Facebook to ban deepfakes, sources say, but new policy may not cover controversial Pelosi video
Facebook to
ban deepfakes, sources say, but new policy may not cover controversial Pelosi
video
Tony
Romm, Drew Harwell and Isaac Stanley-Becker, The Washington Post Published 11:39 pm
EST, Monday, January 6, 2020
Facebook plans to issue new rules as soon as Tuesday that
would ban users from posting computer-generated, highly manipulated videos,
known as deepfakes, seeking to stop the spread of a novel form of
misinformation months before the 2020 presidential election.
But the policy - described by three people familiar with the
matter who spoke on the condition of anonymity to discuss the yet-unannounced
rules - does not ban all doctored videos. It appears, for example, that
Facebook's new guidelines would not have prohibited a deceptively edited clip
of House Speaker Nancy Pelosi that went viral on the social network last year.
Facebook did not immediately respond to a request for
comment late Monday. The change in policy comes as a company executive prepares
to testify at a congressional hearing later this week on "manipulation and
deception in the digital age."
Going forward, Facebook intends to ban videos that are
"edited or synthesized" by technologies such as artificial
intelligence in a way that average users would not easily spot, the people
said, including attempts to make the subject of a video say words that they
never did.
Facebook, however, will not ban videos manipulated for the
point of parody or satire. And it signaled that other lesser forms of
manipulation would not be outlawed, either, though they could be fact-checked
and limited in their spread on the social networking site, the people said.
The policy does not appear to cover the infamously altered
"drunk" video of Pelosi that was viewed millions of times on Facebook
last year. Though her speech was slowed and distorted in the video to make her
sound inebriated, the effect was accomplished with relatively simple
video-editing software. To contrast with more sophisticated computer-generated
"deepfakes," disinformation researchers have referred to these kinds
of videos as "cheapfakes" or "shallowfakes."
A spokesman for Pelosi did not immediately respond to a
request for comment. At the time, Facebook acknowledged that its fact-checkers
had deemed the video "false," but Facebook declined to delete it
because, as a spokeswoman said, "we don't have a policy that stipulates
that the information you post on Facebook must be true."
Nor does the policy seem to restrain other simpler forms of
video deception, such as mislabeling footage, splicing dialogue or taking
quotes out of context, as in a video last week in which a long response Joe
Biden delivered to an audience in New Hampshire was heavily trimmed to make him
sound racist.
Those omissions worried Hany Farid, a digital-forensics
expert at the University of California at Berkeley whose lab has worked with
Facebook on deepfakes. He said the company's new approach is too "narrowly
construed."
"These misleading videos were created using low-tech
methods and did not rely on AI-based techniques, but were at least as
misleading as a deep-fake video of a leader purporting to say something that
they didn't," Farid said in an email Monday night. "Why focus only on
deep-fakes and not the broader issue of intentionally misleading videos?"
The policy does appear to address deepfake videos in which
women's faces are superimposed into pornography without their consent, seen in
an increasing amount of online harassment campaigns. Such videos made up
roughly 96% of all deepfake videos found last year, according to the research
firm Deeptrace Labs.
Facebook and other tech firms last year sponsored a
"deepfake detection challenge," offering prize money to researchers
who could deliver the most refined techniques to automatically detect
manipulated videos. A set of real and manipulated videos was released to
researchers last month, and the challenge is scheduled to end in March.
Siwei Lyu, the director of a computer-vision lab at the
State University of New York at Albany and member of the Deepfake Detection
Challenge's advisory group, applauded Facebook's attempts to clearly pinpoint
altered media, saying, "the line drawn on user-discernible manipulated
videos is operable and useful for implementing this policy."
The language that Facebook is using to delineate its rules
resembles a policy raised at a June 2019 meeting in San Francisco convened by
the Carnegie Endowment for International Peace to discuss how social media
platforms should deal with manipulated media ahead of the 2020 election,
according to a person who was present at the meeting. The person, spoke on the
condition of anonymity because the meeting had been private, said there was
significant debate about what degree of editing is required before something is
declared misleading and whether social media companies should adopt more
sweeping rules against deceptive content.
The policy's provisions for videos "manipulated for
purposes of parody or satire" could potentially lead to thorny debates
over whether a video labeled as "deceptive" was merely intended to
lampoon for dramatic effect. It's unclear, for instance, whether the policy
would ban a deepfake like that of Mark Zuckerberg, created last year during the
aftermath of the Pelosi video scandal, in which the Facebook chief appeared to
gleefully celebrate his control of user data. The creator told The Washington
Post last year that the video was a form of satire and "a cautionary tale
of technology and democracy."
Comments
Post a Comment