►
From YouTube: SOCCYBSEC WG: Black Trolls Matter & Other Conclusions from Analyzing State-Sponsored Disinformation
Description
Date: 06/04/20
Presenter: Dr. Deen Freelon
Institution: UNC-Chapel Hill
Title: “Black Trolls Matter and Other Conclusions from Analyzing State-Sponsored Disinformation"
A
A
So
his
work
has
a
very
unique
flavor
that
is
very
important
in
today's
age,
he's
taught
at
a
variety
of
places,
including
the
american
university
in
washington
d.c,
as
I
mentioned
he's
now
at
unc,
and
he's
also
one
of
the
leads
there
for
this
new
center
they've
developed,
a
new
interdisciplinary
center
on
information
technology
and
public
life,
c-I-t-a-p
or
scitap,
and
so
on.
Today,
he's
going
to
be
speaking
to
us
about
black
trolls
matter
and
and
his
work
of
course
is
extremely
timely
and
we
are
delighted
to
have
him
here.
Dean.
Take
it
away.
B
Okay,
great,
thank
you
very
much
kathleen
and
thank
you,
everyone
out
there
in
zoom
land.
I
heard
that
this
was
supposed
to
be
about
25
minutes,
so
I'm
going
to
try
to
keep
it
to
that.
Although
I
don't
really
have
a
clock
that
I'm
looking
at
so
somebody
about.
If
I'm
going
too
far
over
somebody
just
doing
requirement
questions
and
then
you
can
go
ahead
and
ask
those.
B
I'm
gonna
talk
about
some
studies,
a
bit
more
in
depth
than
others.
It's
gonna
be
a
little
bit
of
a
dump
in
terms
of
the
amount
of
content
that
I'm
throwing
at
you.
But
my
hope
is
that
at
least
something
will
catch
your
interest
and
we'll
be
able
to
discuss
that
as
people's
interests
want.
So
I'm
just
gonna
go
ahead
and
get
into
it
because
you
know
the
q
a
is
always
my
favorite
part
of
these
things.
B
So
I
do
like
to
get
very
rigorous
and
conceptual
when
I
talk
about
them,
so
I
really
like
this
definition,
that
is,
that
was
created
or
written
by
the
high-level
expert
group
on
fake
news
and
this
information
of
the
european
commission
in
2018,
really
great
definition.
I've
used
in
several
publications
now
and
that
is
all
forms
of
false,
inaccurate
or
misleading
information,
design,
preventative
promotion
to
intentionally
cause
public
harm
or
for
profit
right.
B
So
what
I
like
about
this
is:
it
covers
a
wide
range
of
kind
of
problematic
information,
there's
another
term
that
is
used
and
you're
a
senator
for
this
information.
This
content,
of
course
in
2020,
is
often
spread
through
social
media.
That's
not
the
only
way
that
it
can
spread.
That's
one
of
the
main
vectors
for
this
content.
You'll
also
see
what
is
known,
as
you
know,
or
named
as
information
operations,
information
work
there.
B
Propaganda
active
measures
is
another
term,
especially
when
states
are
sponsoring
the
spread
of
this
content,
but
I
do
want
to
also
know
that
this
is
quite
distinct
from
this
information
and
the
main
distinction
between
disinformation.
The
misinformation
is
the
intent.
So
when
you
know
that
something
is
inaccurate
or
misleading,
then
the
intention
is
to
cause
harm
or
to
generate
profit.
That's
what
makes
it
disinformation.
In
other
words,
and
of
course,
if
you
don't
know,
then
it's
misinformation
that
you're
spreading
it
and
then
at
least
potentially
there
could
be
corrective.
B
Thank
you
so
much
so.
The
difference
between
disinformation
and
this
information
is
basically
a
relationship
between
the
individuals
spreading
it
to
the
content
itself
and
how
much
they
know
about
it
and
their
intention
in
doing
that
splitting
so
the
work
that
I'm
gonna
present
today
is
about
the
russian
trolls
and
social
media.
This
is
probably
the
best
known
state
sponsored
disinformation
campaign.
There's
been
a
lot
of
impress
on
it,
and
this
is
really
just
sort
of
a
nice
infographic
that
talks
about
it
right.
What
is
a
troll?
B
I
I
think
you
know
this
is
a
term
that
goes
back
to
at
least
the
1990s,
if
not
before,
but
now
we're
talking
about
trolling
on
the
national
level
paid
trolls
people
who
are
trying
to
insert
contents
of
disinformational
prominence
into
national
political
discussions
for
the
purposes
of
creating.
You
know:
division,
polarization
and
destabilizing
enemies,
so,
in
other
words,
it's
information
as
weapon.
This
is
why
this
information
is
considered
one
form
of
information
work
there,
along
with
you,
know,
hacking
and
fixing
the
rest
of
it
right.
B
So
the
case
that
I'm
looking
at
in
all
four
of
the
studies
I'm
going
to
present
today
is
the
internet
research
agency,
of
which
many
of
you
may
have
already
heard.
This
is
funded
by
the
russian
government,
operates
out
of
st
petersburg
and
they've
been
active
on
a
number
of
social
media
platforms,
including,
but
not
limited
to
twitter,
facebook,
instagram
youtube
and
some
others,
and
in
2017
the
u.s
house
intelligence
committee
runs
the
street
name
for
2752
ira
link
twitter
accounts.
They
also
released
a
bunch.
B
The
facebook
data
twitter
released
a
thousand
more
accounts
for
a
total
of
thirty
eight
hundred
fourteen
partnering
in
accounts.
I
think
the
number
is
a
little
higher
now
this
slides
in
the
age.
So
I
think
the
number
is
actually
about
four
thousand
at
this
point,
but
that's
sort
of
the
basics
on
the
app
on
the
ira,
and
so
I
just
wanted
to
show
a
little
bit
about
what
this
looks.
Like
some
of
you
may
be
familiar
with
some
of
these
examples.
B
Some
of
you
may
not,
but
I'm
going
to
give
you
the
visual
aspect
of
this
right.
So
we
have
some
anti-source
content
over
here
on
the
left
right,
so
you
know
trump
should
put
solos
and
that
guy
was
wanted
list
for
some
reason
and
then
you've
got
the
supposed
war
on
christmas
on
the
right.
If
you
take
a
look
at
the
actual
text,
you
can
see
that
there
are
some
very
distinct
solutions
in
the
writing.
B
The
kinds
of
grammatical
mistakes
that
most
english
native
english
speakers
probably
wouldn't
make
a
bit
of
a
tell
there.
But
then
you
know
hindsight
is
2020
right,
so,
in
other
words,
people
who
are
kicking
themselves
for
not
realizing
this
and
eventually
you
know,
need
to
sort
of
give
themselves
a
break
and
realize
that.
Well,
you
don't
know
that
there's
a
disinformation
campaign
going
on
it's
a
lot
harder
to
to
see
that
without
without
that
knowledge.
B
Also,
these
are
on
right-wing
examples
and
on
the
next
slide,
I'll
show
a
couple
of
left-wing
examples,
and
so
these
are
a
couple
that
were
directed
specifically
at
the
black
community.
So
it's
it's
sort
of
black
left,
the
intersection
of
black
and
white,
so
you
have
sort
of.
I
won't
vote
for
you
right.
B
So
this
is
talking
out
of
the
electoral
system
on
political
grounds,
as
opposed
to
you
know,
can't
get
around
to
work
grounds,
do
not
vote
for
oppressors
same
same
idea,
and
you
can
see
the
intention
of
voter
suppression
right
so
and
voter
suppression
that
specifically
targets
the
black
community
which,
as
you
may
or
may
not
know
disproportionately
those
democrats
for
democrats
in
the
united
states.
So
this
is
one
of
the
main
purposes
of
the
ira
disinformation,
which
is
this
request
in
this
case
the
black
vote.
B
B
So
the
four
studies
I'm
going
to
talk
about
today
there
are
you
know.
Obviously,
within
each
study
there
are
a
number
of
research
questions,
and
so
I've
really
tried
to
break
down
the
central
research
or
most
important
research
question
for
each
study,
and
so
the
first
one,
which
I
actually
is
where
the
title
part
of
the
title
of
this
presentation
comes
from,
is
which
types
of
ira
accounts
receive
the
most
attention
right.
That's
the
black
trolls
matter,
so
obviously
the
title
gives
away
something
of
the
results
there.
B
The
second
question
is
what
audiences
paid
most
attention
to
ira
content
and
how
are
they
connected
from
the
study
that
came
out
in
this
information
review
a
few
months
ago?
To
what
extent
did
I
rate
content
infiltrate
the
news
media?
I
was
gonna.
I
was
a
good
player
on
that
one
in
the
next
one
that
came
out
of
the
international
journal
of
press
politics
and
then,
finally,
what
attitudinal,
because
it
also
has
hazel
effects
there
as
well.
B
A
B
One
is
another
one
where
I
was
kind
of
on
the
tail
end
of
the
author
right
there
very,
very
good
research
I'll
talk
about
that.
Please,
please
I'll,
write
a
few
bit
more
briefly
in
the
first
two,
so
here's
the
citation-
and
I
can
put
this
back
up
all
of
these
are
published,
and
so
you
can
read
these
papers.
They
all
came
out
in
in
2020
or
late
2019's
with
a
pretty
recent
vintage.
B
I'm
gonna
skip
a
whole
lot
of
things,
but
I'm
gonna
start
with
the
categories
that
were
used
in
terms
of
putting
these
russian
trolls
into
categories
based
on
what
they
talked
about.
So
you
have
some
of
these
categories
that
may
be
familiar
to
you
if
you've
seen
this
content
or
if
you've
seen
other
research
on
the
ira,
so
you've
got
trolls
that
were
reaching
out
to
conservatives
right
trolls,
those
that
were
interested
in
folks
on
the
left
side
of
the
spectrum
left
trolls.
B
These
specific
categories
were
adapted
from
the
category
scheme
developed
by
lyndon
walker
of
clemson
university,
one
of
one
of
our
contributions
in
the
studies
we
broke
out,
the
black
trailer
category
from
the
left,
troll
category,
which
lindle
and
walker
didn't
do,
and
that
actually
ends
up
having
major
implications
for
the
finances
of
this
study,
and
so
the
ones
on
the
left
side
of
the
chart
over
here,
as
you
can
see,
are
mostly
there
are
the
primary
political
ones.
The
rest
of
them
are
less
explicitly
political.
B
So
this
is
just
looking
at
unique
users
per
category.
You
can
see
that
non-english
is
by
far
the
largest
category.
There's
there
have
been
some
analyses
of
those
non-investment,
but
the
majority
of
those
are
in
russian
a
bunch
of
unknowns
and
then
among
the
political
categories
you
can
see
that
righteous
is
by
far
the
largest
having
many
more
categories
than
left
troll
and
black
girls
combined
and
then
the
rest
of
the
categories
are
obviously
smaller
in
terms
of
the
numbers
of
accounts
than
that.
B
Next
up
we
have
tweet
totals
per
category,
and
this
actually
shows
some
pretty
big
differences
and
also
shows
that
the
number
of
tweets
per
category
does
not
really
match
up
with
the
number
of
accounts
per
category.
So
you
can
see
that
among
the
categories
that
are
in
english,
newsfeed
actually
has
the
vast
majority
of
tweets
among
the
political
accounts
right.
B
Trolls
developed
the
most
tweets
and
then
left
foreign
far
less
than
right,
troll
and
they're,
not
the
other
categories
are
pretty
low
as
well,
so
so
our
tweet
counts,
definitely
not
even
distributed
across
here
and
that'll
have
some
implications
for
some
of
the
things
I'll
talk
about
later,
and
so
these
are
some
statistical
models
that
I
ran.
They're
actually
model
sets.
I've
talked
more
about
this
during
q
and
a
but
due
to
the
nature
of
the
data
set.
This
is
5.2
million
tweets.
B
For
electrodes
and
the
red
dots
were
the
situations
where
platforms
are
combined
with
electrodes,
and
so
what
you
want
to
focus
on
is
in
the
top
of
this
chart.
You
see
that
the
impact
of
lactose
when
separated
is
much
much
larger
than
leftovers
without
landfills,
as
you
see
here,
and
what
this
dot
this
red
dot
over
here
on
electrodes
shows
is
that
the
vast
majority
of
the
effective
left
trolls
is
actually
an
effect
of
lactose
with
lactose
intolerance,
so,
in
other
words
his
categorization.
B
The
methodological
decision
in
other
ways
to
separate
lactose
and
reptiles,
has
a
huge
impact
on
the
results,
because
when
you
do
separate
leftovers
from
lactose,
you
see
that
the
effect
of
leftovers
is
actually
very
similar
to
that
of
right
holes
right.
So,
in
other
words,
when
they
were
combined,
it
looked
like
left
fold
was
there's
big
driving
factor.
When
you
separate
them,
you
realize
you've
actually
got
driving.
That
repeat
effect.
Looking
at
the
rest
of
these,
you
can
see
like,
for
example,
most
of
the
code.
Various
not
didn't
really
have
any
effects.
B
A
B
Areas
for
entrepreneurs
for
future
research,
so
this
is
for
retweets.
I
also
did
this
for
like
models
likes
as
well
as
requirements,
the
next
line,
but
most
of
them
basically
have
the
same
category,
which
is
when
you
separate,
laterals
and
left
holes.
The
effect
of
electrons
goes
down
a
lot.
It
looks
a
lot
more.
It
moves
a
lot
closer,
so
that
categorization
effect
is
is
pretty,
is
pretty
strong,
but
this
one
then,
of
course
we
see
the
image
and
video
the
system
effect
there.
B
B
The
effect
here
is
a
little
bit
less
clear,
but
you
can
still
see
generally
when
you
take
out
the
left,
the
black
flushing
retros,
the
coefficient
the
electrons
continue
to
decrease
to
move
closer
to
zero.
So
those
are
the
basic
findings
they're
there
and
again
through
the
time.
I
can't
like
get
into
all
of
the
details,
but
something
happens.
Let
me
have
to
talk
about
that
during
here.
There
are
close
thanksgiving
statistical
details
on
that.
B
I
also
did
an
investigation
of
what
is
known
as
false
amplification,
and
so
this
would
be
the
possibility
where
the
ira
would
actually
be
retweeting
replying
to
them
liking
their
own
content.
But
I
should
just
repeat
it
online.
I
couldn't
figure
out
the
like
piece
of
it,
but
I
did
look
at
retweets
and
replies
and
I
found
that
a
little
less
than
three
percent
of
the
ira's
retweets
were
provided
by
ira
accounts
about
1.27
of
replies
were
also
created
by
our
ira
accounts.
B
So
that's
that's
what
what
was
going
on
here.
So
a
very
small
proportion
of
these
refusing
replies
are
actually
provided
by
the
ira.
Now
that
doesn't
mean
that
the
rest
of
them
were
provided
by
real
people,
but
it
does
show
at
least
that
within
the
scope
of
known
ira
accounts,
they
are
we're
not
spending
a
lot
of
time,
juicing
their
own
stats
in
the
world.
B
So
this
is
good
evidence
or
some
evidence
that
many
of
the
basketball-
but
these
other
folks
may
have
been
probably
where
real
people
are
interacting
with
this
country.
That's
also
consistent
with
what
twitter
has
said
about
about
how
folks
who
are
interacting
with
so
now.
I'm
gonna
move
on
to
the
second
study,
and
so
this
one
was
really
about
the
ira's
audiences
and
how
they
were
connected.
The
studies
available
in
the
misinformation
review
issue.
This
is
collaborative
research
with
tanya
lovato
who's
at
dublin
city
university.
B
So
for
this
one
I
also,
I
also
realized
that
I
forgot
to
say
anything
about
the
data
set
for
the
previous
study.
So
if
you
have
questions
about
the
data
center
happy
to
answer
those
questions,
but
this
data
set,
we
downloaded
over
two
million
replies
to
iraq,
foods
over
nine
and
a
half
years.
We
use
network
analysis,
sorry
community
infection
to
find
out
audience
overlapping
details
about
how
the
audiences
were
separated.
B
B
So
what
you're
looking
at
applies
to
ira
accounts,
so
you
have
four
different
right-wing
communities
of
different
sizes,
the
one
that
I
call
right.
Three
has
trump
in
it.
The
other
ones
are
the
two
big
ones
and
two
smaller
ones.
We've
got
one
for
black
interests,
another
slightly
larger
one
for
leveling
interests
and
then
the
the
width
of
the
lines
between
them
are
the
total
number
of
shared
replies
from
indies
who
are
one,
can
use
minimizing
another.
B
So
this
is
a
look
at
ira,
audience
overlap,
and
so
the
y-axis
on
this
is
the
percentage
of
internal
replies
and
so
the
higher
the
percentage
of
internal
requirements,
the
less
engagement
with
other
communities
that
wasn't
so
russian.
Obviously
again,
this
is
sort
of
like
a
reality
check
like
so
the
russian
language,
one
had
almost
100
percent
internal
replies,
meaning
very
little
engagement
with
other
communities
which
make
sense
between
language
barrier,
health
and
diet
again
above
90,
and
that's
a
topical
difference
right.
B
So
you're
talking
about
health
and
diet,
you're,
probably
not
talking
about
politics,
and
so
that
explains
why
there
wouldn't
be
a
whole
lot
of
overlap
there.
You
can
see
that
the
left
and
the
black
ones
that
are
a
little
under
70
were
generally
pretty
insular
a
lot
with
hashtag
gamers,
but
the
most,
the
ones
that
I
gazed
with
external
communities.
The
most
were
the
right
ones,
so
they
were
almost
engaged
with
each
other,
so
they
had
a
lot
of
overlap
between.
A
B
B
It's
on
a
normalized
scale,
centered
on
zero
and
then
of
course,
positive
numbers
are
right.
Our
right
leg,
negative
numbers
are
our
left
wing,
and
so
this
is
again
just
a
reality
check.
You
see
that
all
the
right
wing
averages
these
are
resolved
values
as
well,
execute
to
the
right
with
right
wing.
B
Russia
is
slightly
right
that
one
is
in
the
lowest
valley,
because
it's
not
really
in
english.
News
is
right
there
on
the
center
hashtag
gamer
health
and
diet
left
and
black
activists
all
left
to
center
right.
We
see,
in
other
words,
that
the
ideologies
of
the
individuals
engaging
with
these
russians
with
these
russian
accounts
match
the
famed
identities
of
the
of
the
russians
right.
B
So
so
black
activists
are
talking
to
black
presenting
the
iowa
accounts
and
those
black
activists
themselves
tend
to
mostly
follow
other
left-hand
accounts
of
known
ideology,
and
the
same
is
true
for
writing.
The
constitution
with
white
presidential
accounts,
okay,
so
third,
one
three
or
four.
To
what
extent
did
the
ira
change
people's
minds
again?
B
I
was
a
bit
of
a
good
player
on
this
one
from
p
nas
in
january,
very
interesting
and
rigorous
media
effect
study,
and
so
I'm
just
going
to
go
through
the
primary
I'm
not
going
to
spend
a
whole
lot
of
time
on
this
one.
I
provided
some
of
the
data
for
it,
so
I
got
to
go
out
there,
but
if
you
like
talking
about
it
because
it's
pretty
prominent
venue,
if
you
have
questions
about
this
one
but
I'll-
try
to
answer
whatever
questions
you
may
have
about
this.
B
There
were
democrats
and
republicans
who
use
twitter
at
least
three
times
a
week.
We
also
got
their
twitter
handlers
so
that
we
can
reconstruct
what
kinds
of
ira
content
that
they
may
have
been
exposed
to
so
survey.
Data
linked
with
digital
trace
data
on
twitter,
a
very
unique
design
that
allowed
us
to
analyze
interactions
with
ira
accounts
in
their
impact
with
the
two
waves
on
attitudinal
behavioral
measures.
B
So
a
couple
I'm
going
to
talk
about
a
couple
different
results.
I
don't
talk
about
everything,
but
these
are
predictors
of
ira
interaction,
and
so
this
is.
This
is
a
binomial
regression.
You
can
see
that
the
biggest
predictor
of
ira
interactions
is
the
percentage
of
co-partisans
followed.
In
other
words,
if
you
follow
people
that
share
similar
ideologies
to
so
the
ideology
scores
we
used
here
were
very
similar
to
the
ones
that
were
used
in
previous
study.
B
This
is
the
big
one.
These
are
effects
on
attitudes
and
behavior.
You
can
see
you
see
the
you
know
the
credible
intervals
here.
This
is
basically
causal.
Forest
models,
which
always
goes
right
in
runs.
Don't
totally
understand
them,
but
the
the
chart
is
actually
pretty
easy
to
understand.
Those
credibles
95,
credible
intervals
do
cross
over
that
zero
point,
which
means
there's
no
effect
for
any
of
these.
So
congratulations.
B
Pmask
for
publishing
a
study
with
a
null
effect,
and
so
you
look
at
changing
outposts
and
party
healing
thermometer
changes
in
the
desired
social
distances
also
came
out
before
covered
nineteen's.
This
is,
like
you
know,
actual
social
distance
as
defined
not
related
to
any
kind
of
epidemiological
situation,
changing
ideology,
the
little
conservative
index,
etc,
etc.
Very
little
effect.
This
is
also
consistent
with
what
we
know
about
media
effects.
B
You
know
political
ads,
news,
very,
very
small
effects,
and
so
it's
totally
consistent
with
the
idea
that
the
ira
didn't
really
change
people's
minds
significantly
through
their
social
media
interaction.
The
last
thing
I
said
about
this
board
and
one
is
that
the
ira
can
have
other
types
of
effects
other
than
changing
people's
minds
from
one
direction
to
another,
which
is
something
I'll
discuss
as
I
wrap
up.
B
B
If
they
gave
that
one,
a
volumizing
number
pretty
quickly,
sometimes
don't
really
use
for
that,
but
that's
great.
This
is
already
officially
out.
You
can
check
out
the
international
politics.
So
what
we
did
was
we
looked
at
117
news
outlet,
source
from
the
data
aggregator
platform
media
cloud.
We
looked
at
the
100,
most
repeated
ira
candles
or
storage,
published
stories
published
in
2015
2017.
B
We
found
314
stories
containing
ira
references
published
by
71
hours,
which
is
6.7
of
all
the
outlets,
so
the
ira
was
able
to
really
push
that
message
out
to
the
majority
of
the
top
media
outlets
that
we
looked
at
and
there
is
the
supplemental
material.
This
article
shows
all
of
the
media.
I
would
look
at
so
yeah.
I
was
able
to
get
into
the
majority,
and
so
the
findings
here
is
that
we
categorize
these
traditional
media,
digital
native
outlets.
B
There
are
a
couple
other
ones,
it's
mostly
digital,
innovative
outlets
that
incorporated
the
ira
content
compared
to
visual
media,
about
80
of
those
incorporating
ira
content
where
these
are
the
native
ira
accounts,
mostly
provided
opinion
content.
Sorry,
I
should
say
opinion
content
versus
information
owners
rather
than
anybody,
so
67
percent
providing
opinion
content
rather
than
information
only
and
then
the
ira
accounts
are
mostly
described
as
everyday
people.
B
That's
where
the
box
popular
and
the
top
title
of
that
study
period,
so
they're
presented
as
your
average
everyday
citizen,
which
I'm
sure
is
exactly
what
that
those
disinformation
for,
and
so
this
is
one
example
of
the
media
sort
of
getting
by
the
by
the
ira,
so
yeah
and
the
last
thing
I
want
to
show
for
this.
This
is
actually
something
that
appears
the
supplemental
material
of
this
article,
but
I
wanted
to
pull
it
out
because
I
didn't
think
it
was
anything.
B
This
is
the
ranking
of
the
most
frequently
of
the
news
outlets
that
most
frequently
included
ira
content,
and
I
also
just
sort
of
coated
this
very
briefly
with
an
ideological
slant.
You
can
see
that
there
are,
you
know
you
see
left
right
and
center.
When
I
say
center,
I
simply
mean
not
having
a
self-declared
ideology.
B
I
know
many
people
think
that,
like
the
washington
post
would
not
be
using,
the
mainstream
media
are
liberal,
but
they
are,
they
do
have
a
balance
of
viewpoints
on
their
editorial
pages
and
their
coverage,
and
so
I
actually
had
that
assembly.
So,
if
you
look
at
this,
you
can
see
that
five
out
of
the
ten
were
right
wing
two
out
of
the
ten
were
left
winning
and
then
three
were
centered.
So
it
really
wasn't.
Just
you
know.
B
B
The
ira's
audience
generally
near
explain
that,
in
other
words,
conservatives
will
talk.
Search
representing
ira
accounts,
liberals
or
left-wingers
are
definitely
talking.
Left-Wing
presenting
ireland
accounts,
the
ira
generally
didn't
change
people's
minds
and
finally,
the
ira
senators
used
to
be
on
twitter
by
a
mix
of
netflix
news
media,
so
those
are
ira
specifically
more
generally
for
digital
digital
information.
My
conclusions
are
certain
tactics
and
groups
can
be
disproportionately
effective
informally.
So
the
black
community
is
a
case
point
right
with
the
protests
going
on
right
now.
I'd
be
very
surprised.
B
They
were
not
very
conservative
efforts
to
introduce
disinformation
by
a
space
content,
state
sponsored
entities
and
other
other
entities
into
the
situation,
racial
and
ethnic
type
intentions
here
and
elsewhere,
maybe
especially
by
transportation.
That's
something
that
I
think
future
research
should
really
take
a
hard
look
at
social
media.
This
information
is
not
limited
only
to
social
media
right,
it
spreads,
and
so
you
can
think
of
the
news
media
here.
B
At
super
spreaders
right
in
reference
to
code
19,
but
the
news
media
due
to
their
increased
reach
relative
to
average
users
and
spread
this
content
beyond
social
media.
So
people
who
don't
even
have
a
twitter
account
or
even
a
facebook
account
can
still
interact
with
it
and
still
offensively
affected
by
it
and
then
the
last
conclusion
I
have
from
the
slide
is:
don't
expect
this
information
to
change
people's
minds?
That's
consistent
with
what
we
know
about
the
effects.
B
If
there
are
many
other
kinds
of
effects
that
we
should
explore,
including
just
domain
two
increased
polarization
extremism.
Voter
suppression,
as
we
saw
on
the
visual
slide
earlier
in
the
presentation,
as
well
as
potentially
other
ones
and
I'd,
be
very
interested
to
hear
about
your
thoughts
about
what
those
could
be.
So
I'm
gonna
pause
there.
Here's
my
contact
information
and,
if
you
have
questions
I'd,
be
very
interested
to
hear
those,
and
thank
you
very
much
for
listening.
B
You
know
so
I
think
if
you,
if
you
think
back
to
that
slide
that
I
showed
you
know
earlier,
that
slide
actually
shows
a
couple
different
actually
back
up
to
that
slide.
Just
so
so
I
can
show
you
now.
Obviously
this
is
not
going
to
be
valid
on
any
kind
of
quantitative
level,
but
I
think
it
does
show
some
examples
of
the
rhetorical
strategies
that
are
used.
B
One
is
kind
of
you
know
you
show
a
black
face
and
say
I
don't
vote,
that's
kind
of
an
in-group
sort
of
appeal
that
says:
well,
you
know
black
people
shouldn't
vote
right.
So
it's
not
for
us.
So
don't
do
it
look
out
for
the
group's
interest
and
don't
vote.
That's
one.
You
know
tactic,
do
not
vote
for
oppressors
right.
So
it's
look
at
how
we've
been
oppressed
in
this
country.
B
You
know
sort
of
between
the
truth
or
dumb
and
approach
that
you
that
you've
seen
not
only
a
black
community
but
also
in
sort
of
the
further
reaches
of
the
left
for
many
years,
and
so
that's
not
a
race,
specific
appeal,
but
one
that
I
think
you've
seen
many
cases.
So
those
are
two
examples
of
strategies
now
to
be
able
to
figure
that
out.
B
I
think
you
need
to
find
out
a
way
of
sort
of
isolating
those
kinds
of
those
kinds
of
strategies,
and
I
you
know
just
think
about
in
my
head.
I
would
love
to
have
say
a
content
analysis
to
do
that.
I'm
not
sure
how
well
like
a
machine
learning
algorithm
or
something
highly
technical
would
be
able
to
extract
those
kinds
of
high
level
tactics
from
say,
a
tweet
or
even
from
an
image
that
a
person
will
be
able
to
look
at
or
train
coder.
B
A
Then
this
is
nitin
from
university
of
arkansas.
This
is
a
very
fascinating
talk
and
I'm
just
left
wondering
with
so
many
new
possibilities
that
can
be
done.
That
can
be
studied.
Now
from
your
findings,
I
had
a
question
about
the
research
question
2,
I
believe
which
mentioned
about
audience
ideology
analysis.
B
Okay,
so
so
that's
a
good
question,
so
the
the
panel
came
from
yuko,
which
is
a
very
well
known.
You
know
provider
of
panel
data,
and
so
they
were
responsible
for
and
so
they're
used
widely
throughout
political
science.
So
they've
they
made
actual
contact
with
these
films
and
they
you
know
they
had
to
to
take
in
their
information
to
give
them
their
incentives,
and
so
they
were
given,
like
you
know,
11
equivalent
of
like
11,
and
you
got
points
that
they
can
then
which
they
can.
B
Then
you
know
redeem
for,
like
amazon,
you
know
content
and
that
sort
of
thing.
So
so
this
this
this
panel
company
basically
handled
the
determination
and
they
didn't
simply
look
at
some
of
the
twitter
account
they
rolled
them.
They
actually
got
verification
that
these
were
real
individuals,
and
that
is
really
important.
You
don't
want
bots
to
contaminate
your
analysis,
and
so
I
thought
it
went
with
you
in
terms
of
repeating
this
panel.
A
B
B
A
B
A
B
But
you
know
I'm
not
sure
whether
they
have
changed
their
policies.
Specifically,
we
do
make
recommendations
at
the
end
of
the
paper
that
news
outlets
should
try
to.
You
know,
do
a
little
bit
more
than
just
looking
up
looking
a
tweet
up
and
then
pasting
it
into
their
article,
trying
to
do
a
little
bit
more
due
diligence
and
trying
to
figure
out
or
trying
to
ascertain
that
this
is
not.
B
A
A
I
I
have
a
for
forward-looking
question,
so
I
I
like
your
final
comment
that
just
because
we
found
that
you
know
these
campaigns
did
not
change.
Any
minds
doesn't
mean
that
they
don't
have
an
effect,
and
so,
following
up
on
kathleen's
question
about
voter
suppression,
how
would
we
go
about
measuring
that
in
the
future
like?
Could
we
have
you?
Have
you
given
any
thought
about
either
panel
type
or
or
other
methodologies
to
to
look
at
whether
these
kind
of
campaigns
can
can
have
an
effect
on
on
on
voters?
A
It
seems
to
me,
like
that's
the
low
hanging
fruit
in
terms
of
like,
if
I
was
trying
to
to
carry
out
at
this
information
campaign
and
wanted
to
affect
the
elections,
I
would
never
try
to
change
somebody's
mind.
I
would
just
try
to
make
so
that
certain
groups
are
more
likely
to
go
to
vote
and
others
are
less
likely
to
go
to
vote,
and
your
examples
are
very
clear
about
that.
So
any
thoughts
about
how
how
can
we
measure
impact
of
that
kind.
B
Okay,
that's
a
great
that's
a
great
question.
My
first
thought
about
it
is
you
know
I
I
sort
of
have
two
suggestions
and
they
sort
of
map
onto
you
know
the
traditional
kind
of
research
methods
of
inductive
and
deductive
right.
So
there
is
sort
of
a
deductive
hack
that
you
take.
That
looks
at
the
overall
goals
of
disinformation
in
sort
of
a
very
abstract
and
high-level
way.
So
what
is
it
that
these
disinformation
agents
are
basically
trying
to
do,
and
so
you
can.
B
From
existing
research
on
multiple
disinformation
operators,
so
we
know
so,
voter
suppression
is
licensed,
so
I
said
hold
on
hold
on
a
little
bit
of
specialties.
That's
actually
part
of
the
industry,
but
so,
in
other
words,
you
think
about
this
information
as
a
component
of
information
work
there
so
inflicting
damage
on
the
envy.
How
do
you
put
damage
on
the
enemy?
B
Without
you
know
using
bullets
right,
so
you
can
demoralize
them.
You
can
confuse
them.
You
can
make
them
doubt
whether
the
content
that's
coming
from
their
leaders
is
true
or
not.
You
can
turn
them
against
each
other
divide
and
conquer.
So
all
these
things
are
things
that,
before
you
even
look
at
any
data
you're,
taking
this
deductive
approach
and
you're
being
able
to
look
out
for
this
stuff,
second
piece
is
inducted.
B
Let's
take
a
look
and
see
the
100
most
frequently
retweeted
ira
tweets
and,
let's
try
to
you,
know
divine,
in
other
words,
by
looking
at
them.
What
are
the
tactics
that
they're
really
using
here?
So
the
examples
that
I
gave
that
I
showed
are
sort
of
part
of
that
inductive
approach
right,
we're
saying,
okay.
This
is
a
very
clear
example
of
what
it's
about.
What
else
am
I
gonna
be
trying
to
do?
You
know
some
of
the
the
very
sort
of
derogatory
language
that
they
use
in
reference
to.
B
B
So
obviously,
with
the
planning
that
went
into
it,
it
was
just
very
fortuitous
that
we
had
this
nice
wave
that
the
person
that
occurred,
the
ir
was
still
online.
The
second
wave
happened
and
then
all
the
stuff
went
away.
So
that's
something
that's
very
very
hard
to
plant
here,
and
so
we
were
able
to
capitalize
on
that
to
be
able
to
do
these
very
this
very
nice
effect
study,
but
you
know
so.
B
The
other
piece
of
this
is
you
think,
about
social
media
policies
that
mandate
the
removal
of
this
content
as
soon
as
it's
discovered
it
makes
it
makes
this
makes
discovering
these
effects
very
very
hard,
so
there
may
be
some
sort
of
like
post
hoc,
where
you
know,
I
think
you
probably
would
have
to
get
some
sort
of
participation
from
twitter
or
facebook
or
the
social
media
platform
to
do
this,
because
I
wouldn't
really
want
to
like
rely
on
the
memories
of
interacting
with
these
folks.
Oh
wait.
B
What
was
kind
of
going
through
your
mind
now,
obviously
there's
going
to
be
some
like
mnemonic,
like
degradation
from
like
a
year
later,
but
it
would
at
least
be
able
to
say
we
have
verified
evidence
that
you
communicate
with
these
folks
and
then,
of
course,
you've
got
the
social
desirability
piece
which
is
you
know,
people
don't
want
to
like
people
don't
want
to
exceed
it
like?
I
don't
know
what
they
can.
B
So
that
would
be
another
challenge
in
terms
of
trying
to
get
people
to
talk
about
this
honestly
and
say:
here's
how
I
connected
with
it
and
here's
what
I
was
thinking
when
I
did
it
so
the
memory
piece
and
the
social
desirability
piece,
I
think,
are
two
sort
of
major
hurdles
to
get
over
and
then,
of
course,
the
social
media
delete.
Everything
is
going
to
see
if
these,
I
think
I'll
complicate
this
kind
of
research,
but
very,
very
well,
well
worth
it.
Thank
you
for
that.
For
that
question.
A
B
Yes,
I
do
because
so
we
looked
at
it
so
for
the
recoding,
so
I
you
can
take
a
look
at
lindelof
walker's
original
paper
in
terms
of
how
they
categorize
their
their
accounts,
but
for
us
so
the
only
one
we
categorized
was
the
black
code.
So
what
we
did
is,
I
said
we
looked
at
the
screen
name
and
a
random
sample
of
five
tweets.
B
So
I
think
that
by
using
a
similar
method
across
media
you
could
look
only
and
so
the
other
way
to
enhance
that
would
be.
That
would
be
the
avatars
for
it.
So
the
visual
presentation
which
we
didn't
have,
but
simply
on
the
basis
of
the
screening,
how
that
was
written
and
a
random
sample
of
five
messages.
We
were
able
to
make
that
information
that
half
of
those
classified
originally
as
left
trolls
were,
in
fact
laterals
by
the
definition
of
a
prospect,
so.
A
A
B
Yeah
so,
but
they
were
only
like
they're,
only
like
four
less
than
four
thousand
accounts,
so
whatever
you
do
and
then
the
nice
thing
is,
you
know
with
the
human
coding
you
you
know,
that
gets
a
lot
of
things
that
I
don't
think
machines
are
gonna,
be
able
to
do
aspects
of
slang.
I
mean
you.
Might
I
mean
if
you
train
your
model
well
enough,
it
will
do
it.
B
The
other
problem
is
that,
because
the
data
set
is
so
small,
you
don't
really
have
a
whole
lot
of
training
data,
and
so
you
have
to
rely
on
people's
knowledge
of
you
know.
What
is
what
what
does
black
presenting
mean
in
a
textual
way?
And
so,
if
you
had,
you
know
potentially
hundreds
of
thousands
of
these
accounts,
you
might
be
able
to
train
a
model
well
enough
to
perform
on
it,
but
with
fewer
than
the
four
thousand
and
nine
different
categories,
you're
sort
of
slicing.