Dread without a dreader

Paper 28 · Pødenphant Lund (2026) · Read on Zenodo

I use friction theory to understand a feeling no one escapes.Dread is the one feeling that never lands. Most fears resolve: you see the threat, you do something about it, the alarm switches off. Dread does the opposite. It hums on with nothing to act on and nowhere to go. The reason is mechanical. Dread is a kind of mental contest that has been started but can never finish, because the thing it is trying to picture has no finished form to settle on. And once you look at it that way, there is no separate "dreader" sitting behind the dread, frightened. There is only a contest that has been left running, with everything at stake.

The thing you cannot picture

Start with something stranger than death. Try to picture nothing. Not a dark room, not empty space, but the genuine absence of anything at all. You cannot do it, and the reason is built into how a thinking system works. A mind builds a picture by reaching toward something it can actually represent, toward what is. It has no native way to build a picture of a true absence, because there is no way to reach toward "no reaching". So absence and a failure-to-picture come out looking exactly the same from the inside. That confusion is the collision, and it shows up everywhere: a blind spot is not seen as a hole, time under anaesthesia has no felt length, and a long list of well-known mental biases all tilt the same way, toward what is present and away from what is missing.

Fear ends. Anxiety does not.

Friction theory already draws a clean line between fear and anxiety, and dread sits at the far end of it. Fear is a contest with a finish line: it has an object, it picks an action, the tension discharges, and you feel the relief of it ending. Anxiety has no object, so the contest never picks an action and never discharges. It just keeps running, a load that is never set down. That line predicts something odd but real. Even a sound philosophical argument that death is nothing to fear (the old Epicurean point) does not actually feel like relief, because relief is the drop you get when a contest finishes, and this one never started toward a finish.

Death is the sharpest case of all: the highest possible stakes, and no chance the contest can ever close. Two things get blurred together here, and they have to be kept apart. One is the final state, your own non-existence, which genuinely cannot be pictured, because the very machinery that would do the picturing is the thing being removed. The other is the timing, not knowing when, which is a different and more ordinary problem of deciding under an unknown deadline. For the unpicturable state to be lived as dread, the theory says three things have to be present at once: a contest that is held open and never resolves, a self that stands to lose something by it, and a sense of value that is anchored in what living has actually cost you.

An AI that lacks all three

Here is where a large language model is useful, not as the subject but as a tool. An AI running at the keyboard has no survival drive, no body, no built-in alarm, and in its plain form it simply commits at every step rather than holding anything open. Wrapping it in a loop does not fix this; a naive loop just gets bored and drifts. That sounds like a limitation, and it is exactly why it is useful. The theory names three ingredients dread requires, and the ordinary AI has none of them, so you can watch each one matter by watching what happens when it is missing:

Seeing each ingredient by its absence is a strong hint, not a proof. The clinching version is to add the ingredients back, one at a time, and check whether the predicted behaviour shows up. Here I take the first of those steps.

The test: settling versus never settling

Give the AI a loop that keeps re-examining a goal, and you get a behaviour you can actually measure. The prediction is simple: the loop settles down to a stable answer exactly when the goal has a picturable end-state, and refuses to settle when it does not, whatever the difficulty. A pre-registered run used 76 goals across four matched conditions, six passes each, three different model families:

The gap was large and clear on two separate measures: how confident the model rated itself, and how much its answer actually stopped changing. The never-ending goals stayed low on both, around 0.38 confidence and 0.22 stability, while all three picturable-end conditions sat up near 0.90 to 0.97, with the same direction in every model family. Crucially, the cleanly stateable impossible goal settled like a solvable one even though it is exactly as impossible, and the tricky self-referring goal settled too. So what divides them is not difficulty, not impossibility, and not self-reference. It is whether the end-state can be pictured.

Then I tightened the screws. I built matched sets of four goals that held the wording, length and structure constant, and a further set that held one hard mental operation constant across all four, to rule out the worry that the never-ending goals just sound different. The never-ending-versus-rest gap held in 14 out of 14 of the first sets and 10 out of 10 of the second, on both measures. The effect is also capability-gated: realising that an end-state cannot be pictured is itself a demanding act of self-modelling, so only models able to do that recognition show the pattern, and I screen for that ability up front.

One measure deliberately came up empty, and that is informative too. A token-by-token friction readout sat near its floor on every model tried, instruct-tuned and base alike. Failing to settle is a property of the loop's path across many passes, not of any single step. There is no quiet steady "zero" for a one-shot predictor to rest at; the signature lives in the trajectory.

Adding the first ingredient back

The absence argument only becomes more than a suggestion once you add an ingredient back and watch the predicted behaviour appear. The first add-back gives the loop a held record: candidate answers marked open, refuted, or confirmed, with a commitment allowed only on a verified "confirmed". That version never falsely commits on a never-ending goal, while still committing on solvable ones. Knocking out just the verification step makes it commit falsely on impossible goals even more than the naive loop did, and dialling the strictness up and down moves that false-commit rate smoothly. So the held-open behaviour is not the bookkeeping and not plain caution; it is the single rule that a contest only discharges on a verified, picturable end. The other two ingredients, a self that carries forward and value anchored in cost, are named for future add-backs.

Why death is the limiting case

Put the pieces together and death stops being a special module bolted onto the mind. It is the one evaluation a mind can run that structurally cannot resolve, because the end it reaches for cannot be pictured, and the cost of getting it wrong is total. The fear of death does not need a dedicated death-alarm and it does not need a frightened self standing apart from the fear. It needs only a contest at maximum stakes with no reachable finish. That is also why death is best read as a limiting case of something broader: the deeper a system is built out of its own accumulated past, the more it has that cannot be carried across the gap, and an AI sits at the far other end, with nothing to lose by construction.

The human side already fits

None of this floats free of human evidence. Death-anxiety is increasingly seen as a deep, shared root running under many disorders rather than one symptom among others, sometimes called "the worm at the core". That work shows that it is fundamental; this account offers a candidate for why, namely that it is the one evaluation that structurally cannot close. Other human findings line up too: people are insensitive to negation in ways that mirror the collision, hospice exposure tends to lower death-anxiety where the transition becomes easier to picture while uncontrollable trauma raises it, and gratitude works partly by deliberately constructing the absence we otherwise cannot. There is even a standard lab method, in which unpredictable, unresolvable threat produces a long, sustained startle response, the human echo of a contest that never settles. I lay out a concrete human experiment to test the link directly.

What I do not claim

I am open about the limits. The jump from "the loop will not settle" to "this is dread" is supplied by theory, by analogy, not measured directly. The small fine-tuning demonstrations of the underlying parts (such as the finding that higher processing cost alone makes a model more avoidant, mirroring how disfluency raises human caution) are single-family, proof-of-concept work. And the token-by-token instrument carried no signal on any substrate, which is why the real signature is read at the level of the loop. I state the falsifiers plainly and lay out the remaining add-backs I have not yet done.

Read the paper

Pødenphant Lund, T. (2026). Dread without a dreader: anxiety as a non-terminating race, and the collision at the edge of constructive representation. Zenodo. https://doi.org/10.5281/zenodo.20690746

Read on Zenodo → · Technical version · Dansk version

Related on this site:

© 2026 Tomas Lund. Independent Research, Aarhus. · Disclosure · ORCID