Essay 03
By scm7k
Why what people remove from the internet is more predictive than what they post
In 2015, a group of researchers at the University of Cambridge published a study on the relationship between social media activity and military operations. The finding was not surprising: soldiers posted on Instagram. They posted photos of bases, equipment, meals, landscapes. They geotagged. They used hashtags. They did this consistently, predictably, and in aggregate the posts constituted a real-time map of military deployments that anyone with a scraper and a basic geospatial tool could read.
The military response was predictable too. Operational security briefings. Social media policies. Training modules. Don't post. Don't geotag. Don't reveal your location. The policies worked, to a degree. The most obvious signals diminished.
But the researchers noticed something more interesting than the posts themselves. They noticed the pattern of what happened before operations. Soldiers didn't just stop posting. They scrubbed. Old photos deleted. Location history cleared. Accounts set to private or deactivated entirely. The silence had a shape. And the shape was more predictive than the noise it replaced.
A soldier posting photos of a military base tells you where they are. A soldier who has been posting regularly for six months and suddenly goes dark tells you something is about to happen. The absence is the signal. The deletion is the data.
This insight, that removals are more informative than additions, applies far beyond military OSINT. It is a general principle of information systems, and one that is poorly understood outside the intelligence community.
Consider the journalist's source who stops returning calls 48 hours before a story breaks. The absence of communication is a signal that the source knows something has changed. The analyst who regularly publishes on a topic and suddenly goes quiet. The executive who deletes three years of tweets and updates their LinkedIn to "Exploring new opportunities." The government official whose flight tracker data shows regular travel to a foreign capital, then shows nothing for two weeks, then shows travel resuming.
In each case, the deletion or absence carries more information per bit than any individual post. The reason is asymmetry. Posts are cheap. Anyone can post anything at any time for any reason. The signal-to-noise ratio of additions is low. But deletions are costly. They require awareness that the information exists, judgment that the information is dangerous, and action to remove it. Each deletion is a revealed preference. It tells you what the person considers sensitive. And sensitivity is a function of proximity to something real.
The OSINT community has known this for years. Bellingcat's investigations routinely track deletion patterns across social media platforms. When Russian military personnel deleted geotagged posts from eastern Ukraine in 2014, the deletions were as useful as the original posts. The original posts said "I was here." The deletions said "I shouldn't have told you I was here." The second statement carries more information because it includes a judgment about the first.
Scale this to the modern information environment and the dynamics get interesting.
Social media platforms are deletion machines. Hundreds of millions of posts are removed every day. Most removals are mundane: typo corrections, regretted takes, expired promotions, account cleanups. The challenge is distinguishing signal deletions from noise deletions. This is a classification problem, and it is solvable with sufficient data and the right features.
The features that matter:
Temporal clustering. A single deletion is noise. Five deletions in two hours from accounts in the same professional network is a signal. The clustering tells you that multiple people simultaneously decided that something they had previously considered safe to post was no longer safe.
Account age and consistency. An account that has posted three times a week for four years and suddenly goes silent for ten days is a stronger signal than a new account going quiet. Behavioral baselines matter. Deviation from baseline is the measurement.
Content type. Deletions of personal content (vacation photos, family updates) are noise. Deletions of professional content (industry analysis, company announcements, regulatory commentary) are potential signals. Deletions of content that references specific organizations, events, or decisions are high-value.
Coordination. When multiple accounts in the same industry, geographic area, or professional network delete content within the same time window, the probability that the deletions are meaningful approaches certainty. Coordinated deletion is almost always evidence of coordinated awareness, which is almost always evidence of a real event in progress or imminent.
Asymmetric deletion. When someone deletes posts about Topic A but not Topic B, they are telling you that Topic A has become sensitive while Topic B has not. The selectivity is the signal. It points directly at the sensitive subject.
The intelligence value of deletion pattern analysis is well established in classified environments. What is less well understood is its application in open-source contexts: financial markets, journalism, corporate intelligence, regulatory compliance.
Consider financial markets.
A public company's CEO posts regularly on social media about company milestones, industry trends, product launches. Then, three weeks before an earnings announcement, the CEO deletes six months of posts about a specific product line. The deletion is not inside information in the legal sense. It is publicly observable behavior. Anyone with a scraper and a diff tool can detect it. But the deletion carries information that the remaining posts do not: the CEO considers those posts a liability. The question is why, and the answer usually arrives three weeks later in the earnings call.
This is already being done. Firms specializing in alternative data scrape corporate social media accounts, track modification and deletion patterns, and sell the signals to hedge funds. The practice exists in a legal gray zone. The information is public. The inference is analytical. The edge is real.
Now apply the same principle to prediction markets.
A prediction market prices events based on information. Traders observe the world and express their beliefs as positions. The market aggregates these positions into a price. The price is treated as a probability estimate.
But what if the most informative signal is not what traders post on social media or how they position in the market? What if the most informative signal is what they delete?
An oracle validator who participates in resolution discussions, then suddenly deletes their forum posts and deactivates their account, is telling you something. A blockchain analyst who publishes a thread tracing suspicious wallet activity, then removes the thread, is telling you something. A government official who regularly comments on geopolitical events and goes silent during a crisis is telling you something.
In each case, the deletion is a vote of sorts. A revealed preference about the information environment. The validator who scrubs their posts may have received pressure. The analyst who removes their thread may have been contacted by the entity they were investigating. The official who goes silent may have entered a classified operational environment where public commentary is prohibited.
The deletions are breadcrumbs. They point toward the sensitive nodes in the information network. They tell you where the pressure is, which topics have become dangerous, which actors are aware that something is imminent.
The deeper principle here is about the asymmetry between presence and absence in information systems.
Information theory, as formulated by Claude Shannon, concerns itself primarily with the transmission of signals. The signal is positive. A bit is a 1 or a 0. A message is a sequence of characters. The framework is additive: information is transmitted, received, decoded.
But in practice, in human information systems, the absence of an expected signal is itself a signal. The dog that didn't bark in the Sherlock Holmes story. The satellite that doesn't see what it expected to see. The diplomatic cable that should have been sent and wasn't. These are informational events. They carry bits. They update priors. They matter.
Deletion is a special case of informational absence. It is not merely the lack of a signal. It is the active removal of a signal that previously existed. This is more informative than simple absence, because it requires agency. Someone decided to act. Someone judged that the information was more dangerous in existence than in absence. That judgment is the signal.
In an information environment increasingly dominated by automated systems, the deletion signal may be one of the last reliably human indicators. Autonomous agents generate content. They post. They comment. They publish analyses. But they do not, by default, delete in response to social pressure, legal threat, or operational awareness. Deletion, especially coordinated deletion, is a human behavioral pattern that reveals human awareness of human-world events.
As the information environment fills with machine-generated content (news articles, social media posts, analytical threads, market commentary), the ratio of signal to noise in additions will continue to decline. The cost of generation approaches zero. The cost of deletion does not. In a world where anyone, or anything, can add information, the act of removal becomes the scarce, and therefore valuable, signal.
The implications for journalism are direct. Every journalist has experienced the moment when a source goes dark. The contact who was responsive for months and suddenly stops answering. The official who was on background and is now declining to comment. The analyst who was sharing data and has gone quiet.
These silences are assignments. They tell you where to look next. They tell you that something has changed in the source's risk calculus. They point toward the story.
The deletion pattern is the same signal at scale. A journalist covering prediction markets who monitors the social media accounts of oracle validators, platform executives, and regulatory officials can detect coordinated silence in real time. The silence is the leading indicator. The story that explains the silence arrives later.
This is not surveillance. The information being monitored is publicly posted and publicly removed. The analysis is behavioral, not content-based. The journalist is not reading private messages. They are observing the public pattern of what exists and what no longer exists, and drawing inferences from the delta.
The best investigative journalists have always done this intuitively. The formal version, with scrapers and diff tools and temporal clustering algorithms, is just the industrialization of a craft skill.
There is a recursive dimension to this, naturally. Once deletion pattern analysis becomes widely known, sophisticated actors will manage their deletion behavior. They will add noise. They will delete randomly to mask meaningful deletions. They will automate their scrubbing to remove temporal clustering.
This is the arms race that every signal faces. But the defense against deletion pattern analysis is itself costly and detectable. Random deletions create their own patterns. Automated scrubbing produces unnaturally uniform behavior. The countermeasures generate new signals. The game continues at the next level.
At every level, the fundamental asymmetry holds: creation is cheap, deletion is informative, and the gap between what someone says and what someone removes is where the truth lives.
scm7k is the pseudonymous author of PARALLAX, a speculative fiction thriller set in 2028. The novel's second chapter introduces the deletion detection concept through a blockchain analyst in Berlin who notices that the most informative data about a prediction market conspiracy is not what appeared on-chain, but what disappeared. Chapter 1 is free.