Working note. The outline flags this as the next foundational question to draft, and transport-vs-selection ends by handing it over explicitly. That note landed somewhere provisional: transport and selection are parallel mechanisms, selection is the dominant one in modern digital networks, and want is the prime mover behind both. But it closed by opening a question it couldn’t answer. If selection is what matters, and the criteria at each selection gate are tunable rather than fixed, then what tunes them? The one-word answer it gestured at was “the medium.” This note takes that seriously.
The question in the outline’s sharpest form: which media let core truth survive which kinds of manipulation?
The medium is not the substrate
The intuitive picture has the medium underneath everything — a neutral pipe that content flows through. Paper, airwaves, fiber. On that picture the medium is a transport detail: it affects how fast and how far content moves, and nothing else. Selection happens on top of the medium, as a separate thing.
I think that picture is wrong, and getting it wrong is the reason “the medium” keeps showing up as an afterthought in the earlier chapters. The medium is not underneath selection. The medium is the selection criteria, or close enough that the distinction stops paying rent.
Here’s the cleanest version of the claim, the one I’ll commit to provisionally and then spend the rest of the note testing: every gate in the pipeline has a medium attached, and the gate’s criteria are downstream of what that medium can hold and what it rewards. Change the medium and you change the criteria. You don’t change them as a side effect — changing them is what changing the medium is.
This is the load-bearing version of a point Chapter 1 already made in passing, in the “Where the medium fits” section: the medium is the upstream cause of the selection criteria at a stage, not a downstream effect of them. Chapter 1 said it and moved on. This note is the argument that it can carry weight.
McLuhan, re-asked
“The medium is the message” is in the bloodstream of any book on this topic, and there’s no avoiding it. The trouble with the slogan is that it’s been worn so smooth it can mean almost anything. The version I want is narrow and mechanical.
McLuhan’s claim, stated for this book’s machinery: the medium determines what content can win. Not what content can be sent — what content can win. A medium with a 280-character limit doesn’t merely truncate long arguments; it changes which arguments are competitive at all, and over time it changes which arguments people bother to form. The content that survives is the content shaped to the medium’s gate, and the gate is a property of the medium.
So I’ll re-ask McLuhan as: the medium is the selection criteria. That’s the form the rest of the book can use. It’s testable, it connects to the gate machinery from Chapter 1 and Chapter 5b, and it tells you exactly where to look when something has gone wrong — not at the senders, not at the receivers, but at the medium that set the terms of the contest between them.
Three things the medium does to a gate
The medium acts on a selection gate in three distinct ways, and they’re worth separating because they have different remedies.
One: it sets capacity — what can physically pass. This is the most mechanical and the least interesting, but it’s real and it’s the easiest to see. Information Evolution in Social Media has a clean example: before April 2009, on the platform it studied, “a 160 character limit left little room for the payload of the meme next to the replication instructions,” and textual memes were “effectively suppressed” as a result. A medium parameter — a character count — was functioning as a gate. When the parameter changed, the class of content that could survive changed with it. Capacity is a crude lever but it is unmistakably a selection lever, not a transport one. Nothing was traveling slower; some things simply couldn’t compete.
Two: it sets criteria — what the gate rewards. This is the one that matters. A medium doesn’t just cap what fits; it ranks what’s left. And in modern digital media the ranking is performed by an algorithm, which is the thing I want to name precisely: a recommendation algorithm is not lossy transport and it is not a neutral pipe. It is a selection gate with adjustable weights. Psychology of Virality quotes the platform’s own defense of its ranking — that the feed shows people “what they want to see” — and the defense is more revealing than it means to be. It concedes that the gate has criteria, that the criteria are chosen, and that they could be chosen differently. The same paper notes the other direction is equally available: “social media companies can prioritize people’s stated preferences to design better algorithms.” The criteria are a design surface. Someone is always tuning them. The only question is who, and toward what.
This is also where manipulation enters with a precise definition. Manipulation, in this book’s frame, is tuning a gate’s criteria against truth. Not lying — lying is content. Manipulation is upstream of content: it’s adjusting what the gate rewards so that truth-correlated content loses by construction. A medium “allows a lot of manipulation” exactly when its gate-criteria are (a) cheap to tune and (b) tuned by someone whose interest is not truth. That is a property of the medium, not of the manipulators. The manipulators are downstream.
Three: it shapes want — what receivers come to want. This is the deepest of the three and the one McLuhan was really pointing at. transport-vs-selection established want as the prime mover: someone wants to engage with something, engages, accumulates preconditions, wants to engage more deeply, and so on. A medium feeds that loop. A medium that rewards shallow engagement trains receivers, over thousands of repetitions, to want shallow content — and a want for shallow content is a precondition-set that makes the next shallow thing land even more easily. A medium that rewards sustained engagement runs the loop the other way. The medium doesn’t just gate this round of content; it edits the receiver who will face the next round.
Note the order of operations. Capacity and criteria act on the content. Want acts on the receiver. The first two change what wins today; the third changes who’s judging tomorrow. A medium that has been running its want-loop for a decade is not facing the same population it started with, which is why “just add better content” never works as a fix — the gate has already remade the people behind it. [NEED EVIDENCE HERE]
The honest hedge: the medium selects wants, it doesn’t author them
I want to be careful not to overclaim, because the evidence pushes back in one specific place. Psychology of Virality reports that “people who are hostile online are also hostile offline, as opposed to online anonymity suddenly making non-hostile individuals hostile.” The strong, fun version of medium-determinism — the medium reaches in and rewrites people from scratch — is wrong. The hostility was already there.
So the claim has to be stated at the right strength, with a time-scale caveat I want to be explicit about. At the one-year scale the medium does not author wants out of nothing. It selects among pre-existing wants and amplifies the ones its gate rewards. The raw material — the attraction to threat, the in-group reflex, the appetite for status — is human and old and medium-independent on that scale. What the medium does is decide which of those standing dispositions gets the megaphone. That is still an enormous power; it is a curatorial power over an existing repertoire, not a generative one. The medium is a casting director at one-year scale. Run the same medium for a generation, though, and the curatorial loop has done more than amplify: a cohort that grew up inside a particular medium’s amplification pattern may have wants no prior generation had, because what got amplified for them through their formative years is what they learned to want. The medium may be a co-writer at one-generation scale. The casting-director version is the safe, evidence-backed claim and is what this note rests on; the co-writer version is plausible and matters for any prescription that runs across cohorts, but the note has not earned it cleanly. Part IV will have to commit on which scale matters.
The same paper gives the cleanest illustration of the casting-director role. Before the 2022 invasion, out-group animosity was the strongest predictor of what went viral in Ukraine; after the invasion, in-group solidarity surged into that slot. The underlying repertoire of human wants didn’t change in February 2022. The context changed, and the context shifted which standing want the gate rewarded. Swap “context” for “medium” and you have the mechanism: the medium is the standing context that decides, day in and day out, which of our oldest appetites is currently load-bearing.
Where I land
The question was: which media let core truth survive which kinds of manipulation? Here’s the version of the answer I’ll commit to provisionally.
Core truth survives a medium to the degree that two things hold: the medium’s gate-criteria reward something correlated with truth, and the medium’s want-loop doesn’t erode the receiver’s appetite for truth over time. Both conditions, not either. A medium can reward rigor at the gate and still lose, slowly, if its want-loop is training people to find rigor boring. A medium can cultivate a genuine appetite for depth and still lose if its gate mechanically can’t pass anything long enough to be deep.
Most engagement-optimized media fail both conditions at once, which is why they feel not just bad but progressively bad — the two failures compound. Some media pass both: peer review, long-form writing and audio, apprenticeship, the slow institutions of alternative selection. They pass both at a cost, and the cost is always reach. The brutal version of the trade-off: a medium’s truth-preservation and its reach appear to be in tension, and most of the history of communication technology has been a sequence of trades of the first for the second.
That’s the diagnosis. It also tells you what the prescriptive chapters are for. Part IV is not “find a medium that’s good.” There may be no such thing, if the trade-off is as hard as it looks. Part IV is: can we build media whose gate-criteria correlate with truth without collapsing reach to zero — or, failing that, infrastructure that lets high-truth low-reach media and high-reach low-truth media interoperate without the second eating the first. That’s a sharper target than “fix the algorithm,” and it’s the target this note hands forward.
One more thing falls out of landing here. Manufactured content — the injected content that never passed through measurement — is not a separate problem from the medium problem. Manufactured content is what you get when manipulation (tuning the gate against truth) is cheap and the gate has no truth-correlated criterion to begin with. It’s the same failure, seen from the content side. A medium that allows a lot of manipulation and a medium full of manufactured content are the same medium described twice.
Refinement: capacity is the deepest lever
The section above ranked the three things a medium does — capacity, criteria, want — and called capacity “the most mechanical and the least interesting.” I want to take that back. Capacity is not the shallow lever; it is the deepest one, and I had it upside down.
Here is the correction. A selection gate picks from a set of variants; it never generates that set. Something upstream decides which variants are available to be picked at all — and that is what capacity really is. Not “what can physically pass” as a footnote to transport, but the medium drawing the boundary of the option space: the set of forms an idea is even allowed to take before selection gets a vote. The 280-character limit does not slow the long argument down. It removes it from the menu. Selection then runs, visibly and honestly, on a menu that was rigged before it sat down.
That reframes the medium one more turn. If drawing the option space is itself a criterial act — keep these forms, forbid those — then the medium is selection performed once, in advance, and then frozen into the substrate until it stops looking like a choice. A runtime gate is contestable: you can see it and argue its weights. The medium is not, because it does not look like a gate at all — it looks like the shape of the world. That makes it the most powerful manipulation surface in the picture and the hardest to point at: you never have to suppress the disfavored idea if the option space was built so it can never form. The full version — transport generates the variants, the medium bounds them, selection picks — is worked out in transport-vs-selection.
What this changes for the book
This note resolves an open question that several later chapters were waiting on, so the implications fan out.
Chapter 5b gets its missing engine. The selection chapter argues the gates are tunable. This note says what does the tuning: the medium. 5b can now name the tuner instead of leaving it as a passive-voice “the criteria are set.”
Chapter 10 (political economy) stops being optional and becomes structural. If the medium is the selection criteria, and the criteria are a design surface someone owns, then “who owns the medium” is not a side topic — it’s the whole question of who sets the terms of every contest fought on that medium. Political economy is load-bearing because the medium is.
Chapter 11 (AI) inherits a sharper frame. An LLM is a medium and a gate at once — it both carries content and ranks it. The medium-and-manipulation question applied to AI becomes: what are this medium’s gate-criteria, and whose interest tuned them? That’s a more answerable question than “is AI good or bad.”
Chapter 7 (emotional memetics) gets its hedge. The “raw emotion is the floor” argument is really a claim about the gate-criteria of the dominant current medium. State it that way and Mercier’s challenge — that people have real epistemic defenses — stops being a refutation and becomes a boundary condition: the floor is medium-relative, and a different medium has a different floor.
Chapter 12 (infrastructure) gets its design brief. Build for the two conditions in “Where I land”: gate-criteria correlated with truth, want-loops that don’t erode the appetite for it.
Where I’m still uncertain
A few things this note hasn’t earned yet, for the next pass to pressure-test.
- The two conditions might not be independent. I’ve written gate-criteria and want-loop as two separate failure modes, but the want-loop is partly built out of what the gate has rewarded in the past. They may be one feedback system observed at two time-scales rather than two conditions. If so, “both, not either” is the wrong logical shape and the section needs rebuilding.
- “Correlated with truth” is doing heavy lifting and isn’t defined. It leans on the still-open truth-value question. Until that question has an answer, “gate-criteria correlated with truth” is a promissory note. This note can’t fully discharge it.
- Is the reach/truth trade-off actually fundamental, or just historically contingent? I asserted it as a hard trade-off. But that’s exactly the kind of claim that turns out to be an artifact of the media we happen to have had. If it’s contingent, Part IV is far more hopeful than this note implies. I don’t know which, and the book’s optimism level depends on the answer.
- The casting-director metaphor is now flagged with its time-scale caveat in the prose (casting director at one-year scale, possibly co-writer at one-generation scale). The note rests on the one-year reading because that is what the evidence supports; the one-generation co-writer claim is plausible-but-not-earned and the honest input here is mostly absent — comparative generational studies on want formation under different media regimes would be the right evidence. A Part IV prescription that runs across cohorts will have to commit on which scale matters.