The Goal Wasn't Victory. It Was Division. · Issue #193

Opening

Reader, yesterday a joint research team from KAIST and Germany’s Max Planck Institute released the results of an analysis of 112.66 million Naver News comments. That’s the complete record, spanning 20 years from April 2006 to March 2025, left by 4.05 million users. Out of that dataset, they flagged 23,998 accounts suspected of foreign-linked influence operations, and the study is being formally presented today at USENIX Security, the top-tier conference in the security field.

But what stuck with me more than the detection technology itself is the target list. There’s no left or right on the list of politicians these accounts attacked. Moon Jae-in, Korea’s former left-leaning president, and Yoon Suk-yeol, his conservative successor, were both targeted by hostile comments from the same account pool. Attacks that don’t pick a side — at first glance, that doesn’t add up.

Let me give you the conclusion up front. The goal of these accounts wasn’t victory for any particular camp. It was making us hate each other more — in other words, division itself.

The Same Hand Slapped Both Cheeks

The starting point was 70 foreign-linked accounts that South Korea’s National Security Strategy Institute made public in 2024. The research team expanded the candidate pool to include accounts connected to these 70 by follows, or accounts that repeatedly swarmed the same articles, then applied a staged analysis of comment language and behavioral patterns to narrow the field down to 23,998 suspect accounts. That’s 0.59% of all users. These accounts left more than 15 million comments in total, and of those, the 4.12 million classified as having the character of influence operations became the object of this round of strategic analysis.

Among the “criticize Korea” comments these accounts wrote, the researchers tallied the targets of those that received only upvotes and no downvotes — in other words, the ones that rose to the top of comment threads (per the paper’s main text and appendix tallies). Seven of the top 10 were politicians. Former President Moon Jae-in, 16,651 times; then-presidential candidate Lee Jae-myung, 13,522 times; former President Yoon Suk-yeol, 9,887 times; former Justice Minister Cho Kuk, 6,314 times. This is a cumulative tally over 20 years through March 2025, so the numbers have nothing to do with anyone’s current position. What matters isn’t the ranking but the composition. Progressive and conservative figures sit side by side on the same list. Non-politician targets follow the same pattern. “Daehan Minguk” (the Republic of Korea), “Hell Joseon” (a derisive term for Korea as a hopeless society), and the names of specific political parties fill out the list, and derogatory variant spellings mocking a particular president appeared so often they were tallied as a separate item.

Break it down by administration and the pattern comes into sharper focus. Roh Moo-hyun, Lee Myung-bak, Park Geun-hye, Moon Jae-in, Yoon Suk-yeol — across all five administrations, whichever president happened to be in office at the time always landed in the top 10 targets, regardless of political camp. Even administrations relatively friendly toward the accounts’ presumed country of origin were no exception. Lee Won-jae, the KAIST professor who led the research, explained that these accounts showed a pattern of stirring up conflict by criticizing Korea and its domestic politicians rather than directly praising a foreign country (as confirmed in domestic press coverage). It’s not about taking sides — it’s about starting fights.

For reference, the paper labels the accounts’ presumed country of origin only as a “major neighboring country,” anonymizing the actual name. It’s a deliberate choice to focus on analyzing the tactics rather than getting drawn into a geopolitical dispute, and the research team maintained the same phrasing in domestic interviews. I’m following the same level of restraint in this piece. What matters for judgment isn’t the name of the country — it’s the method.

Accounts Multiply as Elections Approach

The number of newly emerging suspicious accounts tracked the political calendar. In the 30 days surrounding presidential, general, and local elections, new suspicious accounts appeared at an average rate of 34.19 per week — 51% higher than the 22.61 seen in normal periods (per the paper’s tabulation, p<0.01). You might ask whether elections simply bring more new users overall, and the research team asked the same question. Ordinary user inflow rose by only 24%, while the share of suspicious accounts among new users climbed significantly, from 0.82% to 0.91%. That’s not accounts rising along with a general crowd — it’s selective mobilization aimed squarely at elections.

The single largest inflow peak came in May 2017, during the presidential race right after the impeachment. In that period, 635 new suspicious accounts appeared — the most of any stretch in the 20-year dataset. The peak in overall activity volume, however, came the following year, in 2018. In that year alone, “Korea-bashing” comments totaled 675,107 — about 1,800 per day. This overlaps with the period when diplomatic friction over THAAD deployment spilled into economic retaliation, and, as the paper notes, with the point when neighboring countries began institutionalizing organizations dedicated to cognitive warfare¹.

The sheer workload of these accounts isn’t human-paced either. The 70 seed accounts averaged 5,091 comments each. Even writing one comment a day, that would take nearly 14 years — which the research team reads as a sign that automation was involved.

Nor did they move alone. The known 70 accounts followed one another and clustered together: mutual-follow relationships among them were 138 times higher than among ordinary users, and their average following count was 35 times higher (per the paper’s appendix). This means they weren’t isolated individuals, but a single cluster amplifying one another.

Division Is a Winning Business

Why not brag about your own country instead of trashing someone else’s? The answer lies in the platform’s design. Naver is the gateway through which 63% of Korean adults get their news (Korea Press Foundation 2025 survey, cited in the paper). And within that gateway, comment ranking is determined by the empathy ratio — the number of “agree” votes divided by the sum of “agree” and “disagree” votes. Cross 0.5 and agrees outnumber disagrees, which sharply raises the odds of landing at the top.

Of the four rhetorical strategies used by suspect accounts, only one cleared this bar: “criticizing Korea.” Average 0.514. Praising neighboring countries scored 0.362, and praising allies like Russia bottomed out at 0.241. The regression results point the same way. The harsher the criticism of Korea, the higher the empathy ratio climbed; the harsher the praise of neighboring countries, the more it dragged the empathy ratio down. In other words, bragging about one’s own system only breeds resentment among Korean readers, while comments trashing Korean society rise to the top. Under the paper’s regression model, once a comment’s “Korea-criticism” character was detected at even roughly half strength (probability 0.45), it had already crossed into the zone where agrees outnumber disagrees.

Real examples bring this into focus. Comments criticizing Korea tend to frame conscription as state-run slavery, while comments praising neighbors lecture that Korea should learn from its neighbor’s system. The former rise to the top; the latter sink to the bottom.

There’s a theoretical backdrop to this. When Brady’s research team analyzed 500,000 political tweets in 2017, messages carrying moral emotion² spread far more than those without it. Each additional word conveying moral emotion boosted diffusion by roughly 20% on average. Anger and contempt are the emotions that summon the share button. The “moral condemnation” comments in this study press exactly that button. And the final component in this circuit isn’t a bot — it’s the Korean users who click “agree” because it feels satisfying.

The side effects are interesting too. Evading detection means abandoning the telltale rhetoric of moral condemnation — but that rhetoric is precisely the fuel behind top rankings. The moment evasion succeeds, influence is lost. A well-designed detection standard ends up tying the opponent’s hands by its very nature.

This Comment Section Has Two Prior Convictions

Here I was reminded of the British strategy of divide and rule³. Divide first, then rule (Divide and rule)! The same playbook that split Bengal along religious lines in 1905, and in 1909 created separate electorates by religion, embedding conflict into the institutional structure itself. But there’s one crucial difference. Britain’s goal was to rule, so it paid the cost of designing and maintaining the machinery of division itself. These accounts have no interest in ruling. Korean politics supplies the fault lines for free, and the empathy-ranking system amplifies them for free too — all they need to do is pour in fuel. This isn’t division for the sake of governance — it’s division for the sake of weakening, which makes it far cheaper.

And this isn’t the first time this vulnerable spot has been exploited. Ahead of the 2012 presidential election, the National Intelligence Service’s psychological warfare unit was mobilized to systematically post comments and cast approval/disapproval clicks, and former NIS Director Won Sei-hoon had his four-year prison sentence finalized by the Supreme Court in 2018. From December 2016 to April 2018, the Druking gang used a macro program called “King Crab” to flood roughly 76,000 articles with about 1.18 million comments and push through some 88.4 million approval/disapproval clicks. A state institution, a domestic political faction, and now a group suspected of foreign ties — three different hands, striking the same vulnerable spot. The perception that top comments equal public sentiment, and the empathy-ranking system that determines which comments earn that spot.

Divid The difference lies in the direction of the evolution. Both the NIS and Druking had a side. Theirs was directional interference, meant to help a specific candidate win. These accounts have no side. The interference has shifted from making one side win to making both sides fight. The methods have changed too. Druking used machines to press buttons and forge signals, which left traces in server logs — traces that eventually made it to court. These accounts get real users to click the buttons with their own hands. Since every single click is genuine, there’s no log to catch and no clear provision to punish it.

The overlap in timing is even chilling. The peak influx of suspicious accounts the paper identified (May 2017) sits right in the middle of the period King Crab was operating, and it also overlaps with 2018, when activity levels hit their own peak. In the very same comment section, hands with entirely different purposes were moving at the very same time.

How Much Should We Trust These Numbers

We need to calibrate how warmly we embrace this. First, “detection” isn’t “confirmation.” The research team itself labeled all 24,000 accounts as merely “suspected,” not as confirmed state actors. The classification model behind this study is an explainable AI⁴, and its strength is that it produces sentence-level reasoning for why each account is flagged — but that reasoning is a clue for humans to review, not a verdict.

There are also limits to the numbers. The account classifier’s accuracy was reported at a high F1 of 0.94, but the validation sample was small — just 151 accounts — and the “moral sentiment” labels showed disagreement even among the researchers themselves, landing at only moderate inter-rater agreement (0.512). And Naver has pushed back before: when a similarly-aimed study came out in 2024, Naver countered that all the flagged accounts turned out to be Korean users’ accounts (contemporary news coverage). Naver hasn’t yet issued a position on this new study.

On the other hand, there’s evidence favoring the research. When test comments were entirely rewritten by a language model, detection performance held up almost unchanged (see the paper’s appendix experiments) — meaning the model reads rhetorical structure rather than memorizing specific words. The newly identified accounts’ activity timelines also moved in high sync with the 70 previously-known accounts (a monthly correlation of 0.85), a figure notably higher than that of ordinary users exposed to the same articles (0.70). And because the entire dataset is public, this dispute can move toward re-verification rather than remaining a war of words. There’s also a process in place for anyone to check whether their own comments were included and request their removal.

Oswarld’s Lens

I read this data through the lens of market strategy. There’s one thing I learned while studying business strategy: the cheapest way for a challenger to shake up the market leader isn’t to tout its own product — it’s to demolish trust in the market itself. Once you plant distrust across an entire category, you no longer need to prove comparative advantage. That’s also why smearing competitors works so well in the short term when a new product launches.

Twenty years of these accounts follow exactly that structure. Persuading people that “our system is better” is expensive and doesn’t land well — the 0.362 empathy ratio is the report card proving it. By contrast, sowing division with “your society is rotten” is cheap, spreads easily, and even gets voluntarily amplified by readers in the very country being targeted.

So I think we need to change our criterion for judgment. “Whose side is this comment taking?” is now a useless filter — the data showing all five administrations getting hit proves that. As someone who works with data, I’ll add this: with research like this, you should look at the limits of measurement before the conclusions, and this paper actually earned my trust precisely by disclosing those limits itself. Instead, we should be asking: is this comment trying to make someone win, or trying to make us fight each other? And personally, the chilling part for me was this — the final button that executed this strategy wasn’t a bot. It was our own fingers.

Closing

Let me sum up. The 24,000 suspicious accounts filtered out of 20 years of Naver comments didn’t push any particular camp. They took turns attacking politicians on the left and the right, escalating conflict, with inflows rising every election season — and what pushed those attack comments to the top was the empathy-ratio ranking, and our own like button.

So I have just one proposal. When you run into a political-article comment that gives you that cathartic feeling of vindication, before you hit “agree,” ask yourself just once: is this comment trying to make someone win, or is it trying to make us fight each other? If the answer leans toward the latter, that single click of agreement might be someone’s performance metric. I’m not saying suppress your anger. I’m saying: remember that there’s a hand designing the direction of that anger.

Have you noticed a comment pattern recently that made you think, “this looks a bit organized”? Tell me briefly in the comments — which article, what kind of pattern. If enough cases come in that are worth cross-checking against public datasets, I’ll cover it in a future issue.

💬 Leave a comment about any “this looks organized” sighting you’ve had. I’ll consider it as material for a future issue. 📨 If you know someone who reads news comments often, please share this piece with them.

Past issues worth reading alongside this one

I covered this story’s “neighboring country” angle from the industry and supply-chain axis.

References & Further Reading

Primary sources

Jaehong Kim, Hyeonseung Kim, Jiseon Kim, Alice Oh, Thorsten Holz, Wonjae Lee, Meeyoung Cha, “Cross-National Information Attacks: A Two-Decade Analysis of Troll Behavior in Korea”, USENIX Security Symposium 2026 (arXiv:2606.22785). Link ··· This is the backbone of today’s newsletter. The strategic analysis in Chapter 5 is the core, and Appendix Table 11 lists targets by administration.
The research team’s public dataset, Zenodo. Link ··· The raw data behind the 112.66 million comments. Recommended if you want to verify things yourself.
Electronic Times (Jeonja Sinmun), “20,000-Plus Naver Accounts Stoked Korean Political Conflict” (2026.8.12). Link ··· A news report carrying the research team’s comments from their domestic presentation.

Background

Institute for National Security Strategy (INSS), disclosure materials on foreign influence operations (2024). Link ··· The source of the 70 seed accounts that started this research.
Kyunghyang Shinmun, “Suspected Chinese Accounts Show Signs of Organized Comment-Based Opinion Manipulation in Korea-China Rivalry Sectors” (2024.9.29). Link ··· Coverage of a 2024 precursor report. Note that the figures here — 77 on Naver, 239 on YouTube — are a different tally from this paper’s 70 seed accounts, so don’t conflate them.
Law Times (Beomnyul Sinmun), “Former NIS Director’s ‘NIS Comment Manipulation’ Conviction Finalized at 4 Years After 5 Years” (2018.4.19). Link ··· The first finalized conviction record of a state agency hitting the same comment sections.
Newsis, “‘Druking Comment Manipulation Conspiracy’ 2-Year Prison Sentence Finalized” (2021.7.21). Link ··· The source for the scale of manipulation — roughly 76,000 articles and 88.4 million clicks. The figures in the body of this issue are based on this ruling.
William J. Brady et al., “Emotion shapes the diffusion of moralized content in social networks”, PNAS, 2017. Link ··· A landmark study on why moral emotions spread so effectively. This forms the theoretical background for the third section.

📝 Glossary

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis, including business data management and business analytics, while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master's from Korea University's Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

Cognitive Warfare: A strategy that targets not military force but the perceptions and public opinion of another nation’s citizens. NATO defines it as external actors weaponizing public opinion to destabilize policies and institutions. ↩
Moral Emotions: Emotions like anger, contempt, and disgust that come into play when morally condemning others, or admiration when praising them. Many studies find these spread especially well online. ↩
Divide and Rule: A strategy of stoking internal conflict within a subjugated group to prevent solidarity and make control easier. British rule in India is often cited as a prime example. ↩
Explainable AI: AI that provides not just a verdict but also the reasoning behind it. In this study, it highlights the specific phrases within comments that grounded a suspicious classification. ↩