[新聞] MIT研究：deepfake對政治的戕害或許不大
1.媒體來源：Gizmodo
2.記者署名：ByShoshana Wodinsky
3.完整新聞標題：
Deepfakes Maybe Not Quite the Political Apocalypse We Feared, MIT Researchers
Find
麻省理工學院的研究人員發現，Deepfake所帶來的，可能並非我們所擔心的政治劫難那般
4.完整新聞內文:
Ever since we’ve seen deepfakes cropping up across porn, e-commerce, and
literal bank robberies, there’s always been concern that this same tech
could be used to interfere with future elections. Well, according to one new
study, that might be tougher than we thought. Researchers from the
Massachusetts Institute of Technology (MIT) have put out a new report
investigating whether political video clips might be more persuasive than
their textual counterparts, and found the answer is... not really.
自從我們看到 Deepfake 被用於色情、電商和銀行搶案後，人們一直擔心這種技術可能會
被拿來干涉選舉。然而，根據一項最新的研究，那可能比我們想像的更為困難。近期，麻
省理工學院（MIT）針對「政治性質的影片，是否會比文字更具有說服力？」這個問題提
出了一項研究，結果其實不然。
“Concerns about video-based political persuasion are prevalent in both
popular and academic circles, predicated on the assumption that video is more
compelling than text,” the researchers wrote in their paper. This is a point
we’ve heard again and again from lawmakers over the years, ever since
deepfakes first popped up on their radar in mid-2019. When Sens. Rob Portman
(R-OH) and Gary Peters (D-MI) introduced the Deepfake Taskforce Act this past
summer, Portman noted in a statement that deepfakes posed a “unique threat”
to national security.
研究中指出「大眾和學術界都很擔心（Deepfake 的）政治宣傳影片，但這個疑慮成立的
前提是，『影片』這種媒介會比文本更具影響力。」自從 Deepfake 於 2019 年年中問世
以來，我們一次又一次地從立法者那裡聽到類似的論點。去年夏天，當共和黨的 Sens.
Rob Portman 和民主黨的 Gary Peters 攜手推動《Deepfake 專法》（Deepfakes
Taskforce Act）時，Portman 便在聲明中指出，「Deepfake 已經對國家安全構成了『獨
特的威脅』」。
“For most of human history seeing meant believing, but now that is becoming
less and less true thanks to deepfakes,” Portman said at the time. “
Combined with the network effects created by social media, fake videos or
pictures can travel around the world in an instant, tricking citizens.”
「在人類的大部分歷史裡，我們深信著『眼見為憑』，但因為 Deepfake 的出現，這種想
法可能越來越不可靠。」Portman 表示「結合社群媒體建構而成的網路擴散效應，假影片
或假圖片可以瞬間傳遍世界，藉此用來欺騙公民。」
To gauge how effective this tech would be at tricking anyone, the MIT team
conducted two sets of studies, involving close to 7,600 participants total
from around the U.S. Across both studies, these participants were split into
three different groups. In some cases, the first was asked to watch a
randomly selected “politically persuasive” political ad (you can see
examples of what they used here), or a popular political clip on covid-19
that was sourced from YouTube. The second group was given a transcription of
those randomly selected ads and clips, and the third group was given, well,
nothing at all since they were acting as the control group.
為了衡量 Deepfake 這項技術在騙人方面的效果，麻省理工學院的團隊進行了兩組研究，
包含來自美國各地的 7,600 多名受試者。在這兩組研究中，這些受試者被分為三組，給
第一組觀看那些極具政治渲染力的廣告、或是在 YouTube 上跟 COVID-19 有關的政治影
片；第二組則給他們閱讀上面那些廣告和影片的文字稿；第三組則做為對照組，什麼都沒
給他們看。
After that, each member of each group was given a questionnaire asking them
to rate the “believability” of the message they saw or read—specifically,
whether they believed the people in the clip actually made a particular
claim. Then they were asked to rate how much they disagreed with the core
point from whatever persuasive ad they were seeing.
之後，每位受試者都會收到一份問卷，要求他們對剛剛看到、或閱讀到的訊息之「可信度
」進行評分——特別是，他們是否相信影片中的人所提出的主張或政策。然後，他們被要
求針對剛剛看到那些宣傳廣告的核心主題，給出認同或不認同的評分。
The question these MIT researchers were trying to answer was twofold: Was
seeing actually believing, the way Portman (and countless others) have said?
And if it is, how much could someone’s opinion actually be swayed by video,
or by text?
MIT 的研究人員試圖回答兩個層次的問題：
人們看到什麼，就會真的相信什麼嗎（這也是 Portman和無數人的說法）？
如果上述假設為真，那人們的意見真的會因為影片或是文字而動搖嗎？
The result? “Overall, we find that individuals are more likely to believe an
event occurred when it is presented in video versus textual form,” the study
reads. In other words, the results confirmed that, yes, seeing was believing,
as far as the participants were concerned. But when the researchers dug into
the numbers around persuasion, the difference between the two mediums was
barely noticeable, if at all.
結果，該研究指出「整體來說，我們發現當訊息以影片與文本為媒介呈現時，人們更有可
能相信事情是真的」。換句話說，結果證實，就受試者而言，眼見還真的可以為憑。但是
當研究人員深入研究這個「事實」所帶來的說服效果時，影片與文字兩種媒介之間的差異
幾乎不顯著。
As one of the researchers behind the project, Adam Berinsky, noted in a
statement about the work, “[J]ust because video is more believable doesn’t
mean that it can change people’s minds.”
作為該計畫的研究人員之一，Adam Berinsky 指出「即使人們會去相信影片的內容，這也
不意味著影片的內容真的可以改變人們的想法。」
Of course, this study (like all academic studies) comes with a fair share of
caveats. For one, even though 7,600 people is a pretty large sample size, it
might not capture the full range of opinions that every American voter might
have. And as the researchers point out in their piece, the small persuasive
advantage that video has over text might actually be even smaller outside of
a research environment:
當然，這項研究（與所有學術研究一樣）也有相當多的侷限。儘管 7,600 人已經是一個
相當大的樣本數，但仍可能無法涵蓋每個美國選民的意見。正如研究人員在他們的文章中
指出的那樣，在研究環境之外，實際上影片相對於文本的說服優勢可能更小：
"In both of our studies, the text-based treatments were presented in the form
of a detailed transcript containing an exact replication of the audio output
as well as a comprehensive description of key visual cues. In reality,
politically persuasive writing may be structured quite differently (e.g., as
a news article or opinion piece)."
「在我們的兩項研究中，我們（給予第二個實驗組）的文字稿均包含了仔細的聲音資訊以
及影片關鍵幀的詳細描述。但在日常生活裡，政治宣傳的文章可能和影片逐字稿有不同的
結構（例如，新聞文章或評論文章絕對不會是逐字稿）。」
But even if that’s the case, the study notes that information presented over
video has a unique advantage that text simply doesn’t: A video is more
attention-grabbing and can capture more of an audience than a written report
ever could.
不過即使是這樣，該研究仍然指出，透過影片呈現的資訊，還是有相較於以文字為載體所
沒有的優勢：與單純的文章相比，影片可以吸引更多的受眾。
“It’s possible that in real life things are a bit different,” David Rand,
one of the other authors on the study, noted in a statement.
「在現實生活中，情況可能會有所不同」該研究的作者之一 David Rand 指出。
“It’s possible that as you’re scrolling through your newsfeed, video
captures your attention more than text would,” he added. “You might be more
likely to look at it. This doesn’t mean that video is inherently more
persuasive than text—just that it has the potential to reach a wider
audience.”
「當你在瀏覽新聞時，影片可能比文字更能吸引你的注意力，」他補充「你可能更傾向選
擇觀看影片，但這並不意味著影片會比文字更具說服力——只是影片可能會有更高的觸及
率而已。」
In other words: At least as far as this study is concerned, deepfake videos
of a given politician aren’t likely to sway people’s political views more
than a fake news report about that same politician. The only advantage that
video might have is whether you believe what you’re seeing in front of you—
and the number of eyeballs that clip might eventually get.
換句話說，至少就這項研究而言，政治人物的 Deepfake 影片不太可能比假新聞更能影響
人們的政治觀點。 影片可能具有的唯一優勢在於高點閱率，如此而已。
5.完整新聞連結 (或短網址):
https://gizmodo.com/deepfakes-maybe-not-quite-the-political-apocalypse-we-f-1848090462
https://tinyurl.com/3k7z56bv
6.備註:
簡單來說呢，要成立許多政治人物所說「Deepfake 可能被用於偽造政治宣傳或政策，導
致選民傻傻的相信這些錯誤資訊，很可怕！」的前提是
1. 人們真的會因為看到 Deepfake 影片，就去相信影片裡的政治人物有講過那些話。
2. 並且因為看了這些影片，進而改變自己的政治想法。
但研究顯示，即使人們可能會相信親眼看到的影片（前提 1 成立），也不一定會因此改
變自己的想法（前提 2 不成立）。單就政治影響力上，Deepfake 確實很可怕，但不一定
比假新聞還要嚴重，也沒有多數人講的那麼可怕，宛如政宣界的浩劫一般。
不過，撇除作為政治宣傳的工具，Deepfake 還是有性暴力、性騷擾以及詐欺的問題，這
些仍是該技術的隱憂之一。
對這研究有興趣的可以去啃 MIT 的論文：
https://www.pnas.org/content/118/47/e2114388118
https://psyarxiv.com/r5yun/
Wittenberg, Chloe, Ben M. Tappin, adam berinsky, and David G. Rand. 2020.
“The (minimal) Persuasive Advantage of Political Video over Text.”
PsyArXiv. February 19. doi:10.31234/osf.io/r5yun.
--
※ 發信站: 批踢踢實業坊(ptt.cc), 來自: 140.112.217.58 (臺灣)
※ 文章網址: https://www.ptt.cc/bbs/Gossiping/M.1637585043.A.C9C.html
推
11/22 20:44,
2年前
, 1F
11/22 20:44, 1F
推
11/22 20:45,
2年前
, 2F
11/22 20:45, 2F
推
11/22 20:45,
2年前
, 3F
11/22 20:45, 3F
推
11/22 20:45,
2年前
, 4F
11/22 20:45, 4F
推
11/22 20:45,
2年前
, 5F
11/22 20:45, 5F
→
11/22 20:46,
2年前
, 6F
11/22 20:46, 6F
推
11/22 20:46,
2年前
, 7F
11/22 20:46, 7F
→
11/22 20:46,
2年前
, 8F
11/22 20:46, 8F
噓
11/22 20:46,
2年前
, 9F
11/22 20:46, 9F
→
11/22 20:46,
2年前
, 10F
11/22 20:46, 10F
推
11/22 20:47,
2年前
, 11F
11/22 20:47, 11F
噓
11/22 20:47,
2年前
, 12F
11/22 20:47, 12F
推
11/22 20:48,
2年前
, 13F
11/22 20:48, 13F
→
11/22 20:48,
2年前
, 14F
11/22 20:48, 14F
→
11/22 20:48,
2年前
, 15F
11/22 20:48, 15F
噓
11/22 20:48,
2年前
, 16F
11/22 20:48, 16F
推
11/22 20:49,
2年前
, 17F
11/22 20:49, 17F
→
11/22 20:49,
2年前
, 18F
11/22 20:49, 18F
噓
11/22 20:49,
2年前
, 19F
11/22 20:49, 19F
推
11/22 20:50,
2年前
, 20F
11/22 20:50, 20F
推
11/22 20:50,
2年前
, 21F
11/22 20:50, 21F
推
11/22 20:51,
2年前
, 22F
11/22 20:51, 22F
推
11/22 20:52,
2年前
, 23F
11/22 20:52, 23F
→
11/22 20:53,
2年前
, 24F
11/22 20:53, 24F
→
11/22 20:54,
2年前
, 25F
11/22 20:54, 25F
推
11/22 20:54,
2年前
, 26F
11/22 20:54, 26F
推
11/22 20:55,
2年前
, 27F
11/22 20:55, 27F
→
11/22 20:55,
2年前
, 28F
11/22 20:55, 28F
→
11/22 20:55,
2年前
, 29F
11/22 20:55, 29F
→
11/22 20:55,
2年前
, 30F
11/22 20:55, 30F
→
11/22 20:55,
2年前
, 31F
11/22 20:55, 31F
推
11/22 20:56,
2年前
, 32F
11/22 20:56, 32F
→
11/22 20:57,
2年前
, 33F
11/22 20:57, 33F
推
11/22 20:59,
2年前
, 34F
11/22 20:59, 34F
推
11/22 20:59,
2年前
, 35F
11/22 20:59, 35F
推
11/22 21:00,
2年前
, 36F
11/22 21:00, 36F
推
11/22 21:00,
2年前
, 37F
11/22 21:00, 37F
噓
11/22 21:04,
2年前
, 38F
11/22 21:04, 38F
→
11/22 21:04,
2年前
, 39F
11/22 21:04, 39F
還有 27 則推文
推
11/22 21:59,
2年前
, 67F
11/22 21:59, 67F
推
11/22 22:00,
2年前
, 68F
11/22 22:00, 68F
推
11/22 22:11,
2年前
, 69F
11/22 22:11, 69F
推
11/22 22:22,
2年前
, 70F
11/22 22:22, 70F
→
11/22 22:28,
2年前
, 71F
11/22 22:28, 71F
推
11/22 22:31,
2年前
, 72F
11/22 22:31, 72F
噓
11/22 22:38,
2年前
, 73F
11/22 22:38, 73F
推
11/22 22:53,
2年前
, 74F
11/22 22:53, 74F
噓
11/22 22:57,
2年前
, 75F
11/22 22:57, 75F
推
11/22 23:24,
2年前
, 76F
11/22 23:24, 76F
推
11/22 23:41,
2年前
, 77F
11/22 23:41, 77F
噓
11/22 23:45,
2年前
, 78F
11/22 23:45, 78F
推
11/23 00:10,
2年前
, 79F
11/23 00:10, 79F
→
11/23 00:10,
2年前
, 80F
11/23 00:10, 80F
推
11/23 00:13,
2年前
, 81F
11/23 00:13, 81F
噓
11/23 00:21,
2年前
, 82F
11/23 00:21, 82F
噓
11/23 00:25,
2年前
, 83F
11/23 00:25, 83F
→
11/23 00:26,
2年前
, 84F
11/23 00:26, 84F
噓
11/23 00:26,
2年前
, 85F
11/23 00:26, 85F
推
11/23 00:27,
2年前
, 86F
11/23 00:27, 86F
推
11/23 00:31,
2年前
, 87F
11/23 00:31, 87F
推
11/23 00:49,
2年前
, 88F
11/23 00:49, 88F
推
11/23 01:02,
2年前
, 89F
11/23 01:02, 89F
→
11/23 01:16,
2年前
, 90F
11/23 01:16, 90F
推
11/23 01:18,
2年前
, 91F
11/23 01:18, 91F
推
11/23 01:26,
2年前
, 92F
11/23 01:26, 92F
推
11/23 02:12,
2年前
, 93F
11/23 02:12, 93F
推
11/23 02:12,
2年前
, 94F
11/23 02:12, 94F
→
11/23 03:57,
2年前
, 95F
11/23 03:57, 95F
推
11/23 05:37,
2年前
, 96F
11/23 05:37, 96F
→
11/23 05:48,
2年前
, 97F
11/23 05:48, 97F
→
11/23 05:48,
2年前
, 98F
11/23 05:48, 98F
→
11/23 08:33,
2年前
, 99F
11/23 08:33, 99F
→
11/23 08:48,
2年前
, 100F
11/23 08:48, 100F
→
11/23 09:19,
2年前
, 101F
11/23 09:19, 101F
→
11/23 09:50,
2年前
, 102F
11/23 09:50, 102F
→
11/23 10:16,
2年前
, 103F
11/23 10:16, 103F
推
11/23 12:32,
2年前
, 104F
11/23 12:32, 104F
推
11/23 13:13,
2年前
, 105F
11/23 13:13, 105F
推
11/23 23:25,
2年前
, 106F
11/23 23:25, 106F
討論串 (同標題文章)
以下文章回應了本文：
完整討論串 (本文為第 1 之 2 篇)：