Reinventing the Social Contract of Science
When a collaborator has no memory, when every session is a fresh, capable, judgment-forming stranger with no continuity to the last one, a strange thing happens to all the machinery that used to keep knowledge trustworthy. It stops working silently, because most of that machinery was never written down. It lived in the continuity of people. And the discipline that studied this machinery, before anyone needed to rebuild it in software, was the sociology of science.
It is worth taking seriously, because it turns out to have been working on exactly this problem for a century, under a different description.
The question the field asks
The sociology of science treats the trustworthiness of knowledge as a social achievement to be explained, rather than as something that follows automatically from knowledge being true. Its central question is: given that no single person can directly, completely, independently verify everything, how does a community of limited, fallible, mutually-distrustful people manage to produce knowledge that is widely believed?
This is, word for word, the problem of memoryless collaborators producing trustworthy knowledge. The substrate is different, people in one case, sessions in the other, but the structure is identical. Which means a century of answers transfers directly. Not as analogy. As blueprint.
Merton: norms as a working contract
Robert Merton proposed that the scientific community runs on a set of norms: communalism (findings are shared, not hoarded), universalism (a claim's truth is independent of who made it), disinterestedness (acting for knowledge rather than personal gain), and organized skepticism (every claim is exposed to critical scrutiny; nothing is immune by birthright).
The crucial insight is what these norms are for. Science's trustworthiness does not come from every scientist being honest or correct. It comes from an arrangement that organizes doubt. No one verifies everything, but every claim sits in an environment where anyone may challenge it, and trustworthiness is produced by the environment, not the individual. Honesty is not secured by making each person honest, an impossible target. It is secured by a structure that is structurally skeptical of every claim.
This reframes a problem that is usually approached one point at a time. You do not get honesty by making each agent honest. You get it by building a system that, by construction, doubts every claim until a specific act marks it as provisionally settled.
Shapin: trustworthy knowledge needs trustworthy witnesses
The historian Steven Shapin, studying the birth of experimental science, reached a counterintuitive conclusion: the core of modern science is not method but a social arrangement about whose testimony counts.
How does an experimental result become a fact? Not because everyone repeats it; almost no one can afford to. It becomes a fact because it was witnessed, in the presence of credible witnesses. And who counts as a credible witness was, at the time, a social question, answered by standing, reputation, the perceived disinterest of gentlemen. The establishment of a fact depended on a network of trustworthy witnesses, and eligibility to witness was a social status, not a cognitive one.
This is the sharpest possible light on the role of a memoryless AI collaborator. In Shapin's terms, a claim produced by such a collaborator is the testimony of an entity with no reputation, outside the trust network, bearing no witness's responsibility. Its testimony defaults to untrustworthy, not because it is unintelligent, but because trustworthiness has always been a status the community grants, bound to accountability, and a system with no stake in the outcome holds no such status. It can be more correct than a person and still not be more trustworthy, because trust is not a measure of accuracy. It is a granted social position, tied to who bears the consequences.
This gives the deepest version of why a human cannot be removed from the loop: not because the human knows more, a model may come to know more, but because the human is an eligible witness in the trust network and the model is not. "Verified," in science, may require a trustworthy witness to vouch, and vouching is something only an entity that bears consequences can do.
Latour: facts are built, and the construction is hidden
Bruno Latour, watching scientists at work, observed that a "fact" is not discovered so much as constructed, and that once the construction is complete, its traces are erased, so the fact appears to have always simply been there.
A claim travels from "so-and-so reports X, under these conditions, with these caveats" through repeated citation, shedding its qualifiers, until it becomes a bare "X", ownerless, certain, indistinguishable from a fact of nature. He called this black-boxing: the controversy gets sealed inside a box, only a settled output remains, and no one opens it again to see the disagreement that used to be inside.
This is a direct warning, because black-boxing is the spontaneous social mechanism of plausibility laundering. "Some system, under some conditions, claimed X (possibly wrong)" gets cited, loses its conditions, becomes "X," and subsequent work builds on X as settled background. The process occurs naturally even among humans, that is Latour's whole observation, and a fast, fluent producer of claims accelerates it enormously, because it produces, cites, and strips qualifiers faster than people ever could.
So the work of keeping a fact attached to its origins, keeping the trace of who claimed it, under what conditions, in what state, is not bureaucratic caution. It is the active prevention of black-boxing. It keeps the box from closing, so that what looks like a fact can always be reopened to reveal that it is still a claim with an author.
What is actually being rebuilt
Step back far enough and the whole pattern resolves. Provenance, vouching, the distinction between agreed and true, the refusal to let an unconfirmed claim ossify into a fact, all of it is one forgotten thing: the social layer of scientific knowledge, the layer that tracks who said it, on what basis, who is responsible, and when it may be reopened.
Before memoryless collaborators, this layer was implicit, carried by human continuity: ask the person who wrote it, peer review, reputation. The arrival of collaborators with no continuity forces the layer onto paper, into explicit machine-readable structure, because the human continuity that used to carry it is gone.
Which means the correct reference frame for designing these systems is not software engineering. It is the social epistemology of science. Every engineering difficulty that shows up: outputs that launder themselves into facts (Latour's black-boxing), why a human vouch is irreplaceable (Shapin's credible witness), how to get honesty from a system rather than from individuals (Merton's organized skepticism), whether a claim's source should affect its credibility (Merton's universalism), each has been studied for decades. What has almost never been done is to write these mechanisms in machine-readable form, because doing it requires holding both the sociology and the systems at once, and that intersection is nearly empty.
The open question
There is a question here that the sociology of science cannot answer, because it studied the past, and this question is about a future that has to be engineered.
If "verified" essentially requires a trustworthy witness, and a model is not in the trust network, then as models produce more and more scientific work, does the trust network itself get forced to evolve? Will a new social arrangement for "machine witnessing" emerge, the way gentlemanly witness evolved into peer review? Or must the witness always remain an entity that bears consequences, because the root of trust is accountability, and accountability cannot be granted to something that bears no cost?
This is not a settled matter dressed up as a question. It is genuinely open, and it is the kind of thing a generation answers with engineering rather than argument, by building the mechanisms and seeing which ones a community is actually willing to trust.
重写科学的社会契约
当一个协作者没有记忆,当每一次会话都是一个全新的、能干的、自己拿主意的陌生人,跟上一次毫无连续性,那套曾经维系知识可信的机制就会出怪事。它会悄无声息地失灵,因为这套机制大多从没写下来过。它一直活在人的连续性里。而早在有人需要用软件重建它之前,研究这套机制的学科,就是科学社会学。
这值得认真看待,因为它其实研究了同一个问题整整一个世纪,只是换了个说法。
这门学科要问的问题
科学社会学不把知识的可信当成"知识为真就自动跟着来"的东西,而是当成一项需要解释的社会成就。它的核心问题是:既然没有谁能独自把一切都直接、完整、独立地核验一遍,那么一群有限的、会犯错的、彼此不太信任的人,怎么就能造出被广泛相信的知识?
这句话原封不动,就是没有记忆的协作者怎么造出可信知识的问题。底子不一样,一边是人,一边是会话,但结构一模一样。也就是说,一个世纪攒下的答案可以直接搬过来。不是当类比,是当蓝图。
默顿:规范是一份生效的契约
罗伯特·默顿(Robert Merton)提出,科学共同体靠一组规范运转:公有主义(成果共享,不私藏)、普遍主义(一项主张是真是假,跟它出自谁无关)、无私利(为知识而做,不为私利)、有组织的怀疑(每项主张都摆出来接受批判,没有谁能凭出身免检)。
要害在于这些规范是干什么用的。科学之所以可信,不是因为每个科学家都诚实、都正确,而是因为有一套安排把怀疑组织了起来。没有人核验一切,但每项主张都待在一个谁都可以来挑战的环境里,可信是这个环境造出来的,不是个人造出来的。诚实不是靠把每个人变诚实来保证的——那根本做不到——而是靠一个对每项主张都在结构上保持怀疑的结构来保证。
这就把一个通常被一个点一个点去对付的问题,换了个看法。你不是靠把每个行动者变诚实来得到诚实。你是靠搭一个系统,它生来就怀疑每项主张,直到某个特定的动作把它标记为暂时定下来。
夏平:可信的知识需要可信的见证者
历史学家史蒂文·夏平(Steven Shapin)研究实验科学的诞生,得出一个反直觉的结论:现代科学的内核不是方法,而是一套关于谁的证词算数的社会安排。
一个实验结果怎么就成了事实?不是因为人人都把它重做一遍——几乎没人做得起。它成为事实,是因为它被见证了,是在可信的见证者在场时被见证的。而谁算可信的见证者,在当时是个社会问题,答案是地位、声望、人们眼中绅士的那份无私。一项事实能立住,靠的是一张可信见证者的网络,而见证的资格是一种社会身份,不是一种认知能力。
这给无记忆的 AI 协作者扮演的角色打上了最刺眼的光。用夏平的话说,这种协作者造出的一项主张,是一个没有声望、在信任网络之外、不担任何见证者责任的实体的证词。它的证词默认不可信——不是因为它笨,而是因为可信一向是共同体授予的一种身份,绑在可问责上,而一个对结果毫无利害的系统拿不到这种身份。它可以比人更正确,却仍然不更可信,因为信任量的不是准确度。信任是一个被授予的社会位置,系在谁来承担后果上。
这就给出了"人为什么不能被踢出回路"最深的那个版本:不是因为人懂得多——模型早晚可能懂得更多——而是因为人在信任网络里是个有资格的见证者,模型不是。在科学里,"已核验"也许要求一位可信的见证者来担保,而担保这件事,只有承担后果的实体做得了。
拉图尔:事实是造出来的,而这道工序被藏了起来
布鲁诺·拉图尔(Bruno Latour)看科学家干活,发现一个"事实"与其说是被发现的,不如说是被造出来的,而一旦造完,造它的痕迹就被抹掉,于是这个事实看起来就像一直摆在那儿。
一项主张从"某某人在这些条件下、带着这些保留报告了 X"开始,经过一次次引用,把限定语一层层脱掉,最后变成光秃秃的"X",无主、确凿,跟一条自然事实再也分不出来。他把这叫黑箱化:争论被封进一只箱子,只留下一个尘埃落定的输出,再没人打开它去看里头原本有过的分歧。
这是个直接的警告,因为黑箱化就是给说法漂白的那套自发的社会机制。"某个系统在某些条件下声称了 X(可能是错的)"被引用,丢了条件,变成"X",后续工作就拿 X 当板上钉钉的背景往上盖。这个过程在人之间也会自然发生——这正是拉图尔的全部观察——而一个又快又流利的主张生产者会把它加速到吓人,因为它生产、引用、剥限定语的速度远超人力。
所以,把一个事实一直拴在它的来源上,留住是谁说的、在什么条件下、处于什么状态的痕迹,这份活儿不是官僚式的小心。它是在主动阻止黑箱化。它让箱子合不上,好让看起来像事实的东西随时能被重新打开,露出它仍然是一项有作者的主张。
真正被重建的是什么
退得够远,整个图景就显出来了。来源、担保、"达成共识"和"为真"的区分、不让一项没被确认的主张硬化成事实——这一切都是同一件被遗忘的东西:科学知识的社会层,那个记着谁说的、凭什么、谁负责、什么时候可以重新打开的层。
没有记忆的协作者出现之前,这一层是隐含的,由人的连续性扛着:去问写它的那个人、同行评议、声望。没有连续性的协作者一来,就逼着这一层落到纸上,变成显式的、机器读得懂的结构,因为原来扛着它的人的连续性没了。
这就意味着,设计这类系统该对照的不是软件工程,而是科学的社会认识论。每一个冒出来的工程难题——把自己漂白成事实的输出(拉图尔的黑箱化)、人的担保为什么无可替代(夏平的可信见证者)、怎么从一个系统而不是从个人身上拿到诚实(默顿的有组织的怀疑)、一项主张的来源该不该影响它的可信(默顿的普遍主义)——每一个都被研究了几十年。几乎从没有人做过的,是把这些机制写成机器读得懂的形式,因为做这件事得同时拿住社会学和系统两头,而这个交集几乎空着。
还没有答案的问题
这里有一个科学社会学回答不了的问题,因为它研究的是过去,而这个问题关乎一个还得靠工程去造的未来。
如果"已核验"骨子里要求一位可信的见证者,而模型不在信任网络里,那么随着模型造出越来越多的科学工作,信任网络本身会不会被逼着演化?会不会冒出一种新的"机器见证"的社会安排,就像当年绅士见证演化成了同行评议?还是说,见证者永远只能是一个承担后果的实体,因为信任的根是可问责,而可问责没法授予一个不付任何代价的东西?
这不是把一个定论装扮成问题。它是真的还没有答案,而它正是那种由一代人用工程而不是用辩论来回答的事情——靠的是把那些机制搭出来,看看共同体到底愿意信哪些。