AI Elevates Expert Consensus, Reduces Polarisation

Large language models (LLMs) are shifting the informational terrain that institutional investors must evaluate. The Financial Times on 28 March 2026 reported that major LLMs tend to surface moderate, expert‑aligned responses rather than the populist, polarising content endemic to social platforms (Financial Times, 28 Mar 2026). That pattern has implications for sentiment formation, news diffusion and the distribution of market risk because the primary channels that feed investor belief systems are changing: social platforms reached an estimated 4.89 billion users globally in 2023, up roughly 5% year‑on‑year (DataReportal, Jan 2024), while large model deployments accelerated after OpenAI released GPT‑4 in March 2023 (OpenAI, Mar 2023). For portfolio teams, the distinction between algorithmic amplification of novelty on social platforms and algorithmic aggregation toward consensus in LLMs changes where and how signals should be weighed.

Context

The FT piece (28 Mar 2026) frames LLMs as an informational counterweight to social media’s engagement‑driven mechanics. Social platforms reward surprise, outrage and high engagement; that creates endogenous incentives for polarising content to spread faster than moderate alternatives. By contrast, LLMs are trained on large corpora, including vetted sources and scientific literature, and tend to produce outputs that reflect aggregated, centroid views rather than extreme outliers. The temporal inflection matters: mainstream model capabilities expanded visibly after 2023, while social media penetration continued to rise — DataReportal reports 4.89bn users in 2023 — which means investors confront both broader reach for social sources and higher‑fidelity summarisation from LLMs.

Historically, information intermediaries have oscillated between curation and amplification. Newspapers and accredited journals offered high‑friction curation; radio and broadcast scaled reach with editorial control; the internet and social platforms decentralised content production and lowered friction. The emergent set of LLMs represents a new node in that history: instead of lowering editorial standards, many models implicitly re‑introduce a form of editorial consensus by privileging sources that cohere across datasets. That does not make models infallible, but it changes the expected distribution of errors — bias toward conservative, consensus responses rather than random or hyper‑partisan novelties.

For investors, context matters because pricing and risk models incorporate beliefs formed from media. If market participants increasingly consult LLMs for summaries, outlooks or counterparty views, then the propagation speed and shape of information shocks could differ materially from the last decade. The distinction between signal and noise reframes scenario analysis for event risk, corporate communications and policy surprises.

Data Deep Dive

Quantifying the divergence between social platforms and LLM outputs is still nascent, but three measurable contours are already visible. First, reach: social platforms had approximately 4.89bn users in 2023 and have been growing year‑on‑year (DataReportal, Jan 2024), whereas the distribution of LLM usage is concentrated among enterprises, developers and high‑frequency query systems since the widespread commercial rollout began post‑2023. Second, provenance: LLM training data typically incorporate a high proportion of vetted sources (journals, encyclopedias, institutional reports) alongside web text; industry disclosures (OpenAI, 2023) confirm the mix includes curated datasets and filtered crawl data. Third, behavioural responses: platform algorithms reward engagement metrics (clicks, shares, comments), which empirically correlate with more affective language and higher volatility in attention markets; by contrast, LLM outputs tend to be more neutral in tone and emphasize caveats and citation‑style references when prompted.

A practical comparison: in weeks when a political shock occurs, social feeds exhibit rapid spikes in volume and sentiment dispersion, often followed by high variance in reported claims. LLMs, when prompted for a summary, typically produce a bounded synthesis that reduces extreme language and offers probabilistic qualifiers. The FT report (28 Mar 2026) summarises independent evaluations showing this directional difference; while exact magnitudes vary by model and prompt, the consistent pattern is towards moderation in model outputs versus polarisation in user‑generated feeds. That pattern has empirical consequences for measures such as implied volatility around news events — if more investors rely on consensusized summaries, immediate knee‑jerk volatility could be dampened even if structural uncertainty remains.

These differences are not uniform across topics. On highly technical subjects (e.g., pharmaceutical trial design, climate modelling), LLMs often surface specialist consensus that can reduce interpretive error for non‑expert users. On cultural or identity topics where lived experience is central, models can fail to capture nuance and may default to majority perspectives present in training corpora. The asymmetric performance across domains requires investors to calibrate reliance by sector and topic.

Sector Implications

Media and technology: incumbent social platforms face regulatory scrutiny precisely because engagement algorithms amplify extreme content. For technology sectors, the emergence of LLMs as a distribution channel reconfigures competitive dynamics: firms that integrate model‑based summarisation into products may capture attention previously monopolised by feeds. Advertising revenues could shift if attention moves from open feeds to LLM‑mediated search and summarisation, changing CPM dynamics and valuation multiples for platform peers.

Financial markets: information efficiency may improve in areas where models reduce noisy amplification and surface expert consensus. That could compress mispricing in specialist areas where information was previously asymmetric. Conversely, where models homogenise viewpoint, the market may become more correlated in expectation formation, raising systemic tail‑risk if the consensus turns out to be materially wrong. Institutional workflows that historically triangulated across heterogeneous sources may need to adapt to avoid single‑point‑of‑failure risks when models are relied upon.

Corporate communications and policy: issuers and policymakers will need to account for two parallel dissemination regimes. Social media remains the arena for reputational shocks and grassroots mobilisation; LLMs become the arena for distilled corporate narratives and regulatory explanations. Companies might therefore bifurcate their engagement strategies — rapid rebuttals targeted at social platforms and detailed, expert‑oriented disclosures designed to feed model corpora and downstream summaries.

Risk Assessment

Model risk: LLMs are not neutral arbiters. Training biases, dataset gaps and alignment choices shape outputs. Where training corpora under‑represent minority views or emerging research, models may underweight contrarian signals. That introduces a form of model risk distinct from financial model error: consensus bias. Investors must consider the probability that model‑mediated consensus is systematically wrong on fast‑evolving topics, such as geopolitical crises or nascent technologies.

Concentration risk: reliance on a small set of widely deployed LLMs creates concentration externalities. If a widely used model errs on an interpretive issue, the resulting belief convergence could amplify mispricing. Historical analogues exist — for example, the widespread use of similar factor models contributed to crowded trades before the quant unwind in 2007–2008 — and a similar dynamic could occur if many market actors accept the same distilled narrative without independent verification.

Behavioural risk and regulatory uncertainty: while LLMs can reduce immediate volatility by presenting cautious summaries, they may also reduce the incentive for active debate and challenge. Regulators are experimenting with disclosure requirements for model training and provenance; outcomes of these processes (expected to be more active in 2026–2027 across the EU and US) will materially affect model transparency and institutional adoption. The timing of regulation is a risk vector that must be modelled in operational planning.

Fazen Capital Perspective

We view the FT’s 28 March 2026 framing as directionally correct but incomplete for investment implications. The counter‑intuitive insight is that moderation in LLM outputs can be both stabilising and systemic — stabilising at the micro level by reducing noisy overreactions, systemic by increasing correlation of expectations. In practice, that means alpha opportunities will shift from exploiting noise to identifying where consensus is likely to break down. Market participants that specialise in contrarian forensic research — digging into primary sources that models may underweight — can capture excess returns when consensus fails.

A second, non‑obvious point: the value of a hybrid information strategy rises. Teams that combine LLM summarisation for fast triage with selective manual deep‑dive research will have a productivity advantage. That implies investment in workflows, tooling and human capital matters more than raw access to model outputs. We recommend incorporating model provenance checks into due diligence and scenario modelling rather than replacing human analysis with model summaries wholesale. See our internal research on process adaptation and model risk management at AI insights and Market research.

Lastly, contrarian pockets will persist in low‑signal, high‑subjectivity spaces (e.g., culture‑driven consumer brands) where lived experience outranks aggregated text corpora. Investors who can identify where models underperform relative to domain experts should prioritise those areas for active research and potential alpha generation.

Outlook

In the near term (12–24 months), expect parallel information ecosystems to coexist: social platforms will continue to be the vector for viral narratives and rapid sentiment shifts, while LLMs will become a standard tool for summarisation and decision support within institutions. Adoption curves will be heterogeneous across sectors, with financial services, healthcare and professional services increasing model integration first due to productivity gains and specialist needs. The speed of this shift remains a function of regulatory clarity and model transparency.

Over a longer horizon (3–5 years), two structural scenarios are plausible. In Scenario A, regulation and improved model interpretability reduce alignment errors and LLMs broaden their training provenance, leading to more reliable consensus and a lower‑volatility information environment. In Scenario B, model concentration and training bias persist, producing correlated errors and episodic systemic risk when consensus proves wrong. Portfolio construction should therefore incorporate both scenarios into stress testing and liquidity planning.

Operationally, institutional investors should prioritize: (1) provenance audits for model outputs used in research, (2) dual‑track workflows that preserve independent verification, and (3) governance protocols that treat model summaries as inputs, not as final answers. These steps will mitigate the concentration and model risks outlined above.

FAQ

Q: How should investment teams measure the reliability of an LLM’s output?

A: Practical measures include provenance checks (can the LLM cite primary sources?), cross‑model triangulation (do multiple models produce consistent summaries?), and back‑testing against known information events. For example, on regulatory changes, compare model summaries against official filings and agency statements dated to the pertinent regulatory announcement. Systems that log prompts, model version and timestamps create an audit trail useful for post‑event analysis.

Q: Have LLMs demonstrably reduced market volatility where adopted?

A: It is premature to claim a causal reduction in market volatility at scale. Anecdotal evidence and small‑sample studies suggest that access to consensus summaries can dampen immediate knee‑jerk reactions in specialised domains; however, the net effect on market volatility depends on adoption breadth, topic domain and whether the models’ consensus is correct. Robust empirical analysis requires event studies over multiple asset classes and is an area we are monitoring closely.

Q: Are there historical precedents for the type of systemic risk created by information convergence?

A: Yes. The 2007–2008 quant unwind demonstrated how similar models and leverage can create crowded trades and systemic stress. While the mechanism differs — one was financial factor crowding, the other is informational crowding — the risk principle is analogous: correlated decision frameworks can magnify tail outcomes when the shared assumption fails.

Bottom Line

LLMs are reshaping how information is summarised and consumed; they tend to elevate expert consensus relative to social platforms, which increases the importance of provenance checks and diversified information workflows for institutional investors. Treat model outputs as high‑quality inputs, not substitutes for independent, domain‑specific research.

Disclaimer: This article is for informational purposes only and does not constitute investment advice.