LLMs as a Technocratizing Force: How AI Assistants Will Partially Restore Expert Authority Eroded by Social Media

Source: Dan Williams. "How AI Will Reshape Public Opinion." www.conspicuouscognition.com

The Gist

Social media gave everyone a megaphone, which was great for participation but terrible for accuracy—viral outrage and conspiracy theories often drowned out actual experts. AI chatbots like ChatGPT are different because they're built to give you reliable, well-sourced answers rather than whatever gets the most clicks. As more people turn to AI for information instead of scrolling social media, expert knowledge will regain influence in shaping what people believe—not because experts are gatekeeping again, but because AI makes their knowledge easy for everyone to access.

Conclusion

Large Language Models will function as a significant technocratizing force in public discourse by systematically mediating information access in ways that privilege expert consensus and evidence-based reasoning, partially reversing the shift toward populist, engagement-driven opinion formation that social media enabled—though not eliminating democratic participation itself.

Premises

1. Communication technologies structurally shape political discourse by determining the costs of information production, the filtering mechanisms for content, and the epistemic standards applied to public claims—as demonstrated by the printing press enabling the Reformation, broadcast media concentrating editorial authority, and the internet decentralizing it.
2. Social media dramatically lowered barriers to public speech, which democratized discourse but also degraded epistemic quality by rewarding engagement (emotional resonance, novelty, tribal signaling) over accuracy, creating an information environment where viral misinformation routinely outcompetes expert analysis in reach and influence.
3. LLMs are architecturally biased toward expert consensus because they are trained on large corpora that overrepresent peer-reviewed literature, institutional knowledge, and professionally edited content relative to the distribution of content on social media feeds, and they are further aligned through RLHF and safety tuning to produce measured, evidence-based responses rather than emotionally provocative ones.
4. AI companies face strong market and regulatory incentives to maintain factual reliability: enterprise clients demand accuracy, liability concerns discourage misinformation, reputational damage from high-profile errors is costly, and emerging AI regulation in the EU and elsewhere penalizes unreliable systems—creating a structural incentive alignment toward epistemic quality that social media platforms never faced.
5. LLMs dramatically reduce the cost of accessing synthesized, expert-level information by making it available conversationally, on-demand, and in plain language—removing barriers of jargon, paywalls, and information overload that previously limited public engagement with expert knowledge, thereby making technocratic information competitive with populist narratives in accessibility for the first time.
6. As LLMs become embedded in search engines, workplace tools, educational platforms, and personal assistants, they will increasingly mediate the information environment in which people form opinions, functioning as a new layer of epistemic gatekeeping that operates not through editorial discretion but through algorithmic synthesis weighted toward authoritative sources.
7. Technologies that introduce different filtering mechanisms shift political influence accordingly — not by analogy but by mechanism. Filtering determines which claims gain reach; reach determines perceived credibility; perceived credibility shapes political possibility. The printing press, broadcast media, and the internet are illustrations of this mechanism, not the evidence for it. The mechanism itself is near-tautological: if you change what information people encounter, you change the range of views they consider credible. LLMs represent a new filtering paradigm that reintroduces quality…

Assumptions

• Expert consensus, while fallible and sometimes slow to update, represents a more reliable epistemic foundation for policy-relevant claims than unfiltered popular opinion, particularly on empirically testable questions in domains like public health, climate science, and economics.
• A significant and growing share of the public will shift their primary information-seeking behavior toward LLM-mediated interfaces (chatbots, AI-enhanced search, AI summaries) rather than relying exclusively on social media feeds for forming opinions on complex issues.
• The competitive and regulatory pressures on AI companies to maintain accuracy will prove more durable than social media companies' brief flirtation with content moderation, because accuracy is core to the product value proposition of LLMs in a way it never was for engagement-driven social platforms.
• While information quality alone does not mechanistically determine public opinion, the epistemic environment in which people reason significantly shapes the range of views they consider credible, the arguments they find persuasive, and the authorities they defer to—meaning a shift in information quality will produce downstream effects on opinion formation.

Analysis

Overall strength: Moderate. Argument type: Deductive.

Premise Strength

• Communication technologies structurally shape political discourse (Strong) — Well-documented historical pattern with substantial research support
• Social media degraded epistemic quality by rewarding engagement over accuracy (Strong) — Extensively documented phenomenon with clear empirical evidence
• LLMs are architecturally biased toward expert consensus (Moderate) — Training data composition is documented but causal link to consistent expert bias needs more evidence
• AI companies face strong market incentives for accuracy (Weak) — Assumes current incentive structures will remain stable despite rapidly evolving business models and potential shifts toward engagement-driven approaches
• LLMs reduce cost of accessing expert information (Strong) — Demonstrably what LLMs do well - making complex information accessible conversationally
• LLMs will become embedded gatekeepers (Weak) — Highly speculative about adoption patterns and user behavior changes
• Filtering mechanisms determine political influence (Moderate) — Plausible mechanism but overstates deterministic relationship

Potential Fallacies

• Technological Determinism (Premises 1 and 7) — The argument overstates how technology mechanistically determines political outcomes while understating human agency, institutional factors, and adaptive responses that could alter predicted effects
• Hasty Generalization (Premise 3) — Makes broad claims about LLM training data composition and future behavior patterns without systematic empirical support
• Conjunction Fallacy (Overall argument structure) — The conclusion requires multiple independent conditions to hold simultaneously (widespread adoption, maintained accuracy incentives, user behavior changes) but treats this conjunction as more probable than warranted
• Base Rate Neglect (Throughout premises) — Ignores the historically low success rate of technological determinism predictions and the high frequency of unintended consequences from new technologies

Counterarguments

• Premise 3 (High impact) — LLMs may actually amplify existing biases in expert knowledge production and could be manipulated to serve particular interests rather than neutral expertise
• Premise 4 (High impact) — AI companies may eventually pivot to engagement-driven business models similar to social media as markets mature and competition intensifies
• Assumption 1 (High impact) — Expert consensus has historically been wrong on major issues and often reflects elite biases rather than objective truth
• Conclusion (Medium impact) — This represents technocratic capture that concentrates epistemic power in AI companies while reducing democratic participation in knowledge formation

Suggested Improvements

• Empirical Support — Provide systematic data comparing LLM accuracy to social media content across political topics Would strengthen the core claim about LLMs' epistemic superiority
• Causal Mechanisms — Specify more precisely how training data composition translates to consistent expert bias in outputs Would address the gap between architectural features and political outcomes
• Contingency Planning — Address how the thesis would change if AI companies adopt engagement-driven models Would make the argument more robust to changing business incentives
• Democratic Values — Engage more seriously with the tension between epistemic quality and democratic participation Would address concerns about technocratic elitism

Scenario Tests

• AI companies pivot to advertising-based business models that prioritize engagement over accuracy (Challenges) — Would undermine the entire incentive structure the argument relies upon
• Authoritarian governments or corporate interests capture the definition of 'expert consensus' (Challenges) — Could turn LLMs into propaganda tools rather than neutral information sources
• Users develop sophisticated prompt engineering to bypass AI safety measures (Challenges) — Would negate the epistemic benefits by allowing extraction of biased information
• LLMs become primary information source for educated elites while populist movements reject AI-mediated information (Supports) — Could create the predicted technocratic restoration but with increased political polarization

Coherence & Relevance

The argument presents a logically structured chain from technological features to political outcomes, but several links in this chain rest on speculative assumptions about future behavior of companies, users, and institutions. The coherence is strongest when describing current LLM capabilities and weakest when predicting long-term sociotechnical effects.

• Communication technologies shape discourse (Strong) — Connection to specific LLM effects could be more direct
• Social media degraded epistemic quality (Strong) — Well-connected to the problem the argument addresses
• LLMs biased toward expert consensus (Strong) — Critical premise but mechanism from training data to output bias needs elaboration
• AI companies have accuracy incentives (Moderate) — Assumes stability of current business models without justification
• LLMs reduce information access costs (Strong) — Well-connected to how technocratic information becomes competitive
• LLMs will become embedded gatekeepers (Strong) — Speculative about adoption patterns and user behavior
• Filtering mechanisms determine influence (Strong) — Overstates deterministic relationship while understating human agency

View this argument on LogicFirst.ai