Read the original article here.

---

Algorithmic Bias in the Age of "The Dead Internet"

In the unfolding narrative of "The Dead Internet Files," where the digital realm is increasingly populated and shaped by automated systems, bots, and algorithms rather than genuine human interaction, understanding the invisible forces at play becomes critical. One of the most significant of these forces is algorithmic bias.

As bots silently replace us or heavily mediate our online experiences, the algorithms that govern their behavior, the content they generate, the information they curate, and the decisions they make carry inherent biases. These biases, often unintended and opaque, don't just reflect existing societal inequalities; they actively perpetuate, amplify, and create new forms of discrimination and skewed realities within the digital landscape. This resource explores algorithmic bias, its origins, types, impacts, and the challenges of addressing it, all viewed through the lens of an internet dominated by automated processes.

1. What is Algorithmic Bias?

At its core, algorithmic bias is the systematic and repeatable tendency of a computerized system to produce "unfair" outcomes. In the context of "The Dead Internet," this means automated processes like bots, recommender systems, search engines, and AI decision-makers exhibit prejudiced behavior, often favoring or disadvantaging specific groups or categories of users in ways that deviate from a neutral or equitable function.

Definition: Algorithmic Bias describes systematic and repeatable harmful tendencies in computerized sociotechnical systems that create "unfair" outcomes, such as privileging one category over another in ways different from the intended function of the algorithm.

Algorithms, the foundational instructions driving automated online activity, are not neutral. They are designed by humans, trained on data often generated by humans (and increasingly, by other algorithms), and deployed in complex, dynamic environments. As bots and AI entities become the dominant actors in this environment, their embedded biases become the de facto rules of engagement for the remaining human users.

Consider an online space heavily moderated or curated by biased bots. These bots might be more likely to flag content from certain demographics as inappropriate, recommend products or information only to specific groups, or prioritize interactions that align with skewed patterns found in their training data. For a human user navigating this "dead" space, their experience is silently shaped and potentially limited by these invisible algorithmic prejudices.

2. How Bias Enters Automated Systems

Bias doesn't magically appear in algorithms; it's a consequence of the design process, data, and deployment. In the context of a "Dead Internet," these entry points are amplified as automated systems scale processes previously handled by humans, often inheriting and automating human biases along the way.

• Biased Data Collection and Curation: Algorithms learn from data. If the data collected is incomplete, unrepresentative, or reflects existing societal biases, the algorithm will learn and replicate these biases. Imagine bots trained on historical online interactions where certain groups were underrepresented or stereotyped. The bots will then perpetuate this imbalance in their own interactions or content generation.
• Human Design Choices: Programmers and data scientists make decisions about what data to include, how to categorize it, and what metrics to optimize for. These choices, consciously or unconsciously influenced by personal or institutional biases, get encoded into the algorithm's logic.
• Prioritization and Weighting: Algorithms assign importance (weights or priorities) to different pieces of data or rules. Human designers decide which features are more important. If a feature correlated with a protected characteristic (like zip code potentially correlating with race or socioeconomic status) is given high weight based on biased historical outcomes, the algorithm will discriminate based on that correlation.
• Self-Generated Data and Adaptation: Some sophisticated algorithms and bots collect their own data based on interactions. If initial interactions are influenced by human biases or are with other biased bots, the algorithm's subsequent learning can reinforce and amplify these biases in a feedback loop.
• Technical Limitations: Design constraints, computational power limits, or even the seemingly random mechanisms within a system can inadvertently introduce bias (e.g., favoring items appearing first in a list due to non-truly random sorting).

3. Historical Context: The Roots of Algorithmic Bias

The idea that automated systems can embody bias isn't entirely new, though its scale and impact have grown dramatically in the "Dead Internet" era.

• Early Critiques (Joseph Weizenbaum, 1976): AI pioneer Joseph Weizenbaum warned that computer programs, being sequences of human-defined rules, "embody law" and reflect the programmer's assumptions, including biases. Data fed into machines also reflects human decision-making biases. He cautioned against blindly trusting computer decisions without understanding their process, likening it to navigating solely by coin toss – you might arrive, but you don't understand why and the process isn't reliable or rational. This early critique is prescient in a "Dead Internet" where opaque automated processes govern vast spaces.
• Early Automated Discrimination (St. George's Hospital, 1980s): A computer system designed to assess medical school applicants inadvertently encoded historical biases by denying entry to women and applicants with "foreign-sounding names" based on past admissions data. This demonstrated how automating biased human decisions could scale discrimination and make it appear more objective or authoritative.
• Amplification via Machine Learning: As algorithms moved beyond simple rule-following to machine learning on vast datasets, the biases embedded within real-world data became a primary source of algorithmic bias. The scale of data in the modern internet means even subtle biases in the training data can lead to significant discriminatory outcomes when deployed widely by automated systems. The 2018 study by Joy Buolamwini and Timnit Gebru, showing commercial facial recognition errors up to 35% for darker-skinned women compared to <1% for lighter-skinned men, starkly illustrated how data bias manifests in deployed AI.
• "Weapons of Math Destruction" (Cathy O'Neil, 2016): O'Neil highlighted how opaque, automated decision-making processes in areas like credit, policing, and education can entrench and scale discrimination under the guise of objective, scientific processes. Her work underscored how algorithms, even without explicit malicious intent, can create systems of inequality.

4. Types of Algorithmic Bias

Algorithmic bias manifests in various forms depending on its source and how the algorithm interacts with data and its environment.

4.1. Pre-existing Bias

Definition: Pre-existing Bias in algorithms is a consequence of underlying social, cultural, and institutional ideologies and prejudices reflected in the data or the design process.

This type of bias directly translates historical or societal inequalities into code and data. If automated systems learn from a biased past, they will replicate it.

• Example: British Nationality Act Program (BNAP): This program automated citizenship evaluation but encoded the discriminatory legal logic of the time (e.g., differing treatment of children based on parental marital status and gender). Even if the law changed, the algorithm, without modification, would continue to perpetuate the outdated bias. In a "Dead Internet," bots inheriting such logic would apply these historical prejudices in automated interactions or content assessment.
• Example: "Label Choice Bias": Bias introduced by using a proxy metric for a desired outcome, especially when the proxy is itself biased. For instance, using healthcare costs to predict healthcare needs can disadvantage marginalized groups who have lower costs due to systemic barriers in accessing healthcare, even if they are sicker. An automated system using this algorithm would systematically under-allocate resources to these groups based on biased historical cost data.

4.2. Machine Learning Bias

This category focuses specifically on biases that arise within algorithms that learn from data, like the AI models powering many bots and automated systems in the "Dead Internet."

• Language Bias:

  Definition: Language Bias is a statistical sampling bias in language models, often tied to the dominance of certain languages or dialects in training data, leading to skewed representations of topics and perspectives. Large Language Models (LLMs), increasingly used to generate text and interact like bots, are predominantly trained on English data. This can lead to presenting Anglo-American viewpoints as universal truth while marginalizing non-English perspectives. LLMs might also exhibit bias against specific dialects (e.g., African American English), flagging legitimate content as problematic based on linguistic patterns associated with bias in the training data.

• Selection Bias:

  Definition: Selection Bias is the tendency of a model, particularly in multiple-choice settings, to favor certain answer positions or tokens ("A," "B") irrespective of the actual content, often rooted in token probability biases. This is a technical bias within the model's architecture or training data that makes it more likely to pick options based on their presentation rather than their semantic correctness, undermining the reliability of automated evaluation systems.

• Gender Bias: Models trained on data reflecting societal gender stereotypes will replicate them, associating professions (nurse/secretary vs. engineer/CEO) or characteristics with specific genders. An example is models trained on Icelandic data favoring masculine grammatical gender even for female-dominated roles, amplifying societal biases. Bots generating text or making inferences based on these models will perpetuate these stereotypes.
• Stereotyping: Beyond gender, models can reinforce stereotypes based on age, nationality, religion, occupation, etc., leading to harmful generalizations and caricatures in automated outputs.
• Political Bias: Models can exhibit political leanings based on the prevalence of certain viewpoints in training data, potentially generating responses favoring specific ideologies. In a "Dead Internet," bots with political bias could subtly or overtly push particular narratives or suppress others.
• Racial Bias:

  Definition: Racial Bias in machine learning results in outcomes that unfairly discriminate against or stereotype individuals based on race or ethnicity, typically stemming from biased training data. Examples include facial recognition misidentifying individuals from certain racial backgrounds, hiring tools excluding candidates based on race (or proxies), healthcare algorithms underestimating needs of minority patients (due to cost data bias), and mortgage algorithms perpetuating discrimination based on "creditworthiness" metrics tied to historical inequalities. LLMs can also perpetuate racial stereotypes, including covert dialect prejudice (e.g., against African American English speakers).

4.3. Technical Bias

Definition: Technical Bias emerges from the inherent limitations, design constraints, computational power, or technical choices made in creating an algorithm or system.

These biases are rooted in the technical implementation rather than solely in the data or societal context.

• Display Bias: Limiting search results to a certain number per screen (e.g., top 3) technically privileges the top results. Automated systems presenting information in this way introduce bias based on ranking criteria, even if those criteria are technically "neutral."
• Non-Randomness: Algorithms relying on supposedly random number generation might introduce bias if the mechanism isn't truly random, skewing selection towards items at the beginning or end of lists.
• Decontextualized or Misaligned Data:
  • Using irrelevant information for sorting (e.g., alphabetical order for flight results) biases against certain outcomes.
  • Evaluating data collected in one context using algorithms or human reviewers in a vastly different context without necessary external information can lead to misinterpretations (e.g., misidentifying bystanders in surveillance footage evaluated remotely).
  • Formalizing complex human behaviors (like courtroom plea decisions influenced by emotion) into concrete algorithmic steps ignores crucial context and introduces bias.
• Example: Turnitin: This plagiarism detection software is technically biased against non-native English speakers because it's easier for native speakers to alter text slightly to evade detection by breaking up matching strings, a technical limitation of the software's design.

4.4. Emergent Bias

Definition: Emergent Bias arises from the use and reliance on algorithms in new, unanticipated, or evolving contexts, or from the dynamic interaction between the algorithm's output and the real world that feeds data back into the system.

This type of bias is less about the initial design or data and more about how the algorithm behaves and influences outcomes over time in a complex environment like the "Dead Internet."

• Mismatched Contexts: Algorithms designed for one context might fail or become biased when applied elsewhere (e.g., a medical residency matching algorithm designed when few couples applied together becoming biased against partners as more women entered medicine). As automated systems are deployed in diverse and evolving online spaces, their failure to adapt to new norms or knowledge creates bias.
• Unanticipated Users: Algorithms designed for a specific type of user might exclude others (e.g., requiring literacy for an interface). Bias also emerges if end-users rely blindly on the algorithm, even when it's outdated or its original design assumptions about user expertise are incorrect (e.g., immigration officers relying solely on an algorithm that became outdated).
• Correlations Mistaken for Causation: Algorithms finding unexpected correlations in large datasets can lead to discriminatory outcomes without understanding the underlying reasons (e.g., linking browsing patterns to sensitive characteristics or a medical triage program prioritizing sicker asthmatics lower because historical data shows they received immediate, high-level care, leading to higher survival rates within that specific group who received prompt care, a nuance the algorithm missed). In a "Dead Internet," automated systems might make critical decisions based on spurious or decontextualized correlations.
• Feedback Loops:

  Definition: A Feedback Loop (or recursion) in algorithmic bias occurs when the algorithm's output influences real-world data, which is then fed back into the algorithm, reinforcing the initial biased outcome.

  • Predictive Policing (PredPol, COMPAS): An algorithm predicts crime hotspots based on reported crime data. If police presence increases in predicted hotspots, more minor crimes might be observed or reported there, even if actual crime isn't higher. This data is fed back, reinforcing the prediction and increasing police presence further, creating a feedback loop that can amplify existing biases in reporting or policing towards certain neighborhoods (often minority communities). COMPAS, used for recidivism prediction, has been criticized for over-predicting risk for Black defendants; if subsequent arrests (influenced by policing patterns) are fed back, the bias is reinforced.
  • Recommender Systems: Bots recommending content (videos, news) based on user clicks create filter bubbles. The algorithm suggests content, the user clicks (influenced by suggestion), and this action influences the next suggestions. This feedback loop can lead humans to be exposed only to a narrow, potentially biased slice of information, limiting exposure to diverse viewpoints. This is a prime example of automated systems isolating and shaping human users' reality in the "Dead Internet."

5. Impacts of Algorithmic Bias in the "Dead Internet"

The consequences of algorithmic bias are far-reaching, silently shaping our digital experiences and impacting real-world opportunities and interactions, especially when mediated by ubiquitous automated systems.

5.1. Commercial Influences

Automated systems can be biased to favor entities with commercial arrangements. Early flight booking systems favored the airline that created them. Google's founders warned against advertising influencing search results because it would be an "invisible manipulation." In a "Dead Internet" where bots mediate transactions and present information, commercial biases can distort markets, limit consumer choice, and obscure genuine value in favor of incentivized outcomes, all without the human user's awareness.

5.2. Voting Behavior

Studies show that search engine results can significantly shift undecided voters' opinions. Social media algorithms recommending voting information influenced voter turnout. If automated systems (search bots, social media feeds controlled by algorithms) can subtly manipulate information exposure, they can engage in "digital gerrymandering," selectively presenting information to serve an agenda rather than the user's need for balanced information, potentially influencing democratic outcomes in a non-transparent manner.

5.3. Discrimination Amplified by Automation

Algorithmic bias translates existing societal prejudices into code, potentially amplifying discrimination across various dimensions.

• Gender: Search suggestions biased towards male names (LinkedIn), inferring pregnancy for targeted marketing (Target), prioritizing sexualized content for female-related searches (Google), displaying higher-paying jobs to men (job search sites), machine translation defaulting to masculine forms, biased hiring tools (Amazon), and biased music recommendations (Spotify) all demonstrate how automated systems perpetuate gender inequality in digital interactions and opportunities.
• Racial and Ethnic: Algorithmic systems have shown bias in image recognition (Google Photos identifying Black individuals as gorillas, Nikon cameras asking Asians if they were blinking), reflecting biased training data in biometrics. Search results linking common Black names to arrest records, healthcare algorithms underestimating needs of Black patients (due to cost proxy bias), biased mortgage algorithms, and LLMs exhibiting covert racial biases (like dialect prejudice) highlight how automated systems replicate and reinforce systemic racism. COMPAS in law enforcement is a stark example where algorithmic bias can lead to disproportionately harsher outcomes for minority groups.
• Law Enforcement and Legal Proceedings: Risk assessment algorithms (COMPAS) used in sentencing and parole have been shown to be biased against Black defendants. These systems, presented as objective tools, automate historical biases present in crime and sentencing data, potentially leading to unfair legal outcomes. Automated surveillance systems also exhibit racial and gender biases in identification accuracy, meaning automated monitoring disproportionately impacts certain populations.
• Online Hate Speech: Algorithms designed to detect hate speech can be biased, favoring certain groups over others (e.g., Facebook algorithm protecting broad categories like "Muslims" but not subsets like "black children"), or incorrectly flagging content from marginalized groups (e.g., flagging Black users' posts or African American English as hate speech). This means automated content moderation fails to create equitable online spaces for human users.
• Surveillance: Automated surveillance inherently involves algorithmic decisions about "normal" versus "abnormal" behavior and who "belongs." Bias in training data leads to facial recognition systems being less accurate for darker-skinned individuals, meaning automated surveillance is less reliable but disproportionately applied based on biased identification capabilities.
• LGBTQ+: Automated systems have shown bias by linking LGBTQ+ identity to inappropriate categories (Grindr linked to sex offender apps), censoring LGBTQ+ content (Amazon de-listing books), exhibiting gender bias in image searches ("photos of my female friends" suggestions), and failing to accurately recognize transgender individuals (Uber facial recognition difficulties). The ability of AI to infer sexual orientation from images raises significant privacy and safety concerns in an automated world.
• Disability: People with disabilities are often marginalized in AI system design and data, leading to exclusion. The diverse, complex, and sometimes temporary nature of disabilities, coupled with a lack of explicit disability data (partly due to privacy concerns and stigma), means automated systems often fail to serve or actively exclude this population (e.g., voice recognition failing for speech impairments). Algorithmic bias translates existing societal barriers into digital ones.

5.4. Skewed Search Results (Google Search)

Even fundamental interactions like searching are affected. Despite claims of neutrality, search algorithms have historically returned sexist and racist autocompletion suggestions and results, including pornographic content for non-explicit queries related to minority groups. While companies make adjustments, the inherent biases in data and algorithms mean automated search continues to shape perceived reality in potentially harmful ways.

6. Obstacles to Research and Understanding

Studying and addressing algorithmic bias in the vast, complex landscape of the "Dead Internet" faces significant challenges.

• Defining Fairness: There's no single, universally agreed-upon definition of "fairness" in an algorithmic context. Different mathematical definitions (e.g., equality of outcomes vs. equality of treatment) can be mutually exclusive. Deciding what "fair" means for a specific application requires difficult ethical and societal choices, not just technical ones.
• Complexity ("Blackboxing"): Modern algorithms, especially deep learning models and large online systems, are incredibly complex. Even their creators may not fully understand every interaction or permutation of input/output. This "blackboxing" makes it hard to trace why a biased outcome occurred. The "Dead Internet" is a network of interconnected black boxes, making systematic analysis daunting.
• Lack of Transparency (Proprietary Nature): Many powerful algorithms, particularly those used by large tech companies (search engines, social media, etc.), are proprietary trade secrets. Their internal workings are not publicly accessible, preventing external researchers from auditing them for bias. This secrecy is often justified commercially (preventing manipulation) but also conveniently hides potential unethical or biased practices.
• Lack of Data on Sensitive Categories: Data controllers are often hesitant or legally restricted (e.g., GDPR special categories) from collecting explicit data on sensitive attributes like race, gender, sexual orientation, or disability. Without this data, it's difficult to even measure if bias exists for these groups, let alone mitigate it, although methods using privacy-enhancing technologies are being explored.

7. Approaches to Mitigating Algorithmic Bias

Addressing algorithmic bias requires multi-faceted approaches, blending technical, policy, and social interventions.

• Technical Solutions: Developing tools and methodologies to detect, measure, and mitigate bias. This includes techniques applied to training data (cleaning, balancing), algorithm design (building "fairness constraints"), and output analysis (confusion matrices, AI audits). Research explores methods to train models while being "agnostic" to sensitive features.
• Transparency and Monitoring ("Explainable AI"): Increasing the interpretability of algorithms ("Explainable AI" - XAI) so humans can understand how a decision was reached, rather than just trusting the output. Monitoring algorithmic outcomes in deployment is also crucial. While open-sourcing code is a step, true transparency requires understandable explanations and a critical audience willing and able to scrutinize.
• Right to Remedy and Accountability: Establishing legal and regulatory frameworks (like GDPR Article 22 or NYC's bias audit law) that grant individuals the right to challenge automated decisions affecting them and ensuring accountability for harmful algorithmic outcomes. This requires defining responsibility within complex systems.
• Diversity and Inclusion in Design: Actively increasing the representation of diverse individuals (women, minorities, LGBTQ+, people with disabilities) in the teams that design, develop, and deploy AI systems. A more diverse workforce can bring different perspectives, identify potential biases earlier, and build systems that are more inclusive and equitable.
• Interdisciplinarity and Collaboration: Bringing together experts from computer science, sociology, ethics, law, domain expertise (e.g., healthcare professionals for medical AI), and the communities affected by the technology. A human-centered approach that involves stakeholders throughout the design process is crucial for identifying potential harms and ensuring AI serves societal well-being, not just technical optimization or commercial goals. Frameworks like PACT encourage collaboration and power-shifting in AI for Social Good projects.

8. Regulation

Governments are beginning to grapple with regulating algorithmic bias, recognizing its potential for societal harm.

• Europe (GDPR, AI Act): GDPR's Article 22 provides individuals rights regarding solely automated decisions with significant effects, including a limited right to explanation and human intervention. Recital 71 explicitly mentions using appropriate procedures to prevent discriminatory effects based on protected characteristics. The proposed AI Act aims to regulate AI systems based on their risk level, with high-risk systems facing stricter requirements regarding data quality, transparency, human oversight, and bias mitigation.
• United States: Regulation is piecemeal across sectors and laws. Efforts include guidance (Obama administration's AI plan), local laws (NYC's algorithmic accountability bill and bias audit law for hiring), and recent executive orders emphasizing safe, secure, and trustworthy AI development, including mitigating discrimination. The US approach often relies on self-enforcement or existing laws like anti-discrimination statutes being applied to algorithmic outcomes.
• India: The Personal Data Bill draft addresses "harm" from data processing, defining discriminatory treatment resulting from evaluative decisions as a source of harm, suggesting a potential avenue for addressing algorithmic bias within data protection law.

Conclusion

Algorithmic bias is not merely a technical glitch; it is a critical reflection and amplification of societal inequalities embedded within the automated systems increasingly populating and controlling the digital landscape. In the context of "The Dead Internet Files," understanding this bias is paramount. As bots silently replace human presence and mediate our remaining digital interactions, the biases baked into their algorithms dictate what information we see, what opportunities we are presented with, how we are categorized, and whether we are treated fairly. Addressing algorithmic bias requires ongoing vigilance, interdisciplinary collaboration, technical innovation, and strong regulatory frameworks to ensure that as the internet evolves, it does not simply become a large-scale machine for perpetuating the prejudices of the past, but strives towards a more equitable automated future.