So far the picture is of a machine. Ideas move from reality to people through a pipeline of abstractions, and at every stage transport and selection act on what passes: transport re-encodes it lossily, selection decides whether it passes at all. The machine is real, and it is completely morally neutral. That is the problem. A fair reader could finish the transport chapter and the selection chapter and shrug: communication has always been lossy, the compressed form is often the only form that reaches an unprepared receiver, and nothing in either mechanism mentions whether the thing that arrives is true. The machine has no truth-meter. But “something has broken” is a claim about truth, and you cannot derive it from a mechanism that never refers to truth. So here the meter goes in.
The obvious objection is the honest one to start from: simpler isn’t the same as wrong. A short true thing is still true. The answer turns on one idea: every claim travels with a hidden set of instructions for reading it, the scope conditions, the caveats, the “we’re not sure yet,” and compression keeps the claim and throws the instructions away. The claim: compression does not act on truth directly. It strips the decoding key, and whether truth survives depends entirely on which key the receiver reaches for instead. Three outcomes fall out of that one move, and then a story about when each happens.
”Lossy” isn’t the whole worry
Losing detail is not lying. A photograph at low resolution is still a photograph of the thing it was pointed at, and if the only worry were dropped detail, the honest response would be a shrug: of course the pipeline drops detail, that is what abstraction is for. “Compression loses information” and “compression produces falsehood” are different sentences, and the two are easy to conflate when the second is the one that needs earning. A book-length alarm needs more than “things get blurry.” It needs the blur to sometimes point the wrong way. So the question is not whether compression loses information but what kind of loss it is: when does the compressed form stay a true, low-resolution image of the thing, and when does it quietly become an image of something else?
Truth value isn’t in the claim
The move rests on the three-layer message. Any message has an inner message (the claim, the finding, the number) and an outer message, the decoding key: the scope conditions, the qualifications, the “this is provisional” that tells a receiver how to read the inner one. A lossy hop strips the outer message, not the inner one. The claim is small and survives; the key is what falls away. Now ask where the truth value of a claim actually lives. It is tempting to say it lives in the claim: “X” is simply true or false. But findings don’t work that way. “Coffee raises mortality risk” is not true or false on its own. It is true under one decoding (in this cohort, at this intake, by this much, pending replication) and false under another (for you, a lot, settled). Same five words. The truth value was never in the words; it was in the key.
The load-bearing sentence: truth value is not a property of a claim; it is a property of the fit between the claim’s decoding key and the key its receiver actually uses. That is where truth value enters the model. Not with the inner message, not with transport, but at the join, the instant a receiver decodes a surviving claim with whatever key they happen to hold. And that is why compression is dangerous in a way blur is not: it strips the key and leaves the claim, so it never strictly degrades a truth value, it removes the thing the truth value was relative to and hands the bare claim to a receiver who supplies a key of their own. What happens next is three things.
Three things compression can do to truth
A receiver’s substituted key comes from the medium they’re standing in; the medium is where a receiver’s default decoding equipment comes from. So the outcome turns on the gap between the key the idea was encoded with and the default key the receiving medium supplies. Three sizes of gap, three regimes.
Preservation: the keys are close. Sometimes the stripped outer message was elaboration, not correction: detail that sharpened the claim without changing its direction. Strip it, and the receiver’s default key decodes the bare claim to roughly where the full key would have. “The Earth goes around the Sun” has had nearly its entire outer message removed and is still true; “Bitcoin is internet money” loses the whole settlement-layer apparatus and still lands somewhere roughly right. Preservation is when the stripped key was elaboration rather than correction, and the bare claim still decodes to roughly the right place. This is compression behaving the way the naive view assumes all communication behaves: lossy but honest. The regime is real and common, and conceding it cleanly is what earns trust on the other two.
Inversion: the keys point opposite. Sometimes the stripped outer message was load-bearing and corrective, a qualification running against the receiver’s default reading. Take “compound X is carcinogenic in mice at doses orders of magnitude above any human exposure.” The inner message, X is carcinogenic, is small and survives every hop. The outer message, in mice, at megadose, which for a human at ordinary exposure implies almost nothing, is the part that won’t compress and the part the meme stage strips. The receiver decodes the survivor with their default key, carcinogenic means avoid, and the belief produced is X will hurt you when the finding may have meant X is fine for you. That is not a blurry version of the finding; it is its opposite. Inversion is when a true inner message, decoded with the wrong key, yields a false belief, and it is the dangerous regime precisely because nothing was lied about. The claim is intact, checkable, and checks out; the error rides entirely in a key no one can see was swapped. The nutrition whiplash (eggs ruin you, eggs are fine, eggs ruin you) is this regime in public: the findings refined by inches, the compressed forms flipping between confident opposites, because a confident opposite is what survived the trip.
Orthogonality: the receiver isn’t using a key at all. Sometimes the compressed form is selected for a reason unrelated to its content. It is no longer decoded as a claim; it is read as a flag, a marker of which side is showing it and which side passing it on makes you. “Climate change is real” can carry the inner message of atmospheric physics, and in some rooms it still does; in most it has become a team color, and nobody in the exchange is decoding the inner message with any key at all. Orthogonality is when a claim stops being decoded as a claim: its truth value is intact and is no longer the axis anyone is operating on. The operative question has quietly changed from is this true to are these my people, and a claim does not drift here on its own. It is pulled in by a gate rewarding something (identity, belonging, the safety of the in-group) that is simply not on the truth axis.
Two notes on the three. They are outcomes, not intentions: none needs a liar, inversion happens to scrupulously honest findings and orthogonality to perfectly true claims. And preservation and inversion are two settings of a single dial (how large the key-gap was), while orthogonality is a different kind of event (the gate changed which axis it scores on); the honest structure underneath may be two-plus-one, and I come back to it at the end. One more thing, the hinge for Part IV: a receiver lands in one regime, but a population does not. Hand the same compressed claim to a million people and their default keys are not identical, so the claim is preserved, inverted, and rendered orthogonal all at once, in different heads. The width of that spread is set by how much was stripped: the more of the outer message you remove, the more keys the bare claim supports, the further the readings scatter. Compression doesn’t only move the expected decode, it widens the variance, and that interpretive latitude is the seam a large network eventually tears along, which the myths note follows.
Selection picks the regime
If compression strips the key and leaves the claim, then at each hop it produces not one output but a menu: the surviving inner message paired with whichever keys are lying around. Transport does not choose among the menu; transport scrambles. Run the telephone game and what comes out is not reliably inverted and not reliably tribal, just off, a random walk from the input. *Information Evolution* measured it: a phrase in the middle of a meme’s popular variant surviving in 70% of copies, the opening in 59%, the closing in 50%, decay with a shape (edges go first) but no semantic direction. Transport manufactures noise, not false belief.
Direction is the signature of selection. Inversion and orthogonality are not failures of compression; they are compressed variants that a selection gate preferred. Inversion wins when a gate keeps reaching, off the menu, for the alarming variant, and the inverted variant is very often the alarming one, because stripping a scope condition converts a narrow finding into a broad threat, and threat beats calibration at any gate tuned for arousal. Orthogonality wins when a gate reaches for the variant that best sorts in-group from out-group, and a claim emptied of its key makes an excellent flag, because a flag is meant to be content-free. Preservation wins when the gate reaches for correspondence, which some gates do and most engagement-tuned gates do not. This pays a debt the medium note left open: it claimed core truth survives a medium to the degree two conditions hold, gate-criteria that reward something truth-correlated, and a want-loop that does not erode the receiver’s appetite for truth, and flagged “truth-correlated” as a promissory note to be paid here. Here is the payment: condition one is the condition for landing in preservation rather than inversion, and condition two is the condition for the receiver still holding a truth-shaped key at all rather than sliding into orthogonality. The two conditions are one-to-one with the two failure regimes.
Harari’s truth and order, relocated
This is also the right home for Harari’s truth/order axis. His argument in Nexus is that information networks are pulled between truth (an accurate model of the world) and order (the shared story that lets a large group coordinate), and that optimizing hard for one compromises the other. The three regimes sharpen it: order-over-truth is not a global setting on a network, it is the orthogonality regime seen at scale. A claim serving order rather than truth is a claim being decoded as a flag, and that happens one gate at a time, whenever a criterion rewards coordination over correspondence. That relocation makes the axis addressable: if truth-versus-order is a property of a whole network’s design, the only intervention is to redesign the network, which is a sentence and not a plan; if it is the aggregate of millions of per-gate choices, the intervention has a target, the gates and what they reward. The intersubjective-truth note adds the mirror image: reification, the same border crossed the other way, an intersubjective truth processed as if objective, an agreement wearing the costume of nature (“money just is valuable,” “the law is the law”). Orthogonality takes a claim that had an answer and makes it unanswerable; reification takes an arrangement that could be renegotiated and makes it look fixed, which is the camouflage a captured agreement-making process wants most.
The edge of the map: manufactured content
One honest limit. All three regimes assume something true entered the pipeline: an original key to strip, an original encoding to depart from. Purely manufactured content (an invented statistic, an AI-generated claim measured against nothing) has no original outer message, because there was no act of encoding, so the question here is malformed for it: you cannot ask whether compression preserved or inverted a truth value that was never in the pipeline. Most real cases, though, are mixtures, and the regimes apply to the measured component. Astrology is the clean example: the planetary positions are real measurements the regimes can score, while the interpretive framework bolted onto them was never measured and sits off the map. So the three regimes diagnose what the pipeline does to whatever genuine finding it was given, and they do not, by themselves, catch a component that was never true and never even pretended to be measured. Chapter 5b showed why a gate can’t tell the two apart (it scores fitness, not provenance); Chapter 10 is who profits from that blindness.
Where I land
The moral-neutrality gap was that the machine had no truth-meter, so the alarm didn’t follow from the mechanism. Here is the meter: compression has three truth-outcomes, and which one occurs is set by the gap between an idea’s key and the key the receiving medium supplies, with the selection gate choosing which variant off the menu travels. Compression itself is not the villain; the preservation regime is real and ordinary, and most of what compresses, compresses honestly. And here is the alarm, which now follows from the structure instead of sitting on top of it: the breakage this book is about is not that information gets compressed; it is that the dominant medium’s gates systematically reward the compressions in which truth inverts or stops mattering over the ones in which it survives, because alarm out-competes calibration and a flag out-competes a claim at any gate tuned for engagement. Hand such a gate a true finding and a menu of compressed variants and it reliably passes the confident-and-false or the tribal-and-empty one. Not always. Reliably, in the sense the misinformation-spread literature keeps measuring, false and outraging content traveling faster and wider than its corrections (Vosoughi, Roy, and Aral 2018; the out-group-animosity work of Rathje and colleagues 2021). That is the whole distance between “communication is lossy” and “our information environment manufactures false and tribal belief at scale,” which is the book’s actual claim.
That is the brief for Part IV. The prescriptive chapters are not chasing a medium with no compression; there is no such thing, and compression was never the enemy. They are chasing gates that reward correspondence and want-loops that keep receivers holding a truth-shaped key. The medium note’s two conditions, now cashed against the two failure regimes, are the design target.
Where I’m still uncertain
- The three regimes may really be two-plus-one. Preservation and inversion are two readings of one quantity (the key-gap) and slide continuously into each other; orthogonality is a different event (the gate changed which axis it scores on). If so, “three regimes” is a presentational convenience hiding a cleaner structure: one axis (key-gap, preserve ↔ invert) plus one switch (is the gate scoring truth at all).
- “Approximate truth” has no tolerance yet. Preservation rests on how load-bearing the stripped key was, and I have not located the threshold where a low-resolution-true image stops being true, nor shown the boundary is sharp rather than a gradient. Some claims survive compression to almost any depth (heliocentrism); some collapse the moment their qualifications come off (a regression coefficient stripped of its confidence interval). Closing this depends on complexity-as-outer-message-size becoming an operational measure, which is downstream work.
- The scheme here is about objective claims, by decision. The scheme assumes the inner message has a generator the decode can be scored against. The intersubjective-truth note carries the other case: intersubjective truths are still truths (the generator differs, not the truth-ness), and the regimes translate rather than transfer, convergent / divergent / hollow constitution. The residual risk is at the boundary, deciding which generator a given claim answers to, whose hard cases are open there.
- Inversion with a manipulator versus without one. I claim inversion needs no liar, that an engagement-tuned gate produces it from honest findings by itself, and I believe it. But the line between an emergent inversion and a deliberately engineered one is blurry, and until Chapters 5b and 10 draw it, the word systematically in the landing claim carries more weight than it has strictly earned.
← Interlude: The Medium · Chapter 6: Where It All Gets Fucked Up →