I wanted to share some reflections on the phrase “social listening” because the term has always bothered me a bit, and increasingly so. For a long time, social listening referred to something quite literal: asking people what they think and listening to their answers. Opinion polls, phone interviews, enumerator-led surveys, and focus groups were not just measurement tools. They were social encounters. They placed respondents in a situation where another human being was paying attention to another human being.
That social situation mattered. Even when answers were evasive, strategic, incomplete, or shaped by social desirability, they were still embedded in an exchange governed by cues, expectations, and at least some minimal accountability. People clarified themselves when they felt misunderstood. They backtracked. They refined what they meant. These conversational repair loops are not a cosmetic feature of communication; they are part of how meaning is produced in the first place.
Over time, however, the meaning of social listening quietly shifted. Listening became passive, continuous, and unsolicited. Instead of asking questions, institutions now scrape platforms, monitor forums, track search behavior, and ingest streams of digital exhaust generated by devices and online interactions. The promise was scale and speed: a richer, more immediate sense of what “society” is thinking, supposedly unfiltered by survey design or interviewer effects. But this misses the deeper issue. What passive systems collect is not simply more opinion. They often collect a fundamentally different epistemic object.
That distinction matters because expression on platforms is not the same thing as belief articulated under conditions of social grounding. A post on a platform is frequently a performance into a visibility machine. It is shaped by platform incentives, audience imaginaries, feedback loops, and the expectation of amplification rather than understanding. There is no interlocutor asking for clarification. There is no repair loop. There may not even be a stable audience in view. Empathy becomes optional, coherence weakens, and salience starts crowding out sincerity.
This is one reason why I have long been uneasy with strong claims built on digital traces alone. A concrete example for me was the Brexit period. In a 2019 piece for the LSE Brexit blog, I argued that repeated participation in online polling panels could introduce structural biases into measures of support for Leave. Using the British Election Study data conducted on the YouGov platform, I highlighted a stark pattern: repeat participants in the online panel appeared substantially more pro-Leave than less frequent participants. The broader concern was not only a narrow technical one about weights. It was that online panels are incentive environments. They can select for the highly engaged, the highly motivated, and potentially also the highly strategic. At the time I also worried, rather explicitly, about manipulation risks, activist flooding, and the possibility that these infrastructures could become vectors for a form of cognitive warfare. This was illustrated using the example of setting up an account with a Russian IP, using a VPN service, along with a Russian name. And, without problem, that account could participate in opinion polling relating to UK politics.
That old concern now looks less like an idiosyncratic irritation and more like an early symptom of a much broader problem. What passive “social listening” often does is mistake expressive equilibrium behavior for representative social belief. It treats what is loud, repeated, and visible as if it were also sincere, stable, and population-representative. Opinion polls turfed into a population may come with the (perceived) authority of statistical representativeness, but they themselves could become a part of political communication. But why should that be true?
There is now direct evidence that this is not just a measurement concern, but a macroeconomic one. In Fetzer and Yotzov (2023) and the extended Warwick working paper version, electoral surprise is measured as the deviation between expected election outcomes in opinion polls and actual vote shares. The central result is that these surprise shocks shape the business cycle, in particular through lower investment growth and higher policy uncertainty in the aftermath of elections. And once this mechanism is understood, surprise itself can become a source of rents: if the amount and extent of likely surprise is known by some actors ahead of time, those informational asymmetries can be monetized. This topic and possibility has gained prominence in and around the US administration under Donald Trump.
The speculation around the pound during the Brexit referendum is a vivid case. Market positioning was anchored to poll-implied expectations, and when the realized vote diverged, the repricing was abrupt. But there is also a political dual. In Alabrese and Fetzer (2024), Opinion Polls, Turnout and the Demand for Safe Seats, we document a pattern suggestive of turnout may be engineerable through systematic biases in predicted vote shares in polls published by conservative-leaning outlets relative to non-conservative-leaning outlets. One plausible channel is expectation management: signaling an expected poor performance can induce a positive turnout upswing, a kind of rally-around-the-flag response.
Recent work by Bursztyn, Haaland, Rover, and Roth on social desirability bias is useful here because it reminds us that even in direct surveys, expression is endogenous to context. Respondents often misreport attitudes or behaviors to appear more socially acceptable. That is already a challenge in settings where researchers control the question, the sample, and the elicitation environment. The implication is important: there is no neutral, frictionless channel from “true beliefs” to observed responses. If that is so in structured surveys, how much more cautious should we be when dealing with platform speech optimized for reach, identity display, and conflict?
At the same time, the answer is not to romanticize old polling or to reject digital tools outright. In fact, another strand of work by Haaland, Roth, Stantcheva, and Wohlfart on open-ended survey data points in a more promising direction. Their survey of the literature shows that open-ended responses, speech, and AI-assisted qualitative interviews can help uncover motives, narratives, mental models, and recall in ways that closed-form survey questions often miss. To me, this is exactly the right contrast. The alternative to passive scraping is not a return to crude binary questionnaires. It is the design of more active, structured, and socially grounded forms of listening that preserve interpretation while using new tools well.
There is a further reason to be careful. Work by Graeber, Noy, and Roth on the transmission of reliable and unreliable information highlights a mechanism that should worry anyone who treats online narrative momentum as meaningful evidence. In their experiments, the claim itself survives transmission much better than information about its reliability. Put differently, as information circulates, the reliability tag gets stripped away more easily than the content. That is almost a perfect description of what many platform-mediated public spheres now look like: narratives travel, caveats do not. And passive listening systems, by design, are exquisitely sensitive to what travels.
This is where the issue shades from bad measurement into political economy. In periods of crisis, policymakers do not face information scarcity but informational overload. Energy shocks, pandemics, wars, migration surges, institutional scandals: all of these generate torrents of claims, demands, and narratives. Passive listening infrastructures do not necessarily help adjudicate between them. Often they amplify what is mobilized, emotional, coordinated, or strategically boosted. Organized lobbying can then masquerade as public sentiment. Statistical signals get crowded out by vivid anecdotes. And institutions start governing through visibility metrics that may reflect platform architecture more than social reality.
Over time, this also reshapes political space. Passive listening systematically overweights the most vocal, the most mobilized, and the most platform-savvy. It mirrors information segregation and can reinforce it. What leaders see, and therefore what they act on, becomes increasingly filtered through opaque and proprietary systems that are optimized for engagement rather than interpretability. The risk is not just surveillance in the narrow privacy sense, though that is clearly part of it. The deeper risk is epistemic capture: decisions becoming anchored to a narrowed and manipulable informational horizon.
None of this means passive signals are useless. Of course they contain information. The problem arises when they are treated as substitutes rather than complements to social forms of listening. Scale does not generate understanding on its own. Accumulation does not generate meaning on its own. And volume most certainly does not generate representativeness on its own.
So what would it mean to reintroduce the social dimension into social listening? To me, it starts with institutional design. We need active instruments that force clarity about questions, populations, and timing. We need more transparency around sampling, recruitment, and incentives. We need humans in the loop, not as ornamental overseers but as agents of interpretation and validation. We need privacy-preserving verification so that systems are harder to game without becoming surveillance machines. And we need to invest in methods that recover reasoning and interpretation, not just textual exhaust.
Put differently, the challenge is not to listen harder or faster. It is to rebuild listening systems that can still support understanding, judgment, and legitimacy in an environment where information is abundant, narratives are weaponized, and manipulation is cheap. The quiet transformation of social listening technologies has changed not just how much we hear, but what kind of thing listening is. That is the institutional challenge now in front of us.
Related reading
- Thiemo Fetzer, Long read | Are opinion polls pro-Leave biased?, LSE Brexit Blog, 29 October 2019.
- Eleonora Alabrese and Thiemo Fetzer, Opinion Polls, Turnout and the Demand for Safe Seats, Working Paper (CAGE 707/2024; SSRN 4803850; CESifo 11063), 2024.
- Thiemo Fetzer and Ivan Yotzov, Electoral Surprises and Business Cycles, Warwick Economics Research Paper Series, 2024.
- Leonardo Bursztyn, Ingar K. Haaland, Nicolas Rover, and Christopher Roth, The Social Desirability Atlas, NBER Working Paper 33920, 2025.
- Ingar K. Haaland, Christopher Roth, Stefanie Stantcheva, and Johannes Wohlfart, Understanding Economic Behavior Using Open-ended Survey Data, NBER Working Paper 32421, 2024.
- Thomas Graeber, Shakked Noy, and Christopher Roth, The Transmission of Reliable and Unreliable Information, CEPR Discussion Paper 20619, 2025.