07 Oct 2025 12 min read Algorithmic Accountability

The Consent of the Connected: Auditing How Networks Manufacture Belief

I. The Moment We Lost the Map

On a cold spring morning in 2025 a journalist in London opens her social‑media feed. Within seconds she is guided through a carefully engineered tour of the world—a stream of posts, headlines and short‑videos. What she sees, and what she doesn’t see, are not the results of deliberate curiosity. They are the output of recommender algorithms whose logic remains largely invisible to the public. The digital environment she inhabits filters more content in a day than any editor ever could, deciding which topics, voices, and political angles deserve her attention. Belief is no longer built—it is brokered by networks.

This observation is not a polemic: it is the product of recent policy and research debates. Studies by organisations such as the Oxford Internet Institute show how social connections and recommender designs influence whether users engage with fact‑checking posts[1][2], and by extension how easily misinformation travels. Regulators have responded. The European Union’s Digital Services Act (DSA) creates a transparency regime that allows vetted researchers unprecedented access to the internal data of very large platforms[3]. The UK’s Ofcom, preparing to enforce the Online Safety Act, has warned that badly designed recommender systems can disseminate illegal or harmful material[4].

If the twentieth century was defined by mass propaganda, the twenty‑first is defined by algorithmic persuasion. The central question for democracies is whether citizens can meaningfully consent to the beliefs they adopt when those beliefs are curated by opaque recommendation engines. To restore autonomy, we propose a new layer of infrastructure for epistemic integrity: an Information Audit Kit that makes informational ecosystems measurable, auditable and accountable. This article sets out why such a kit is needed, how it would work, and how it can align with existing policy frameworks like the DSA, the UK’s AI assurance roadmap and the NIST AI Risk Management Framework.

II. The New Supply Chain of Belief

Platforms such as YouTube, TikTok and Instagram operate what we might call a belief supply chain. Data about our behaviour—likes, shares, watch time—is collected, fed into machine‑learning models, and used to generate personalised feeds that maximise engagement. Each link in this chain shapes the information that users encounter. The DSA acknowledges this power imbalance. Its data‑access regime allows researchers to study how design choices in recommender systems may push users toward polarising content[5]. The Act’s implementation emphasises the systemic risks posed by platforms, including threats to democracy and fundamental rights[6].

Outside Europe, researchers have relied on adversarial audits to understand how these systems work. The Ada Lovelace Institute’s survey of audit methods identifies six technical approaches: code audits, user surveys, scraping audits, API audits, sock‑puppet audits and crowd‑sourced audits[7]. Each method reveals different facets of algorithmic behaviour: code audits allow inspection of the underlying objectives; user surveys capture lived experience; scraping and API audits collect data on what platforms show users; sock‑puppet audits simulate users to record recommendations; and crowd‑sourced audits enlist real users to donate data[7]. These techniques show how engagement‑optimising algorithms can reinforce biases and create filter bubbles that isolate users from diverse viewpoints.

The problem is not that these systems exist—personalisation can reduce information overload and connect users with relevant content—but that their operation and impact remain largely invisible. With limited access to platform data, regulators and researchers cannot easily verify whether content moderation policies are being followed or whether harmful content is being systematically amplified. As a result, trust in digital media erodes. Full Fact, a UK fact‑checking charity, emphasises that a better information environment is essential to rebuild public trust[8][9]. To move forward, we need to quantify how healthy our information diets are and how recommender systems influence them.

III. The Anatomy of Epistemic Health

Epistemic health refers to the quality of the information diets that shape what we know and believe. Like physical health, it has multiple dimensions: diversity, balance and credibility. Making these dimensions measurable is the foundation of any audit regime.

1. Diversity: exposure entropy and evenness

Diversity captures how wide the field of view of an information diet is. Borrowing from ecological metrics, we can quantify diversity using Shannon entropy. The US National Institute of Standards and Technology (NIST) explains that diversity indices summarise how evenly members of a population are distributed across categories[10]. The Shannon diversity index measures uncertainty: its maximum value occurs when all groups have the same frequency[11]. Dividing this index by its maximum yields the Shannon equitability (or evenness) index, which ranges from 0 to 1; a value of 1 means all groups are represented equally[11]. Applied to media exposure, a high equitability index would mean that a user’s feed draws from a balanced mix of sources, topics and viewpoints. Conversely, a low index would indicate that a user’s consumption is dominated by a few sources or narratives.

2. Contradiction rate: productive challenges

Diversity alone is insufficient if the information diet never challenges existing views. We define contradiction rate as the percentage of recommended content that credibly questions a user’s prevailing beliefs or exposes them to opposing perspectives. Operationalising this metric requires combining content classification with stance detection and fact‑check labels. Researchers at Oxford have shown that engagement with fact‑checking posts increases when social ties encourage it, suggesting that exposure to corrections can shift beliefs[1]. A high contradiction rate indicates a feed that supports deliberative resilience, while a low rate signals an echo chamber.

3. Credibility and bias: inequality of exposure

Credibility measures whether content is trustworthy and grounded in evidence. One way to quantify the inequality of credible versus non‑credible exposure is by adapting the Gini coefficient—commonly used to measure income inequality—to information diets. The Gini index ranges from 0 (perfect equality) to 100 (perfect inequality)[12]; a high Gini value indicates that exposure is concentrated in a small set of sources or claims[12]. Applied to credibility, a high information inequality coefficient would show that a small number of unreliable sources dominate a user’s consumption. Combined with metrics like the false‑claim prevalence rate, this forms an overall Recommender Bias Index: a measure of how fairly a platform allocates attention across credible and non‑credible content and across different user cohorts.

4. Claim‑cluster diversity and harm deflection ratio

To capture thematic richness, we can group content into claim clusters—sets of messages conveying similar factual propositions—and measure how many distinct clusters a user encounters. The claim‑cluster diversity metric counts the number of such clusters in a feed, weighted by size. Meanwhile, the harm deflection ratio compares the proportion of harmful content that the platform successfully suppresses to the proportion that reaches users. This ratio can be grounded in the harm categories defined in platforms’ community guidelines and complemented by external fact‑checking organisations such as Full Fact and the European Digital Media Observatory (EDMO).

By combining these indicators—exposure entropy, contradiction rate, information inequality, claim‑cluster diversity and harm deflection—we can model epistemic health as we model economic or environmental health. The NIST AI Risk Management Framework emphasises that trustworthy AI requires risk management practices organised around functions such as GOVERN, MAP, MEASURE and MANAGE[13], with governance as a cross‑cutting function[14]. Our indicators provide the MEASURE function for information ecosystems.

IV. The Information Audit Kit: Turning Transparency into Trust

Having defined what we want to measure, we can design a standardised auditing kit. The kit comprises tools, protocols and governance structures that allow independent auditors, researchers, regulators and even platforms themselves to evaluate epistemic health. This section introduces the architecture, workflow and governance of the kit.

A. Architecture and workflow

Collect: Audits begin with data collection.
Data donation and crowd‑sourced audits: Building on the success of projects like Mozilla’s RegretsReporter—where volunteers donated their YouTube recommendation data—citizens can install browser extensions to share anonymised information about their feeds. Full Fact emphasises that providing links to all sources and ensuring transparency is essential for public trust[8].
Scraping and API access: For baseline measurements, auditors can use scraping tools or official platform APIs to retrieve content displayed to different types of users. The Ada Lovelace Institute notes that scraping audits collect data by automatically scrolling through pages and are particularly useful for describing the proportion of certain kinds of content[15]. API audits provide programmatic access and are easier to automate[16].
Sock‑puppet audits: Perhaps the most revealing method, sock‑puppet audits create simulated user profiles to systematically observe how recommendation engines respond to specific behaviours[17]. Researchers control these personas—varying demographic cues, initial viewing histories and interactions—to capture how algorithms tailor content.
Simulated user journeys and LLM‑generated content: The audit kit can incorporate synthetic personas powered by large language models (LLMs) that interact with platforms in controlled ways. LLMs can also generate controlled news diets and content to test how recommender systems handle synthetic media, linking to emerging guidelines on synthetic media transparency (e.g., watermarking and provenance).
Official data access: Under the DSA, vetted researchers can apply to access internal platform data[6]. The process involves submitting a research application, undergoing vetting by the national Digital Services Coordinator and, once approved, working with platforms to obtain the necessary data[18]. If platforms refuse, they must justify the refusal and propose alternative means[19]. The DSA also envisages an independent advisory mechanism to assist coordinators[20], though its exact structure is still being defined.
Measure: With data in hand, auditors apply the metrics described above. They calculate exposure entropy and equitability across sources and topics, track the contradiction rate by comparing new content to user baseline beliefs, compute the information inequality coefficient by adapting the Gini formula[12], and measure the prevalence of harmful or fact‑checked misinformation. Machine‑learning classifiers and topic modelling help cluster content and identify claim diversity.
Validate: Raw measurements must be validated through independent audit labs. Here, the UK’s Centre for Data Ethics and Innovation (CDEI) provides a blueprint. The CDEI’s roadmap emphasises developing an ecosystem of tools and services to assure that AI systems work as intended, akin to auditing or kitemarking in other sectors[21]. Assurance services complement regulation by enabling companies to demonstrate compliance and trustworthiness[22]. Independent information‑audit labs can be accredited under standards such as ISO/IEC 42001. This standard requires organisations to establish, implement and continuously improve an AI management system, emphasising fairness, non‑discrimination and respect for privacy. ISO/IEC 23894 further guides risk identification and mitigation[23]. Auditors would be trained to conduct sock‑puppet and scraping audits, interpret metrics, and produce public reports.
Disclose: Finally, the results must be disclosed. Platforms would publish Epistemic Health Reports—similar to ESG or financial reports—detailing the diversity, contradiction, bias and harm metrics of their recommender systems. These reports would be auditable and comparable over time. Regulators could use them to assess compliance with the DSA’s systemic‑risk obligations and the UK’s Online Safety Act, while users and publishers could use them to judge whether platforms are upholding their values.

B. Governance and interoperability

To function effectively, the audit kit must operate within a governance framework that balances transparency, privacy and competitive concerns. Several existing frameworks provide guidance:

NIST AI RMF: The RMF’s four functions (GOVERN, MAP, MEASURE, MANAGE) emphasise that AI risk management requires a holistic approach[13]. Our kit operationalises the MEASURE function and feeds back into GOVERN (through independent labs) and MANAGE (through platform improvements).
ISO/IEC 42001: As the first certifiable AI management system standard, ISO 42001 requires organisations to maintain trustworthy, transparent and accountable AI. Adopting this standard can provide a global quality system for information audits.
Ofcom’s evaluation guidance: Ofcom’s research emphasises that recommender systems must be regularly evaluated using methods such as A/B testing, model debugging and user surveys[4]. The audit kit formalises these evaluations and expands them to include external audits.
CDEI assurance ecosystem: The CDEI envisions an AI assurance market where independent auditors provide confidence in the effectiveness, trustworthiness and legality of AI systems[21]. Information audit labs would operate within this ecosystem, certified by national regulators and possibly overseen by an international consortium.

C. A prototype simulation

Imagine deploying the kit to assess the information diets of three simulated users: Asha, a 21‑year‑old climate activist; Ben, a 40‑year‑old conservative parent; and Clara, a non‑political sports enthusiast. Each persona interacts with a platform using predetermined behaviours for two weeks. Data donation, scraping and sock‑puppet techniques log every recommended item.

Asha sees a broad mix of climate news, activism content and policy debates. Her exposure equitability index is high (0.78), and she encounters contradictory views (e.g., pro‑nuclear arguments) 12 % of the time. Her information inequality is low, indicating balanced exposure.
Ben receives a narrow stream of politically aligned content. His exposure equitability is low (0.32), his contradiction rate is just 3 %, and his information inequality is high—mirroring a filter bubble.
Clara experiences a feed dominated by sports commentary. While her content is not political, an audit reveals that platform recommendations occasionally inject unrelated sensational content to drive engagement, lowering her overall epistemic health score.

The audit kit flags Ben’s low diversity and contradiction rates as areas of concern and suggests algorithmic adjustments (e.g., introducing a baseline level of cross‑cutting content). It also recommends that the platform review how harmful sensational content is introduced into non‑political feeds.

This prototype illustrates that epistemic health metrics can diagnose problems and suggest remedies, making information governance proactive rather than reactive.

V. The Credibility Dividend

Why should platforms, publishers and regulators invest in such an audit regime? The answer lies in the credibility dividend: transparency yields tangible benefits.

1. Platforms: risk reduction and competitive advantage

For platforms, independent audits offer a risk‑management tool. They pre‑empt regulatory sanctions by demonstrating compliance with the DSA and national laws. The DSA obliges platforms to provide non‑personalised recommendation options and greater transparency[24]. Complying voluntarily and proving it through audits reduces legal uncertainty. Moreover, platforms that can show a high epistemic health score may attract users wary of misinformation scandals. Just as environmental, social and governance (ESG) data is now used by investors, information integrity metrics could become a competitive differentiator.

2. Publishers: trust as a premium currency

Publishers stand to gain as well. In a fragmented media landscape, trust is scarce. Full Fact’s mission to build a better information environment underscores how impartial, transparent fact‑checking can restore credibility[8]. By partnering with audit labs, publishers can certify that their content appears in balanced recommendation mixes, reassuring readers. High epistemic health scores could justify premium subscriptions or advertising partnerships, much like how “sustainability” labels create brand value.

3. Regulators and society: accountability and democratic resilience

For regulators, the audit kit provides enforceable metrics. The DSA’s transparency and data‑access provisions rely on vetted research to identify systemic risks[3], while the UK’s Online Safety framework emphasises continuous evaluation[4]. Audits give regulators evidence to intervene when recommendations drive users toward harmful or illegal content.

Society benefits through restored democratic resilience. Oxford research shows that social connections can increase engagement with corrections[1]. If platforms optimise not only for engagement but also for epistemic health, they can counteract polarisation and rebuild a shared factual foundation. In the information economy, credibility becomes a new kind of capital adequacy ratio—a measure of a platform’s solvency in the marketplace of ideas.

The Information Audit Kit is not merely a technical tool; it is the blueprint for a new social infrastructure. To succeed, it must be supported by a Global Epistemic Integrity Consortium—a network of universities, civil‑society organisations, standards bodies and regulators. Institutions such as the Oxford Internet Institute and Full Fact can provide research expertise and neutrality. Regulatory bodies like Ofcom and the European Commission can mandate audit disclosures. Standards organisations (ISO, NIST) can align metrics and processes internationally. Independent audit labs, accredited through the CDEI’s assurance ecosystem, would carry out the work. Funding could come from a mix of platform levies (as envisaged in some Online Safety Act proposals), philanthropic foundations and public research grants.

This consortium would develop Generally Accepted Audit Practices (GAAP) for information integrity, analogous to financial audit standards. It would issue guidelines, accredit auditors and maintain an open repository of epistemic health reports. Crucially, it would be accountable to the public, with transparent governance and periodic reviews.

The pathway forward

Creating such infrastructure will not be easy. Platforms may resist external scrutiny. Researchers must protect user privacy and avoid inadvertently reinforcing harmful content. Regulators must coordinate across jurisdictions and update rules as technology evolves. Yet the costs of inaction are greater. Without measurement, there is no accountability; without accountability, there can be no trust.

We stand at a turning point. Just as accounting standards disciplined capitalism, information audit standards can discipline digital capitalism. The DSA has opened the door by mandating data access for vetted researchers[3], while the UK’s CDEI is building an assurance ecosystem[21]. The NIST and ISO frameworks offer structured risk management[13]. The task now is to integrate these strands into a cohesive, global regime. When we can finally see how networks shape our minds, we will recover the right to our own consent—and strengthen the democratic project for the algorithmic age.

[1] [2] OII | Social Ties More Effective Than Shared Politics in Garnering Engagement With Online Corrections

https://www.oii.ox.ac.uk/news-events/social-ties-more-effective-than-shared-politics-in-garnering-engagement-with-online-corrections/

[3] [5] [6] [18] [19] [20] A guide to the EU’s new rules for researcher access to platform data - AlgorithmWatch

https://algorithmwatch.org/en/dsa-data-access-explained/

[4] Evaluating recommender systems in relation to illegal and harmful content

https://www.ofcom.org.uk/online-safety/safety-technology/evaluating-recommender-systems-in-relation-to-the-dissemination-of-illegal-and-harmful-content-in-the-uk

[7] [15] [16] [17] Technical methods for regulatory inspection of algorithmic systems | Ada Lovelace Institute

https://www.adalovelaceinstitute.org/report/technical-methods-regulatory-inspection/

[8] [9] Who we are – Full Fact

https://fullfact.org/about/

[10] [11] Shannon Diversity Index

https://www.itl.nist.gov/div898/software/dataplot/refman2/auxillar/shannon.htm