保育團體用 AI 找火、數象，資料中心的電力與用水也得算進去

Reuters 於 11 月 7 日刊出對 15 個環境、保育及發展組織的採訪。受訪組織均表示已在某些工作中使用 AI，也對資料中心的環境影響表達關切。這是對受訪組織的觀察，不能據此推算整個保育部門的採用率。較值得關注的是，從偵測到派員處置，哪些步驟真的變快，哪些成本只是移到雲端？

In a report first published on 7 November, Reuters said all 15 environmental, conservation and development organisations it interviewed used AI in some form and were concerned about the environmental impact of the data centres behind it. The interviews describe the experience of those organisations; they are not a survey of the entire sector. The next question is whether a faster detection leads to a useful response in the field, and at what cost.

柬埔寨洞里薩湖周邊，保育國際（Conservation International）試行火災偵測與警報，協助社區巡守隊縮短從辨識火點到現地回應的時間。在肯亞，Save the Elephants 與牛津大學的研究則評估無人機在特定飛行條件下如何用於觀察象群。兩個案例的驗收方式不同：前者看通報是否導致行動，後者還要看影像能否用於研究，以及飛行是否干擾動物。

Around Cambodia’s Tonle Sap, Conservation International has piloted fire detection and alerts to give community patrols more time to respond. In Kenya, Save the Elephants and the University of Oxford studied how drones could observe elephant families under particular flight conditions. The use cases call for different checks. A fire alert needs a response, while an aerial survey needs usable observations and a way to check whether the flight disturbed the animals.

保育團隊可能記得一次成功撲滅的火情，卻未留下未接獲警報的火情、誤報或空跑紀錄。沒有這些紀錄，難以比較導入前後的效果。資料中心一端也有類似問題：供應商的全年平均用電或用水，無法直接回答某個模型、某個機房區域或某個任務的資源消耗。機房冷卻用水與發電端用水，還需分開辨識。

A successful response is easy to remember; a missed detection, a false alarm or a dispatch that yielded nothing is harder to account for. Without those records, teams cannot compare the new system fairly with the old one. On the computing side, a provider’s annual footprint does not identify the resources used by a particular model or job. Water used to cool equipment and water associated with generating its electricity are also different measures.

以森林火警為例，可先建立一條能回查的工作紀錄：保存原始衛星熱點、模型版本與判定門檻，接著記下警報發送、接收、派遣和現場確認的時間。通訊中斷、人工覆核推翻模型判斷、巡守隊無法到場，也應留在同一紀錄中。野生動物調查則可抽樣以人工標註複核模型計數，並保留飛行高度與動物行為觀察。

A fire pilot can retain the original satellite detection, the model version and alert threshold, and then record when the alert was sent, received, acted on and confirmed. Connectivity failures, human overrides and cases in which crews could not respond belong in the same record. Wildlife survey teams can sample model counts against human labels and retain flight settings and observations of animal behaviour.

保育國際的專案紀錄說明，較快的火警可協助巡守人員爭取處置時間；Save the Elephants 的研究則提供特定場域的無人機觀察依據。這些資料支持個別用途，尚不足以證明 AI 對保育整體有淨環境效益。另一端，國際能源總署（IEA）在 2025 年《Energy and AI》估計，全球資料中心用電由 2024 年約 415 TWh，在基準情境下到 2030 年將增至約 945 TWh。這是整體資料中心情境值，不能直接歸給單一 AI 服務或保育專案。

Conservation International’s project account shows how earlier alerts can give patrols more time. Save the Elephants’ field research supports aerial observation under the conditions studied. Neither establishes a sector-wide net environmental benefit from AI. In its 2025 Energy and AI base case, the IEA estimated global data centre electricity use at about 415 TWh in 2024 and projected about 945 TWh in 2030. Those figures cover data centres as a whole. They cannot be assigned directly to an individual AI tool or conservation project.

國際能源總署指出，資料中心用電的相關排放取決於供電結構。美國勞倫斯柏克萊國家實驗室的研究也區分機房冷卻用水與發電相關用水。對雲端採購而言，若比較兩個方案，應先問是否用同一工作量、機房區域及期間衡量；對保育資助者而言，則應問有多少警報經過確認，又有多少促成實際行動。

The IEA links emissions from data centre electricity use to the power mix. Lawrence Berkeley National Laboratory also distinguishes water used directly for cooling from water used in producing electricity. For cloud buyers, that makes comparable workloads, hosting regions and reporting periods important. For conservation funders, it makes confirmed alerts and documented responses more useful than a raw count of detections.

國際保育機構將模型接到巡護流程，台灣採用類似工具時，雲端供應商提供的資料粒度會影響採購方能否同時評估現場成效與環境成本。SNN 編輯分析｜揭露前證據基礎設施：保育專案主持人應在擴大布建前，把衛星警報、巡護派遣與現地結果對照，公布誤報率及確認時間，並與既有做法比較。雲端採購與資安團隊應在招標或續約時，要求提供工作量紀錄、運算區域與期間、可取得的用電和冷卻用水資料，以及資料保留條款，若供應商只能提供估算就標明邊界。資助方或驗收單位應在季末付款前，抽查模型版本紀錄、標註樣本與人工覆核結果，再判斷是否擴大採用。以上是台灣應用建議，不表示台灣已有這些法定要求。

As conservation organisations connect models to field operations, Taiwan buyers of similar tools depend on their cloud vendors for information that can link computing costs to results. SNN editorial analysis: Three records-based checks form the Pre-Disclosure Evidence Infrastructure for this decision. Project team leads should reconcile satellite alerts, patrol dispatch logs and field confirmations before expanding a pilot, using the previous workflow as a baseline for false alarms and response time. Cloud procurement and security teams should ask before tender or renewal for workload records, computing region and period, available electricity and cooling-water data, and retention terms. Where the supplier can provide only estimates, their boundaries should be explicit. Funders and acceptance reviewers should sample model change logs, labelled observations and human overrides before quarterly payment or a decision to scale. This analysis does not establish a legal requirement for deployments in Taiwan.

後續可追蹤三項可核驗變化：火警是否更常在仍可處置時被確認、野生動物辨識是否在不同季節維持準確，以及供應商能否提供足以比較的工作量與資源資料。若工具在某些場域表現較佳，下一步是說明哪些條件讓它奏效，避免把一地成效直接外推到其他生態系。

The useful tests are whether more fires are confirmed while crews can still respond, whether wildlife detection remains reliable across seasons, and whether providers can supply comparable workload and resource data. A successful pilot should also explain the conditions under which it worked, so that a result from one ecosystem is not assumed to hold elsewhere.