Democracy By Design
1
Democracy
By Design
Published September 2023
A Content-Agnostic Election
Integrity Framework for
Online Platforms
Democracy By Design
2
Overview
With over 50 countries set to hold elections amidst tectonic geopolitical and
technological shifts, 2024 will present unprecedented challenges for democracy and the
information gatekeepers who support it.
Those challenges are compounded by increasing volatility of the content policy
landscape, as partisan debates over where platforms should draw the lines grow more
bitter, and repressive regimes engage in draconian censorship under the guise of
countering disinformation.
But there is a clear path forward to avoid these pitfalls, build consensus, and protect both
freedom of expression and free and fair elections – a path to build Democracy By Design.
Below, we outline a set of high-impact, content-agnostic election integrity
recommendations for online platforms – readily actionable interventions rooted in their
own product design and policy toolkits, and in many cases, empirically backed by their
own research.
The Democracy By Design framework includes three major planks, under which there
are specic recommendations and background to support them:
• Bolstering Resilience, which focuses on ‘soft interventions’ that introduce
targeted friction and context to mitigate harm;
• Countering Election Manipulation, which outlines bulwarks against evolving
threats posed by malign actors and automated systems; and
• Paper Trails, which highlights key transparency measures needed to assess
systemic threats, evaluate the ecacy of interventions, and foster trust.
The result is a framework meant to avoid political landmines and broadly resonate
across nations with distinct laws and cultures, and platforms with incongruous
architecture and resources – a consensus roadmap to enhance systemic resilience
against election threats.
Democracy By Design
3
1. Employ well-designed interstitials when users engage with things like old articles
or state-controlled media. Election authorities provide voters with ocial information
about casting their ballots; polling places have check-in stations and safeguards against foul
play; voting machines prompt voters to double-check their ballots before submitting them.
Platforms should similarly work to protect and inform voters by injecting targeted friction and
context into their systems via well-designed interstitials (click-through screens) and labeling.
Examples of such tools – some of which have already been implemented across various
platforms – include:
• Pop-ups asking users if they want to read an article before sharing, or alerting users if
articles are old
• Clearly labeling accounts of government ocials and using interstitials to alert users who
try to read or share content from state-run media
• Placing posts that contain misleading or unveried election information behind click-
through warning labels that include clear context and facts
• Appending distinct verication badges to the ocial accounts of local, state, and national
election authorities
Bolstering System Resilience | Targeted Friction
and Context to Mitigate Threats
Election integrity vulnerabilities on social platforms often stem from their own
architecture: The features designed to make platforms frictionless and maximize
engagement – from recommendation algorithms to reshare buttons – can serve to warp
discourse and undermine democracy.
Extensive research from experts and tech companies themselves indicates that content-
agnostic soft interventions that introduce targeted friction and context – bolstering the
resilience of their systems and better informing users – can signicantly mitigate threats.
Twitter state-affiliated media label Twitter prompt to read before RT Facebook interstitial when sharing old article
Democracy By Design
4
2. Utilize virality circuit breakers to automatically ag fast-spreading posts and
trigger a brief halt on algorithmic amplication. Not all high-reach content comes
from users with huge followings – especially now, as platforms increasingly rely almost
entirely on algorithms in determining what to surface with little friction. Inspired by
automatic triggers used in nancial markets to prevent panic selling at moments of high
volatility, experts have proposed that platforms introduce circuit breakers to automatically
ag posts that are beginning to gain virality, temporarily pause algorithmic boosting, and
present users an interstitial about the post’s virality and status upon click-through or
reshare. It’s a proposal that’s been lent credence by tech companies themselves, with Meta
reportedly testing the concept and Snap going a step further, as an executive testied that
all posts on their Spotlight platform were speed-bumped and checked before reaching 25
unique viewers. Proper thresholds and processes will of course vary by platform and content
type, so companies should outline their policies and publish pertinent aggregate data in
transparency reports.
3. Restrict rampant resharing during election season by removing simple share
buttons on posts after multiple levels of sharing. Frictionless resharing is a staple
of social platforms – and a key driver of toxicity. Internal Meta research showed users are
4x more likely to encounter falsehoods in a reshare of a reshare than in the News Feed in
general, and concluded aggressively limiting these ‘deep reshares’ would be “an eective,
content-agnostic approach to mitigate the harms.” Meta also placed limits that proved
eective on how many times WhatsApp messages could be forwarded after mass-sharing
exacerbated unrest in India and Brazil. Platforms should remove share buttons on posts
after multiple levels of reshare, and/or test other mechanisms that enhance reshare friction
in a targeted manner during election season, with careful consideration of the impact on
legitimate advocacy campaigns.
4. Implement clear strike systems to deter repeat oenses, curtail the outsized
impact of malign actors, and better inform users. This is not meant to dictate
specic content policies; strike systems should be based on violations of platforms’ own
standards. Most platforms already utilize some form of strike system to levy sanctions
on repeat oenders in recognition of the disproportionate harm they drive, but both the
policies and their application are typically vague and/or obscured from users – often
under the argument that it’s impossible to clarify such rules without helping bad actors
game the system. And even approaches that have more explicitly addressed this threat –
like Twitch’s Harmful Misinformation Actor policy, and Twitter’s bygone 5-strike Civic
Integrity policy – have focused chiey on when to suspend the worst actors. Platforms
should develop and implement transparent strike systems that clearly outline escalating
‘soft interventions’ to limit the impact of repeat oenders, such as restricting resharing,
curtailing algorithmic amplication, and placing posts behind click-through warning
labels with context. This approach steers clear of the false choice between censorship
and inaction, and would demystify enforcement decisions, deter habitual rule-breaking,
and defang the malign actors who pose the greatest election integrity threats. A detailed
example of what such a strike system might look like for a given platform, developed by
Accountable Tech, can be found here.
Democracy By Design
5
1. Prohibit – in policy and practice – the use of generative AI or manipulated
media to:
a. Falsely depict election irregularities. In recent elections, we’ve seen users
intentionally and unintentionally share images and videos that have been taken out of
context in ways that appear to show election irregularities (i.e. burning ballots), undermining
faith in the legitimacy of the democratic process. Using tools like reverse-image search in
previous election cycles, researchers were often able to trace the provenance of content
relatively quickly and neutralize false narrative – but if new AI-generated hoaxes of this
nature ood the zone, without the same capacity to identify them, the harms to democracy
could be severe. Platforms should establish clear policies prohibiting the false depiction of
election irregularities and mechanisms for meaningful enforcement and deterrence.
b. Fraudulently misrepresent the speech or actions of public gures in video,
audio, or images. The deepfake threat has loomed for years, but generative AI has
drastically lowered the barrier to entry for anyone – be it political opponents or malign
actors – to convincingly put words into public gures’ mouths and fool voters or hijack
the discourse. While many platforms already have baseline policies addressing synthetic
or manipulated media, they are often overly vague. Platforms should expound upon
them, clearly detailing – including example cases – the scope of these policies, and
should clarify that any AI-generated or -manipulated content likely to harmfully deceive
voters about public ocials’ speech or actions is strictly prohibited. (Such a policy might
narrowly exempt, for example, clearly labeled parody or satire.)
c. Create personalized political or issue ads that microtarget voters with distinct
content generated by using their personal data. Even before the recent explosion in
generative AI, concerns have been ballooning over the manipulative nature of hyper-targeted
personalized advertising – particularly in the political sphere. The EU is currently debating
strict new limitations on how political ads can be targeted. Google has similarly restricted
political ad targeting categories to age, gender, and general location. The threats are even more
harrowing when considering the potential for generative AI ad tools to be deployed to serve
bespoke ads to exploit individual voters based on their behavior, identity, or even inferences
about their current mood. With the dominant players all already touting new generative AI ad
products, it’s critical that platforms strictly prohibit the use of such tools to create personalized
Countering Election Manipulation | Safeguards
Against Malign Actors and Automated Systems
Malign actors have weaponized social platforms to meddle in elections, attack democracy,
and erode our shared reality. Now their capacity for manipulation has been turbocharged
by new technology, including powerful algorithms and generative AI tools tailormade for
high-impact, low-cost inuence operations.
Platforms must do everything in their power to thwart unlawful eorts to interfere with
elections or individuals’ free exercise of their right to vote – including eorts to intimidate
voters or mislead them on how to participate – and to counter manipulation more broadly.
Democracy By Design
6
political or issue ads that microtarget voters with distinct content generated by using their
personal data. Additionally, they should maintain robust ad transparency libraries that shed
light on ad targeting and variations, including cataloging any ads removed for violations.
2. Implement strong provenance standards and require clear disclosures within any
political or issue ad that features AI-generated images, video, or audio. Given the
breakneck speed at which generative AI and synthetic media tools are being incorporated into
every phase of content creation, all participants in the online information ecosystem have an
urgent responsibility to empower users by providing as much transparency as possible about the
provenance and authenticity of content, disclosing when and how video, images, or audio has
been generated or manipulated. Leading technology and media companies are already working to
establish global, interoperable protocols to do just that, but the ecacy of such eorts will depend
on widespread adoption; platforms should commit to working collaboratively and implementing
strong provenance standards across the digital sphere, and require clear disclosures to be
included within any political ad utilizing AI-generated or -manipulated media.
3. Prompt users as election season
1
starts, outlining the main parameters of core
algorithmic systems and giving users autonomy to adjust ranking criteria so their
recommendations are not based on proling or optimized for engagement. One of
the greatest threat vectors for large-scale manipulation and harm, as shown by platforms’ own
internal research, is the black box recommendation and curation systems that have increasingly
usurped users’ autonomy in determining what content they see. This threat has come into sharp
focus as nations grapple with fears that the Chinese Communist Party could eectively weaponize
the powerful TikTok algorithm to reshape public discourse and undermine democracy in critical
moments without detection. Echoing bipartisan US legislation and proposals from European
leaders, platforms should prompt users as election season starts, outlining the main parameters
of core algorithmic systems and giving users autonomy to adjust ranking criteria so their
recommendations are not based on proling or optimized for engagement, along with easily turning
o non-essential discovery tools. For example, users might opt to revert to a reverse-chronological
feed of accounts they follow, or to prioritize content from authoritative news sources or diverse
viewpoints. All user choices should be respected for the duration of election season, at a minimum.
1
For the purposes of this framework, “election season” refers to the period beginning 60 days before Election Day and
ending after the peaceful transfer of power, and should be considered a baseline, as elevated risk may extend further.
Paper Trails | Fostering Trust through Meaningful
Transparency
Just as machine voting systems are backed by paper trails to ensure our elections are
trustworthy and secure, platforms that shape the information ecosystem must nally
open up the black box and start showing their work. The status quo has yielded distrust
from all sides, with partisans suspecting foul play, neutral observers unable to make
informed evaluations, and non-English speakers left behind.
Platforms must begin disclosing meaningful data to voters, independent researchers, and
election ocials regularly in order to engender trust and substantiate integrity measures.
Democracy By Design
7
1. Clearly and accessibly detail in one place all election-related (or applicable
platform-wide) policies and approaches. Platforms have developed complex and
constantly evolving approaches to countering election-related threats. But it’s hard to keep
track of all the policies, which can be rolled out via disparate press releases or posts from
spokespeople. And enforcement can be uneven and opaque, sowing distrust and confusion
among aected users. Platforms should ensure that all pertinent policies are distilled into
one easily accessible and intelligible election policy hub. And users should be notied if they
are subject to any enforcement action in clear language explaining the rule that is being
applied and its implications, and given the opportunity to appeal.
2. Release regular transparency reports during election season – broken down
across widely spoken languages – detailing high-performing and violative
content, enforcement, and resource allocation. While many platforms now release
transparency reports, they tend to be quarterly or biannual English-language-centric
products with company-selected categories of data and metrics – not standardized, frequent,
or granular enough to illuminate the state of the information ecosystem around elections, nor
the ecacy of interventions to mitigate them. Platforms should commit to releasing reports
as frequently as possible during election season that include the following information:
• A snapshot of the week’s highest-performing election-related content – the posts,
advertisements, accounts, URLs, and groups that generated the most engagement
• Aggregate data about enforcement of all election-related policies, including specics
about the measures detailed in previous sections and their eectiveness
• Details about the number of full-time sta and contractors working on election integrity,
broken down by department role and primary languages spoken
3. Provide independent researchers with direct access to platform data to inform
studies, threat analysis, and systemic impact assessments. The widespread distrust
of tech companies’ impact on our elections, our discourse, and our democracy will persist so
long as platforms continue to be black boxes, with the public kept in the dark on everything
from the basic inputs of their recommendation algorithms to the enforcement of their
content policies. While bipartisan legislation in Congress to remedy this has not yet passed,
the EU’s Digital Services Act is set to nally give researchers in Europe access to some of
that data so they can study the systemic impacts of the products, services, and features at
the center of our information ecosystem. Companies must grant direct and meaningful
access to platform data – in a responsible and privacy-preserving manner – to independent
researchers in the US and globally such that they can analyze and help deter election threats
in real-time and summarize their ndings for the broader public.
Democracy By Design
8
BOLSTERING
SYSTEM
RESILIENCE
Targeted Friction
& Context to
Mitigate Threats
COUNTERING
ELECTION
MANIPULATION
Safeguards Against
Malign Actors &
Automated Systems
PAPER
TRAILS
Fostering Trust
Through Meaningful
Transparency
Democracy By Design | Summary
Election integrity vulnerabilities on social platforms often stem from their own architecture:
The features designed to make platforms frictionless and maximize engagement – from
recommendation algorithms to reshare buttons – can also serve to warp discourse and
undermine democracy.
Extensive research from experts and tech companies themselves indicates that content-agnostic
soft interventions that introduce targeted friction and context – bolstering the resilience of their
systems and better informing users – can signicantly mitigate threats. Platforms should:
• Use well-designed interstitials when users engage with things like old articles or
state-controlled media
• Utilize virality circuit breakers to automatically ag fast-spreading posts and
trigger a brief halt on algorithmic amplication
• Restrict rampant resharing during election season by removing simple share
buttons on posts after multiple levels of sharing
• Implement clear strike systems to deter repeat oenses, curtail the outsized
impact of malign actors, and better inform users
Malign actors have weaponized social platforms to meddle in elections, attack democracy, and
erode our shared reality. Now their capacity for manipulation has been turbocharged by new
technology, including powerful algorithms and generative AI tools tailormade for high-impact,
low-cost inuence operations.
Platforms must do everything in their power to thwart unlawful eorts to interfere with elections
or individuals’ free exercise of their right to vote – including eorts to intimidate voters or mislead
them on how to participate – and to counter manipulation more broadly. Platforms should:
• Prohibit, in policy and practice, the use of generative AI or manipulated
media to:
• Falsely depict election irregularities
• Fraudulently misrepresent the speech or actions of public gures in video, audio,
or images
• Create personalized political or issue ads that microtarget voters with distinct
content generated by using their personal data
• Implement strong provenance standards and require clear disclosures within
any political or issue ad that features AI-generated images, video, or audio
• Clearly present users with the main parameters of algorithmic systems as
election season starts, and prompt them to opt in
Just as machine voting systems are backed by paper trails to ensure our elections are trustworthy and
secure, platforms that shape the information ecosystem must nally open up the black box and start
showing their work. The status quo has yielded distrust from all sides, with partisans suspecting foul
play, neutral observers unable to make informed evaluations, and non-English speakers left behind.
Platforms must begin disclosing meaningful data to voters, independent researchers, and election
ocials regularly in order to engender trust and substantiate integrity measures. Platforms should:
• Clearly and accessibly detail in one place all election-related (or applicable
platform-wide) policies and approaches
• Release regular transparency reports during election season – broken down
across widely spoken languages – detailing high-performing and violative content,
enforcement, and resource allocation
• Provide independent researchers with direct access to platform data to
inform studies, threat analysis, and systemic impact assessments
