Three spotlights, one red cup
Place a red cup on a crowded desk. A transformer receives the sentence, “Move the red cup beside the notebook”, and assigns strong attention from move and cup towards red. A macaque sees two shapes inside one receptive field while a cue biases neural responses towards the red one. A person reaches for the cup while experiencing it as the same cup seen from one side.
All three descriptions invite the same verb: attend. The verb then encourages a quiet slide. A weighted token becomes a selected stimulus; a selected stimulus becomes an experienced object; an experienced object becomes evidence of awareness. The sequence feels natural because each case contains a form of directedness. Yet the cup lets us pull the cases apart.
Change its colour from red to blue. Hide it briefly behind the notebook. Ask the person to imagine it rather than perceive it. Replace the transformer's learned attention map with a different map that leaves the output unchanged. Each change preserves some relations and destroys others. The question is not whether all three systems “focus”. The question is which causal relation survives the same intervention.
Accessible reading: the figure shows three unlike geometries rather than three renamed boxes. Only the bottom three landmarks are aligned. The intentional field contains relations absent from the other two.
One word enters three laboratories
The term attention names a computation in one laboratory, a family of organism-level and neural effects in another, and a feature of lived or mental organisation in ordinary and philosophical language. The word is useful only after its level is declared.
In a transformer, an attention head creates a weighted mixture. A query vector is compared with key vectors; the resulting scores determine how much each value vector contributes to an updated representation. Nothing in the equation says that one vector is a cup, that the cup exists, or that the system takes itself to be perceiving rather than imagining. The mechanism can nevertheless be causally important. Change its scores, mask a key or replace a value, and later computation may change.
In a biological system, selective attention is not one scalar or one anatomical switch. Spatial, feature-based and object-based paradigms expose different forms of priority. Neural accounts include biased competition, gain modulation, normalisation, synchronisation and recurrent control. The organism's task, body, history and possible actions shape what is selected. A visual neuron's response can therefore change even when the retinal stimulus at its classical receptive field is held fixed.
Intentionality answers another question: how can a perception, thought, memory, fear or hope be of or about something? Brentano revived intentional directedness as a mark of mental phenomena. Husserl then analysed the correlation between an act and its object as intended, including the mode of the act and the horizons through which the object is given. A remembered cup, a perceived cup and a doubted cup can concern the same object without being the same kind of act.
Selection answers “what influences processing now?”
Intentionality answers “what is this act or state about, in what mode, and under what conditions could it be fulfilled or mistaken?” The two questions can interact in a human mind. They are not identical. Priority is a relation among candidates; intentionality is a relation to an object as intended.
Three questions prevent the slide
Before comparing domains, ask three questions in order. What was selected? This identifies candidates and the variable that changed their relative influence. What was the selection used for? This locates the downstream task, action or report. What remained the same object through the change? This tests whether the system preserved reference rather than merely following a feature. Most loose analogies answer only the first question, then borrow the language of the third.
Consider hearing your name in a noisy room. The sound can be physically salient, selected for further processing and used to turn your head. Those are separable claims. A flashing advert may be more salient yet fail to alter your action. Your name may alter action without becoming the only content of experience. The person named can also remain the same across different voices, spellings and memories. Salience concerns stimulus contrast; selection concerns differential influence; access concerns availability to some process; intentionality concerns what the state or act is about. Their correlations are empirical, not definitional.
The same discipline applies to machines. A token can receive a large weight and contribute little because its value is redundant. A low-weight token can matter through another head or earlier layer. A retrieved document can dominate the final answer even though no single attention head visibly concentrates on it. Conversely, a model may route strongly towards a phrase while the configured system's persistent target is supplied by a case identifier outside the model. The unit that selects need not be the unit that owns the target relation.
Match relations, not nouns
A defensible bridge does not say “query is intention” or “gain is attention weight”. It says that, under a declared intervention, a change in context alters relative influence over candidates and produces a measurable downstream difference. The source and target may realise that relation through different variables, timescales and system boundaries. The comparison strengthens when the same counterfactual ordering survives. It weakens when only the vocabulary or heatmap resembles.
The replacement test
Replace attention while preserving the sentence
Suppose a trained classifier assigns the correct label to “The medicine is not safe”. Its attention display highlights not. Now replace the learned distribution with a fixed router, or with a different distribution constructed to preserve the same output. If the label remains unchanged across the test set, what survived?
The behaviour survived. Perhaps information about negation had already entered contextual representations, or another layer compensated. The original heatmap did not earn the claim that it was the sole reason for the answer. More importantly, neither heatmap earned the claim that the system experienced negation or was directed towards a proposition in the phenomenological sense. Output preservation across a routing replacement lowers a mechanism claim before any mental claim is considered.
This test exposes the hidden assumption behind loose comparisons: that visible concentration reveals the essence of attention. It does not. A spotlight picture can be generated for token weights, neuronal gain, gaze, report frequency or subjective salience. The picture shows a distribution. It does not tell us what caused it, what the selected item is for the system, or whether an object remains the same when its features change.
Published results support a narrow conclusion
Transformer attention is a defined computation. Visual attention has replicated behavioural and neural effects, with causal interventions in some circuits. Attention weights as explanations remain task- and architecture-dependent, and the literature contains explicit disagreement about which tests are decisive. Phenomenological intentionality is a descriptive and philosophical structure, not a neural measure. The present bridge is therefore a bounded structural comparison, not an equivalence proof.
flowchart LR A["Three spotlights, one red cup"] --> B["One word enters three laboratories"] B --> C["Selection leaves a causal<br/>fingerprint"] C --> D["The object survives the swap"] D -. "re-test" .-> A classDef first fill:#0a1424,stroke:#7B4CD8,stroke-width:2px,color:#e6f3ff classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff class A first class B,C,D step
Selection leaves a causal fingerprint
A router, not a witness
Take one query position in one attention head. Let q be its query vector. Let k_j and v_j be the key and value vectors at candidate position j. The head computes a compatibility score and normalises the scores:
The dimension d controls score scale. The temperature τ controls sharpness if explicitly varied. The weight α_j is not a quantity of awareness. It is the coefficient by which v_j contributes to z at that head and layer. Residual connections, other heads and later feed-forward transformations continue the computation.
Three scores become one mixture
Suppose three candidates have scores [2, 1, 0]. At temperature one, their softmax weights are approximately [0.665, 0.245, 0.090]. If their scalar values are [10, 4, -2], the mixture is 7.45. Raising the first score increases the first value's contribution. This is a clean causal role.
Now change the values to [8, 8, -2]. Many different weight distributions can produce similar mixtures because the first two candidates carry redundant information. A large weight can therefore be neither necessary nor uniquely explanatory. The equation explains routing under declared representations; it does not identify semantic ownership or objecthood.
Accessible reading: three score boxes feed three weight boxes, which scale values and sum to 7.45. A note below warns that alternative weights can preserve the result.
The same answer through a different map
Construct two attention distributions that differ sharply but produce the same output because contextual representations are correlated. A viewer sees two unlike heatmaps and wants to infer two unlike reasons. Yet the model's next-token distribution barely moves. The causal variable was not the picture alone. It was the picture acting on a particular set of values inside the remaining network.
This does not make attention useless. It changes the test. Patch or ablate the routed contribution, compare matched alternatives, inspect downstream effects and repeat across examples. A machine-attention claim becomes stronger when intervention, not visual salience, carries it.
Competition inside a living loop
Biological attention also changes differential influence, but its candidates are not token positions and its control is not exhausted by a single normalisation. Desimone and Duncan's biased-competition account begins from limited processing: simultaneously presented objects compete for neural representation, while task and stimulus factors bias the competition. Reynolds and Heeger's normalisation model showed how a family of apparently different attentional response patterns can arise from the relation among stimulus drive, an attention field and a suppressive pool.
A deliberately simplified numerical model makes the intuition visible. Let two stimuli each supply drive six. Let the normalisation constant be two. With equal gain, the first response is 6/(2+6+6)=0.429. Give the first stimulus gain two while the second remains at one. Its response becomes 12/(2+12+6)=0.600, while the competitor falls to 6/20=0.300. Selection changes because gain and competition interact.
Accessible reading: two equal peaks share a suppressive pool. A task bias increases the right peak's gain, producing a higher response for it and a lower relative response for the other peak.
Biasing one control site changes another area's response
Moore and Armstrong stimulated sites in frontal eye fields below the threshold for producing an eye movement. Visual responses in retinotopically corresponding V4 sites were enhanced. The result matters because it perturbs a control-related region and observes a selective sensory consequence. It supports a causal route from bias to differential processing. It does not establish that the stimulated animal experienced the target in any particular way, and it does not make frontal stimulation mathematically identical to a transformer's query-key score.
The shared selection contract
We can now state the bounded overlap without metaphor. A selection system has: a candidate set; a context or bias; an allocation variable; a downstream process; and interventions that can alter the allocation while other relevant factors are controlled. Machine and biological attention instantiate this contract differently. A valid comparison predicts the direction of change under matched interventions, not equality of components.
| Contract field | Transformer head | Biological selection | What may be compared | What must not transfer |
|---|---|---|---|---|
| Candidates | Key-value positions or features | Locations, features, objects or actions | More than one potential influence | A token is not automatically a perceived object |
| Context | Query, mask, layer state and learned parameters | Task, salience, expectation, bodily state and control signals | Context changes relative priority | A query vector is not an intention or a subject's act |
| Allocation | Weights and routed value contribution | Gain, synchrony, competition, sampling and access | Differential influence can be measured | One coefficient is not the whole biological phenomenon |
| Consequence | Changed representation and later output | Changed neural response, detection, action or report | Intervention can test causal dependence | Similar behaviour does not identify one mechanism |
Compare intervention signatures, not numerical units
Machine weights, firing rates, response gain and behavioural accuracy live in different measurement spaces. Their absolute values cannot be lined up honestly. The transferable object is a pattern of counterfactual dependence. Let a selection episode be written as S = ⟨C, b, a, y⟩. C is the candidate set, b the context or bias, a the allocation state and y the downstream variable. A bounded comparison asks whether interventions on b and a preserve an ordering such as:
Suppose a synthetic language model test moves a candidate's routed share from 0.30 to 0.65 after a cue, and patching out that routed contribution lowers the correct-answer probability by twelve percentage points. A biological task might move detection from 72 to 86 per cent under a spatial cue, while perturbing a control pathway removes most of the benefit. The magnitudes are not commensurable. The common result is that context changes relative allocation and an independent intervention changes the candidate's downstream effect in the predicted direction.
This stricter criterion excludes several false friends. A cue and an outcome can covary because both are driven by a third state. An attention map can change while residual paths preserve the answer. Behaviour can improve through decision criterion or motor preparation rather than sensory selection. In each case the descriptive pattern remains, but the proposed bridge loses causal specificity. Mediation tests, matched controls and restoration experiments are therefore more valuable than a higher correlation between two visualisations.
The tuple does not make the mechanisms identical
The notation deliberately omits substrate, learning history, bodily stakes and phenomenal character. It is a comparison interface, not a common ontology. Passing it permits the claim that two systems implement a tested selection relation. It does not permit component-by-component identity, shared purpose or shared experience.
A system with unlimited downstream capacity
Imagine a biological organism able to process every visual candidate fully and act on all of them without conflict. Biased selection would lose much of its functional pressure. Now imagine a transformer with enough compute to mix every value at every position. It still uses normalised weights because that is how its layer is defined. The comparison therefore cannot rest only on “limited resources”.
What survives is narrower: both systems can implement context-sensitive differential influence. The biological phenomenon is embedded in organismal scarcity and action; the machine computation can persist as an architectural operation even when the functional story about scarcity is weakened. A shared relation need not have a shared reason for existing.
The object survives the swap
What directedness adds
Intentionality is often reduced to a thin arrow from a subject to an object. That arrow hides the work. Consider the cup again. You see only one profile at a time, yet the cup is given as having unseen sides. You may walk around it and experience successive appearances as appearances of the same thing. You may misjudge its material, discover the error and correct the judgement while retaining the object. You may later imagine the cup without taking it to be perceptually present.
These relations are not captured by saying that cup-related features received high weight. The same object can remain intended while salience moves from colour to handle to weight. Conversely, the same salient colour can belong to a different object. Husserl's noesis-noema vocabulary is one way of marking the correlation between an act and the object as intended, while preserving differences among perceiving, remembering, judging, doubting and wishing. The noema is not a key vector waiting to be scored, and the noesis is not a query vector.
Accessible reading: a cup occupies the centre, surrounded by four profiles and four act modes. A dashed horizon ring indicates anticipated but unseen aspects.
A compact notation can keep these obligations visible without pretending to explain them. Write an intentional structure as I = ⟨m, o, p, h, c⟩. Here m is act mode, such as perceiving or imagining. o is the object as intended. p is the current profile or content. h is the horizon of further possible presentation. c is a correctness or fulfilment condition. The tuple is a ledger, not a metaphysical reduction.
Object constancy is more than stored similarity
Imagine two factory cups produced from the same mould. They have the same colour, dimensions and logo. One is yours; the other belongs to a colleague. A feature embedding can make them nearly indistinguishable. Yet a promise to return your cup concerns a particular history. The relevant relation may depend on continuous tracking, location, possession, a labelled record or a chain of interaction. Similarity helps recognition, but it does not by itself fix reference.
Now duplicate the digital record. Both copies contain the same description and the same photograph. Only one remains connected to the live object through a sensor stream and effect receipt. A system that follows the photograph may answer fluently while acting on the wrong item. A system that follows the live identity key can tolerate changed lighting, a new handle or temporary occlusion. This is why the feature-swap test is not a decorative puzzle. It asks whether the proposed target is constituted by current resemblance or by a relation that survives changing profiles.
There is also a boundary lesson. If an external tracker supplies stable identity while the model merely converts selected evidence into language, then object constancy belongs to the configured system, not to the model weights or one attention head. Removing the tracker should destroy identity-sensitive performance while leaving local fluency partly intact. Removing the model may destroy interpretation while the world-state record remains correct. Reference can be distributed across state, perception, action and evidence even when the final sentence is produced in one place.
This functional minimum still falls short of phenomenal intentionality. A database row can preserve an identifier; a controller can maintain a target; a person can experience one cup through changing appearances. The first two show reference-supporting organisation. The third includes first-person givenness on many theories. The protocol therefore grades the claim instead of erasing the difference.
Perceiving the cup and imagining the cup
Keep the object description constant: “the chipped red cup on my desk”. In perception, present sensory evidence can constrain position, lighting and reachability. In imagination, the same object may be intended while no present sensory profile is taken as its occurrence. A system that stores the same embedding in both cases but lacks a mode tag may produce identical verbal content and incompatible action.
Adding a typed mode can improve functional discipline. It can prevent an imagined object from authorising a physical action. That is a meaningful engineering gain. It supports a claim about mode-sensitive object-directed control, not a claim that the system enjoys imagined or perceptual experience.
The feature-swap break
Swap the colours but keep the cups
Two cups, A and B, sit on a table. A is red and left; B is blue and right. The task is “track cup A”. After a screen blank, A returns blue and right while B returns red and left. A feature selector follows red. A location selector follows left. An object tracker uses continuity cues and follows A.
Which one is intentional? The experiment cannot answer that metaphysical question. It can expose a structural minimum. A system that loses the target whenever a salient feature moves has not preserved the object relation that the comparison requires. Object constancy is a break test: failure blocks the move from selection to object-directed control.
Accessible reading: cup A begins red on the left and ends blue on the right. Feature and location selectors switch to cup B, while an identity tracker stays with A.
Why the tempting Q, K, V translation fails
A popular comparison maps a transformer's query to noesis, keys to noema, values to sensory matter, and attention score to fulfilment. The mapping is memorable and under-specified. Query, key and value are learned projections of internal representations. The key is not the object as intended. The value is not automatically a raw sensory substrate. The query does not encode whether the system perceives, imagines, doubts or desires. A high dot product has no built-in correctness condition relative to an absent or misperceived object.
The comparison preserves one limited relation: a current state can be differentially related to candidate contents. It breaks at act quality, object identity, horizon, fulfilment and first-person givenness. The break is productive. It tells an engineer what must be added and tested before making an object-directed claim: persistent identity, typed mode, expectation, contradiction handling and correction. It also tells a philosopher what not to infer: implementing those functions would still leave phenomenal and metaphysical questions open.
Accessible reading: two systems both answer A. One routes by redness, the other tracks object A. Scrambling weights and swapping identity affect them differently.
Passing the structure does not choose a metaphysics
A physicalist can treat mode, object constancy and error correction as realised by physical processes. A biological naturalist can argue that living organisation and sensorimotor stakes remain necessary. A functionalist can ask whether the tested role is multiply realisable. An emergentist can place object-directedness at a configured-system level. Idealist or consciousness-primary views can treat selection as shaping the contents or manifestation of awareness rather than producing awareness. Neutral-monist, dual-aspect, panpsychist and process views can accept relational organisation while disagreeing about its underlying nature. The experiment types a structural claim; it does not adjudicate ontology.
Optional depth: how the metaphysical alternatives read the same positive result
Physicalism: a positive result supplies candidate physical-functional structure for object-directed cognition. Biological naturalism: the result may be functionally interesting but insufficient without the relevant living dynamics. Functionalism: preserving the full role across interventions strengthens a multiple-realisation claim. Emergentism: the relevant property may belong to the configured whole rather than any attention head.
Idealism or consciousness-primary philosophy: the mechanism can organise appearance and action while awareness remains ontologically prior. Neutral monism or dual-aspect theory: the same organisation may admit physical and experiential descriptions without reduction of either. Panpsychism: selection could integrate or structure basic experiential aspects rather than create them. Process philosophy: object-directedness may be an achieved stability across events, not a static representation. No reading is established by the laboratory result alone.
stateDiagram-v2 [*] --> Claim state "Selection answers “what<br/>influences processing now?”" as Claim state "Three questions prevent the slide" as Test state "Match relations, not nouns" as Revision Claim --> Test: expose mechanism Test --> Decision: evidence survives Test --> Revision: rival survives Revision --> Claim: narrow or rebuild Decision --> [*]
Make the bridge earn its claim
The practical contribution is a two-stage protocol. The first stage looks for shared landmarks of selection. The second tries to break the proposed move towards intentionality. The protocol does not produce a consciousness score. It produces a typed claim and a receipt showing which interventions the claim survived.
The landmark-and-break protocol
Landmark L1, candidate multiplicity: more than one item could influence the downstream process. L2, context sensitivity: a cue, task or state changes relative priority while candidates are held fixed. L3, differential allocation: priority is embodied in a measurable routing, gain, access or sampling difference. L4, causal consequence: intervention on that allocation changes an appropriate downstream variable.
Passing L1 to L4 supports a selection homology. It does not yet support object-directedness. Break test B1 swaps features and positions while preserving identity. B2 inserts interruption or occlusion. B3 holds the object constant while changing act mode, such as perceive versus imagine. B4 introduces a false expectation and asks whether the system can register error, preserve the object and revise the profile or judgement.
Turn the comparison into a discrimination experiment
A useful protocol must make rival organisations predict different failures. Use at least three matched systems: a feature router, an identity tracker and a mode-aware controller. Give them the same observations and require the same overt response. Cross two factors. In the cue-aligned condition, the target's colour and location remain stable. In the identity-conflict condition, those features move to a distractor. Separately, keep the stream continuous or insert an interruption that forces recovery from state.
Measure four quantities. Selection elasticity is the change in allocation after a cue. Identity retention is accuracy when features conflict with target lineage. Mode leakage is the rate at which imagined, predicted or retrieved content is treated as currently perceived fact. Correction latency is the time or number of steps needed to register contradiction, preserve the target and update the mistaken profile. These measures answer different questions and should not be averaged into one attention score.
Three systems can tie on the easy condition and separate under breaks
Suppose all three systems score about 92 per cent when colour, location and identity agree. After the feature swap, the feature router falls to 46 per cent, the identity tracker remains at 84 per cent and the mode-aware controller remains at 82 per cent. After occlusion, the tracker falls to 61 per cent unless its persistent state is restored. In an imagination-versus-perception trial, the first two authorise the same physical action in 28 per cent of imagined cases, while the mode-aware controller does so in 3 per cent. These values are illustrative, but the pattern shows why the breaks are diagnostic.
The easy condition establishes competence, not organisation. The identity conflict distinguishes current resemblance from target continuity. Occlusion tests whether continuity survives absent input. Mode leakage tests whether identical content is used under different commitments. Correction latency tests whether error updates a profile or silently replaces the object. A convincing study reports all four, including nulls and recovery paths.
Pre-register the expected dissociations and the lowering rule. If the identity tracker does not outperform the feature router under conflict, lower the claim to feature-sensitive selection. If mode tags do not alter action permission, lower the claim to content classification. If contradiction causes target replacement rather than profile correction, do not claim fulfilment-sensitive object control. A negative break test is informative because it locates the missing relation.
Use the narrowest claim that survived
If L1 to L4 pass, say context-sensitive selection with causal downstream effects. If B1 and B2 also pass, say object-directed control candidate. If B3 and B4 pass under typed modes and correctness conditions, say intentional-structure candidate. Do not infer phenomenal awareness, subjectivity or moral status from any tier.
Accessible reading: a dashed boundary encloses a human goal, retriever, prompt compiler, model attention, tool policy and readback. Arrows show that target identity can be maintained outside model weights.
An incident-review assistant tracks a case, not a token
A review assistant receives a case identifier, retrieves an authorised timeline, asks a model to summarise the failure and drafts a proposed remediation. A heatmap over the narrative may highlight the phrase “payment timeout”. Yet case identity is maintained by the workflow key and evidence ledger. The model's attention may route relevant text, while the configured system's object-directed control comes from typed state, provenance and readback.
Remove the case key but keep the narrative. The model may still produce a plausible summary and attach it to the wrong incident. Keep the key but perturb one attention head. The system may recover through redundant representations. The two interventions reveal different ownership. The right claim is distributed: internal routing occurred in the model; persistent reference belonged to the configured system.
Accessible reading: a staircase rises through five claim levels. Four gates interrupt the ascent, and a red shortcut attempting to bypass them falls away.
Object-directed controllers will show a double dissociation
A controller built around persistent object identity should resist feature swaps better than a pure similarity router. A similarity router should remain more sensitive to query-key perturbations than the object controller when external state preserves identity. Observing that double dissociation would strengthen the proposed separation between selection and object-directed control. Failure to observe it under well-matched tasks would weaken the framework or show that the supposed object controller is still feature-bound. A rival explanation is ordinary redundancy: either system may appear invariant because another component silently restores the answer.
Landmark-and-break assessor
Configure a toy architecture, run the six tests and receive the narrowest claim permitted. The controls are not a consciousness scale. They expose which assumptions are required to move from routing to object-directed and mode-sensitive control.
| Test | Result | Meaning |
|---|
A positive highest-tier result permits the phrase “intentional-structure candidate” for the tested functional organisation. It cannot establish consciousness, experience, subjectivity or moral status.
Accessible reading: selection homology requires four landmarks, object-directed control adds identity and persistence, and intentional structure adds mode and error. Awareness is outside the matrix.
How to run the protocol in research
Begin with a preregistered mapping. Name the candidate system boundary, the selected variable, the proposed object and the downstream consequence. Build at least one matched rival that can produce the same output through feature matching, lookup or external state. Then apply cue changes, routing interventions, feature swaps, occlusion, mode changes and contradiction. Record nulls rather than redefining the mapping after seeing results.
A positive selection result permits a mechanistic comparison at the declared grain. A negative result may mean the routing is epiphenomenal, the intervention missed the actual pathway, or the task did not require selection. A positive object-constancy result permits investigation of reference and mode. A negative identity result blocks an intentionality analogy at that boundary, even if performance remains high. No positive result establishes experience. No negative result proves its absence.
Technical depth: minimal typed protocol and Python reference checker
The smallest useful implementation needs a candidate manifest, a set of interventions, expected qualitative outcomes and an allowed-claim function. The browser lab above implements that function. The Python version below can be placed in a test repository and extended with measured values.
from dataclasses import dataclass
from typing import Literal
Claim = Literal[
"descriptive resemblance",
"selection homology",
"object-directed control candidate",
"intentional-structure candidate",
]
@dataclass(frozen=True)
class Evidence:
multiple_candidates: bool
context_sensitive: bool
measurable_allocation: bool
causal_consequence: bool
identity_under_swap: bool
persistence_under_occlusion: bool
typed_mode: bool
error_preserves_object: bool
def allowed_claim(e: Evidence) -> Claim:
landmarks = (
e.multiple_candidates
and e.context_sensitive
and e.measurable_allocation
and e.causal_consequence
)
if not landmarks:
return "descriptive resemblance"
if not (e.identity_under_swap and e.persistence_under_occlusion):
return "selection homology"
if not (e.typed_mode and e.error_preserves_object):
return "object-directed control candidate"
return "intentional-structure candidate"
# Invariant: no return value licenses consciousness or phenomenal awareness.sequenceDiagram participant Q as Question participant M as Mechanism participant E as Evidence participant D as Decision Q->>M: Three spotlights, one red cup M->>E: Selection answers “what influences processing now?” E-->>M: compare rival explanations E->>D: Intervene on selection before naming intentionality D-->>Q: revise scope or act
Intervene on selection before naming intentionality
Stop asking whether a system “has attention” as though the phrase named one property. Ask which candidates competed, what supplied the bias, where allocation was implemented, and which downstream variable changed when allocation was intervened upon. That question can be answered for a transformer head, a cortical circuit or a configured agent system without importing a theory of experience.
When the claim concerns intentionality, add the harder tests. Swap features while preserving identity. Interrupt the stream. Hold the object constant while changing the mode from perceive to imagine or judge. Introduce a false expectation and require correction without losing the target. A system that passes these tests has earned a stronger functional description. It has not earned a declaration of consciousness.
The practical decision is to treat attention as a causal selection claim and intentionality as an object-preservation claim with mode and correctness obligations. This changes model interpretation, experiment design and public language. Heatmaps become hypotheses for intervention. Biological analogies become relation-specific. Phenomenological comparisons become sources of break tests rather than decorative translations.
The red cup no longer tempts us to collapse three spotlights into one. It becomes an instrument. Change its colour, hide it, imagine it, misjudge it and move the system boundary. What survives tells us exactly how far the bridge goes.
Glossary
- Attention weight
- A coefficient controlling the contribution of a value representation inside a specified attention operation.
- Biased competition
- An account in which simultaneously represented stimuli compete and task or stimulus factors bias their influence.
- Intentionality
- The directedness or aboutness of a mental act or state towards an object, property or state of affairs.
- Act mode
- The manner in which an object is intended, such as perceiving, remembering, imagining, judging, doubting or desiring.
- Horizon
- The implicit range of further profiles, expectations and possible fulfilments associated with an intended object.
- Object constancy
- Preserving target identity across changes in feature, location, viewpoint, interruption or partial absence.
- Selection homology
- A bounded comparison in which context-sensitive differential influence and its causal consequence survive matched tests.
References
- Bahdanau, D., Cho, K. and Bengio, Y. (2015). Neural machine translation by jointly learning to align and translate. ICLR. Open paper.
- Vaswani, A. et al. (2017). Attention is all you need. Advances in Neural Information Processing Systems. Open paper.
- Jain, S. and Wallace, B. C. (2019). Attention is not explanation. NAACL-HLT. ACL Anthology.
- Wiegreffe, S. and Pinter, Y. (2019). Attention is not not explanation. EMNLP-IJCNLP. ACL Anthology.
- Serrano, S. and Smith, N. A. (2019). Is attention interpretable? ACL. ACL Anthology.
- Abnar, S. and Zuidema, W. (2020). Quantifying attention flow in transformers. ACL. ACL Anthology.
- Desimone, R. and Duncan, J. (1995). Neural mechanisms of selective visual attention. Annual Review of Neuroscience, 18, 193–222. PubMed record.
- Reynolds, J. H. and Heeger, D. J. (2009). The normalization model of attention. Neuron, 61, 168–185. Open full text.
- Moore, T. and Armstrong, K. M. (2003). Selective gating of visual signals by microstimulation of frontal cortex. Nature, 421, 370–373. Publisher page.
- Egly, R., Driver, J. and Rafal, R. D. (1994). Shifting visual attention between objects and locations: evidence from normal and parietal lesion subjects. Journal of Experimental Psychology: General, 123, 161–177. PubMed record.
- Koch, C. and Tsuchiya, N. (2007). Attention and consciousness: two distinct brain processes. Trends in Cognitive Sciences, 11, 16–22. PubMed record.
- Wu, W. (2011). Attention as selection for action. In Mole, C., Smithies, D. and Wu, W. (eds), Attention: Philosophical and Psychological Essays. Open manuscript record.
- Watzl, S. (2017). Structuring Mind: The Nature of Attention and How It Shapes Consciousness. Oxford University Press. Publisher page.
- Brentano, F. (1874/1973). Psychology from an Empirical Standpoint. Internet Archive record.
- Husserl, E. (1913/1982). Ideas Pertaining to a Pure Phenomenology and to a Phenomenological Philosophy, First Book, trans. F. Kersten. Internet Archive record.
- Jacob, P. (2024 revision). Intentionality. Stanford Encyclopedia of Philosophy. Reference entry.
- Zahavi, D. (2025 revision). Edmund Husserl. Stanford Encyclopedia of Philosophy. Reference entry.