Who Watches the Birdwatchers? Sociotechnical Vulnerabilities in Twitter’s Content Contextualisation

Abstract

At the start of 2021, Twitter launched a closed US pilot of Birdwatch, seeking to promote credible information online by giving users the opportunity to add context to misleading tweets. The pilot shows awareness of the importance of context, and the challenges, risks and vulnerabilities the system will face. But the mitigations against these vulnerabilities of Birdwatch can exacerbate wider societal vulnerabilities created by Birdwatch. This article examines how Twitter presents the Birdwatch system, outlines a taxonomy of potential sociotechnical vulnerabilities, and situates these risks within broader social issues. We highlight the importance of watching the watchers, not only in terms of those using and potentially manipulating Birdwatch, but also the way Twitter is developing the system and their wider decision-making processes that impact on public discourse.

A Description of the Taxonomy

Key

• Scale [A] Individual, [B] Type of individual, [C] Community, [D] Public discourse, [E] Systemic/principles.

• Timeframe [1] Immediate, [2] Short, [3] Mid, [4] Long, [5] Persistent.

Targets

• Pilot: vulnerabilities during the US and subsequent pilots

• Implementation: the transition to global context;

• Tweets: user interactions and vectors for attacks;

• Notes: user interactions and vectors for attacks;

• Note ranking: user interactions and vectors for attacks;

• Twitter management: internal decisions create/mitigate vulnerabilities;

• Regulation: external (legislative) decisions prevent/permit vulnerabilities.

Attacks

• A 1 Posts: tweets and notes, including many directly abusive tactics;

• A 2 Followers: tools in extended abuse;

• A 2 Bots: automating abuse to scale up coordinated attacks;

• A 1 Text-as-image: an attack on whether the Birdwatch system is triggered;

• A 2 Iterating tweets: as above;

• A 1 Varying content or categories: as above;

• A 1 Abusive or harmful content: as above;

• C 3 Data: includes data poisoning of ranking system;

• B 2 Algorithm: gives differential visibility or credibility to tweets/notes;

• C 3 Third party: external attacks exploit vulnerabilities;

• E 4 Design: internal flaws create vulnerabilities;

• C 3 Injection: includes user interactions (likely) and breached security (less so);

• C 3 Manipulation: includes user interactions and breached security;

• E 4 Data structures: design flaws enable manipulation of data/the algorithm;

• D 4 Faux transparency: data availability risks obscuring underlying structures;

• D 4 External validation: public scrutiny and PR as tool for credibility;

• D 5 Lobbying: pressure on regulators to prevent external constraints;

• D 5 Loopholes: flaws in regulation (e.g. loose definitions) enable harmful design;

• D 5 Self-regulation: Birdwatch is part of continued efforts to avoid regulation.

Harms

• A 2 Coordinated attacks: combining attacks/accounts increases scale/impact;

• B 3 Weaponisation: systematic targeting of certain groups/communities;

• A 1 Abuse: effects (emotional, physical) against specific individuals(/groups);

• B 3 Verification: shift for Twitter; harms marginalised groups with need for ID;

• E 4 Policies/enforcement: precedent of unequal application/lack of context;

• C 5 Access/exclusion: design, policies, implementation; method & type of harm;

• D 3 Game rankings: vulnerabilities in practice and in credibility;

• D 4 Ranking system: vulnerabilities to public discourse in Birdwatch design;

• D 4 Avoiding moderation: not moderation; community not platform;

• D 5 Avoiding regulation: visible action to placate regulators;

• C 3 Exploitative labour: reliance on users; lack of protection; uneven burden;

• E 5 Avoid scrutiny: systemic avoidance or deflection of external audit/criticism;

• E 5 Systemic/narrative: structural impact on society; influence over debates;

• E 4 Context shift: marginalisation of geospatial/cultural/etc. communities.