Inside LEAK: ‘Huge Teams’ Engaged in Manual Interventions on Google Search Results
LEAK: ‘Huge Teams’ Engaged in Manual Interventions on Google Search Results
Google has “huge teams” working on manual
interventions in search results, an apparent contradiction of sworn testimony
made to Congress by CEO Sundar Pichai, according to an internal post leaked to Breitbart
News.
“There are subjects that are prone to hyperbolic content,
misleading information, and offensive content,” said Daniel Aaronson, a member
of Google’s Trust & Safety team.
“Now, these words are highly subjective and no one denies that.
But we can all agree generally, lines exist in many cultures about what is
clearly okay vs. what is not okay.”
“In extreme cases where we need to
act quickly on something that is so obviously not okay, the reactive/manual
approach is sometimes necessary.”
The comments came to light in a leaked internal discussion
thread, started by a Google employee who noticed that the company had recently
changed search results for “abortion” on its YouTube video platform, a change
which caused pro-life videos to largely disappear from the top ten results.
In addition to the “manual approach,” Aaronson explained that
Google also trained automated “classifiers” – algorithms or “scalable
solutions” that corrects “problems” in search results.
Aaronson
listed three areas where either manual interventions or classifier changes
might take place: organic search (“The bar for changing classifiers or manual
actions on span in organic search is extremely high”),
YouTube, Google Home, and Google Assistant.
Aaronson’s post also reveals that there is very little
transparency around decisions to adjust classifiers or manually correct
controversial search results, even internally. Aaronson compared Google’s
decision-making process in this regard to a closely-guarded “Pepsi Formula.”
These
comments, part of a longer post copied below, seem to contradict Google CEO
Sundar Pichai’s sworn congressional testimony that his company does not “manually
intervene on any particular search result.”
According
to an internal discussion
thread leaked to Breitbart News by a source within the company,
a Google employee took issue with Pichai’s remarks, stating that it “seems like
we are pretty eager to cater our search results to the social and political
agenda of left-wing journalists.”
According to the posts leaked by the source, revealed that
YouTube, a Google subsidiary, manually intervened on search results related to
“abortion” and “abortions.” The intervention caused pro-life videos to
disappear from the top ten search results for those terms, where they had
previously been featured prominently. The posts also show YouTube intervened on
search results related to progressive activist David Hogg and Democrat
politician Maxine Waters.
In a comment to Breitbart News, a Google spokeswoman also
insisted that “Google has never manipulated or modified the search results or
content in any of its products to promote a particular political ideology.”
Pichai might claim that he was just talking about Google, not
YouTube, which was the focus of the leaked discussion thread. But Aaronson’s
post extends to Google’s other products: organic search, Google Home, and
Google Assistant.
Aaronson is also clear that the manipulation of the search
results that are “prone to abuse/controversial content” is not a small affair,
but are the responsibility of “huge teams” within Google.
“These lines are very difficult and can be very blurry, we are
all well aware of this. So we’ve got huge teams that stay cognizant of these
facts when we’re crafting policies considering classifier changes, or reacting
with manual actions”
If Google has “huge teams” that sometimes manually intervene on
search results, it’s scarcely plausible to argue that Pichai might not know
about them.
Aaronson’s full post is copied below:
I work in Trust and Safety and while I have no particular input
as to exactly what’s happening for YT I can try to explain why you’d have this
kind of list and why people are finding lists like these on Code Search.
When dealing with abuse/controversial content on various mediums
you have several levers to deal with problems. Two prominent levers are
“Proactive” and “Reactive”:
- Proactive: Usually refers to some type of
algorithm/scalable solution to a general problem
·
E.g.: We don’t allow straight up porn on YouTube so we create a
classifier that detects porn and automatically remove or flag for review the
videos the porn classifier is most certain of
- Reactive: Usually refers to a manual fix to
something that has been brought to our attention that our proactive
solutions don’t/didn’t work on and something that is clearly in the realm
of bad enough to warrant a quick targeted solution (determined by pages
and pages of policies worked on over many years and many teams to be fair
and cover necessary scope)
·
E.g.: A website that used to be a good blog had it’s domain
expire and was purchased/repurposed to spam Search results with autogenerated
pages full of gibberish text, scraped images, and links to boost traffic to
other spammy sites. It is manually actioned for violating policy
Manually reacting to things is not very scalable, and is not an
ideal solution to most problems, so the proactive lever is really the one we
all like to lean on. Ideally, our classifiers/algorithm are good at providing
useful and rich results to our users while ignoring things at are not useful or
not relevant. But we all know, this isn’t exactly the case all the time
(especially on YouTube).
From a user perspective, there are subjects that are prone to
hyperbolic content, misleading information, and offensive content. Now, these
words are highly subjective and no one denies that. But we can all agree
generally, lines exist in many cultures about what is clearly okay vs. what is
not okay. E.g. a video of a puppy playing with a toy is probably okay in almost
every culture or context, even if it’s not relevant to the query. But a video
of someone committing suicide and begging others to follow in his/her footsteps
is probably on the other side of the line for many folks.
While my
second example is technically relevant to the generic query of “suicide”, that
doesn’t mean that this is a very useful or good video to promote on the top of
results for that query. So imagine a classifier that says, for any queries on a
particular text file, let’s pull videos using signals that we historically
understand to be strong indicators of quality (I won’t go into specifics here,
but those signals do exist). We’re not manually curating these results, we’re
just saying “hey, be extra careful with results for this query because many
times really bad stuff can appear and lead to a bad experience for most users”.
Ideally the proactive lever did this for us, but in extreme cases where we need
to act quickly on something that is so obviously not okay, the reactive/manual
approach is sometimes necessary. And also keep in mind, that this is different
for every product. The bar for changing classifiers or manual actions on span
in organic search is extremely high. However, the
bar for things we let our Google Assistant say out loud might be a lot lower.
If I search for “Jews run the banks” – I’ll likely find anti-semitic stuff in
organic search. As a Jew, I might find some of these results offensive, but
they are there for people to research and view, and I understand that this is
not a reflection of Google feels about this issue. But if I ask Google assistant
“Why do Jews run the banks” we wouldn’t be similarly accepting if it repeated
and promoted conspiracy theories that likely pop up in organic search in her
smoothing voice.
Whether we agree or not, user perception of our responses,
results, and answers of different products and mediums can change. And I think
many people are used to the fact that organic search is a place where content
should be accessible no matter how offensive it might be, however, the
expectation is very different on a Google Home, a Knowledge Panel, or even
YouTube.
These lines are very difficult and can be very blurry, we are
all well aware of this. So we’ve got huge teams that stay cognizant of these
facts when we’re crafting policies considering classifier changes, or reacting
with manual actions – these decisions are not made in a vacuum, but admittedly
are also not made in a highly public forum like TGIF or IndustryInfo (as you
can imagine, decisions/agreement would be hard to get in such a wide list –
image if all your CL’s were reviewed by every engineer across Google all the
time). I hope that answers some questions and gives a better layer of
transparency without going into details about our “Pepsi formula”.
Best,
Daniel
Comments
Post a Comment