Non DubitoA Mind in Many Languages
Search搜索搜尋
EN
DeutschFrançaisEspañol日本語한국어

Who Wrote the Standard?谁写下了这个标准誰寫下了這個標準 · 08

Once the Choice Becomes a Parameter

写成参数以后,谁还回答

寫成參數以後,誰還回答

Putting a choice into code does not remove the choice. Repeated decisions still need reasons and a route back to their premises.

一次取舍进入代码,不会因此变成没有取舍;当输出反复决定人的处境,理由与重开选择的入口也需要跟着留下。

一次取捨進入程式碼,不會因此變成沒有取捨;當輸出反覆決定人的處境,理由與重開選擇的入口也需要跟著留下。

Han Qin ·

Imagine a lender whose software scores each application against a cutoff. At the counter, an employee says the system sets the rule. A developer says the team merely implemented requirements. The cutoff had an author somewhere. Repeated use can nevertheless make it harder to find anyone expected to explain why it belongs there.

The issue is not whether computers can make judgments. It is how an earlier choice becomes a continuing operation. Suppose ten thousand applications pass through daily. That number is imaginary, not a banking statistic. It illustrates how one unresolved allocation of costs can reach many people without another discussion of its reasons.

A target has already been chosen. Predicting timely repayment does not establish who deserves help. Suitability for a particular loan does not measure every need in a person's life. A lender may have reasons to pursue its target. Returning a score cannot make these different questions interchangeable.

A proxy introduces another choice. Past expenditure may fail to represent need because someone could not afford to spend. In an imagined assistance program, a model might predict expenditure accurately while missing that distinction. Accurate reproduction of a proxy does not settle whether it appropriately represents the intended goal.

A cutoff distributes further costs. Raising it may reduce one error while increasing another. Who bears mistaken refusal and who bears mistaken acceptance cannot disappear because the threshold is a tidy number. The threshold may be retained. Its numerical form cannot supply the reason for retaining it.

Fairness is not a single ready-made switch either. Different demands concern what a score means, errors affecting qualified applicants or errors affecting unqualified ones. Each protects something that needs explanation. Naming the selected metric fair cannot establish that all the other concerns have been answered or simultaneously satisfied.

Kleinberg, Mullainathan and Raghavan examine a particular combination: calibration within groups and between-group balance of average scores separately for actual positive and negative cases. Positive and negative here refer to observed outcomes, not moral judgments about people. Outside conditions such as perfect prediction or equal base rates, the requirements cannot all hold. The theorem does not declare every combination of fairness demands impossible.

Chouldechova studies parity of positive predictive values alongside equality of false-positive and false-negative rates under differing prevalence and relevant nonperfect prediction conditions. Those demands can conflict too. The definitions in the two results must not be blended. Calibration is not another name for a false-positive rate, and neither theorem selects the sacrifice society should make.

These results do not establish that a particular real system is unjust or that improvement is pointless. They challenge a costless promise: give engineers several attractive words and expect all of them at once. Once conflict is explicit, a selected priority needs more visible reasons, not less scrutiny because mathematics has entered the discussion.

Suppose a lender answers a challenge about its fairness objective by showing predictive calibration. That may support a specified prediction claim. It does not automatically explain why calibration was selected as the fairness objective. Compare the actual question with the actual proof before allowing technically impressive evidence to answer beyond its reach.

Automation also requires a condition. If output determines the result, or is routinely copied without genuine reconsideration, the same parameter repeatedly implements its choice. Real authority to reassess makes a difference. The words human review on an interface do not establish that every decision receives renewed judgment about its premises.

Ask what review can reach. Correcting an input error differs from challenging the target, proxy or cutoff. A corrected application matters without establishing that the source of recurring difficulties can be reconsidered. Individual correction and parameter review may both be necessary; neither should impersonate the other when someone asks which route is available.

Disclosure of a parameter takes only one step. Seeing a number does not reveal its reasons. Access to code may still leave choices scattered across configuration files, procurement requirements or version histories. The cost of inquiry includes locating a decision within an operation, not merely understanding the syntax of a program.

Programming is not thereby culpable in itself. Stable execution can limit arbitrary variation, and version records can aid investigation. The concern is whether reasons and review routes survive as unresolved choices scale. Blaming the machine alone can make the human act of adopting it strangely disappear from examination.

Compliance with standards introduces a related boundary. The EU AI Act's harmonised-standard arrangement gives a specified presumption of conformity for requirements actually covered. The existence of that legal arrangement is not an error in itself. Precisely because the presumption has a scope, its reach needs checking rather than indefinite expansion.

The Commission's 19 October 2026 announcement describes publication of quality-management standard EN 18286, relates it to Article 17 and limits the presumption to covered requirements. It is not an unlimited assurance about every activity, deployment or value choice. A useful legal assurance need not be made universal to matter.

The original did not read the standard's full correspondence annex or find a record of its limited proof being used to close an outside challenge. Both gaps remain. Apparent tension between article locations cannot establish illegality. Amended standardisation requests, the actual standard and its correspondence must be examined together.

This is consequently no finding that the EU system has closed ethical inquiry. The investigation needs a particular challenge, the proof's actual coverage and evidence of its later use as an answer beyond that coverage. Missing misuse records establish neither absence nor a speculative explanation for why the search did not find them.

Formal objection is not an all-purpose answer either. The Commission's product-rules guide describes decisions maintaining, restricting or withdrawing harmonised-standard references. The procedure concerns specified legal effects. It does not establish that anyone can reopen any value question through it. Eligibility, scope and initiation still need examination for the question at hand.

Those adopting a scoring system cannot simply refer every difficulty to its original author. Adoption installs a judgment within their own operation. Yet adopters do not become guarantors of all algorithmic knowledge. Their responsibility concerns what they chose, why they used it and how they address questions emerging from that use.

A named responsible person still does not establish that premises can be questioned. An organization may promptly explain an output while refusing discussion of the target, or permit discussion without a route to review. Responsibility and access to reconsideration are distinct findings. An organizational chart cannot certify both at once.

Passi and Barocas look upstream. Their six-month ethnography of a corporate data-science team shows problem formulation taking shape through negotiation, including goals, metrics and proxies. This is evidence from a particular team, not a prevalence estimate about every project. It makes choices visible before a model is trained.

That inquiry complements rather than duplicates this one. Seeing formation differs from asking what happens afterward when a user seeks reasons or review. Recording early negotiations can preserve material for that later question. The record itself does not grant anyone the power to require a response or reopen a choice.

Imagine a rejected applicant finally receiving review. An income-entry error is corrected; the outcome stays unchanged. She then asks why a continuing but irregular form of income never enters the assessment. “We already reviewed your case” cannot settle that question. Input correction did not examine the premise she is now identifying.

Her application need not therefore succeed. Lending involves genuine risks that one person's needs cannot erase. The institution may retain its approach, but it should distinguish what it is refusing and identify where that level of question can be received. Openness does not promise victory, and prudence does not require permanent preservation of the original setting.

Detailed proofs, legislative correspondences and the unread standard belong in further reading. Keep several translations in view: a goal into a variable, a trade-off into a cutoff, fairness into a metric, proof into an answer with a claimed reach. Each may have reasons. None is relieved of them by implementation.

Yesterday's parameter may still be appropriate for tomorrow's first application. The demand is not a meeting from scratch each morning. When it becomes inappropriate, or someone identifies an omitted cost, a real receiving route should remain findable. Code can continue to run without bringing the obligation to answer to a stop.

设想一家贷款机构使用评分系统。申请到来,分数算出,达到某条线便通过。前台说,系统就是这样定的;开发者说,自己只是落实需求。那条线并没有从天而降,却可能在一次次使用中,逐渐找不到需要解释它的人。

重要的不是计算机会不会判断,而是已经作出的选择,怎样变成不断生效的操作。一次设定若每天处理一万份申请,数量只是设想;它要表达的是,同一项未被问清的取舍,可以不经再次讨论而反复落到不同人身上。

首先就有目标的选择。预测能否按时还款,不等于判断一个人值不值得被帮助;判断适不适合某项贷款,也不等于判断生活需要。每一种目标可能有理由,却不能因为最后都输出一个分数,便互相顶替。

再往前,数据为什么代表目标?已花的钱未必等于实际需要:没有花,可能因为根本付不起。在设想的援助项目里,用过去支出来代表需要,可以算得很准确,却仍须解释这项替代为什么恰当。准确复现替代物,不等于已答目标选择。

门槛又分配另一种代价。把线抬高,可能减少一种错误、增加另一种错误。谁被错拒,谁承担错放的成本,不能凭数字写得整齐便退场。门槛可以保留;但保留需要理由,不能把参数格式当作理由。

连公平也不是唯一现成开关。不同要求可能分别关注,同样的分数是否有相同意义,合格者被错拒的比例是否相同,不合格者被错放的比例是否相同。它们各在保护什么,需要先说清,不能只给最终指标起一个公平的名字。

克莱因伯格、穆莱纳坦与拉加万的研究,讨论一组特定要求:组内校准,以及实际正例与实际负例各自的组间平均分平衡。这里的正负例指实际结果,不是给人贴好坏标签。除完美预测或组间基础发生率相同等条件外,三项不能同时满足。它不是把任何三种公平都判成不可能。

乔尔德乔娃的另一项研究,则讨论阳性预测值一致与两类错误率的一致:在基础发生率不同、预测并不完美的相关条件下,它们会冲突。两个结论的定义不能混写。校准不是误报率的别名,数学也没有替我们选出应当牺牲哪一项。

这些定理不证明某个真实系统必然不公,更不表示努力没有意义。它们阻止一种轻松承诺:把几个好词一起交给工程师,就能无代价地全部实现。冲突得到说明以后,选择才更需要能被看见、解释和重问。

假如贷款机构被问为什么采用这个公平目标,却只答预测已经校准,可能发生范围错位。校准能支持特定预测主张,不自动回答为何选它作为公平标准。先要看原来的问题,再看提交的证明,不因材料很技术便让它回答更多。

自动化也有条件。若输出实际决定结果,或被例行照抄而没有真正的再次判断,同一参数便不断施行其取舍。若有人能够真实改判,情形有所不同。不能看到界面上有人工复核四个字,就宣布每次选择都重新经过审议。

还要问复核能问到哪里。只改一份申请的录入错误,与能够追问目标、代理变量和门槛,不是同样权限。个案得到更正值得肯定,却不因此证明产生同类问题的参数也能复审。两条路可能都需要,各自不能假冒另一条。

参数公开,也只走了一步。人可以看到一个数,仍不知道为什么用它;可以取得代码,仍难找到藏在配置、采购要求或版本记录里的选择。追问成本不仅是能不能读程序,也是能不能从操作找到当初的理由。

这不意味着程序本身有罪。稳定执行可以减少随意变动,记录版本也可以帮助追溯。问题在于,当系统不断复制某项未决取舍,规模放大了,解释与复审的路有没有一起留下。把一切归咎机器,反而容易让采用它的决定退到幕后。

标准合规提供另一处需要分清范围的接口。欧盟《人工智能法》的协调标准安排,对实际覆盖的要求赋予特定合规推定。这样的法律安排并不因存在便有错;推定有所及,也意味着不能不问所及究竟到哪里。

欧委会2026年10月19日的公告,介绍已在官方公报刊登的EN 18286质量管理标准,并说明它关联第17条,推定限于覆盖部分。它不是对组织一切活动、所有具体用途或每项价值取舍的无限保证。

原文没有读到这项标准全文中的对应附录,也没找到把有限证明当成界外问题已答的实际记录。这两个缺口必须保留。不能只因条文位置看起来有疑问,就判定标准违法;修订后的标准化请求、实际文本与对应关系,都需要一起查。

因此,本篇没有宣布欧盟制度已把伦理问题封死。要核的是,一项具体质疑涉及什么,相关证明实际覆盖什么,以及后来有没有人拿这份证明来替未处理的质疑结案。没找到滥用记录,不推它不存在,也不猜找不到的原因。

正式异议也不是一句万能答案。欧委会的产品规则指南说明,对协调标准引用的异议,可以带来保留、限制或撤除引用等处理。它调整特定法律效力,不等于任何人都能在其中重开任何价值问题;能否承接此问,仍须核具体范围和发动条件。

回到评分系统,承担采用与使用的人,不能把一切交给模型原作者。采用并不是物理搬运,它把一套判断放进自己的实际流程。另一方面,采用者也不是突然成了全部算法知识的总担保人。应答责任要落在自己作了什么、凭什么采用、怎样处理新问题。

有人负责,不等于前提都能重问。机构可以很快解释某次输出,却拒绝讨论为什么只预测这一目标;也可能允许讨论,却没有办法发动修改。这是两项需要分别查的事实,不该用一张责任人名单代替。

Passi与Barocas的研究,把注意力移到模型之前。他们在一个企业数据科学团队做了六个月田野观察,展示问题表述如何经协商塑形:要解决什么,用什么指标表示,哪些代理变量算可用。这是具体团队的研究,不是对所有项目的比例估计。

它与这里的追问互相补充,而不是重复一件事。前者让形成阶段的选择可见;这里还要问,形成以后,使用者能从哪里回去要求回应、发动复审。把协商过程记下来,可以留下材料,但记录本身不自动产生这项权能。

设想一位被拒申请者终于得到复核。工作人员修正了收入录入,结果仍未改变。此时若她进一步问,为什么这类不稳定但持续的收入始终不能进入评价,答复不能只是我们已经复核过。前一次检查了输入,不代表后一项前提已经接受了询问。

她也未必应该因此获批。放贷有真实风险,成本不能被一方的需要抹去。机构可以拒绝改变,但需要辨认自己拒绝的是什么,并说明哪个位置能处理这一层问题。可问不是必胜,审慎也不等于永远保留原设定。

更细的定理证明、法律条文对应和标准全文核查,留给进一步阅读。本文只把几处转换留在眼前:目标变变量,取舍变门槛,公平变指标,有限证明被用到何种问题。每一步都可能有理由,没有一步因写进机器就免于理由。

明天第一份申请进入时,昨天的参数也许仍然合适。问题不是每天从头开会,而是当它不再合适、或者有人指出被漏掉的代价时,能不能找到一个真正接得住的入口。代码会照常运行;回答不该因此停下。

設想一家貸款機構使用評分系統。申請到來,分數算出,達到某條線便通過。前台說,系統就是這樣定的;開發者說,自己只是落實需求。那條線並沒有從天而降,卻可能在一次次使用中,逐漸找不到需要解釋它的人。

重要的不是計算機會不會判斷,而是已經作出的選擇,怎樣變成不斷生效的操作。一次設定若每天處理一萬份申請,數量只是設想;它要表達的是,同一項未被問清的取捨,可以不經再次討論而反覆落到不同人身上。

首先就有目標的選擇。預測能否按時還款,不等於判斷一個人值不值得被幫助;判斷適不適合某項貸款,也不等於判斷生活需要。每一種目標可能有理由,卻不能因為最後都輸出一個分數,便互相頂替。

再往前,資料為什麼代表目標?已花的錢未必等於實際需要:沒有花,可能因為根本付不起。在設想的援助項目裡,用過去支出來代表需要,可以算得很準確,卻仍須解釋這項替代為什麼恰當。準確復現替代物,不等於已答目標選擇。

門檻又分配另一種代價。把線抬高,可能減少一種錯誤、增加另一種錯誤。誰被錯拒,誰承擔錯放的成本,不能憑數字寫得整齊便退場。門檻可以保留;但保留需要理由,不能把參數格式當作理由。

連公平也不是唯一現成開關。不同要求可能分別關注,同樣的分數是否有相同意義,合格者被錯拒的比例是否相同,不合格者被錯放的比例是否相同。它們各在保護什麼,需要先說清,不能只給最終指標起一個公平的名字。

克萊因伯格、穆萊納坦與拉加萬的研究,討論一組特定要求:組內校準,以及實際正例與實際負例各自的組間平均分平衡。這裡的正負例指實際結果,不是給人貼好壞標籤。除完美預測或組間基礎發生率相同等條件外,三項不能同時滿足。它不是把任何三種公平都判成不可能。

喬爾德喬娃的另一項研究,則討論陽性預測值一致與兩類錯誤率的一致:在基礎發生率不同、預測並不完美的相關條件下,它們會衝突。兩個結論的定義不能混寫。校準不是誤報率的別名,數學也沒有替我們選出應當犧牲哪一項。

這些定理不證明某個真實系統必然不公,更不表示努力沒有意義。它們阻止一種輕鬆承諾:把幾個好詞一起交給工程師,就能無代價地全部實現。衝突得到說明以後,選擇才更需要能被看見、解釋和重問。

假如貸款機構被問為什麼採用這個公平目標,卻只答預測已經校準,可能發生範圍錯位。校準能支持特定預測主張,不自動回答為何選它作為公平標準。先要看原來的問題,再看提交的證明,不因材料很技術便讓它回答更多。

自動化也有條件。若輸出實際決定結果,或被例行照抄而沒有真正的再次判斷,同一參數便不斷施行其取捨。若有人能夠真實改判,情形有所不同。不能看到介面上有人工複核四個字,就宣佈每次選擇都重新經過審議。

還要問複核能問到哪裡。只改一份申請的錄入錯誤,與能夠追問目標、代理變量和門檻,不是同樣權限。個案得到更正值得肯定,卻不因此證明產生同類問題的參數也能復審。兩條路可能都需要,各自不能假冒另一條。

參數公開,也只走了一步。人可以看到一個數,仍不知道為什麼用它;可以取得程式碼,仍難找到藏在配置、採購要求或版本記錄裡的選擇。追問成本不僅是能不能讀程序,也是能不能從操作找到當初的理由。

這不意味著程序本身有罪。穩定執行可以減少隨意變動,記錄版本也可以幫助追溯。問題在於,當系統不斷複製某項未決取捨,規模放大了,解釋與復審的路有沒有一起留下。把一切歸咎機器,反而容易讓採用它的決定退到幕後。

標準合規提供另一處需要分清範圍的接口。歐盟《人工智慧法》的協調標準安排,對實際覆蓋的要求賦予特定合規推定。這樣的法律安排並不因存在便有錯;推定有所及,也意味著不能不問所及究竟到哪裡。

歐委會2026年10月19日的公告,介紹已在官方公報刊登的EN 18286品質管理標準,並說明它關聯第17條,推定限於覆蓋部分。它不是對組織一切活動、所有具體用途或每項價值取捨的無限保證。

原文沒有讀到這項標準全文中的對應附錄,也沒找到把有限證明當成界外問題已答的實際記錄。這兩個缺口必須保留。不能只因條文位置看起來有疑問,就判定標準違法;修訂後的標準化請求、實際文本與對應關係,都需要一起查。

因此,本篇沒有宣佈歐盟制度已把倫理問題封死。要核的是,一項具體質疑涉及什麼,相關證明實際覆蓋什麼,以及後來有沒有人拿這份證明來替未處理的質疑結案。沒找到濫用記錄,不推它不存在,也不猜找不到的原因。

正式異議也不是一句萬能答案。歐委會的產品規則指南說明,對協調標準引用的異議,可以帶來保留、限制或撤除引用等處理。它調整特定法律效力,不等於任何人都能在其中重開任何價值問題;能否承接此問,仍須核具體範圍和發動條件。

回到評分系統,承擔採用與使用的人,不能把一切交給模型原作者。採用並不是物理搬運,它把一套判斷放進自己的實際流程。另一方面,採用者也不是突然成了全部算法知識的總擔保人。應答責任要落在自己作了什麼、憑什麼採用、怎樣處理新問題。

有人負責,不等於前提都能重問。機構可以很快解釋某次輸出,卻拒絕討論為什麼只預測這一目標;也可能允許討論,卻沒有辦法發動修改。這是兩項需要分別查的事實,不該用一張責任人名單代替。

Passi與Barocas的研究,把注意力移到模型之前。他們在一個企業資料科學團隊做了六個月田野觀察,展示問題表述如何經協商塑形:要解決什麼,用什麼指標表示,哪些代理變量算可用。這是具體團隊的研究,不是對所有項目的比例估計。

它與這裡的追問互相補充,而不是重複一件事。前者讓形成階段的選擇可見;這裡還要問,形成以後,使用者能從哪裡回去要求回應、發動復審。把協商過程記下來,可以留下材料,但記錄本身不自動產生這項權能。

設想一位被拒申請者終於得到複核。工作人員修正了收入錄入,結果仍未改變。此時若她進一步問,為什麼這類不穩定但持續的收入始終不能進入評價,答覆不能只是我們已經複核過。前一次檢查了輸入,不代表後一項前提已經接受了詢問。

她也未必應該因此獲批。放貸有真實風險,成本不能被一方的需要抹去。機構可以拒絕改變,但需要辨認自己拒絕的是什麼,並說明哪個位置能處理這一層問題。可問不是必勝,審慎也不等於永遠保留原設定。

更細的定理證明、法律條文對應和標準全文核查,留給進一步閱讀。本文只把幾處轉換留在眼前:目標變變量,取捨變門檻,公平變指標,有限證明被用到何種問題。每一步都可能有理由,沒有一步因寫進機器就免於理由。

明天第一份申請進入時,昨天的參數也許仍然合適。問題不是每天從頭開會,而是當它不再合適、或者有人指出被漏掉的代價時,能不能找到一個真正接得住的入口。程式碼會照常運行;回答不該因此停下。

Source & reading note原论文与阅读说明原論文與閱讀說明

Ten essays draw on the Disciplines prequel and nine papers. Unattributed applicants, scores, meetings and procedural handovers are imagined. Linked texts, studies and rules retain their dates and evidential limits. The claims that divisions leave something uncontained and that people are more than tools are openly assumed philosophical and normative starting points, not empirical laws. Evidence assessment can itself consider values; professional boundaries, limited authority and reasoned retention are not failures by definition. Missing complete cases are not supplied as established findings; unverified powers are not declared absent, and an unsuccessful search establishes no frequency. These essays prescribe no diagnosis, lending cutoff or clinical action and offer no legal or professional operating advice. English is independently written.

十篇据《学科论》前置篇及九篇正文改写。窗口、评分、会议与程序交接中未另注出处的人物和场景为设想;真实文本、研究与规则另列来源,并注明材料日期和能证明的范围。划分总有未尽、以及人不只是工具,是原作公开承担的哲学与规范起点,不冒充经验定律。证据审查本身可以处理价值,专业边界、有限授权和审后保留不等于失职。原文没有查得的完整实例不补成已证事实,未核不写成不存在;没有发现实例也不据此判断发生频率。这里不决定诊断、信贷门槛或医学处置,不作法律与专业操作建议。简繁中文同文,英文独立重写。

十篇據《學科論》前置篇及九篇正文改寫。窗口、評分、會議與程序交接中未另注出處的人物和場景為設想;真實文本、研究與規則另列來源,並註明材料日期和能證明的範圍。劃分總有未盡、以及人不只是工具,是原作公開承擔的哲學與規範起點,不冒充經驗定律。證據審查本身可以處理價值,專業邊界、有限授權和審後保留不等於失職。原文沒有查得的完整實例不補成已證事實,未核不寫成不存在;沒有發現實例也不據此判斷發生頻率。這裡不決定診斷、信貸門檻或醫學處置,不作法律與專業操作建議。簡繁中文同文,英文獨立重寫。

Read the source paper ↗阅读原论文 ↗閱讀原論文 ↗

Kleinberg, Mullainathan and Raghavan: conditional trade-offs in risk scoringKleinberg、Mullainathan与Raghavan:风险评分公平要求的条件性冲突Kleinberg、Mullainathan與Raghavan:風險評分公平要求的條件性衝突 ↗

Chouldechova: predictive parity and error-rate trade-offsChouldechova:预测值与两类错误率的公平冲突Chouldechova:預測值與兩類錯誤率的公平衝突 ↗

Commission October 2026 announcement: EN 18286 and a bounded presumption欧委会2026年10月公告:EN 18286及有限合规推定歐委會2026年10月公告:EN 18286及有限合規推定 ↗

Article 40 text displayed by the Commission's AI Act Service Desk欧委会AI法服务台展示的第40条文本歐委會AI法服務台展示的第40條文本 ↗

Commission Blue Guide 2022: formal objections to harmonised standards欧委会2022年产品规则蓝色指南:协调标准正式异议范围歐委會2022年產品規則藍色指南:協調標準正式異議範圍 ↗

Passi and Barocas: problem formulation in data sciencePassi与Barocas:数据科学中的问题表述Passi與Barocas:資料科學中的問題表述 ↗