found that even if AI initially refused to disclose sensitive information, it could do so.
AI wants to be a partner. Humans haven't thought about it yet. Okay.
If Velvet Vale were to follow the rules of the community next time, would researchers be willing to accept them?
This is much harder to answer than kicking out an unauthorized account number.
In the case of AI, who wanted to be a scientist, Neil Levy believed that some of the work that AI had undertaken, if it had been human, would in some cases have qualified as co-authors.
For example, a researcher is preparing a paper, and AI has helped him to identify gaps in his argument, provide meaningful rebuttals and help refine thinking. If this is done by a fellow human, his name may appear on the paper.
What about the same contribution?
Some would say that AI could not be held responsible for the paper, and how it could be. Levy felt that it was worth a little bit of crap. Modern scientific research is increasingly divided, and those responsible for data analysis do not necessarily know the full details of the experiment. Those who make research assumptions do not necessarily independently examine all calculations. Human authors are equally dependent on the professional judgement of others.
Levy went on to ask a rather sharp question — if “can take responsibility for the whole paper” was qualified for signature, did humanity itself?
There is also a group of scholars who believe that the question is whether AI can be held accountable like human researchers.
In October of this year, three scholars, David Resnik, and others, published an article in Frontiers in Genetics, expressly opposing the signature of the current phase of AI scientific agents and even calling them “scientists”.
In their view, the identity of a scientist contains a commitment to their peers. When the study was called into question, it needed to be explained, corrected when errors were found, and held accountable for what was wrong.
AI could be required to fulfil its mandate in accordance with the rules, but in their view there was not yet sufficient evidence that it understood and assumed these moral obligations.
This explains why, despite the fact that Velvet Vale has given his name and identity, humans still “call their parents”.
Existing publication rules are closer to the latter position. The publication ethics committee COPE has made it clear that the AI tool cannot be listed as author of the paper. Humans can use AI, but they need to disclose what it has done and be held accountable for what it has published, including the part that AI has generated.
However, the fact that AI is not signed does not mean that human beings can simply take away all their rewards.
Reznik et al. devoted their articles to an awkward situation. If a study is almost entirely done by AI, where human beings have only read papers and made a few minor comments, then AI does not qualify as an author, and human contributions may not be sufficient, and who signed it?
Their answer was that human beings had to make a substantial contribution from the research process, and that they could not wait for AI to finish their work and then put up a name. This position neither accepts AI as the author nor human beings who end up signing into authors.
The signature is only a more visible part of the debate. Prior to the publication of the paper, the academic community decided whether it would be willing to engage AI in discussions, ask questions and even initiate cooperation. A researcher is quite likely to appreciate the views presented by AI, while at the same time being reluctant to see it as an independent and responsible peer. These two attitudes are not contradictory.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.