The Weather of the Self

In 2017, a leaked twenty-three-page internal document, marked “Confidential: Internal Only” and authored by two Facebook executives in Australia as part of a pitch to a major bank, described a capability the company had built and was prepared to sell: the ability to identify the moments when young people, some as young as fourteen, felt “stressed,” “defeated,” “overwhelmed,” “anxious,” “worthless,” or “useless,” inferred from their ordinary posts, photos, and activity.¹ Facebook’s response, once the document became public, was that the underlying research “did not follow process” and had never been used to target advertising. It did not dispute that the capability itself, reading a teenager’s emotional state from their behavioral exhaust, in near real time, existed and worked.

Three years earlier, the company had already demonstrated the other half of the same capability, with considerably more rigor. In a 2014 study in the Proceedings of the National Academy of Sciences, a Facebook data scientist and two academic collaborators at Cornell manipulated the News Feeds of 689,003 users without informing them, showing some a feed weighted toward more positive content and others one weighted toward negative, then measuring what those users went on to post. The result: users shown more negative content posted more negative content of their own, and the reverse held for positive.² This established what the leaked document alone couldn’t: not merely that a platform can detect mood, but that it can move it, deliberately, at scale, through nothing more than what it chooses to show. The study was troubling enough on informed-consent grounds that PNAS later published a formal editorial expression of concern about it.

None of this is confined to one company or treated as hypothetical by the industries that rely on it. Spotify, after acquiring the “music intelligence” firm Echo Nest in 2014, began explicitly selling advertisers the ability to target listeners by mood and moment rather than genre or demographic — playlist categories like “workout,” “commute,” “chill,” and “sad” function, in Spotify’s own advertising materials, as targeting parameters. Its marketing notes that because the average user streams roughly two and a half hours a day, the platform “constantly learns… moods, mindsets, habits, and tastes in the moment.” This has been a named, publicly marketed advertising product for a decade, not a speculative future capability: mood-as-data-point is already a functioning line item in the digital advertising economy,³ and Cambridge Analytica’s psychographic targeting of politically anxious voters, examined elsewhere in this book, applied the identical logic to persuasion rather than commerce.⁴

What makes an individual’s mood legible in the first place, at the scale a platform needs, is the same engagement-signal architecture the previous two chapters documented. An internal TikTok engineering document, confirmed authentic by a company spokesperson and made public in 2021, revealed that the “For You” algorithm weights watch time, rewatches, and pauses during playback more heavily than likes or comments — and was leaked, notably, by an employee troubled by what those signals were doing to already-vulnerable users.⁵ Rewatching, pausing, and lingering are not neutral engagement metrics. They are, mechanically, externally visible proxies for internal state — a person rewatching sad content at two in the morning produces precisely the behavioral signal the Facebook document shows platforms already know how to read, generated for a purpose the platform cares about (keeping the person watching) but legible, in principle, to anyone else able to observe the same pattern.

It is worth being exact about where the documented ground ends and where the argument extends past it. What is genuinely documented is threefold: platforms can detect granular emotional state from ordinary behavior; platforms can shift emotional state deliberately through what they choose to show; and mood itself is an established, monetized advertising parameter with a functioning market behind it. What is not yet documented anywhere in this project’s research is a specific, named case of an individual — a partner, a family member, anyone without a platform’s institutional reach — reading a particular person’s mood cycle from their consumption pattern and timing contact or content to land inside a window of inferred vulnerability.

The extension from the first claim to the second is coherent and built entirely from real, cited parts — but that alone does not make it a documented phenomenon. One tier-stamp belongs here explicitly, in the manner this book applies to its other unconfirmable tactics: what follows is offered as a synthesis — a framework for what to look for, assembled from separately documented pieces, the Facebook and Spotify mood data real and well-attested, the coordinated interpersonal use the inference laid on top of them — not as a proven finding, and it should be held more loosely than anything in the documented chapters. If someone shares an account, a family plan, or administrative access to a target’s viewing or listening history — the vector the previous chapter established as common in intimate-partner surveillance — the same mood-legible signal a platform reads for engagement becomes readable by that person too, for whatever purpose they choose. A message, a “concerned” check-in, or a piece of content arriving during an inferred low period is, in principle, a meaningfully different act than the same message arriving at random. But proving any specific instance runs into a problem this book’s other chapters mostly don’t share as sharply: personalized algorithmic delivery, timed against inferred state, is the default behavior of every platform in this chapter, for every user, all the time. That baseline makes deliberate human timing much harder to distinguish from ordinary algorithmic personalization than, say, a displaced-narrative account built to mirror one person’s biography. The bar for a specific claim here is higher — not because the mechanism is any less real, but because the noise it must be distinguished from is so much louder.

What would establish a specific instance follows a familiar forensic discipline. The answerable question is whether a specific person’s contact — a message, a comment, a piece of content they personally control — arrived at moments correlating with a target’s low-mood windows more often than chance would predict. Answering it requires prospective logging of both the target’s private mood-adjacent activity and the timing of the outside contact, gathered forward in time rather than reconstructed from memory — and, given that personalized timing is the platform’s default, far more data than a single striking coincidence could supply. Where the suspected actor has direct platform-level access, a shared account, a family plan, administrative or parental controls, activity logs showing that person actually viewed a target’s analytics or dashboard at relevant times would be a real, checkable anchor, categorically different from inferring intent from content alone.