← Back to list

The Hardest Part of Social Listening Isn’t the AI #1

After studying how these platforms are built, I came away convinced the industry is solving the wrong problem in public and the right one…

Ayush Agrawal · 2026-07-21 08:03 · 84 claps · 14.1 min read
#artificial-intelligence #machine-learning #osint #socmint #data-engineering
Open on Medium ↗
Wiki topics: ML · Machine Learning AI · AI · General EDU · Education & Learning 🔒 · Cybersecurity 🔧 · Data Engineering

The Hardest Part of Social Listening Isn’t the AI #1

After studying how these platforms are built, I came away convinced the industry is solving the wrong problem in public and the right one in private

The Iceberg: Models Above, Difficulty Below

The Iceberg: Models Above, Difficulty Below

If you asked most engineers what makes a social listening platform hard to build, they’d say the AI. Sentiment models, topic clustering, maybe some large language model doing summarization. That’s where the demos live, so that’s where people assume the difficulty lives.

I spent a long stretch studying this space — the vendors, the government buyers, the open-source tooling, the legal constraints, the academic literature — and I came away with a different conclusion. The models are the easy part. The hard parts are getting the data at all, and being able to prove your output is true. Neither of those is an AI problem, and both are getting worse, not better.

This piece is an attempt to explain why the industry behaves the way it does in 2026 — not what social listening is, but why building it well is so much harder than it looks.

What social listening actually is (briefly), and why the definition hides the difficulty

Social listening is the practice of collecting public online conversation — posts, comments, forums, news, reviews — and turning it into something a human can act on: what people are saying about a brand, whether a rumor is spreading, whether a crisis is forming, what the mood is around a policy.

That definition sounds like a data pipeline plus some NLP. And a decade ago, it roughly was. You could pull a healthy sample from Twitter’s API, run sentiment on it, and put it on a dashboard. Whole companies were built on that assumption.

The reason the industry is hard today is that almost every assumption baked into that sentence has quietly broken. The data got expensive or disappeared. The “sentiment” turned out to be the least valuable thing you could compute. And the buyers who care most — the ones with real budgets and real stakes — need guarantees that current AI simply cannot give.

Let me take these one at a time.

Three markets wearing the same name

The first thing that surprised me is that “social listening” is not one market. It’s at least three, and they’re solving genuinely different problems. People conflate them constantly, which is why so much advice about the space is useless.

Enterprise / brand intelligence. This is the marketing-and-PR world: share of voice, campaign measurement, influencer tracking, catching a reputation problem before it trends. The buyer is a CMO or a comms lead. The metric is engagement and ROI. The tools are multi-tenant SaaS, and they compete on breadth, dashboards, and — lately — how much generative AI they can bolt on.

OSINT and investigative tooling.This is a different animal entirely. Here the unit of analysis is often a person or a network: link analysis, account attribution, dark-web and messaging-app coverage, turning a lead into a case. The buyer is an analyst, an investigator, a security team. The workflows look like case management, not marketing dashboards.

Government and public-sector monitoring.This overlaps with both but has constraints neither shares: procurement rules, security accreditation, data-residency requirements, and — crucially — an obligation to stay on the right side of a bright legal line. A government watching public conversation for crisis response is doing something categorically different, legally and ethically, from a brand tracking hashtags.

One Name, Three Markets

One Name, Three Markets

Why does this matter to an engineer? Because architecture follows the buyer. A tool built for brand marketers assumes multi-tenant cloud, permissive data handling, and “good enough” accuracy. Try to sell that same tool into a public-health agency or a regulator-facing team and it fails not on features but on constraints it was never designed to satisfy. The teams that get burned are the ones who assume a great enterprise product can be lightly reskinned for a high-stakes buyer. It can’t, and the reason is structural.

There’s a subtle consequence here worth sitting with: the capability that makes a tool attractive to an investigative buyer — resolving individuals, attributing accounts — is often the exact capability that makes it unusable for a lawful population-level use case. The two aren’t points on a spectrum you can slide along. They’re different design centers. A lot of vendors quietly straddle this, and it’s the source of most of the controversy in the field.

The data problem is the real problem

Here’s the part that changed how I think about the whole industry.

For most of social listening’s history, the hard-won asset was the analytics. Everyone had roughly the same firehose; you differentiated on what you did with it. That has completely inverted. Today the analytics are close to commoditized, and the scarce, defensible, fragile asset is access to the data itself.

A short, incomplete timeline explains how we got here better than any argument:

2018 — the “APIcalypse.”After Cambridge Analytica, Facebook and Instagram sharply restricted their APIs. A generation of research and tooling that assumed open platform access simply stopped working. Academics still cite this as the moment public-platform data became, in practice, unavailable to most outsiders.

February 2023 — Twitter/X ends free API access. The free tier vanished with about a week’s notice, replaced by paid tiers and an enterprise offering priced far out of reach for most researchers and small vendors. Studies that had run for years on Twitter data were stranded overnight. Published research using Twitter data measurably declined afterward.

August 2024 — Meta shuts down CrowdTangle. CrowdTangle was the de facto standard tool for studying public Facebook and Instagram content — used heavily by journalists and researchers. Meta replaced it with a far more restricted Content Library available to a narrower set of approved researchers, over public objections. The most widely used window into two enormous platforms closed.

2024–2026 — the licensing era. Reddit signed data-licensing deals reportedly worth tens of millions per year with large AI companies, then tightened access for everyone else. Other platforms followed the pattern: data is now something you license, increasingly for AI training, not something you query.

A Decade of Closing Doors (Data-Access Timeline)

A Decade of Closing Doors (Data-Access Timeline)

If you connect these dots, a pattern emerges that goes back further than people remember — Twitter buying its own data reseller and cutting off third parties happened as early as 2015. The platforms have repeatedly pulled data access back in-house, and the trend has only accelerated now that the platforms’ own AI ambitions make their data more valuable to keep than to sell.

For anyone building in this space, three practical lessons fall out of this:

1. Platform coverage is not a feature you build once. It’s a liability you carry forever. Any source can reprice or disappear with little warning. If your architecture assumes a given platform will always be reachable through a given channel, you’ve built a single point of failure into a system that’s supposed to run for years. The durable design treats every source as temporary: an abstraction layer with per-source adapters, redundancy across sources for each information need, and active monitoring of “coverage health” as an operational metric, not a footnote.

2. “We cover 25 platforms” tells you almost nothing. Here’s something I didn’t appreciate until I dug in: many nominally-competing vendors draw from the same upstream sources. A platform’s official partner program may feed a dozen “different” products. So a buyer who picks three vendors thinking they’ve diversified their data supply may in fact be sitting on one channel that can be cut once. Brand diversity is not source diversity. If you’re evaluating vendors, the question that matters isn’t “how many platforms do you cover” — it’s “for each platform, how do you get the data: official partnership, licensed aggregator, scraping intermediary, or your own scraping?” Most vendors don’t love that question, which tells you how important it is.

Many Dashboards, One Pipe

Many Dashboards, One Pipe

3. Scraping is not the escape hatch it looks like. The case law is genuinely mixed. Scraping public, logged-out data has survived some legal challenges in the US, but the same disputes have also produced liability under contract and other theories, and terms-of-service breach remains a live risk that platforms litigate. “It’s public, so we can take it” is not a strategy; it’s an unpriced legal exposure.

The upshot: build-versus-buy in this industry is not really a software decision. It’s a data-access decision wearing a software costume.

Where AI actually helps — and where it quietly doesn’t

Now the part everyone wants to talk about. Generative AI has genuinely changed parts of this stack. But the useful version of that statement is much narrower than the marketing.

Where large language models earn their place:

Summarization and “ask your data.” Compressing thousands of posts into a readable brief, or letting an analyst query a corpus in natural language, is a real and welcome improvement. This is the most defensible LLM use case in the field.

Clustering and topic discovery. LLM-assisted topic modeling — using embeddings and then LLMs to label and consolidate clusters — is genuinely better than the older bag-of-words approaches at surfacing themes a human wouldn’t have thought to search for.

Assistive, Not Autonomous (Where AI Helps)

Assistive, Not Autonomous (Where AI Helps)

Multilingual reach. Modern multilingual models let you work across languages you don’t have bespoke models for. This is real, with an asterisk I’ll come back to.

Semantic search. Retrieval over embeddings beats keyword search for “find me things like this,” which is most of what analysts actually want.

Where classical NLP and older methods still win, and where I think a lot of teams over-reach:

Event and trend detection at scale. Detecting that something is happening from a burst in the stream is largely a statistical problem, and mature, cheap, fast methods handle it well. Throwing an LLM at every post to do this is slower, more expensive, and no more accurate.

High-volume classification. For narrow, well-defined labels at firehose scale, a fine-tuned smaller model is usually faster, cheaper, and more measurable than a general LLM. Measurability turns out to matter enormously (next section).

Anything adversarial. Bot and coordinated-behavior detection is a cat-and-mouse problem where the adversary adapts. There’s no model you train once and trust forever, and the field’s own benchmarks show detectors degrade as manipulation evolves.

And then there’s the category that gets the most marketing and has the least substance right now: agentic workflows — autonomous multi-step AI that supposedly investigates and reports on its own. When you look at the public benchmarks for agentic systems, the completion rates on realistic multi-step tasks are not close to what you’d want for anything consequential, and several widely-cited agent benchmarks have been shown to be gameable. “Agentic monitoring” in 2026 is, for the most part, rule-based alerting and LLM summarization chained together and given a more exciting name.

The honest mental model I landed on: LLMs are a superb assistive layer and a poor autonomous one. They help a human analyst move faster. They do not, yet, replace the analyst — and the reason is trust, which is the deepest problem in the field.

The trust problem: you can’t verify the thing that matters

This is the insight I keep coming back to, because it’s the one that’s genuinely under-discussed.

There’s a comfortable assumption that “AI compliance” is a solved-ish problem — you get your certifications, you document your model, you’re good. But it helps to split verification into layers, because they are not equally solvable:

Process verification — is your pipeline controlled, access-managed, auditable? This is solved. Security frameworks and management-system standards audit exactly this, and a third party can check it.

Data verification — is your input representative, unpoisoned, authentic? This is partial. Platform sampling is opaque and non-random; bot contamination is adversarial; you can diagnose problems but you can’t fully certify the input.

Model verification — is the model calibrated, robust, unbiased? This is hard, and for a closed vendor model that you can only access through an API, rigorous auditing is provably limited — you can’t inspect what you can’t see.

Output verification — is this specific generated report actually true? This is the one that matters for high-stakes use, and it is unsolved.

Four Layers of Verification (Only One Is Solved)

Four Layers of Verification (Only One Is Solved)

Sit with that last one. Every technique we have for checking whether an AI-generated claim is true is itself a fallible model: factual-consistency scorers, retrieval-grounding checks, hallucination detectors, “LLM-as-judge” setups. They all have meaningful error rates. Open-domain automated fact-checking tops out well below the reliability you’d demand for anything consequential. And LLM-as-judge — the popular shortcut where one model grades another — has documented, serious reliability problems: it’s sensitive to answer position, it favors its own family of models, and its rankings shift depending on the benchmark.

Here’s the part that should give any builder pause: no major AI-assurance framework actually verifies outputs. They certify process. They ask you to document your system and self-declare metrics. None of them can hand you a stamp that says “the claims this system produced are true.” For a brand-sentiment dashboard, that’s fine — the cost of a wrong sentiment score is low. For a system informing a public-health response or a government’s understanding of an unfolding event, it is not fine at all.

Which is why, in every mature, serious deployment I looked at, there is a human in the loop making the final call — not as a nice-to-have, but as the only available mechanism for output assurance. The World Health Organization’s open-source epidemic-intelligence work, for instance, keeps human review before anything is published. That’s not technological timidity. It’s an accurate reading of what the technology can and cannot guarantee.

The design principle that follows is uncomfortable for a lot of AI product roadmaps: claim only what you can defend. A responsible system in this space presents AI output as grounded (every statement traceable to source evidence), hedged (confidence expressed, not asserted), and adjudicated (a human signs off). It does not present generated analysis as verified fact, because it can’t.

One more trap: “aggregate” doesn’t mean “safe”

A quick technical aside that I think is genuinely counterintuitive, and that burned assumptions I didn’t know I had.

There’s a natural belief that if you only report aggregates — trends, topic volumes, sentiment distributions — you’ve side-stepped the privacy problem, because you’re not tracking individuals. That belief is wrong in a way that matters.

Naive aggregation leaks. If you cluster posts and summarize each cluster, those summaries will happily reproduce identifying strings — usernames, distinctive phrasing, quotable content — verbatim, unless you actively engineer against it. Research on this has found identifying information surviving in the large majority of naively-generated cluster summaries. And the legal bar for genuine anonymization is stricter than most engineering techniques actually meet; even rigorous differential-privacy deployments come with measurable accuracy and fairness costs.

So “we only do aggregate analysis” is a design goal, not a property you get for free. Staying genuinely aggregate — thresholding, suppression, leakage testing — is real engineering work that, from what I can tell, very few products actually do. If you’re building anything that touches personal data at population scale, this is the unglamorous work that separates a defensible system from a lawsuit waiting to happen.

Why regulation is an architecture problem, not a paperwork problem

Engineers tend to file “compliance” under “someone else’s job, later.” In this domain that instinct is expensive.

Take GDPR. It doesn’t just say “be careful with personal data.” It shapes what your system is allowed to compute. Inferring someone’s political opinions, health status, or ethnicity from their posts can pull you into the special-category-data regime even if the user never stated those things — because you derived them. That’s not a policy you can add at the end; it’s a constraint on your feature set and your data model from the first commit.

The EU AI Act sharpens this further. It doesn’t ban “social listening” by name. What it does is attach obligations to specific capabilities — individual-level emotion inference, biometric categorization, predictive scoring of people. So whether your system is low-risk or high-risk depends on design choices about output granularity, not on what you call the product. Two systems with the same marketing description can land in completely different regulatory categories based on whether they resolve individuals.

And for government deployments specifically, the constraints multiply: data-residency and sovereignty requirements, security accreditation, procurement frameworks. None of these can be retrofitted. A system architected for permissive commercial use and then “hardened” for a regulated buyer generally has to be rebuilt, because the constraints touch the data model, the hosting, the logging, and the human-oversight design — the load-bearing parts.

The lesson I’d offer any founder eyeing the high-stakes end of this market: compliance is architecture. If it’s not in the design from day one, you don’t have a compliance gap, you have a rewrite.

Where this is heading over the next five years

I’ll keep this grounded and avoid crystal-ball territory. A few trends are well-supported by public evidence:

Data access keeps getting harder, not easier. The platforms increasingly see their data as fuel for their own AI, which makes external access a cost to them rather than a revenue line. Expect more licensing, more captivity, and more sudden changes. The open-protocol corner (Mastodon, Bluesky) is a genuine bright spot for durability — open, stable, no gatekeeper — but its audience share is small, so it’s a resilience play, not a coverage play.

Regulation becomes a design input, then a procurement criterion. As the AI Act’s higher-risk obligations phase in, “how do you assure your outputs” stops being a nice slide and becomes a checklist item — and, in Europe, increasingly a sovereignty-weighted procurement criterion. Teams that treated compliance as architecture will move faster than teams that treated it as documentation.

The AI conversation shifts from capability to verifiability. We’re near the top of the “look what the model can do” curve for this domain. The differentiation that’s left — and the genuinely hard research — is in proving what the output is worth: independent benchmarks, calibrated confidence, provenance you can audit. Whoever builds credible verification for this space has something nobody currently sells.

Multilingual gets the asterisk it deserves. Multilingual models give you breadth, but breadth is not depth. For the higher-value analytical tasks — narrative tracking, event detection, stance — the benchmarks that let you measure quality often don’t exist for anything but English and a handful of high-resource languages. “Supports 90 languages” almost never means “measured in 90 languages.” For non-English deployments, the honest position is that you can’t claim accuracy you have no way to measure, and closing that gap is real, fundable work.

What I’d want you to take away

If I compress months of study into a few sentences, it’s this:

Social listening looks like an AI problem and is mostly a data-access and trust problem.

The analytics have commoditized; the data supply chain is the fragile, concentrated, under-examined part.

LLMs are a strong assistive layer and a weak autonomous one, and the honest architecture keeps a human making the final call.

You cannot independently verify AI outputs to a high-stakes standard today, and no framework does it for you — so systems that matter must ground, hedge, and adjudicate rather than assert.

Compliance is architecture. The interesting version of this market — the lawful, high-stakes, population-scale version — is a build-and-research problem, not a shopping trip.

None of this is a reason not to build here. It’s a map of where the actual difficulty lives, which is usually the most useful thing you can have before starting.

So here’s the question I’ll leave you with, because I genuinely don’t think it has a settled answer yet: if the outputs of an AI system can’t be independently verified, what is the right way to put that system in front of someone who has to make a real decision based on it? Every serious team in this space is answering that question right now, whether they’ve noticed it or not. The ones with the best answer will build the products worth trusting.

If you work on any of this — ingestion at platform scale, evaluation and verification, multilingual NLP, or the governance side — I’d be curious how you’re handling the output-trust problem. That’s the part I’m least sure anyone has solved.


메타데이터
post_id
c293cda144a0
slug
the-hardest-part-of-social-listening-isnt-the-ai-c293cda144a0
url
https://medium.com/@agrawalayush730/the-hardest-part-of-social-listening-isnt-the-ai-c293cda144a0
canonical_url
https://medium.com/@agrawalayush730/the-hardest-part-of-social-listening-isnt-the-ai-c293cda144a0
author_url
https://medium.com/@agrawalayush730
status
ok
fetched_at
2026-09-11 05:42:06