Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:02
have you ever shared something online that turned
0:04
out? to be true i in
0:07
fact not that long ago, i shared a photo
0:09
of an adorable horse named sugar who
0:11
supposedly pretends to be asleep to avoid
0:14
being that photo was shared
0:16
over
0:16
40,000 times on to and
0:18
i i thought, well, it must be true,
0:20
but as it turns out, sugar
0:23
was just taking a nap
0:26
okay so mistakes happen
0:28
but misinformation at scale can
0:30
also be really dangerous especially
0:32
when it relates our democracies
0:35
my disinformation a speech new highs
0:37
because it's not just a human problem it's
0:39
an algorithmic problem too one
0:42
that exploited by murky political forces
0:44
in your country and in mine in
0:47
all of our elections in different languages
0:49
in different political contacts our
0:51
platform is stupid enough to fight it
0:54
these companies have not allocation of resources
0:56
to protecting all actions at once
0:59
much less societies between elections
1:01
because if you're paying attention to a country
1:04
three weeks before the vote that in the die
1:06
is cast malicious people
1:08
have said a year creating
1:10
you know soccer fans eat is that the slowly
1:13
politicize
1:14
that's the harm masashi he
1:16
used to be a data engineer on facebooks
1:18
integrity team working on election
1:21
more hear more from him in a bit
1:33
it episodes on the perils and promise
1:36
of artificial intelligence or the
1:38
internet and him realize
1:40
we're leading ai builders and policy
1:42
folks who make who make more trustworthy
1:45
in the special season that doubles as mozilla
1:47
twenty twenty two internet help report
1:49
this time it's a i elections
1:52
and disinflation how can we
1:54
create healthier
1:55
the nation
2:02
reading this information is about
2:05
unraveling mysteries it's
2:07
about spot in clues and pattern
2:09
second lead you to
2:10
source
2:11
let's meet someone
2:13
who
2:14
on dismantling descend from than across
2:16
more than twenty countries and now
2:19
that an employee of twitter facebook
2:21
or tic tac just
2:23
and aren't time the founder and chief
2:25
executive of code for africa this
2:27
is africa's largest network of city the
2:30
an open data group that includes
2:32
investigative reporters that setting
2:34
groups and data scientist
2:36
whether it be as things around climates
2:38
or around religion nor around kind
2:41
of reproductive issues and gender
2:43
right
2:44
you can hire sophisticated
2:46
small agile seems
2:48
they're not able to build campaigns
2:50
for you including creating
2:53
sake accounts that kind of
2:55
reference each other and look like and and
2:57
function like coherent communities
3:00
i'm and sec people in and
3:02
build quite a bit of momentum
3:05
and it's from south africa he's
3:07
currently based in tbilisi georgia he
3:09
wants us to understand that there is a global
3:12
industry of dissent and for higher
3:14
worth tens of millions of dollars
3:17
like a well oiled machine at these networks
3:20
make fake content and lots of it
3:22
more importantly they create artificial
3:25
surges of attention on topics
3:27
the you spot and network social
3:29
media accounts out
3:31
content in a way that's designed to game
3:33
social media algorithms which
3:36
in turn it amplify these messages
3:38
eventually humans and media
3:40
organisations begin to genuinely engage
3:43
with that and this is how social media
3:45
is weaponized
3:48
like you have a military industrial complex
3:50
there is a can of disinformation
3:53
industrial complex and only
3:55
way we gain to descendants is
3:57
by d minute sizes
3:59
this information networks become more powerful
4:02
the influence the messages we here and
4:05
how we interact with one another the
4:07
corrupting democratic discourse throwing
4:10
them out of the shadows and taking down their networks
4:12
threatens their business model that
4:14
shines a light on those who better to politically
4:17
for misinformation
4:18
south africa we've seen the
4:20
developments of these kinds of networks
4:23
using xenophobia
4:25
as as can as a rallying
4:28
call we south africans believe
4:30
that nigerians and kenyans
4:32
and zimbabweans mozambicans are taking jobs
4:35
that to that's a long to them and
4:37
hitting on very many of the same trigger points
4:39
to julia in conversations in the us so
4:42
elsewhere or eastern europe they're
4:44
using of playbook that it's been proven
4:47
to work elsewhere the dressing
4:49
it in local language and and
4:51
often generating manipulated media
4:53
to supports local plans
4:56
certain set of this information adds
4:59
fuel to the fire in countries where there
5:01
are already electoral tooth religious
5:03
insurgencies
5:04
and foreign mercenaries
5:06
trying to fact check everything is whack a mole it's
5:09
very worthwhile and we need to do it and
5:11
it does not scale because
5:14
we cannot operate at the same level that these
5:16
machine generated hate machines
5:18
do
5:19
oh for africa coordinates fact checking
5:21
teams in more than a hundred and seventy news
5:23
comes across africa they're journalism
5:26
becomes training data for machine learning tool
5:28
we operate across twenty one countries
5:31
and there's probably and twenty
5:33
one countries probably half a billion people
5:36
we're the largest organization in
5:39
the space in africa take this
5:41
information would and were only ninety three people
5:44
and you have an outsized impacts i mean the
5:46
sec checking that are sick chicken sandwiches only
5:48
city people produces may be two
5:50
thousand section so yeah which is not
5:52
a big number at that in turn has
5:55
a multiplier impacted just say spoke
5:57
labels or removes either
5:59
six million posts per year
6:02
based on those two thousand sections
6:04
that could acknowledges that physicist a drop
6:06
in a sea of disinformation
6:09
why is a small non profit in africa
6:12
or him to clean up the platforms of the world's richest
6:15
internet companies
6:16
what's your social media companies to be doing
6:19
to use sites to this problem
6:21
as scale i mean this is clips
6:23
but they should be more collaboration not just
6:25
between social media companies
6:28
themselves or technology companies
6:30
the platform should be doing more sharing
6:33
and more joint problem solving to
6:35
solve a problem that ultimately they've created
6:39
these big platforms should be collaborating
6:42
with each other sharing intelligence
6:45
mitigating this information from spreading from
6:47
one channel to the other
6:48
they'll never be able to be everywhere the
6:51
time and so they
6:53
need to figure out productive
6:55
sustainable ways
6:57
collaborating with a wider sense
6:59
as sits also com of watch dogs
7:02
in the media in fact checking
7:05
in kind of political and
7:07
watchdog ecosystems that
7:09
, point as important african
7:12
countries are divorced and culture
7:13
the and spoken languages in
7:16
nigeria alone over five
7:18
hundred languages are spoken hold
7:20
for africa's team speak local languages
7:23
they , local history and geopolitics
7:26
they're building science data sets of me
7:29
it had had from the region and they
7:31
use a i to supercharge
7:33
their work the machine learning tools
7:35
that we use and the the natural language
7:37
processing tools
7:39
the to use to
7:40
not just track the use of
7:42
specific terms but understand
7:45
emerging narratives of conversations
7:47
and the site geist that almost
7:50
makes people
7:52
okay people to be susceptible
7:57
that's a coma societies
7:59
we
7:59
be blind to all of that if we
8:02
didn't have and his
8:04
machine learning to be able to analyze
8:06
millions of online articles
8:09
and how spots the trains or the outliers
8:12
so we use a lot of tools
8:16
the platforms do collaborate
8:18
with and even pay local organizations
8:21
certain has only the largest
8:24
fact checking networks are invited by platforms
8:26
to partner with them he believed the
8:28
work code for africa and other researchers
8:30
do developing and sharing a i
8:32
tools and create a positive ripple
8:34
effect when power smaller group
8:37
we need to figure out ways of cascading
8:39
down these ecosystems and creating
8:42
and as layered interactions
8:44
wage grassroots media
8:46
have access to resources and see
8:48
techniques and tools they
8:51
will concede into canada and
8:53
it was cascading up series
8:55
of more complex organizations
8:57
who can start using a eyes
8:59
the kinds of technologies that will never be
9:02
a in the reach of people operating done it's
9:04
really simple
9:05
that is big platforms like facebook credit
9:07
for the worth they're doing and misinformation even
9:10
though he thinks they could be doing more
9:12
they at least are doing things that we
9:14
can see and we can criticize
9:16
the problem of the closed channels
9:19
the signals of what sets were
9:21
it's invisible the talk social
9:29
dark souls
9:31
that sounds much spooky or than what i see
9:33
and my what's have sat with my family but
9:36
messaging apps are seized sectors of
9:38
political and electoral disinclination
9:40
and the countries from our
9:42
and to end updated that make them
9:44
more secure it also
9:46
makes them harder to study from the outside specially
9:50
when companies are not transparent
9:53
part of what was happening and radiating with the
9:55
the with a lot of attention on facebook and twitter
9:57
and all that have had some health
9:59
not come to a global
10:02
attention but for many of us ala
10:04
like an indian or than other countries that one
10:05
the dominant blackballed
10:08
that to run i'm a pro backer she's
10:10
the resource lead and cofounder of the open
10:12
source projects cattle that's
10:14
a community of technologists and researchers
10:16
in india they build machine learning
10:19
tools and datasets to understand and
10:21
respond to misinformation they
10:23
began their work three years ago
10:25
trying to crack a puzzle how
10:28
can they help people verify information
10:30
on what's that
10:31
the company was
10:33
on what up a lot of the content is in
10:35
audio visual format or you don't
10:37
necessarily have you already use of images
10:40
and voice recording that a shared in the black
10:42
phone we knew that if you wanted
10:44
to do any automation
10:46
the it around like linking
10:48
it to have fact checked report or
10:50
on thinking it to contend that has
10:52
been said in the past you have
10:54
to be able to work with the language
10:57
that content with being said in as
10:59
well as
11:00
the modality which have weathered like image video
11:03
on your for that's where
11:05
the started with oh beating the
11:07
the machine learning to have to just process
11:09
the content in the modality and the language
11:12
of that i used in in india and
11:14
every other told archiving a lot of the
11:16
contents we started archiving gone from that was
11:18
located on shot out
11:20
they started creating a data set off a condom
11:22
that had been fact checked in india
11:24
they collected those fat sacks using a technique
11:27
called scraping and the whatsapp
11:29
messages were from large public group
11:31
then they developed a suitable depository
11:34
for cross referencing sat messages with
11:36
the fact sex the
11:38
process of doing this work they were stolen
11:41
by the close connection between hate speech
11:43
and misinformation
11:45
you're collecting of feeding data
11:47
from some of these other thought and that while we
11:49
were just thought of shocked at how much heat with
11:51
there was on the platform and
11:53
though we're thinking about why platform
11:56
had not address these issues
11:58
what do that before
11:59
the opened some of the data set for a cd
12:02
thought or journalistic storytelling we wanted to filter
12:04
out the he'd be to we found other than a place on
12:06
the we need to the point in moderation reported
12:09
even the theater and we didn't have the booth to do it
12:12
the paddles like expanded to include tools
12:14
and plug ins for moderating content in their
12:16
own open datasets in hindi camel
12:19
and indian english but why
12:21
haven't the platforms address these issues
12:24
with so much now as and resource in
12:26
the hate speech and disinclination there's
12:28
there's still so much of it one
12:31
reason is language
12:33
i couldn't be more platform the that
12:35
they're just not geared to hand
12:38
the indian language and very well
12:43
let's talk about content moderators the
12:45
people whose job it is to see the most gruesome
12:48
content on the internet part
12:50
, their job involves labeling content
12:53
so becomes training data for automated
12:55
content moderation this is important
12:57
work we know from facebook whistleblowers
13:00
like francis how good that only a small
13:02
proportion of moderation happens outside
13:04
of the u s even though ninety percent
13:06
of facebook users live elsewhere in
13:09
other words most languages are under
13:11
resourced and even for english
13:13
there are many different dialects even
13:16
though a lot of indian social media users
13:18
when he was english they with you that in you very distinct
13:21
way with way lot of mixing and matching
13:23
with regional languages
13:25
and using words in ways that you
13:27
wouldn't use that an american english
13:32
here that requested
13:34
which country
13:35
the most facebook users
13:38
the answer is india the us
13:40
and indonesia but the number
13:42
of users that are placed doesn't always mean
13:45
that a platform is going to be more accountable for them
13:47
things that are done to protect elections in one place
13:50
never habit and others and
13:52
that's true for all platforms
13:55
should it be the case that
13:57
accompany has at least a million
14:00
people speaking a language that no
14:02
one in the company understands the
14:05
very least you need to have some people who
14:07
understand the language and we're paying attention to
14:09
it as their job
14:11
harm a fatty is the cofounder of
14:13
the nonprofit integrity institute in the u
14:15
s it's a new member organization
14:17
for people who work at integrity teams at
14:19
social media platforms her
14:22
work at facebook as a data into their ended
14:24
up technical tools to protect elections
14:27
i silly accurate infant
14:28
listen to voters
14:29
the also looked at real time dashboard
14:32
and what they called a war room to identify
14:34
spikes in misinformation these
14:37
can be caused by parker
14:39
well known at that the
14:41
we share button is really dangerous
14:43
or the retrieval yeah and
14:46
if a thing is we shared many times very
14:48
likely to be bad you could
14:50
imagine putting in place
14:53
sort of lake stevens for that are
14:56
, to sort of make it harder to photo
14:58
either quoted or between a retreat a retreat
15:00
that's not technically very hard the
15:03
real challenge is like arguing about weiss be
15:05
allowed to launch and
15:07
almost ninety percent the job can be to sort of diplomacy
15:10
around being allowed to do the research on
15:12
i diplomacy zahara means
15:14
they need to negotiate for changes
15:17
integrity teams are considered costs center's
15:20
at odds with other teams focused on growth
15:22
you're always to be that including a person
15:25
who points out why the easy thing
15:27
or the growth making saying may not be a good idea
15:30
it's a really awkward position to be
15:32
the company like facebook a lot
15:34
of different teams were com policies and moderation
15:37
integrity work is just one piece of the
15:39
puzzle
15:40
i think it's fair to say that in general integrity
15:43
teams are new or
15:45
and less well resourced then we would like
15:48
he said he's especially proud of
15:50
the work his team at facebook dead i'm
15:52
a twenty eight team brazilian presidential
15:54
election and us midterm elections
15:57
now that there were missed out but
16:00
he describes intense periods of work
16:03
where they grew in terms of skills and capacity
16:06
we're pulling fourteen hour shifts and
16:08
dislike windowless stinky
16:10
room you know you are like
16:13
build a tool on day one show
16:16
up on day two for your shift and
16:18
so when had upgraded it i d
16:20
five so had to the whole new tool i
16:22
did a better job you know by the
16:24
seven so i might have actually remember documentation
16:26
about how to use it
16:28
the harsh that it was rewarding when different parts
16:30
of the company would pull together
16:33
different teams around the company really
16:35
cared about it and you're able to pull the men and
16:37
say we really need you to
16:39
teach us how to use this tool that you built that we can use
16:41
it or we really need use
16:43
it seems your apps are
16:45
your products temporarily so
16:47
that it is safer and they would do it because
16:50
he really wanted to do the right thing
16:53
companies are secretive about what they do
16:55
or don't do
16:56
a fight this information saw
16:59
her says this this collaboration
17:01
with outside groups difficult even
17:03
when they could really help the
17:05
integrity institute will serve as a bridge
17:07
to the inside
17:08
every company's probably different one
17:11
we the companies are different is in the way that the you think
17:13
about the outside world how control
17:15
the are for their workers to talk to the
17:17
outside world and a
17:19
level of paranoia or
17:22
being locked down in one company freely
17:24
can surprise you if you come
17:27
from a different company harder
17:29
what the heck institute is trying to do in
17:33
really speak to that and six that and
17:35
we say that we're representing integrity workers
17:38
though that we are the one place where
17:40
and years academics and and
17:42
the restless
17:43
the world can come talk to us and amazon
17:45
seventy to channel workers are
17:47
the technique professionals for our members
17:49
into their day jobs
17:54
let me introduce you to one more person
17:57
rashid six center coordinates global contribute
18:00
the to a crowdsourced data set up
18:02
online hate speech called shape base
18:04
through an initiative called the citizen linguists
18:06
lab this is all run
18:09
by the senses or project and canada but
18:11
rossi works remotely from bangalore
18:13
india
18:14
the sitter remember slab is
18:16
the for any one across the
18:18
board that wants to contribute
18:21
and amplify and
18:23
of meant odd database
18:24
we in many cases
18:27
might lack the social and cultural
18:29
and linguistic context of things
18:32
and who better to contribute than
18:35
the locals who live in particular
18:37
setting
18:39
hippies now covers and ninety eight languages
18:42
going across many countries they
18:44
work with hundreds of universities to research
18:47
the impact of hate speech and information
18:49
particularly leading up to to elections
18:52
right he explains the connection like this
18:55
hate speech not the gun
18:57
but misinformation pulls the trigger
19:00
hate speech in itself might
19:02
not contribute to offline violence
19:04
but it kind of said
19:05
the tone and the environment else
19:07
backgrounds hostility towards a particular
19:10
community or ethnicity and
19:12
then the woman in the form
19:14
or malignant information on full
19:17
information that circulate around
19:20
that can best lead to jacksonville
19:22
and
19:23
sentinel project has documented this dynamic
19:25
in
19:25
whereas countries
19:27
in kenya the democratic
19:29
republic of congo sudan
19:32
sri lanka the me and man
19:34
their mission is to prevent mass atrocities
19:37
too early warning and cooperation the
19:40
global repository of hate speech enables
19:42
them to perform automated sightings of
19:45
the offensive term across the internet nearly
19:48
a million of them they labels rooms
19:50
that a way that lets them take the time
19:51
you're a conflict
19:53
every contribute as call a citizen linguist
19:56
and they can also help
19:57
by offering the assessment of the
19:59
offense
19:59
the midst of a particular dome which is
20:02
then calculated with all other inputs
20:04
this , us to sort of crowdsource
20:07
sentimental analysis and
20:09
acts as one part of the output ah
20:12
for the system so the
20:14
offensiveness rating kind of helps us to
20:16
understand or social understand political
20:19
environment
20:21
the data can be accessed by local and
20:24
right groups free of charge i'll
20:26
be platforms can pay for access
20:28
it's a resource to help moderate online
20:30
conversations is designed to keep
20:32
people safe
20:33
in real life
20:38
we can pardon elections taking place around
20:40
the world it is vital that
20:42
platforms get a handle on misinformation
20:45
this isn't something that anyone company can
20:47
handle alone the can't be
20:49
solved in secrecy with content moderation
20:51
algorithms are underpaid and
20:53
unprotected moderators we
20:56
need companies to practice meaningful transparency
21:00
so they can collaborate better with each other and
21:02
local groups this would empower
21:04
researchers to uncover harmful this information
21:07
that transcends platform as languages
21:09
and yet ecosystems and those
21:11
policies that platforms create for transparency
21:14
and safety during elections they
21:16
shouldn't just apply in some countries
21:18
they should apply as
21:19
the were
21:22
this is i around in original podcast
21:24
mozilla a nonprofit behind firefox
21:28
this season of iran doubles as the internet
21:30
self report you
21:32
can be more about our guests and a i
21:34
i visited internet health report that
21:36
it
21:37
i'm bridget had thanks so much for listening
21:40
for more on what can be done look at
21:42
mozilla minimum election standards for flat
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More