Generative AI Is Poised to Worsen the E-Waste Crisis生成式人工智能或將加劇電子垃圾危機

Generative AI could saddle the planet with1 heaps of more hazardous waste.

生成式人工智能可能產生大量有害垃圾，加重地球負擔。

Every time generative artificial intelligence drafts an e-mail or conjures up an image， the planet pays for it. Making two images can consume as much energy as charging a smartphone; a single exchange with ChatGPT can heat up a server so much that it requires a 16-ounce bottle’s worth of water to cool it. At scale， these costs soar. By 2027 the global AI sector could annually consume as much electricity as the Netherlands， according to one 2023 study. And recent research in Nature Computational Science identifies another concern： AI’s outsize contribution to the world’s mounting heap of electronic waste. The study found that generative AI applications alone could add 1.2 million to five million metric tons of this hazardous trash to the planet by 2030， depending on how quickly the industry grows.

每次用生成式人工智能編寫一封郵件或生成一張圖片，地球都要付出相應代價。生成兩張圖片所消耗的能源相當于給手機充一次電；與ChatGPT進行一次對話，需要1瓶16盎司瓶裝水的水量來冷卻服務器產生的熱量。隨著規模提升，這些成本將激增。根據2023年的一項研究，到2027年，全球人工智能領域的年耗電量可能與荷蘭全國相當。《自然-計算科學》雜志近期刊登的一項研究指出另外一個問題：全球大量累積的電子垃圾有相當大比例來自人工智能。研究顯示，到2030年，僅生成式人工智能應用一項就會給全球新增120萬噸至500萬噸此類有害垃圾，具體數值取決于行業發展速度。

Such a contribution would add to the tens of millions of tons of electronic products the globe discards annually. Cell phones， microwave ovens， computers， and other ubiquitous digital products often contain mercury， lead， or other toxic substances. When improperly discarded， they can contaminate air， water and soil. The United Nations found that in 2022 about 78 percent of the world’s e-waste wound up in landfills or at unofficial recycling sites， where laborers risk their health to scavenge2 rare metals.

這將無異于雪上加霜，因為全世界已經每年廢棄數千萬噸電子產品。手機、微波爐、電腦和其他無處不在的數碼產品常含有汞、鉛或其他有毒物質，不當丟棄會污染空氣、水源和土壤。聯合國數據顯示，在2022年，全世界約78%的電子垃圾最終會進入垃圾填埋場或非正規回收站，工人在那里冒著健康風險搜集稀有金屬。

The worldwide AI boom rapidly churns through physical data-storage devices， plus the graphics-processing units and other high-performance components needed to process thousands of simultaneous calculations. This hardware lasts anywhere from two to five years， but it’s often replaced as soon as newer versions become available. Asaf Tzachor， a sustainability researcher at Israel’s Reichman University， who co-authored the new study， says its findings emphasize the need to monitor and reduce this technology’s environmental impacts.

席卷全球的人工智能熱潮正急速消耗實體數據存儲設備，以及用來處理數以千計同時計算的圖形處理器等高性能組件。這些硬件的使用壽命為2至5年，但往往一有新型號上市就立刻更換。這項新研究的聯合作者、以色列賴赫曼大學可持續性研究員阿薩夫·察喬爾表示，研究結果凸顯了監管和減輕人工智能技術對環境沖擊的迫切需求。

To calculate just how much generative AI contributes to this problem， Tzachor and his colleagues examined the type and volume of hardware used to run large language models， the length of time that these components last and the growth rate of the generative AI sector. The researchers caution that their prediction is a gross estimate that could change depending on a few additional factors. More people might adopt generative AI than the authors’ models anticipate， for example. Hardware design innovations， meanwhile， could reduce e-waste in a given AI system—but other technological advances can make systems cheaper and more accessible to the public， increasing the number in use.

為了計算生成式人工智能對這一問題的影響，察喬爾及同事研究了運行大語言模型所使用的硬件類型和數量、這些組件的使用壽命及生成式人工智能行業的發展速度。研究人員提醒，他們的預測只是粗略估算，可能會因一些額外因素發生變化。例如，使用生成式人工智能的用戶規模可能超過研究模型預期。同時，硬件設計的創新有可能減少特定人工智能系統產生的電子垃圾——但其他技術進步則可能降低系統成本，使其更易普及，從而導致使用者數量增加。

This study’s main value comes from its attention to AI’s broad environmental impacts， says Shaolei Ren of the University of California， Riverside， who studies responsible AI and was not involved in the research. “We might want these [generative AI] companies to slow down a bit，” he says.

加利福尼亞大學河濱分校的任紹磊表示，該研究的主要價值在于關注人工智能對環境的廣泛影響，他說：“我們或許應該讓這些（生成式人工智能）公司把發展節奏放緩一點。”任紹磊致力于研究負責任的人工智能，他并未參與上述研究。

Few countries mandate the proper disposal of e-waste， and those that do often fail to enforce their existing laws on it. Twenty-five U.S. states have e-waste management policies， but there is no federal law that requires electronics recycling. In February 2024 Senator Ed Markey of Massachusetts introduced a bill that would require federal agencies to study and develop standards for AI’s environmental impacts， including e-waste. That bill， the Artificial Intelligence Environmental Impacts Act of 2024 （which has not passed in the Senate）， proposes a reporting system， but it would be voluntary， and AI developers would not be forced to cooperate. Some companies， however， claim to be taking independent action. Microsoft and Google have pledged to reach net zero waste and net zero emissions， respectively， by 2030; achieving these aims would most likely involve reducing or recycling AI related e-waste.

只有極少數國家強制要求妥善處理電子垃圾，而即使制定了相關法律的國家，也常常不能有效執行。美國有25個州出臺了電子垃圾管理政策，但聯邦層面并未立法要求回收電子產品。2024年2月，馬薩諸塞州參議員埃德·馬基提出一項議案，要求聯邦機構就包括電子垃圾在內的人工智能環境影響展開研究并制定標準。這項名為《2024人工智能環境影響法》（尚未在參議院通過）的議案提出一套環境影響報告制度，但將是自愿性質，并不強制要求人工智能開發商配合。不過，一些企業聲稱正在自主采取措施。微軟和谷歌分別承諾在2030年前實現凈零廢棄物和凈零排放；要實現這些目標，很可能需要減少或回收人工智能相關的電子垃圾。

Companies that use AI have numerous options to limit e-waste. It’s possible to squeeze more life out of servers， for instance， through regular maintenance and updates or by shifting worn-out devices to less intensive applications. Refurbishing and reusing obsolete hardware components can also cut waste by 42 percent， Tzachor and his co-authors note in the study. And more efficient chip and algorithm design could reduce generative AI’s demand for hardware and electricity. Combining all these strategies would reduce e-waste by 86 percent， the study authors estimate.

采用人工智能的公司有多種手段減少電子垃圾。例如，通過定期維護和更新，或將耗損設備換給性能要求較低的應用，可以延長服務器的使用壽命。察喬爾及合著者在研究論文中指出，翻新及重復使用老舊硬件也可以減少42%的電子垃圾。此外，效能更高的芯片及算法設計可以降低生成式人工智能的硬件和電力需求。論文作者估計，綜合運用以上策略可以減少86%的電子垃圾。

There’s another wrinkle as well： AI products tend to be trickier to recycle than standard electronics because the former often contain a lot of sensitive customer data， says Kees Baldé， an e-waste researcher at the U.N. Institute for Training and Research， who wasn’t involved with the study. But big tech companies can afford to both erase those data and properly dispose of their electronics， he points out. “Yes， it costs something，” Baldé says of broader e-waste recycling， “but the gains for society are much larger.”

此外還存在一個問題：聯合國訓練研究所電子垃圾研究員凱斯·巴爾德指出，由于包含大量客戶敏感數據，人工智能產品通常比普通電子產品更難回收。巴爾德本人并未參與該項研究。不過，他同時指出，大型科技公司有能力清除這些數據并妥善處理電子設備。巴爾德談及更廣泛的電子垃圾回收時表示：“這當然需要付出一些成本，但社會收益要大得多。”

（譯者為“《英語世界》杯”翻譯大賽獲獎者）

1 saddle sb with sth使（某人）承擔（重任或難題）。

2 scavenge撿破爛。