This document summarizes a presentation about how social media and artificial intelligence may be undermining democracy. Micro-targeting of political ads using vast user data allows messages to be tailored to small groups in a way that reinforces polarization. Automated bots and fake accounts spread both real and fake messages virally. Deep learning AI will make influence techniques even more sophisticated and difficult to detect. Studies show these tools have been used to shape elections internationally and in the US. The future may include advanced deepfakes and increased digital voter suppression efforts. Recommendations for 2020 include better detection of disinformation across platforms and improved media literacy education.
1. AI, Social Media and
Political Polarization:
Is Our Democracy at Risk?
Dave Hess &
Jim Isaak
Sept 27 & Oct 4th 2019
OLLI - Concord
2. Social media and artificial intelligence (AI) are reshaping our political and electoral processes – and
not for the better. Increasingly sophisticated algorithms searching through massive data sets enable
political parties, candidates, and nation states to micro-target advertising with messages tailored
to small, carefully selected segments of the population that are most susceptible to a particular
message, thereby reinforcing the political dispositions of the recipients. Even more perniciously, this
technology provides insights into how to tailor that message to most effectively influence those
recipients. Online “bots” (automated accounts) can disseminate such messages – both real and
indistinguishably fake ones – to millions of online recipients instantaneously – all known only to
those who are receiving them. And these methodologies will become even more sophisticated and
undetectable as AI advances to “deep learning”. This technology is contributing to, if not to a
significant extent causing, the further political polarization of our population as people increasingly
receive only political messages tailored to and reinforcing their own political predispositions to the
exclusion of other perspectives and points of view. We will ask: What are algorithms? Bots?
Machine learning? Deep learning? How do they work? How do they shape, influence, manipulate,
distort and polarize political discourse? And how have they been used historically to shape political
thought and influence the outcome of elections as, for example: by ISIS in Iraq ; in the Brexit vote in
England; and in our own 2016 presidential election (to mention but a few). And what does the
future hold?
9/30/2019 https://is.gd/AIPolitics
3. A few Technology Concepts
• Algorithms?
• Big Data?
• Digital Footprints?
• Deep learning?
• Trolls, Sockpuppets and Bots (oh my!)
And why do you care?
4. What are algorithms?
Recipes
Jims Fudge
• In a double boiler:
– Melt 6 Tbsp butter
– Stir in
• ½ cup cocoa
• 4 Tbsp milk
• 1 tsp vanilla
• 1 lbs powdered sugar
until smooth
– Pour into buttered pan
– Cool in frig until hardened
5. Actual Political Algorithm circa 1970
• For each precinct(I) in town
– R=0, D=0, U=0
– For each voter in precinct
• if voter is republican then R=R+1
• If voter is democrat then D=D+1
• Else U=U+1
– Next Voter
– Precinct(I) = {R,D,U}
• Next Precinct
• Sort Precincts (highest R first,
lowest D second)
• Print priority list for
republican_voter_contact
• Candidate objective:
prioritize precincts to
visit door-to-door
with maximum
republican contact
density
6. But, a typical program
• Contains thousands or millions of instructions
• Often broken into “sub routines” or
“methods” which can be reused in many
programs
[send_email(emailaddr, msg) might be an example]
• An associate tells me that there are 100,000+
identifiable “flaws” in MS Windows, but in
sections of code so old no-one is prepared to
fix it (or perhaps the source code is lost)
7. Software/Program objectives
• Traditional – programmer implements a model that solves the problem
• Emerging AI – Deep Blue (Chess-1996) (“brute force learning”)
• Trained Artificial Intelligence – (2010)
programmer provides data set, examples and evaluation criteria and
the program develops the model (AlphaGo, language recognition/translation)
Watson (Jeopardy 2011, medical diagnosis, xRay evaluation 2017)
Google/Facebook ad insertion, headline selection, recommendations
job candidate selection, credit evaluation, real estate ad presentation
• Goal driven AI – 2017
Programmer provides model for evaluation and defines output set,
program develops model (Alphazero, facial analytics)
• General AI -- ??? --
Program revises actions, and perhaps goals independently
(a significant % of people think it is human or super human)
8. Big Data?
• Disk storage is now cheap – often measured in
terabytes even on consumer computers
• 15 terabytes = total Library of Congress (text)
• Tools exist to process diverse data formats – so
queries can be made that span different
databases, and even “unstructured” data
• Facebook collects 500+ terabytes/day (2017)
• Youtube uploads 500 hours of video/minute (2019)
Don’t delete anything, you never know ….
9. Big Data + Deep Learning
• Can infer information about individuals that
they might not “know” or admit*
– Political orientation, property holdings
– Sexual orientation, credit card transactions
• Key “hot buttons” that will motivate them you
to act (buy, vote, …) or
not act (voter suppression)
*Key papers by Michel Kosinski, 2012+ working with Facebook now at Stanford
10. Digital Footprints?
• When you connect to the web/net that is your
first “foot print”
– User from this IP address, MAC id, date/time
– Store cookie #123 on user’s system
• Go to a web site/service/server
– #123 at site/url, date/time/query string
…duration
• Email content (written, read), actions
• “click thru”, “like”, even “hover” in some cases
11. “But I never use Facebook”
• Even if you never use Facebook, they can track
you … initially they may not know your name,
phone number, address, personality, age,
income level, political orientation, sexual
orientation, facial image, ……but
“we only share your personal data with our trusted
business partners” (typical privacy statement)
• Which either means a contract, or simply sale
of your data (what is a business partner?)
12. Advertising Business Model
• Facebook, Google, et al use this model
• Revenue comes from selling ad placements,
and views
– Advertisers are the customers/revenue source
We are the product
• The greatest revenue comes from
– Maximum selectivity – “Micro targeting”
– Maximum attention opportunities
(“engagement” = Getting you to stay on their site)
13. Channels
• A new “ad” or “news feed” item can be placed
on each page as you click though the web
• A “campaign” targeting you (or users like you)
can span from web site to web site
• In theory it can span from the web to cable TV
“ad insertion” directed to your TV,
• Certainly it can touch the paid content inserted
in streaming videos, online “radio” etc.
14. Fake net users/sockpuppets
• Part of the allure of the web is anonymity
• Folks often do not use their real names
• And can have many accounts with various
services –Facebook, Gmail, Hotmail, Instagram
• Some of this may be “legitimate”
• Others may be malicious
I could have 500 folks who “Like” me on Facebook,
498 are my own fake accounts: “sockpuppets”
15. Trolls (not just under bridges)
• Persons or organizations operating a number of
fake user ID’s pushing an agenda, independent of
Truth, Justice or the American Way
• Often using outrage and fear to gain attention and
get their content to go viral
(i.e. shared by initial users to their lists, and from
those folks to their lists, etc.)
• Create fake user “pages”, also fake “news sites”,
feed content to blogs/news-sites that are not
trying to be objective/accurate/etc.
16. Not every abuser is a troll
• You can get paid by Google, Facebook, etc. for
putting their ads on your site
• So, if you create a “fake news site”, and get some
good outrage/fear going, you can get rich –
some kids in Macedonia have done this
• “The Pope has endorsed Donald Trump”
(a real instance of a fake news meme that went viral)
• And the real trolls can point to this as part of
their campaigns
17. Bots?
“On the Internet, no one knows you are a dog”
This is half true – it is fairly easy to create “fake accounts”
(gmail, Facebook, etc.) … particularly if you have an army
of folks working at this (literally)
Facebook has tried to purge fake accounts (millions!), but
many have years of apparent activity
Bots are captive computers that can be triggered to convert
a troll post into a “trending” item by clicks, LIKEs, sharing,
re-posting, etc.
(turning off your computer reduces the risk of your computer
participating in a zombie bot-net)
Click farms are large numbers of accounts, with either
automated or humans that “click” on targeted content to
trigger “trending”, etc.
18. The IRA is not Irish
• Internet Research Agency is a Russian troll
farm (lots of employees, millions of fake
accounts) used to promote their agenda over
a number of years.
• One way to build web-credibility is to have a
long established site/id that gradually builds a
network … may use “real names” and bios also
19. IRA Trolls triggering US events (NYU)
In 2016, the technique entailed phony
IRA social media personas or groups
first announcing events, such as dueling
anti-Muslim and pro-Muslim demonstrations
that the Russians successfully
organized in May 2016 in Houston
20. LikeWars (2018) – some notes
• When Trump announced, 58% of his followers were
overseas, many were click farms or bots
• 30+% of the tweets in the BRIXIT window were from
botnets; ditto 2016 election, ditto Brazilian election
• “Tennessee GOP” was one fake user id,
managed by the Russian Trolls
it trended at #7 just before the 2016 election
• “Angie Dixson” is a Russian troll inflaming tension prior
to the Charlottesville white supremacist rally
“And the beat goes on”
21. The really sophisticated promoters
• Combine many components
– Web sites for (fake) organizations
– Sock puppet “Friends”
– Email lists and postings
– Real world “events”
– Fake news stories and sites
– Paid placements with micro targeting
– “Clickbait” ideally trending and going viral
Trolls, sockpuppets and bots – oh my!
22. Psych Traits Social Media Exploits
1. Conformation bias
2. Availability bias
3. Emotive bias
4. Repetition bias
5. Visual bias
6. Belonging/”herd” bias
7. Social acceptability
23. Attention “grabbers” (Emotive bias)
Fear and Outrage
• “Race to the bottom of the brainstem”
(Tristan Harris, ex Google Ethics manager)
– Likely to trigger a “click thru”
– Likely to be shared
• The “paid placements”
(ads, “news feed positioning”, etc.)
headlines are selected knowing this
• “Trending” is used to present current “hot” content to
target communities to build engagement
Truth, accuracy, credibility, etc are NOT a factor
24. Implementing a Social Media Campaign
1. ID Macro Audience
2. Goals (participation, persuasion, activate, GoTV, discourage)
3. Collect “big data”
4. In depth surveys (10,000+) “Computational linguistics”
5. Segmentation
6. Mine bid data – “Look alike builder” (Facebook)
7. Micro-target segments
tailored, cheap, confidential “under radar”
8. Conversion Rate Optimization .. Most effective slogan
A/B testing, constantly fine tuned (200,000+ posts)
25. “This ad uses dynamic creative, a process whereby
advertisers upload multiple image and text
options and the best performing combination for
the audience is automatically created.”
• Facebook Ad Archive note
(on a Frosted Flakes Ad, listed as a national issue)
• i.e. You have no idea what version is being presented to
any specific individual or community
26. Micro Targeting
• The more specific the target individual(s):
– Fewer # to “buy” (lower cost)
– More likely the desired impact
• Ideal: a single individual as he/she approaches
their decision (currently “looking”) --- which
requires
– real time monitoring of “everyone” and
– knowing who they are in detail
This is now automated
27. Looking Forward to 2020?
Are we there yet?
Slides posted, along with syllabus,
Resource list and links to related content
https://is.gd/AIPolitics
28. Perspective Difference
• How did specific initiatives impact the 2016
Trump Campaign? …
– Cambridge Analytica – not a factor
– Russia – minimal impact
• Impact on national discourse
– Cambridge Analytica – created visibility for issues
– Russia – Driving polarization at every opportunity
with “Strategic Intent” and significant financial,
staffing and expertise invested
29. Voter Suppression (NYU)
Facebook has said that in the weeks before the
2018 midterm election it found and removed
45,000 pieces of voter suppression content.
“Few behaviors strike as directly at the
heart of democracy as confusing or
bullying people who are entitled to vote.
The social media companies will have
to remain vigilant on this front.“
[An interesting 1st amendment question]
30. 2020 Political Algorithm
• Select set of critical precincts in swing states
(or targeted congressional districts)
• Identify all persons likely to vote
• For each voter:
– If on-our-side then flood with ads that maximize probability of voting
(terrible things “they” will do – tied to individual’s hot button issues
and matched to their personality)
– If not(on-our-side) then flood to minimize probability of voting (at
least for viable opponent)
– Track each paid placement/click-bait to see which
individuals are reading/sharing, learn what works, more
finely profile individuals for next point of
placement/contact. – in real time, per target individual
31. Social Media makes Ads Visible
(sort of)
• Facebook / Instagram Ad Archive
• Google / Youtube ad Transparency
• Twitter Transparency site
32.
33. Google Transparency Report May 2018 to Sept 2019
TRUMP MAKE AMERICA GREAT AGAIN COMMITTEE
$6,800,800
SENATE LEADERSHIP FUND $5,124,600
CONGRESSIONAL LEADERSHIP FUND $4,159,200
Need to Impeach $3,596,500
NRCC $3,453,900
DEDICATEDEMAILS.COM $2,800,900
NRSC $2,695,000
PRIORITIES USA ACTION & SMP $2,569,200
DONALD J. TRUMP FOR PRESIDENT, INC. $2,412,700
TOM STEYER 2020 $2,255,900
WARREN FOR PRESIDENT, INC. $2,223,500
INDEPENDENCE USA PAC $2,208,900
PETE FOR AMERICA, INC $2,029,200
BETO FOR TEXAS $1,944,000
ONE NATION $1,740,100
KAMALA HARRIS FOR THE PEOPLE $1,430,600
34.
35.
36. Analysis and Reports
• 2018 US Senate received 2 reports on 2016 Russian
Interference:
The IRA and Political Polarization in the United States,
(from Oxford (UK) Computational Propaganda Project, Dec 2018, 47 pages)
The Tactics & Tropes of the Internet Research Agency,
(from New Knowledge, a U.S. Web security firm; Dec 2018, 101 pages)
• Aug. 2019 – NYU did a report based on these, the
Muller Report and subsequent analysis:
Disinformation and the 2020 Election:
How the Social Media Industry Should Prepare
Links on website: https://is.gd/AIPolitics
37. From NYU report
In its 2019 Worldwide Threat Assessment,
the Office of the Director of National
Intelligence predicted that next year,
Russia and other American adversaries
“almost certainly will use online influence
operations to try to weaken democratic
institutions, undermine U.S. alliances and
partnerships, and shape policy outcomes.”
38. NYU recommendations for 2020
recommendations of additional steps social media companies should take to
prepare for 2020, including:
• Detect and remove deepfake videos: Realistic but fraudulent videos have
the potential to undermine political candidates and exacerbate voter
cynicism.
• Remove provably false content in general: The platforms already remove
hate speech, voter suppression, and other categories of content; the
report recommends that they add one more.
• Hire a senior content overseer: Each company needs an executive with
clout to supervise the process of guarding against disinformation.
• Improve industry-wide collaboration on disinformation: For example, when
one platform takes down abusive accounts, others should do the same
with affiliated accounts.
• Teach social media literacy in a more direct, sustained way: Users have to
take responsibility for recognizing false content, but they need more help
to do it.
40. NYU Report - Expectations
• Deep fake videos of “candidates”
• Digital Voter Suppression – major partisan goal
• Solicitation into “real world” rally’s/demonstrations
• For profit firms hired to generate disinformation
• Domestic dis-info more prevalent than foreign
• Abuse via Instagram
Images playing bigger role than simple text
• WhatsApp – encrypted messaging, hard to monitor
– WhatsApp has served as a powerful vehicle for disseminating
false information during recent presidential elections in both
Brazil and India.
41. From NYU
The March 2019 Mueller report fleshed
out our understanding of how IRA operatives,
posing as grassroots U.S. activists,
mobilized Americans to participate in
dozens of rallies in 2016 and thereafter.
By expanding on this strategy in 2020,
the Russians would accomplish one of
their main goals—translating influence
online into real-world discord
42. 2019
• Instagram has 1 billion users
– Images & video
That Instagram has outperformed Facebook as a Russian engagement machine may
‘indicate its strength as a tool in image-centric memetic warfare,’
according to a report commissioned by the Senate Intelligence Committee
• 2.38 billion for Facebook
• 2 billion for YouTube
– video
• 330 million for Twitter
• WhatsAPP has 1.5 billion users globally,
300 million of them in India and another 120 million in Brazil.
In the U.S. it has fewer than 60 million adult users.
43. From NYU – re Iran
• Some of the phony Iranian accounts impersonated real
Republican political candidates who ran for office in
2018.
• The Iranian network promoted antiSaudi, anti-Israeli,
and pro-Palestinian themes. It also expressed support
for the 2015 nuclear deal, from which President Trump
unilaterally withdrew. Some of the phony personas—
using names such as “John Turner” and “Ed Sullivan”—
had letters published on these themes in the New York
Daily News, Los Angeles Times, and other mainstream
media outlets.
44. • Misleading content created by U.S.
citizens can be difficult to distinguish
from ordinary political expression, which
enjoys First Amendment protection from
government regulation or prosecution
• Private, LLC’s protected by Citizens United are
buying placements – unlike PACs they do not
have to disclose revenue sources
(opportunities for foreign and domestic abuses)
45. Removing Fake Accounts
Facebook uses improved AI to delete automated
fake accounts by the billions—2.19
billion in just the first three months of 2019. Most
of these accounts were blocked within minutes of
their creation, preventing them from doing any
harm, according to the company. (NYU)
Twitter has challenged and taken down millions of
fake accounts for similar reasons. And YouTube
has done so as well, if to a lesser degree
46. Demoting/flagging untrue content
• Facebook is considering banning
and removing all deep-fakes
• generally, the platforms don’t remove content
deemed to be false
they typically reduce the prominence of the
untrue material, based on the notion that their
executives and employees shouldn’t be “arbiters
of the truth.”
• Facebook, for example, “down-ranks” falsehoods
in users’ News Feeds by 80%, labels the content
as untrue, and provides context on the topic. (NYU)
48. Archiving Ads for Analysis
• Facebook, Twitter and Google/YouTube now
require that ad purchasers confirm
their identities and locations within the U.S.
• All three have created searchable
political-advertising databases, which
allow anyone to look for patterns in
this kind of spending.
• The bulk of the distribution is via un-paid channels
– shares, retweets, etc.
• Fake assertions of “paid for by” is a problem
49. Account suspended on Twitter
• Goldstein, who is making her second run for
Congress .. posted a statement on Twitter that
linked Cleveland's 66 percent illiteracy rate to
Tuesday's passage of the legislation meant to
grant equal access to employment, housing, and
public accommodations to lesbian, gay, bisexual,
transgender, and queer people.
• "If CCC had LITERATE inner-city church attending
Black voters following this issue=entirely opposite
outcome," her tweet said.
50. Oxford research 2018
• During a 30-day period before the U.S. midterms, 25%
of Facebook and Twitter shares related to the election
contained “junk news”
• Disinformation sharing is down as a %,
but not in absolute numbers
• individual junk news stories can “hugely outperform
even the best, most important professionally produced
stories, drawing as much as four times the volume of
shares, likes, and comments.”
51. Information Wars: How We Lost the Global Battle
Against Disinformation and What We can Do About It
(2019 – R.Stengel*)
1. Prohibit foreign entities from buying online advertising to
influence U.S. elections
2. Make source & buyers of all political ads transparent, informing
users why they were specifically targeted
3. Provide more privacy for personal data
4. Remove verifiably & provably false information
5. Get media organizations to agree not to use stolen information
from hacks or cyber-meddling
6. Require campaigns to disclose foreign contacts & persons seeking
to influence elections
7. Appoint a senior government office to address disinformation &
election interference
* Richard Stengel was Under Secretary of State for Public Diplomacy in 2016, started
Ukraine Task Force => Russian Information Group =>Global Engagement Center
52. Congressional Discussions
• Honest Ads Act &
Deceptive Practices and Voter Intimidation Act
• Various Privacy protection initiatives
• Discussion of anti-trust actions against big
tech companies
• Discussion of first amendment implications
• Regulating social media companies
such as removing some “section 230”
protections from responsibility for content
53. What’s Next?
• Real time monitoring, by AI technology to determine
what is effective at influencing each individual’s
behavior
• AI “Nudging” … but perhaps with highly effective
results
• Obvious application: selling a product
• Next: selling a political candidate
• Then: brainwashing a sufficient amount of the
population
You can fool some of the people all of the time, and all of
the people some of the time … but can you fool a
sufficient number of the people enough of the time?
54. Some resources https://is.gd/AIPolitics
• McNamme, Roger; "Zucked: Waking Up to the Facebook Catastrophe“; (2019)
• Singer & Brooking; LikeWar: The Weaponization of Social Media (2018)
• Friedman, Tom; Thank You for Being Late, 2016
• Harari, Yuval Noah; 21 Lessons for the 21st Century; 2018
• Linked from website:
– Presentations on Cambridge Analytica
– Presentations on Facebook & Psychographics
– Jan 2019 IEEE Institute - emergence of Persuasive Technology
– US Senate Reports from Dec. 2018 on Russian interference
– NYU 2019 study and social media recommendations for 2020,
and paper analyzing US online political advertising
– Sept 2019 CSPAN/Georgetown U. panel on Social Media & First Amendment