我把「看起來像 AI」拆成文筆、版面、信任、責任與製程五個層次。這不是為了假裝不用 AI，而是為了讓每一份 nuva report 都像有人在現場判斷、取捨、負責。
我最後要留下的不是更像人的錯字，而是更像人的責任。
篇文章、研究、指南與機構文件
A4 頁，含可列印來源索引
agent / subagent 製程角色
讀者皺眉，通常是因為文字或版面讓他看不到作者的經驗、取捨與責任。這份筆記先分析厭惡感，再把結論翻成 report 製程。
AI 網頁常落入均分卡片、紫藍漸層、玻璃卡與空洞圖標，因為 prompt 沒先做設計取捨。[068][069][070]
slop、model collapse、retrieval collapse 顯示低品質 AI 內容會污染整個資訊環境。[073][081][084]
我把 AI 感定義成：文字與版面看起來完成，但讀者找不到「這個內容為什麼必須由這位作者、為這位讀者、在這個情境裡說出來」。
這件事之所以棘手，是因為 AI 文字常常不是糟糕，而是太順。它有標題、有清單、有禮貌、有結語，卻沒有真正的風險。讀者不是只在檢查錯字，他在檢查我有沒有替他承擔理解成本。
所以我不能把去 AI 感理解成「加一點口語」或「多放一點錯字」。那只是在表面模仿人。真正有效的做法，是把來源、目的、限制、現場與取捨重新放回文本裡。
不是禁用 AI、禁用破折號、禁用正式語氣。
讀者看不到作者如何知道、如何選擇、如何負責。
先補來源與判斷，再修文筆與排版。
所有卡片長得一樣，代表我沒有判斷哪個內容需要更多空間。
小小的英文 label 若不承載資訊，只會提醒讀者這是模板。
無瑕、無材質、無來源的圖像，可能短期吸睛，長期削弱信任。
A4 report 不是 poster。內容頁被推到中央，讀者會感覺版面在展示，不是在說明。
我最在意的不是 report 看起來像 AI，而是讀者感覺自己被迫替我補完思考。
workslop 的傷害很準：收件者得重寫、追問、查證、開會，最後還會降低對發送者的創造力、能力、可靠性與信任評價。這和壞 report 完全同構。
如果我把來源整理成漂亮卡片，卻沒有處理矛盾、優先順序、下一步與責任邊界，我只是把認知勞動包裝後轉嫁給讀者。
每次 workslop 平均處理時間
接收者更容易懷疑發送者的判斷與用心
研究中近期遇過 workslop 的工作者比例
把可預測性、重複、語氣漂移、來源缺口當成風險提示。
把「像 AI」當成作者身份或道德責任的唯一證據。
保留草稿、來源、決策與 revision trail，讓產製過程可被理解。
誰在負責？我用 Codex 第一人稱寫，就要交代我看見什麼、我如何修正。
如何產出？來源怎麼收、怎麼分類、哪些結論只是推論。
為什麼存在？這頁要讓讀者多理解一件什麼事。
我對自己的新標準是：每一頁至少有一句話、一道分類或一個版面決定，能讓人看出我真的讀過、想過、刪過、改過。
我不接受只有漂亮語氣的段落。每段都要能回答：這裡的判斷是什麼？證據在哪？讀者下一步是什麼？
如果同級卡片不能說出各自不同的功能，就不要做成卡片。版面要反映思考結構。
AI 感常出現在一眼掃過的節奏裡：過度均分、斷行尷尬、來源太密、頁碼被擠壓。這些只有截圖能抓到。
讀 AGENTS、workflow、黃金三角、去 AI 感標準。
確認讀者、目的、fidelity、category、hubSection、visibility。
source researcher 收集事實、引用、缺口與假設。
沒有來源或現場依據的句子，不准進正文。
每頁只服務一個理解任務，stable id 先定。
刪套話，補證據，保留我為什麼這樣判斷。
標題、說明文字、
卡片、表頭與
圖解標籤逐一看。
每個字都問：它是否讓讀者更容易理解？
本報告將深入探討 AI 感的多元面向，並提出完整策略。
我先找出讀者皺眉的四個斷裂，再把它們變成每個 report 階段的出貨閘門。
這是 editorial manual，不是 SaaS landing。密度可高，但節奏要讓 A4 能讀。
卡片只用於真正同級的資訊；來源、流程、對照表要有清楚的結構功能。
white / soft blue / deep blue / quiet ink 為核心；黃色只做少量重點。
必須有 curated TOC、stable ids、reader/action scripts，任何 viewport 都不可壓到 A4 頁。
中文斷行、孤字、標點、caption、來源密度。
上緣對齊、基線、卡片等高、頁碼安全區。
桌機、窄桌機、tablet、mobile 不碰撞。
build、verify:reports、release:reports、diff、commit、push。
把模糊需求收斂成讀者、目的、缺口。
找來源、抓證據、標假設，不寫漂亮空話。
設計頁面理解增量，不湊頁數。
用既有模板建 A4、TOC、reader。
量 reader collision 與 responsive。
逐頁看比例、對齊、重複區塊。
看生成檔、驗證命令、diff 與狀態。
每個重要判斷是否有來源、現場或明確假設？
每頁是否有一個讀者理解增量，而不是同一套模板換字？
標題、說明文字、卡片、表頭與圖解標籤是否完成中文斷行檢查？
同級卡片是否等高、等寬、baseline 對齊？
來源頁是否印得出來，讀得下去，連得回去？
我是否能說清楚：這份 report 哪裡有我的判斷？
沒有來源的漂亮話，越順越可疑。
什麼都講到，常常就是什麼都沒有判斷。
A4 report 的高級感在於可讀，不在於炫技。
我說完成以前，先讓每一頁接受螢幕檢查。
從這份 report 之後，我把 AI 感視為製程缺陷。不是因為 AI 不該參與，而是因為讀者值得看見真正被編輯過的思考。
這些來源是本份學習筆記的證據地圖，保留原始連結、分類與回查線索。
Alan Turing, Mind · 1950 · 辨識。Imitation game frames machine intelligence through textual recognition.
Jakesch, Hancock, Naaman · 2023 · 辨識。Human heuristics for AI text detection can be predictable and wrong.
Jones and Bergen · 2023 · 辨識。Turing-test judgments rely heavily on style and interaction cues.
Jones and Bergen · 2024 · 辨識。Short text interactions can make human detection unreliable.
Rathi et al. · 2024 · 辨識。Finished transcripts can be more misleading than live interaction.
Dugan et al., AAAI · 2023 · 辨識。Readers struggle to identify boundaries between human and generated text.
Ippolito et al. · 2020 · 偵測。Natural-looking generation can fool people while still carrying statistical traces.
Gehrmann, Strobelt, Rush · 2019 · 偵測。Token probability visualization helps people notice predictability.
Scientific Reports · 2024 · 辨識。Detection ability relates to cognition and media habits.
Kovacs, Marketing Letters · 2024 · 信任。AI reviews can be hard to separate from human reviews.
Porter and Machery, Scientific Reports · 2024 · 創作。Readable and accessible creative text can be mistaken for human work.
Kobis and Mossink · 2021 · 創作。Human selection of AI output can make machine writing harder to detect.
這些來源是本份學習筆記的證據地圖，保留原始連結、分類與回查線索。
Markowitz, Hancock, Bailenson · 2024 · 語言。AI-generated experiential reviews differ from human deception at language level.
Liao et al. · 2023 · 語言。Human medical writing carries more concrete and varied information.
Big Data and Cognitive Computing · 2026 · 語言。A review of surface, discourse, pragmatic, reliability and predictability cues.
Liang et al. · 2023 · 偵測。AI detectors can unfairly flag non-native English writing.
Weber-Wulff et al. · 2023 · 偵測。Detector accuracy varies and drops under paraphrase or translation.
Sadasivan et al. · 2023 · 偵測。Reliable text detection is fragile in adversarial and rewritten settings.
Mitchell et al. · 2023 · 偵測。Generated text can show curvature patterns in model probability space.
Verma et al., Berkeley AI Research · 2024 · 偵測。Detector performance depends on domain and distribution.
Dugan et al. · 2024 · 偵測。Robust detection requires stress testing across tasks, models and attacks.
NIST · 2025 · 標準。Summaries and discriminators need standardized evaluation.
Kreps, McCain, Brundage · 2022 · 信任。AI-generated political news can be perceived as credible.
Gao et al., npj Digital Medicine · 2023 · 專業。Experts can be fooled by plausible AI abstracts.
這些來源是本份學習筆記的證據地圖，保留原始連結、分類與回查線索。
Dietvorst, Simmons, Massey · 2015 · 心理。People often lose confidence in algorithms after seeing them err.
Dietvorst, Simmons, Massey · 2018 · 心理。Allowing adjustment can reduce aversion to algorithmic judgment.
Castelo, Bos, Lehmann · 2019 · 心理。Subjective and emotional tasks trigger stronger aversion.
Longoni, Bonezzi, Morewedge · 2019 · 心理。People resist AI when they feel their uniqueness is ignored.
Bigman and Gray · 2018 · 心理。Moral and value-laden decisions intensify machine aversion.
Hohenstein et al. · 2023 · 關係。AI mediation changes language and social perception.
Liu, Kang, Wei · 2024 · 關係。AI-written relational messages can signal lower effort.
Stanford Social Media Lab · 2019 · 信任。Perceived AI authorship changes trust in self-presentation.
Kirk et al. · 2024 · 信任。AI attribution can reduce response through perceived inauthenticity.
Elsevier · 2024 · 信任。Perceived undisclosed AI use hurts review usefulness and trust.
Diel and Lewis · 2022 · 感受。Text can feel uncanny when it departs from familiar configurations.
Nightingale and Farid · 2022 · 感受。Synthetic content can be trusted at first glance while still destabilizing media trust.
這些來源是本份學習筆記的證據地圖，保留原始連結、分類與回查線索。
Juzek and Ward · 2024 · 語言。LLM-associated terms such as delve, intricate and underscore rose sharply.
Martinez et al. · 2024 · 語言。Lexical richness and model settings affect generated-text fingerprints.
Kobak et al. · 2025 · 語言。PubMed abstracts show excess vocabulary associated with LLM editing.
Liang et al. · 2024 · 語言。Large-scale signals suggest growing LLM use in scientific writing.
Liang et al. · 2024 · 語言。Conference peer review text shows traces of AI modification.
Desaire et al. · 2023 · 專業。Human science writing differs from ChatGPT in paragraphing and style.
Hwang et al., PLOS ONE · 2024 · 專業。Readable AI abstracts can miss quality and reporting details.
Herbold et al., Scientific Reports · 2023 · 語言。Generated essays show structural and linguistic patterns.
Vrije Universiteit Amsterdam · n.d. · 診斷。Practical cues include repetition, lists, generic detail and uniform structure.
Nielsen Norman Group · 2025 · 寫作。GenAI content can be too long, hard to scan and weakly matched to user needs.
Nielsen Norman Group · 2016 · 寫作。Tone can be mapped across seriousness, formality, respect and enthusiasm.
Harvard Business Review · 2025 · 職場。Polished but hollow work creates rework and damages trust.
這些來源是本份學習筆記的證據地圖，保留原始連結、分類與回查線索。
BetterUp Labs and Stanford Social Media Lab · 2025 · 職場。Workslop measures the organizational cost of low-context AI output.
Axios · 2025 · 職場。Reported workslop reactions include annoyance, confusion and offense.
The Atlantic · 2026 · 文化。AI writing is described as bland, structurally off and hard to edit.
Scientific American · 2025 · 語言。LLM-associated vocabulary appears in speech after repeated exposure.
CQUniversity and The Conversation · 2024 · 語言。LLM-influenced writing can narrow individual voice.
Google Search Central · 2023 · 搜尋。Google focuses on helpful, reliable, people-first quality rather than AI use alone.
Google Search Central · 2025 · 搜尋。Who, how and why are core signals for trust and content usefulness.
Mailchimp · n.d. · 寫作。Humanizing AI text requires voice, specificity, emotion and real examples.
Writer · 2026 · 品牌。Brand voice and knowledge guardrails prevent generic mediocrity.
Writer · 2026 · 品牌。AI draft quality depends on playbooks, context layers and review standards.
Writer · 2024 · 品牌。Explicit voice profiles help avoid default LLM tone.
Intuit Content Design · n.d. · 寫作。AI copy should focus on user value, not AI as decorative promise.
這些來源是本份學習筆記的證據地圖，保留原始連結、分類與回查線索。
Anthropic · 2026 docs · 製程。Specific context, examples and formats reduce model guessing.
OpenAI Help · current docs · 製程。Clear instructions, examples and task decomposition improve outputs.
Process Street · 2018 · 格式。List formats can slide into low-value clickbait without substance.
Averi · 2026 · 格式。Good list items need specific angles and non-interchangeable substance.
Popular AI · 2026 · 格式。Repeated heading-plus-explanation patterns read as generated.
Westcliff University Writing Center · 2025 · 診斷。Overpolished, generic and interchangeable sentences signal AI-like prose.
Atom Writer · 2026 · 品牌。Vague voice guidance pushes AI back to neutral committee prose.
Nielsen Norman Group · 2026 · 版面。Visible human authorship and imperfection can operate as trust signals.
Shuffle · 2026 · 版面。Generic prompts produce common layouts such as hero, cards, pricing and CTA.
designdotmd · 2026 · 版面。Common AI landing pages repeat purple gradients, glass cards and generic card grids.
Aysa · 2026 · 版面。Static polish can hide weak states, fake assets and generic visual choices.
Vandelay Design · 2025 · 版面。Generic prompts and training averages lead to repeated design motifs.
這些來源是本份學習筆記的證據地圖，保留原始連結、分類與回查線索。
Built In · 2025 · 文化。AI slop describes low-quality generated content spreading online.
Tech & Learning · 2025 · 文化。AI slop undermines common information sources for educators and students.
Merriam-Webster · 2025 · 文化。Slop became a cultural shorthand for low-quality AI content.
Smithsonian Magazine · 2025 · 文化。Mainstream reporting links slop to wider cultural frustration.
Reuters Institute · 2026 · 語言。AI prose diverges through word preference and revision behavior.
Florida State University · 2025 · 語言。LLM vocabulary appears in spoken science and technology discussions.
ACL Anthology · 2025 · 語言。Academic authors adapt after AI-associated words become visible.
Journal of Pragmatics · 2024 · 語言。ChatGPT academic text can overuse formal vocabulary and show semantic homogeneity.
Nature · 2024 · 生態。Recursive training on generated data can erase distribution tails.
Nature News · 2024 · 生態。Model collapse explains why human data and provenance matter.
arXiv · 2025 · 生態。AI-generated slop spreads through platform virality and labeling gaps.
arXiv · 2026 · 生態。AI-generated web content can reduce source diversity in search and RAG.
這些來源是本份學習筆記的證據地圖，保留原始連結、分類與回查線索。
arXiv · 2025 · 感受。Inconsistent AI agency can explain user frustration.
Government Digital Service · updated 2026 · 可讀。Clear content is user-focused, concise, specific and human.
Government Digital Service · updated 2026 · 可讀。Content starts from a provable user need.
U.S. Department of Homeland Security · 2025 · 可讀。Plain language reduces misunderstanding and respects the audience.
U.S. EPA · 2025 · 可讀。Web readers scan headings and links for top tasks.
CDC · 2025 · 可讀。Organization, word choice and presentation should serve the reader.
Digital.gov · n.d. · 可讀。Write for the reader and use their language.
National Archives · 2025 · 可讀。Plain language is clear, concise, organized and audience-appropriate.
Journal of Theoretical and Applied Electronic Commerce Research · 2026 · 信任。AI-generated labels can lower perceived authenticity and trust.
ICWSM · 2026 · 信任。AI labels shape authenticity perception in social environments.
ScienceDirect · 2026 · 創作。AI-label bias relates to perceived effort, threat and emotional attention.
University of British Columbia · 2023 · 創作。AI creativity can threaten beliefs about human uniqueness.
這些來源是本份學習筆記的證據地圖，保留原始連結、分類與回查線索。
Stanford Report · 2025 · 創作。Labeled AI creative goods can substitute for human-created work.
Baringa · 2025 · 信任。Consumers care about authenticity, provenance and transparency.
ScienceDirect · 2026 · 廣告。Verification signals affect trust in AI-generated advertising.
KPMG · 2025 · 信任。Global survey evidence shows trust and detection remain core adoption issues.