部署之後：為何高風險 AI 需要持續證據系統

歐盟執委會的高風險 AI 研究指出，部署不是符合性工作的終點。系統開始營運後，資料、版本、使用者、整合方式與預定用途都可能改變，原有證據因而逐步失去代表性。 高風險 AI 的制度問題並不在符合性文件是否完成，而在系統上線後，文件能否繼續代表實際狀態。資料、模型、介面、部署場域、使用者行為與預定用途會逐步改變；任何一項變更都可能讓原風險評估、測試與責任配置失去部分有效性。 本期 Analysis 因而不把 高風險 AI 部署後的證據衰減與持續治理 視為單一技術或政策更新，而是把官方主錨點可證明的制度事實、SNN 的分析推論與仍待驗證的結果分層處理。讀者必須能看見主張從何而來、推論在哪裡開始，以及哪些結論目前不能由來源直接支持。 為使摘要能直接支援決策，本段同時回答五個問題：已發生什麼、由哪一第一方紀錄支持、影響透過什麼機制傳導、目前仍缺少哪些結果證據，以及下一個可推翻或強化判斷的檢查點。任何未被來源明示的因果關係都保留為編輯推論，不以肯定語氣包裝。

The Commission study indicates that deployment is not the end of conformity. Once a high-risk AI system operates, its data, version, users, integrations and intended purpose may change, weakening the representativeness of earlier evidence. The institutional problem of high-risk AI is not whether conformity documents were completed, but whether those documents continue to represent the operating system after deployment. Data, models, interfaces, deployment environments, user behaviour and intended purpose can change incrementally. Any one of those changes may reduce the continuing validity of the original risk assessment, testing and allocation of responsibility. This Analysis therefore does not treat evidence decay and continuous governance after deployment of high-risk AI as a self-contained technical or policy update. It separates the institutional facts supported by the official anchor, SNN editorial inference and outcomes that remain unverified. The reader should be able to see where the source ends, where interpretation begins and which conclusions the present evidence cannot support. To make the lead decision-ready, it answers five questions together: what has occurred, which first-party record supports it, through what mechanism the effect may travel, which outcome evidence is still missing, and what next observation could strengthen or overturn the judgement. Any causal relationship not stated by the source remains an editorial inference and is not converted into a factual claim through confident wording.

高風險 AI 的供應商、整合商與部署者分別控制不同環節。傳統治理多在上線前完成分類、風險評估、技術文件與符合性判斷，之後才進入市場後監測。 AI Act 將供應商、進口商、經銷商與部署者置於不同責任位置，實際系統又可能由基礎模型、應用、客製設定與客戶流程組合。每個主體掌握的資訊不同。若價值鏈只交換合規聲明而不交換版本、用途與變更事件，後端使用者無法判斷手上的證據適用哪一配置。 為了維持分析密度，制度背景必須同時標示規則制定者、執行者、資料擁有人、覆核者與受影響市場。這些角色可能由不同組織或部門承擔；文件發布、系統上線、企業採用與結果交付也不是同一證據狀態。 時間與權限也必須分開記錄：公告日不等於生效日，試點不等於普遍採用，技術規格不等於法律義務，企業自願導入也不等於監理核准。把這些節點放在同一時間軸，才能判斷一項制度變動何時真正進入資料、合約、投資或揭露流程。

Providers, integrators and deployers control different parts of the AI lifecycle. Conventional governance concentrates classification, risk assessment, technical documentation and conformity decisions before release, followed by post-market monitoring. The AI Act places providers, importers, distributors and deployers in different responsibility positions, while a real system may combine a foundation model, application, custom configuration and customer process. Each actor controls different information. If the value chain exchanges only a conformity statement without version, purpose and change events, a downstream user cannot determine the configuration to which the evidence applies. Institutional context must identify the rule setter, implementer, data owner, reviewer and affected market. Those roles may sit in different organisations or functions, and publication of a document, deployment of a system, enterprise adoption and delivery of an outcome are different evidence states. Time and authority must also be separated. An announcement date is not an effective date; a pilot is not general adoption; a technical specification is not a legal obligation; and voluntary enterprise use is not regulatory approval. Putting these events on one timeline shows when an institutional development actually enters data, contract, investment or disclosure processes and which actor is authorised to make that transition.

真正的風險不只在系統失效，也在文件仍存在、卻已無法完整描述實際運作狀態。責任分散與漸進式修改，使治理邊界及應由誰重新評估變得不清楚。 Evidence Decay 不等於原文件造假，而是證據與實際系統之間的距離隨時間增加。模型權重不變，輸入人口可能已改變；軟體版本不變，使用情境可能超出預定用途；監測發現偏差，矯正措施也可能未回寫風險檔案。文件仍存在，卻不再完整支持當前制度判斷。 上游證據若沒有穩定識別、形成時間、適用邊界與版本，最終輸出即使格式一致，覆核者仍可能無法重建形成過程。真正風險不是單一欄位缺失，而是錯誤主體、過期方法、推定關係或未核准版本在傳遞中被當成確定事實。 具體失效模式至少包括主體配對錯誤、資料人口不完整、邊界前後不一致、方法或係數未版本化、例外未留下理由、核准與發布時間倒置，以及下游重用時脫離原用途。每一種失效都可能讓合理的單筆資料，在彙總後形成無法防禦的結論。

The risk is not limited to system failure. Records may remain available while no longer describing the operating system. Distributed responsibility and incremental modifications make governance boundaries and reassessment duties harder to determine. Evidence Decay does not mean that the original document was false. It means that the distance between evidence and the actual system increases over time. Model weights may remain unchanged while the input population shifts; software may remain stable while use exceeds intended purpose; monitoring may identify bias while remediation is not reflected in the risk file. The documents exist but no longer support the present judgement completely. When upstream evidence lacks stable identity, formation time, applicable boundary and version, a standardised output may still be impossible to reconstruct. The material risk is not one missing field. It is the silent conversion of the wrong entity, an expired method, an inferred relationship or an unapproved version into an apparent fact as information moves downstream. Concrete failure modes include incorrect entity matching, an incomplete data population, inconsistent boundaries, unversioned methods or factors, exceptions without rationale, approval occurring after publication, and downstream reuse outside the original purpose. Each failure can turn a reasonable individual record into a conclusion that cannot be defended after aggregation, comparison or machine-assisted interpretation.

可行作法是建立持續證據系統，把系統版本、用途、資料、責任主體與治理邊界連結起來；重大變更發生時，判斷哪些證據仍有效、哪些需要補充或重新評估。 持續證據系統應為系統、模型、資料集、部署、用途、角色、風險、控制、事件與矯正建立穩定身分。每次更新觸發 evidence impact assessment，明確判斷哪些測試、文件與核准仍有效、哪些需要補充或重作；重大修改與一般維護的判斷理由也必須保留。 可執行的控制單位應是受治理的證據物件：每項重要主張連接原始來源、計算或判斷方法、組織與時間邊界、責任人、控制狀態、例外、核准及版本。當任何元件改變時，系統保存差異與影響範圍，不以覆寫舊檔取代變更紀錄。 營運上可建立最小控制集：主張登錄、證據擁有人、來源快照、方法識別、適用期間、控制頻率、例外門檻、覆核與核准，以及允許的下游用途。高判斷或高財務影響項目採更嚴格的覆核層級；低風險資料則以自動化完整性檢查降低重工。

A continuous evidence system should bind system versions, purposes, data, responsible actors and governance boundaries. When material change occurs, it should identify which evidence remains valid, which needs supplementation and which requires reassessment. A continuous evidence system assigns stable identities to systems, models, datasets, deployments, uses, roles, risks, controls, incidents and corrective actions. Every update triggers an evidence impact assessment that identifies which tests, documents and approvals remain valid and which require supplement or repetition. The reasoning used to distinguish substantial modification from ordinary maintenance is also preserved. The implementable control unit is a governed evidence object. Each material claim links to its primary source, calculation or judgement method, organisational and temporal boundary, accountable owner, control state, exception, approval and version. When any component changes, the system preserves the difference and affected uses instead of overwriting the earlier basis. A minimum operating control set includes a claim register, evidence owner, source snapshot, method identity, valid period, control frequency, exception threshold, review, approval and permitted downstream use. High-judgement or high-financial-impact items receive a stronger review tier. Lower-risk records use automated completeness and consistency checks so that governance effort is concentrated where a wrong claim would change a decision.

研究彙整 544 份公開諮詢回覆、三場專家工作坊的 166 次參與紀錄及後續問卷，反覆出現預定用途、角色責任、價值鏈資訊交換與重大修改等執行疑問。 歐盟研究能支持諮詢規模、實作疑問與角色、預定用途及重大修改等主題的重要性；AI Act 則提供法律架構。兩者都不能證明所有高風險系統會以相同方式衰減。本文的 Evidence Decay 是用來描述證據適用性變化的分析概念，不替代法定分類或個案符合性判斷。 證據判讀以本期官方主錨點為起點，再以獨立第一方或具方法透明度的來源確認背景與邊界。來源能支持的是已發布的制度事實與明確數據；是否代表全面採用、因果改善或跨市場普遍有效，必須另有實作結果才能成立。 每一項關鍵句都應進入 claim ledger，標示其為官方事實、直接量測、估算、企業承諾、已交付結果或 SNN 編輯推論，並記錄來源所能支持的精確範圍。若兩筆來源衝突，保留衝突與處理理由；若資料尚未出現，明確標成待驗證，而不是以相近案例補成確定答案。

The study draws on 544 consultation responses, three expert workshops with 166 recorded participations and a follow-up survey. Intended purpose, role allocation, value-chain information and substantial modification repeatedly emerged as implementation concerns. The Commission study supports claims about consultation scale and the importance of intended purpose, roles, value-chain exchange and substantial modification; the AI Act provides the legal architecture. Neither establishes that every high-risk system decays in the same way. Evidence Decay is an analytical concept describing changed applicability and does not replace legal classification or case-specific conformity judgement. Evidence assessment begins with the official anchor and uses independent primary or method-transparent sources to test context and limits. The sources support stated institutional facts and explicit figures. Claims of comprehensive adoption, causal improvement or universal cross-market effectiveness require separate implementation evidence. Every material sentence should enter a claim ledger and be classified as official fact, direct measurement, estimate, corporate commitment, delivered outcome or SNN editorial inference. The ledger records the precise scope that each source supports. Conflicting evidence is retained with the resolution rationale; absent evidence is marked pending rather than filled with a convenient analogue from another entity, period or jurisdiction.

這表示符合性將從單次結論轉為持續維護。監測只能發現變化，治理還必須辨識邊界如何變動、責任由誰承擔，以及先前證據能否繼續支持制度判斷。 治理效果因此取決於組織能否把監測訊號轉成證據狀態變更。只蒐集效能與事件指標，卻不更新風險、責任與使用邊界，會形成資料很多但制度判斷停滯的情況。董事會與監管者需要看見的不只是系統是否被監測，也包括哪些原結論因此被維持、限制或撤回。 這種資訊密度的目的不是增加篇幅，而是縮短主張到決策之間的查核距離。董事會、投資人、監理者與營運團隊應能辨識哪些內容是事實、估算、承諾、進度或結果，並在情境改變時更新判斷，而不是重新尋找散落的檔案與口頭說明。 責任分工因此必須落到決策權：資料擁有人維護來源，方法擁有人控制計算，業務單位說明使用情境，內控或確信人員測試可重建性，核准者承擔最終採用責任。例外若沒有到期日、補救人與影響範圍，就會從暫時處置變成永久的證據債務。

Conformity therefore becomes a maintained condition rather than a one-time conclusion. Monitoring may detect change, but governance must also determine boundary shifts, responsibility and whether previous evidence still supports institutional reliance. Governance effectiveness therefore depends on whether monitoring signals change evidence status. An organisation that collects performance and incident indicators without updating risk, responsibility and use boundaries has more data but a static institutional judgement. Boards and regulators need to see not only that the system was monitored, but which prior conclusions were maintained, constrained or withdrawn as a result. The purpose of this information density is not length for its own sake. It is to shorten the verification distance between claim and decision. Boards, investors, regulators and operational teams should be able to distinguish fact, estimate, commitment, progress and outcome, then update the judgement when conditions change without reconstructing the case from scattered files and oral explanation. Accountability therefore attaches to decision rights. The data owner maintains the source, the method owner controls calculation, the business function defines the use case, internal control or assurance tests reproducibility, and the approver accepts responsibility for final use. An exception without an expiry date, remediation owner and impact scope stops being temporary treatment and becomes persistent evidence debt.

SNN 編輯分析：歐盟 AI Act 的持續治理要求，將透過歐洲客戶與跨境產品責任影響台灣半導體、資通訊、製造、醫療與金融業。資料集、模型版本、系統整合、用途、部署場域或責任角色改變後，原證據是否仍有效必須重新判斷；台灣 AI 落地不能停在上線前文件。 台灣半導體、資通訊、醫療、金融與製造企業向歐洲提供 AI 元件或部署系統時，可能同時扮演供應商、整合商與部署者。企業應讓模型版本、訓練或輸入資料、客戶設定、部署用途、事件與矯正跨組織傳遞，並區分台灣自願治理框架與歐盟法律義務，避免一份初始測試報告被無限延用。 對台灣市場的意義應沿實際傳導路徑判斷：國際制度或客戶要求先進入融資、採購、合約、供應商資料與確信流程，再影響台灣企業的系統與控制。這不等於外國規則自動成為台灣法律義務；本地企業要做的是辨識適用情境、保留跨語言對應並建立可授權覆核的證據鏈。 台灣企業可把傳導測試落在具體物件：公司與法律主體、廠區、產品、批次、供應商、合約、融資工具及揭露欄位。中英文名稱、內外部分類與不同框架之間應連到同一主張識別，同時保留商業機密、個資與權限邊界，避免可追溯被誤解為全面公開。

SNN editorial analysis: EU AI Act continuous-governance expectations reach Taiwan semiconductor, ICT, manufacturing, healthcare and finance through European customers and cross-border product responsibility. Changes to datasets, model versions, integrations, uses, deployment environments or responsible roles require renewed evidence-validity decisions. Taiwan semiconductor, ICT, medical, financial and manufacturing companies supplying AI components to Europe or deploying systems may act as providers, integrators and deployers simultaneously. Model versions, training or input data, customer configuration, use, incidents and remediation need to travel across organisational boundaries, while Taiwan voluntary frameworks remain distinct from EU legal obligations. One initial test report cannot be treated as permanently valid. For Taiwan, relevance should be traced through an actual transmission path. An international rule or customer requirement first enters finance, procurement, contract, supplier-data and assurance processes, then changes local systems and controls. It does not automatically become Taiwan law. Companies need to identify the applicable scenario, preserve bilingual mappings and make the evidence chain reviewable under controlled access. Taiwan companies can perform the transmission test on concrete objects: the company and legal entity, facility, product, batch, supplier, contract, financing instrument and disclosure field. Chinese and English names, internal and external classifications and different reporting frameworks should resolve to the same claim identity. Traceability must still preserve commercial confidentiality, personal data and access boundaries; it does not require unrestricted publication.

未來高風險 AI 的可信治理，將取決於證據能否隨部署、更新、事件與矯正行動持續回流。編輯審核應特別核對官方研究、AI Act 適用範圍與各角色責任。 後續應追蹤歐盟指引、標準、重大修改判準與市場後監測實務。低後悔行動是選擇一個已部署 AI 系統，回溯最近六個月的所有資料、設定、用途與責任變更，逐一判斷它們是否影響原測試及核准，並為未被記錄的變更建立補正與再評估機制。 後續追蹤應分開最終文件、技術指引、採用範圍、執行控制、監督活動與可觀察結果。現階段的低後悔行動，是挑選一項高風險主張做端到端重建測試，記錄缺少的身分、來源、方法、責任與版本；這是治理準備，不是合規保證或結果預測。 監測不只依日曆，而應由事件觸發：最終規則發布、技術指引變更、適用範圍擴大、主管機關執法、企業採用數據或實際成果出現時，都建立新版本並重評原判斷。舊結論不刪除，而是標示當時依據、被何種新證據修正，以及哪些決策需要重新檢視。

Future high-risk AI governance will depend on evidence flowing back from deployment, updates, incidents and corrective actions. Human review should verify the official study, the AI Act scope and the responsibilities assigned to each actor. Future evidence should follow EU guidance, standards, substantial-modification criteria and post-market monitoring practice. A low-regret action is to select one deployed AI system, reconstruct six months of data, configuration, purpose and responsibility changes, assess the impact on every original test and approval, and create remediation and reassessment controls for changes that were not recorded. Future monitoring should separate final text, technical guidance, adoption scope, operating controls, supervision and observable outcomes. A low-regret step is to select one high-risk claim for an end-to-end reconstruction test and record missing identity, source, method, accountability and version. That is governance preparation, not a compliance guarantee or forecast of results. Monitoring should be event-triggered as well as calendar-based. A final rule, amended technical guidance, expanded scope, supervisory action, adoption data or observed outcome creates a new version and a reassessment of the earlier judgement. The prior conclusion is not erased. It retains its original basis, identifies the new evidence that changed it and states which decisions or downstream uses now require review.