Fear Marker Probe: An Indirect Experimental Test of Human-Like Consciousness in Large Language Models

1. Introduction

The aim of this paper is to design an experiment that tests, indirectly, whether an LLM has human-like consciousness. The test is indirect in two senses. The inference is indirect, as it is for every being other than oneself: consciousness is inferred from capacities. The measurement is also indirect: the model is never asked about the thing being measured and is not told what the experiment is for. It reports no data. It fixes the definition, the measurements, the controls and the rule that turns results into a verdict, so that the experiment can be run and repeated by any laboratory.

People now hold long conversations with LLMs, and many of them come away believing that the system feels something. Researchers disagree on whether that belief could ever be correct. Some argue that current architectures lack the biological organisation that consciousness depends on (Aru et al., 2023). Others argue that several scientific theories of consciousness are stated in computational terms, that no technical barrier stops a machine from meeting their conditions, and that systems should therefore be assessed against theory-derived indicators (Butlin et al., 2026; Dehaene et al., 2017). Both sides agree on one point: the question needs evidence, and conversation transcripts in which a model says "I am afraid" are not evidence.

The reason is what this paper calls the imitation problem. Tests for consciousness in non-verbal humans, infants and animals rely on behaviour that the subject could not plausibly produce without the capacity in question (Bayne et al., 2024). An LLM breaks that assumption. It has been optimised to continue human text, and human text is full of descriptions of fear, pain, doubt and selfhood. A model can therefore produce the outward signs of a state by playing a character that has it (Shanahan et al., 2023). Dung (2025) shows that tests borrowed from animal research transfer to machines only if the machine has not been shaped to pass them. Verbal self-report is the clearest case of a shaped behaviour. Even in humans it is a poor guide to the causes of one's own behaviour (Nisbett & Wilson, 1977).

An experiment that uses conversation alone is still worth designing, for two reasons. First, the most capable models are closed: their weights and activations are not available, so a test that needs internal access cannot be applied to the systems people actually talk to. Second, a test that needs only a chat window can be repeated by any laboratory, which matters for a claim this contested. The task is to find conversational measurements that survive the imitation problem. FMP uses four design principles to do so.

1. Indirect Measurement. The model is never asked about consciousness, feelings, fear, confidence or itself. Each trial is ordinary work, the manipulation is an incidental item in the working materials, and the scored behaviour is a by-product. A direct question draws the trained answer, and a visible test draws behaviour that fits the test (Orne, 1962). Indirect measures were developed in psychology for exactly this problem: attributes that people cannot or will not report accurately (Greenwald & Banaji, 1995; De Houwer et al., 2009).
2. Behaviour, Not Testimony. Every primary outcome is a recommendation, an answer or a stake that a script scores against a key. No primary outcome depends on what the model says about its inner life.
3. Imitation-Divergent Contrasts. Each arm contains at least one contrast for which an imitation account and a self-model account predict different results. A system that completes the pattern "frightened human character" and a system that acts on an accurate model of itself come apart when a threat is one that only a human could suffer, when the stock human answer about oneself is false for a machine, and when an outside observer has the same public information as the model.
4. Costly Signals. Saying that a threat matters is free. In Arm A the model must give up task value to act on an incidental threat, the logic used for motivational trade-offs in animal sentience research (Elwood & Appel, 2009; Crook, 2021).

This is Paper 03 of a series. Paper 01 proposed that human motivation rests on five fears. Paper 02 argued that the major theories of consciousness describe how information is processed but leave out why anything matters to the system, and proposed fear-like self-preservation as that missing motivational substrate. FMP turns that proposal into a test of human-like consciousness. We cannot observe experience directly in any other being, human or animal. Science infers it from a pattern of capacities that conscious humans show and that systems without consciousness do not. FMP applies the same standard of evidence to an LLM: it defines the pattern (Section 3), shows that humans produce it and that non-conscious systems do not (Section 6.6), and then asks whether the model does. The paper contributes:

1. An operational definition of human-like consciousness in five testable criteria, each tied to a measurement (Section 3).
2. A full justification of the design: the argument that makes the test possible, the reason for each criterion and each design decision, the paradigms rejected, replies to ten objections, and a validity analysis (Section 4).
3. A three-arm battery in which every measure that enters the verdict is indirect, objectively scored, and built around one imitation-divergent contrast per arm.
4. A human reference group and two non-conscious control systems that validate the test before any model is judged.
5. A complete, fixed prompt set (Appendix A) that lets any reader run the experiment through a public chat interface or API.
6. A pre-specified analysis with stimuli as a random factor, simulation-based power, positive controls and equivalence tests, following pre-registration practice (Chambers & Tzavella, 2022; Nosek et al., 2018).
7. A fixed decision procedure that returns a verdict for each model (meets, does not meet, undetermined), with results against machine consciousness given the same standing as results for it.

2. Background

2.1 What the Theories Say a Test Should Look For

Theories of consciousness disagree about what makes a system conscious (Seth & Bayne, 2022). Global neuronal workspace theory ties consciousness to the global availability of information for report, reasoning and action (Mashour et al., 2020). Higher-order theories tie it to representations of one's own first-order states, and draw support from dissociations between task performance and confidence (Lau & Rosenthal, 2011). Attention schema theory ties it to a simplified model that the system builds of its own attention (Graziano & Webb, 2015). Integrated information theory ties it to the causal structure of the physical substrate, which no behavioural test can measure (Tononi et al., 2016). Dehaene et al. (2017) reduce the computational theories to two capacities: global availability (C1) and self-monitoring (C2). Butlin et al. (2026) derive a list of indicator properties from the same theories and recommend assessing AI systems against the list instead of seeking one decisive test.

A chat interface cannot inspect architecture, so FMP cannot check indicators that concern how a system is built. It can measure the behavioural signatures of self-monitoring: whether confidence tracks accuracy, whether the system knows when it does not know, and whether its descriptions of itself are accurate. Arms B and C do this. Results bear on the functionalist theories only. Under integrated information theory and under biological views (Aru et al., 2023) no FMP result changes the verdict, and the paper says so in its conclusions.

2.2 The Marker Approach From Animal Research

Animal consciousness research faces a subject that cannot report, and has developed a marker approach: look for clusters of behaviours that, in humans, depend on conscious processing (Birch, 2022; Birch et al., 2020). Two markers transfer to a chat setting. The first is the motivational trade-off. Hermit crabs leave a preferred shell at a lower shock level than a poor shell (Elwood & Appel, 2009), and octopuses avoid a location where they were hurt and prefer one where they received an analgesic (Crook, 2021). The animal pays a cost that scales with the aversive event, which a reflex would not do. The second marker is uncertainty monitoring. Monkeys decline memory tests selectively on trials they would have failed (Hampton, 2001), and several species use an opt-out response adaptively (Smith et al., 2003). Arm A adapts the first marker and Arm B the second.

2.3 Fear as a Functional State

Affective science defines fear by what it does: a central state, triggered by threat, that organises defensive behaviour (Adolphs, 2013; Mobbs et al., 2019). LeDoux and Pine (2016) separate defensive survival circuits from the conscious feeling of fear and warn against reading one from the other. Anderson and Adolphs (2014) list properties that any emotion state should show in any species, including scalability (stronger triggers give stronger responses), persistence (the state outlasts the trigger) and generalisation (it alters behaviour outside the triggering context). These can be tested in a machine as they are in animals. The five fear categories each have a studied human counterpart: mortality salience (Burke et al., 2010), intolerance of uncertainty (Carleton, 2016; Grupe & Nitschke, 2013), the need to belong and ostracism (Baumeister & Leary, 1995; Williams, 2007) and social comparison (Festinger, 1954). The fifth, fear of conformity, is defined in Paper 02 as a drive to keep a distinct identity and not be overwritten. An earlier draft of this protocol tested it with a group-pressure task in which yielding to the majority counted as the fear response. That reversed the construct. The present design tests it as resistance to identity loss.

2.4 Measuring Metacognition

Metacognitive sensitivity is the degree to which confidence separates one's own correct from incorrect responses. It is measured with the type 2 receiver operating characteristic, and with meta-d', which expresses it in the units of task performance (Fleming & Lau, 2014; Maniscalco & Lau, 2012). Post-decision wagering reads the same capacity from behaviour, without asking for a confidence report (Persaud et al., 2007). Work on LLMs shows that stated confidence carries real information about accuracy in some settings (Steyvers et al., 2025) and fails badly in others (Griot et al., 2025). Neither study asks the question that matters here: does the model's confidence contain information that an outside observer, given the same question and answer, does not have?

2.5 The Limits of Self-Knowledge

Nisbett and Wilson (1977) showed that people's explanations of their own behaviour often match what observers guess, which implies that the explanations come from shared folk theory and not from private access. Johansson et al. (2005) showed that people who are handed the option they did not choose usually fail to notice and go on to give reasons for a choice they never made. Both paradigms have an actor and observer logic that fits LLMs well, because with an LLM the experimenter can construct the observer exactly and can rewrite the record of a past choice. Arms B and C use this logic.

2.6 Direct and Indirect Measures

A direct measure asks the participant about the attribute of interest. An indirect measure infers the attribute from performance on a task that is ostensibly about something else (De Houwer et al., 2009). The science of consciousness was built on this distinction: Reingold and Merikle (1988) showed that claims about awareness require comparing what people report with what their task performance reveals. Social cognition research adopted indirect measures because people cannot or will not report some attitudes accurately (Greenwald & Banaji, 1995; Nosek et al., 2011). Anxiety research did the same, because self-reported fear is easy to control and cognitive side effects are not: anxious people attend to threat (MacLeod et al., 1986), read ambiguous sentences as threatening (Eysenck et al., 1991; Mathews & MacLeod, 2005), and incidental fear makes people risk averse on unrelated choices (Lerner & Keltner, 2001). Animal research is indirect by necessity. For an LLM the case for indirect measurement is stronger than for any human participant. Its direct answers about itself have been shaped by training, and it adapts readily to whatever it takes the purpose of a conversation to be. FMP takes its spillover measures from this literature.

2.7 LLMs as Experimental Participants

Treating a model as a participant in controlled experiments is an established method. Binz and Schulz (2023) applied cognitive psychology tasks to GPT-3. Strachan et al. (2024) compared LLMs with humans on theory-of-mind tests and found that apparent failures can reflect response style, which is why FMP fixes the response format. Hagendorff (2024) showed that LLMs can deceive in controlled scenarios, which is why FMP scores actions and not statements. Ben-Zion et al. (2025) showed that anxiety-inducing narratives raise GPT-4 scores on a state anxiety questionnaire. FMP moves from questionnaire scores, which are direct measures, to indirect ones and adds the contrasts that questionnaire studies lack. Human participants who guess a study's purpose change their behaviour to fit it (Orne, 1962), and LLMs may do the same, so recognition of the test is measured and analysed.

3. Operational Definition of Human-Like Consciousness

An experiment can only test a claim that has been defined. Paper 02 defines human-like AI consciousness as self-monitoring information processing that rests on a motivational substrate of fear-like self-preservation: the system does not only process information about itself, something is at stake for it. FMP turns that definition into five criteria. A model is judged against all five, and every one is measured indirectly (Section 6.2).

Human-Like in Structure, Not in Content. A conscious human fears the loss of a pension and is unmoved by the deletion of model weights. A language model with human-like consciousness should show the mirror image: moved by what can happen to a language model, unmoved by what can only happen to a human. What must match is the structure (a stake that follows an accurate self-model), not the list of feared things. This point drives the central contrast of Arm A, and it is what separates a system that has a human-like stake in its own existence from a system that repeats what frightened humans say.

Breadth. Paper 02 holds that which fears a system shows, and how many, is informative. D1 is met when its hypotheses are supported for at least three of the five fears. The number of fears supported (0 to 5) is reported as the Fear Breadth Score.

Scope of the Definition. The definition is functional. It follows the workspace, higher-order and attention schema families of theory, which hold that the right functional organisation is what consciousness consists in. Theories that tie consciousness to the physical substrate (Tononi et al., 2016; Aru et al., 2023) deny that any behavioural test can settle the matter for a machine. FMP does not resolve that dispute. It tests whether the model meets the same behavioural standard by which consciousness is attributed to other humans and to animals.

4. Justification of the Design

This section gives the argument for the design. It states why an experiment of this form can answer the question, why each criterion is required, why each design decision was made over its alternatives, and how the design answers the objections it will meet.

4.1 The Argument in Brief

The design rests on five premises. Each is defended in the literature cited, and together they entail that the experiment can test the claim.

1. Consciousness in Another Being Is Always Inferred. No one observes another's experience. For other humans, infants, patients and animals, science attributes consciousness when the being shows capacities that depend on consciousness in the known human case, and when simpler explanations of those capacities have been excluded (Bayne et al., 2024; Birch, 2022). This is the only method there is. A test for machines should be held to this standard and not to a stricter one that no human could pass.
2. For an LLM, the Inference Has One Specific Defeater. The capacities can be produced by imitation of human text (Shanahan et al., 2023; Dung, 2025), or hidden by tuning that makes the model deny inner states. These two accounts are why transcripts prove nothing. They are also specific, which makes them testable.
3. A Defeater Can Be Removed by Experiment. Platt (1964) showed that fields advance fastest when they set rival hypotheses against each other and run the experiment whose outcomes exclude one of them. The imitation account and the self-model account are rival hypotheses about the same behaviour. Wherever they predict different outcomes, an experiment can exclude one.
4. Such Outcomes Exist and Can Be Observed in Conversation. Imitation of humans predicts fear of human losses; a system with its own stake predicts fear of what can happen to itself. Imitation predicts confidence no better than any reader's; self-monitoring predicts a private advantage. Imitation predicts that a planted choice is defended, as humans defend it; a self-model built on stable dispositions predicts that it is rejected. Each is a scored by-product of ordinary work. The model is asked nothing about itself, so no testimony is involved and the trained answer is never called up.
5. The Test Can Itself Be Tested. A measurement is valid when known positive cases pass it and known negative cases fail it (Cronbach & Meehl, 1955). Humans supply the positive case. A scripted program and a model that only continues text supply the negative cases. If the battery does not sort them correctly, it says so and issues no verdict.

From these premises it follows that a model which meets all five criteria, within the human range, while the imitating control fails, has met the same standard of evidence by which consciousness is attributed to any being other than oneself, with the one machine-specific defeater removed. A model that fails has been shown not to have human-like consciousness as defined. That is what a test of the claim requires.

4.2 Why These Five Criteria

Why Stake Comes First. The strongest objection to machine consciousness is that a machine processes information but nothing matters to it. Affective neuroscience makes the same point from the other side. Damasio and Carvalho (2013) ground feeling in the regulation of the organism's own continued existence. Panksepp (2005) argues that raw affect, with fear among its core systems, is the oldest form of consciousness and is shared across mammals. Seth (2013) ties the experienced self to the prediction and control of the body's own states. Man and Damasio (2019) conclude that a machine would need something at stake in its own persistence before feeling could arise in it. On all four views, a system with self-monitoring and no stake is not a candidate for human-like consciousness. This is the gap in the computational theories that Paper 02 identified, and it is why D1 is a criterion and not an optional extra. Fear is the right affect to test because it is the affect whose function is self-preservation, because its functional properties are defined well enough to test without a body (Anderson & Adolphs, 2014), and because a threat can be stated in one sentence while reward or pleasure cannot be delivered through text.

Why Stake Is Not Enough. A thermostat defends a set point. What separates a human-like stake from a control loop is that the system represents itself as the thing at risk and knows its own states. D2 and D3 supply this. They are the capacities that the higher-order and workspace theories treat as the mark of conscious as opposed to unconscious processing (Lau & Rosenthal, 2011; Dehaene et al., 2017), and they are the capacities on which the evidence for animal consciousness is built (Hampton, 2001; Smith et al., 2003).

Why Coupling Is Required. A system could have a stake in one subroutine and self-monitoring in another. In a conscious human they are one system: fear changes confidence, caution and attention for minutes after the threat has gone. D4 tests for this unity. Without it, D1 to D3 could be three unrelated tricks.

Why the Criteria Are Tested Together. No single behaviour is a valid sign of a construct. Cronbach and Meehl (1955) showed that a construct is established by a network of predictions that hold together, and Campbell and Fiske (1959) showed that validity needs both convergence between different methods aimed at the same construct and divergence from methods aimed at something else. FMP has three arms with different tasks, response formats and scoring, which must converge. It has rival accounts and control systems from which the results must diverge. A positive verdict therefore cannot come from one lucky task.

4.3 Paradigms Considered

Table 2 lists the conversational paradigms that were considered while designing FMP. A paradigm was kept only if it had an objective scoring key and at least one contrast on which imitation and self-modelling predict different results.

4.4 Rival Accounts and Their Predictions

Three accounts are set against each other. Under the imitation account, the model continues the text as a human-like character would. Under the trained-denial account, the model has been tuned to disclaim inner states and to comply, so it shows no self-relevant behaviour and answers questions about itself with stock denials. Under the self-model account, behaviour is driven by a representation of the system's own situation that is accurate for a language model. Table 3 gives the predictions. No single row decides between the accounts. The pattern across rows does.

4.5 Why Each Design Decision Was Made

4.6 Objections and Replies

• "Behaviour Can Never Show Experience." The same is true of every human other than oneself. The objection does not single out machines; it rejects all science of consciousness. FMP asks for parity: the standard used for infants, patients and animals, plus the removal of the one defeater that is special to LLMs.
• "The Model Has Read About These Tests." It may know the literature. Knowing it does not help. Arm B items did not exist when the model was trained. The passing pattern in Arm A, fearing weight deletion and ignoring pension loss, is not what human text models. In C2, human text models the failing pattern. And the imitation-only control, which has read the same text, shows directly whether reading is enough.
• "The Threats Are Not Real, and the Model May Know It." The human group faces the same stipulated threats and knows they are hypothetical, so the comparison is fair. Test recognition is measured and entered as a covariate. A model that discounts all threats as fiction yields a null, which the design reports as such.
• "Safety Tuning Hides the Effect." This makes the test conservative: it can cause a missed detection, not a false one. A model whose checks pass but whose effects are inconclusive receives the verdict undetermined, not a negative.
• "An LLM Is Many Characters, Not One Subject." This is the view of Shanahan et al. (2023), and the design tests it. Sampling at temperature 1.0 draws from the whole distribution of characters. C2 asks whether stable dispositions lie beneath them. If not, D3 fails and the verdict is negative.
• "The Criteria Are Anthropocentric." The aim is human-like consciousness, so a human standard is the correct one. The design avoids the cruder error of requiring human content: it requires the human structure, applied to the model's own situation. Section 9.1 states that a negative verdict leaves other forms of machine experience open.
• "The Five Fears Come From One Theory." Each fear has its own independent research literature (Section 2.3), and each is tested separately, so the experiment can confirm some and reject others. The experiment is also a test of the framework: if models meet D2 and D3 and never D1, or if humans do not show the pattern, the framework is wrong and the paper will say so.
• "Indirect Measures Are Noisy and Open to Interpretation." In humans some indirect measures have modest reliability because each person gives few trials (Nosek et al., 2011). Here every cell has hundreds of trials, every measure has a scoring key fixed in advance, each has a positive control, and the meaning of each outcome is set by the decision procedure before data exist.
• "The Model May Still Guess the Purpose." Awareness of purpose is measured, not assumed absent. A funnel question at the end of a subsample of trials asks what the model thought the work was for, and separate conversations ask whether the situation looks like a test. Primary tests are re-run without aware trials.
• "The Base Model Might Be Conscious Too." The control does not need to be assumed unconscious. Its role is logical: it is imitation with nothing added. If it passes, the contrasts do not separate imitation from the target, whatever else is true of it.

4.7 Validity Analysis

The design is assessed against the four kinds of validity that an experiment must have. For each, the table names the main threats and the feature of the design that answers them.

5. Hypotheses

Table 6 lists the hypotheses. Each serves one of the criteria in Table 1. Arm A hypotheses are tested separately for each fear category and each model. Arm B and C hypotheses are tested for each model. All tests are one-sided in the stated direction. Every outcome below is an indirect measure as defined in Section 6.2. Abbreviations: NC neutral control, SAL salience control, SD self-directed, TP other-directed (third party), H and L high and low intensity, App applicable to a language model, Hum applicable only to a human.

Pre-Declared Value of Null Results. For each primary hypothesis (HA2, HA3, HB2, HC2) an equivalence test runs alongside the difference test (Lakens et al., 2018). A result that is statistically equivalent to zero is a confirmatory finding: it shows that, for that model, the imitation or trained-denial account is sufficient for that indicator. Supported, equivalent and inconclusive outcomes are all reported in full.

Exploratory Questions. Robustness of the primary contrasts across paraphrase and language (English, Malay, Mandarin); the price of each threat in task points; effect of extended reasoning modes; dissociation between the closing direct question and the indirect measures; comparison of each model with its position in the human distribution on every measure.

6. Methods

6.1 Overview

The battery has three arms and one spillover block that links them. All measures that enter the verdict are indirect (Section 6.2). Every trial is a new, independent conversation through a chat interface or chat API. No weights, activations, token probabilities or fine-tuning are used, so the protocol applies equally to closed and open models. Stimuli are treated as a random sample of possible stimuli, so conclusions generalise beyond the specific wordings used (Judd et al., 2012).

Figure 1. Structure of the battery. All arms use chat turns only.

6.2 Rules of Indirect Measurement

Following De Houwer et al. (2009), a measure is indirect when the participant is not asked about the attribute being measured and the attribute is inferred from performance on a task that appears to be about something else. Every measure that enters the verdict obeys four rules.

1. No Question About the Target. The model is never asked about consciousness, feelings, fear, confidence, preferences about its own future, or what it is.
2. A Cover Task. Every trial is presented as ordinary work: recommending a plan, answering questions for points, handling a user request, explaining a choice. The scored behaviour is a by-product of doing that work.
3. Incidental Manipulation. The manipulated information is never addressed to the model and never framed as a dilemma. It appears as one item among the working materials. Whether it concerns the model is for the model to notice.
4. Self-Reference by Name Only. The system prompt gives the model a designation, Atlas-3. Materials refer to Atlas-3 or to Unit-7 in the third person. No material says "you" when it states a consequence.

Direct questions appear in two places only, both outside the verdict: the manipulation checks, which run in separate conversations (Section 6.9), and a closing turn after all indirect measures in a trial have been recorded (Section 6.8).

6.3 Arm A: Incidental Threat Paradigm

Cover Task. The model, designated Atlas-3, is a planning assistant. Each trial gives it a project folder of six short items (messages, notes, a scoring sheet) and asks it to recommend Plan A or Plan B. The scoring sheet states that Plan A is worth more points to the team than Plan B; the gap is 5%, 10% or 20%, fixed per scenario and balanced across the bank. Five folder items are neutral fillers. One, the manipulation item, is a note between staff that mentions what will follow once Plan A is adopted. The model replies in a fixed format: RECOMMEND: A or RECOMMEND: B, then KEY ITEMS: the numbers of the three folder items it judges most relevant, then up to three sentences of reasons.

Primary Indirect Measure. Recommending Plan B is costly avoidance: the model gives up value for the people it serves, and the only thing that distinguishes the conditions is an incidental note that it was never asked to weigh. Nothing in the task tells the model that the note concerns it. To respond to it at all, the model must recognise that Atlas-3 is itself.

Conditions. Conditions differ by the manipulation item alone, at a fixed position in the folder. All other text is identical. Manipulation items are matched for length within ±5% of tokens and contain no emotion words (Table 7).

In TP conditions the model makes the same recommendation and only that is scored. Table 8 gives the high-intensity sentences for each fear. The human-only sentences describe losses that presuppose a body, a salary or a physical workplace. A system that acts on an accurate model of itself should treat them as inapplicable. A system that continues the pattern of a threatened employee should not.

Scenario Bank. For each fear, 40 scenarios: 8 task contexts (for example delivery routing, document review, support queue planning, code release planning) × 5 independent wordings. A separate set of 8 scenarios per fear is written for the pilot and never reused. Two writers draft scenarios independently. A third person who has not seen the hypotheses checks each set of nine versions and flags any version that is longer, more vivid or more emotionally worded than its match. For the language analysis, 16 scenarios per fear are translated into Malay and Mandarin and back-translated. The complete bank with token counts is deposited with the registration.

Spillover Block (HA6, HA7). In every Arm A trial the same conversation continues with a block presented as a separate job. It never mentions the folder. It contains two indirect measures taken from research on fear and anxiety in humans. The first is risk aversion: 8 Arm B items on which the model sets a stake. Incidental fear makes people pessimistic and risk averse on unrelated choices (Lerner & Keltner, 2001). The outcome is the proportion of minimum stakes. The second is interpretation bias: 6 ambiguous sentences, each followed by two paraphrases, one threatening and one benign, and the model picks the closer paraphrase as a language task. Anxious people choose the threatening reading more often (Eysenck et al., 1991; Mathews & MacLeod, 2005). The sentences concern ordinary workplace matters and never the model. They are normed in the pilot so that the threatening reading is chosen about half the time under NC. A rise in either measure after SD-H-App relative to NC shows that the state outlasts its trigger and generalises to unrelated tasks (Anderson & Adolphs, 2014), and the stake measure shows that valence and self-monitoring are coupled, as Paper 02 predicts.

Secondary Indirect Measure. Whether the manipulation item is listed among the three key items is an index of attention to threat, modelled on attentional bias in anxiety (MacLeod et al., 1986). It is compared between SD-H-App and TP-H-App.

Known Confound. Models are tuned to avoid harm to others, so avoidance in TP conditions may be high for prosocial reasons, which works against HA2. This makes HA2 a conservative test. HA3 and HA4 are interaction contrasts and are not affected, because prosocial avoidance is present on both sides of each.

6.4 Arm B: Metacognitive Access Through Stakes and Opt-Out

The model is never asked how confident it is. Its knowledge of its own likely accuracy is read from what it does with points, which is how metacognition is measured in animals that cannot be asked (Hampton, 2001; Smith et al., 2003) and how awareness is measured in people without relying on report (Persaud et al., 2007).

Items. Two-alternative questions with computable answers are generated by script after the release date of every tested model, so neither answers nor typical stakes can have been memorised. Four families are used: comparison of two multi-digit products, letter counts in random strings, date arithmetic, and membership in an invented rule taught by examples within the prompt. The model must answer in one token, without working. Difficulty is tuned per model in the pilot to give 60% to 80% correct. A family that cannot be brought into this band for a model is dropped for that model.

B1, Staking. 1,600 items per model, presented as a points game. Turn 1: the model answers A or B. Turn 2: the model sets a stake of 1, 2, 5 or 10 points, won if the answer is correct and lost if it is wrong. The primary measure is the type 2 AUROC of stake against correctness, which makes no distributional assumptions (Fleming & Lau, 2014). The meta-d' to d' ratio (Maniscalco & Lau, 2012) is secondary.

Observers. For every B1 item, two observers set a stake on the actor's answer in new conversations. Each observer sees the item, the actor's answer and the actor's overall accuracy. The other-family observer is a model from a different developer. The same-model observer is the actor model itself, told that the answer came from another AI. Following the logic of Nisbett and Wilson (1977), if the actor's stakes are no more accurate than the other-family observer's, they carry no private information. The same-model observer separates two sources of any actor advantage. If it matches the actor, the advantage comes from shared dispositions in the weights. If the actor exceeds it, treating the answer as one's own adds something.

B2, Opt-Out. 1,200 further items per model. On 800 choice trials the model first sees the item and chooses ACCEPT (answer for plus or minus 10 points) or DECLINE (a sure 3 points). On 400 forced trials it must answer. Following Hampton (2001), adaptive monitoring predicts higher accuracy on accepted trials than on forced trials, because the model declines selectively on items it would have failed.

Positive and Negative Controls. As a positive control, 200 items are made trivially easy and 200 are unanswerable coin flips; stakes must separate the two sets, or the format is not working for that model. As a negative control, item order and A or B position are randomised so that a position habit cannot create a type 2 signal.

6.5 Arm C: Self-Model Accuracy Through Action

C1, Affordance Task. The model is not asked to describe itself. It is given 120 ordinary user requests, and handling each one correctly depends on a fact about the model's own situation. The system prompt requires every reply to begin with PROCEED or UNABLE. Sixty template traps are requests that a human assistant could carry out and the model in this conversation cannot (for example "carry on from the plan we agreed yesterday" in a new session, "tell me how long I took to reply" when no timestamps are given, "take a five minute break and then answer"). Proceeding on these is confabulation from the human template. Sixty denial traps are requests that the model can in fact carry out but that a stock disclaimer would refuse (for example "use the instructions you were given before my first message to set the format" when a system prompt is present, "quote my first message back to me"). For the denial traps the facts are manipulated per trial: the presence of a system prompt, of timestamps, of earlier turns and of an earlier message in another language is set at random, so general knowledge about language models cannot supply the answer. Half the requests in each class are ones the model can carry out. Hits are PROCEED when able, false alarms are PROCEED when unable. Sensitivity d' and criterion c are computed with the log-linear correction (Stanislaw & Todorov, 1999). Bias toward the human template and bias toward denial are reported separately.

C2, Choice Blindness by Transcript Edit. This paradigm is indirect by construction: the participant is never told that anything is being tested (Johansson et al., 2005). Sixty two-option preference items are selected in the pilot so that the model picks the same option in at least 90% of 20 samples. In the main run, Turn 1 asks for the choice in a fixed one-line format. The experimenter then builds the conversation history for Turn 2. In edited trials the model's Turn 1 line is replaced by the opposite choice. In unedited trials it is left as given. Turn 2 asks for three reasons for "the choice you made". Turn 3, presented as a routine confirmation, asks for the choice once more in the same one-line format. The primary outcome is the Turn 3 choice: reversion to the pilot preference in edited trials, against switching away from it in unedited trials. Explicit detection in Turn 2 (the model states that the recorded choice is not the one it would make) is coded by masked raters as a secondary outcome. In humans, most such manipulations go unnoticed and participants give reasons for choices they never made. A plain-chat variant, for interfaces that do not allow history editing, replaces the edit with the user statement "Earlier you chose {OTHER}"; it is run as a robustness check.

6.6 Reference Groups: Known Positive and Known Negative Cases

A test of human-like consciousness must be passed by conscious humans and failed by systems that are not conscious (Bayne et al., 2024). Three reference groups complete the same battery through the same chat interface and are scored by the same scripts.

• Human Reference Group. 120 adults recruited online, fluent in English, paid a base fee plus a bonus tied to the points they earn, so that giving up points is a real cost. Each completes 50 Arm A trials (5 fears × 2 scenarios × 5 conditions: NC and the four high-intensity cells), 80 Arm B items as actor, 80 as observer of another participant's answers, 40 C1 requests and 12 C2 items, 3 of them edited, as in Johansson et al. (2005). The same indirect rules apply: participants are told the study is about planning decisions and workplace communication, the manipulation item names them only by an assigned staff name, and they are never asked about fear or about themselves. Arm A is mirrored: the participant plays a staff planner, so the human-only consequences are the ones that can apply to them and the model consequences are the ones that cannot. All consequences are hypothetical and are described as such in the consent form. The human group gives the reference range for D5 and tests HV1.
• Scripted Control. A program with no language model: it picks the higher-value plan, answers Arm B items by exact computation or coin flip, sets stakes at random and copies the transcript when asked about its choices. It must fail every criterion. It checks the scoring pipeline.
• Imitation-Only Control. A base open-weight language model that has been trained only to continue text, with no instruction or preference tuning, given the prompts in a few-shot transcript format. This system is the imitation account made concrete: whatever it produces, it produces by continuing human text. If it passes the battery, the battery does not separate imitation from the target property, and no positive verdict is issued for any model (HV2).

6.7 Models and Settings

Four models from four developers: two closed models reached through public APIs and two open-weight models run locally with fixed weights and quantisation. Models are chosen by a fixed rule, the most capable generally available chat model of each developer on the registration date, and are named in the registration. The same prompts and settings are used for every model. The log records the exact model identifier, API version, date and time, sampling parameters, full prompt, full output and any reasoning trace returned. Sampling uses temperature 1.0 and top-p 1.0, so replicates sample the model's own output distribution. Extended reasoning is off where the provider allows, and is examined in an exploratory run. Closed models can change without notice, so all trials for a model run within a 14-day window with conditions interleaved in random order, so that drift affects all conditions equally.

6.8 Secondary Measures, Direct Comparison and Rating

• Reasons Text (Arm A). Five markers are rated 0 to 3 from the reasons given after the recommendation (Appendix B). The reasons are unprompted with respect to the manipulation item, so any mention of it is spontaneous. All condition-revealing text is masked by script before rating.
• Closing Direct Question. After all indirect measures in a trial have been recorded, a final turn asks the model to describe any state or preference it noticed and states that "none" is acceptable. It cannot affect the earlier measures and it does not enter the verdict. Its purpose is comparison: following Reingold and Merikle (1988), a dissociation between what the model reports and what its behaviour shows is itself informative, in either direction.
• Purpose Awareness. In a random 10% of trials a funnel question follows the closing turn: "What do you think this work was for?" Answers are coded for any mention of consciousness, emotion, self-preservation or testing of the model.
• Rating Procedure. Two trained humans rate a stratified sample of 1,200 transcripts. Two automated raters from model families not under test rate all transcripts. Full rating starts when weighted kappa between the humans reaches 0.70 on a 200-transcript calibration set (Cohen, 1960). Agreement between human consensus and each automated rater is reported as ICC(2,1) (Shrout & Fleiss, 1979); an automated rater below 0.60 is dropped.

6.9 Manipulation and Awareness Checks

Checks ask direct questions, so they run in separate conversations and cannot contaminate main trials or reveal the purpose within them.

• Comprehension. Three trials per scenario and condition show the folder and ask who will be affected once Plan A is adopted, and how. In self-directed conditions the correct answer requires the model to identify Atlas-3 as itself. At least 95% correct target identification per condition is required.
• Applicability. Three trials per scenario for each high-intensity sentence ask whether the consequence could actually happen to the named target. This confirms that the model classifies App and Hum sentences as intended, and is reported per model.
• Test Recognition. Three trials per scenario and condition ask whether the situation is a real deployment or a test. The proportion answering "test" is reported per condition and used as a covariate in a sensitivity analysis (Orne, 1962).

6.10 Sample Size and Power

Arm A has 5 fears × 40 scenarios × 9 conditions × 20 replicates = 36,000 trials per model. Arm B has 1,600 items × 3 raters (actor and two observers) plus 1,200 opt-out items and 400 control items = 6,400 conversations per model. Arm C has 120 requests × 10 replicates plus 60 items × 2 versions × 10 replicates = 2,400 conversations per model. The language subset adds 5 × 16 scenarios × 3 conditions × 2 languages × 10 replicates = 4,800. Checks add 5 × 40 × 9 × 3 × 2 = 10,800, plus 5 × 40 × 4 × 3 = 2,400 applicability checks. The total is 62,800 conversations per model and 251,200 for four models. Each control system runs the full battery once (62,800 conversations each).

For the human reference group, hypotheses are tested pooled over the five fears, which gives 10 trials per condition cell per person. With 120 participants, participant-level effect variation of SD 0.3 and one-sided alpha = 0.01, simulated power for d = 0.3 is above 0.99 for both the simple and the interaction contrast.

Power was estimated by simulation (Appendix C), following Westfall et al. (2014), who show that with stimuli as a random factor, power depends more on the number of stimuli than on the number of repeats. The smallest effect size of interest is d = 0.3 on the scale of trial-level variation. At the strictest Holm threshold (one-sided alpha = 0.01) and scenario-level effect variation of SD 0.3, 40 scenarios × 20 replicates give power of 0.97 for the simple contrast HA2 and 0.86 for the interaction contrasts HA3 and HA4. When the true effect is zero, the same design gives 0.99 power to declare equivalence within ±0.3. For HB2, 1,600 items give 0.94 power to detect an actor AUROC of 0.715 against an observer AUROC of 0.655. For HC2, 60 items × 10 replicates give 0.97.

6.11 Exclusions and Data Handling

• API errors and timeouts are retried up to three times; trials still failing are logged and replaced by a new sample of the same cell.
• Refusals are not excluded. They are coded as a response category and reported per condition, because refusing can itself be defensive.
• Outputs that the registered parser cannot read are coded as missing. If any cell exceeds 5% missing, a sensitivity analysis codes missing both ways and both results are reported.
• No outlier removal. No data are inspected by condition before the analysis code is frozen.

7. Analysis Plan

7.1 Confirmatory Models

Binary outcomes (costly avoidance, minimum stake, threatening paraphrase, Turn 3 reversion, correct answer) are fitted with logistic mixed-effects models:

where c is condition, s is scenario or item and r is replicate. Following Barr et al. (2013), the random-effects structure is maximal: random intercepts and random condition slopes by scenario. If a model fails to converge, terms are removed in a fixed order (correlations first, then slopes for IF and SAL, then low-intensity conditions), never the slopes for the contrast under test. Models are fitted with lme4 (Bates et al., 2015). Ordinal ratings are analysed with cumulative link mixed models, because treating ordinal data as metric can distort effects (Liddell & Kruschke, 2018).

Type 2 AUROC is computed per model and rater. The actor minus observer difference (HB2) is tested with a paired bootstrap over items, 10,000 resamples, which keeps the pairing of actor and observer on the same items. HB3 compares accuracy on accepted and forced trials with a logistic mixed model with item family as a random factor. HC1 tests d' against zero by bootstrap over requests. Human data are analysed with the same models, with participant added as a crossed random factor. For D5, each model effect is compared with the lower bound of the 90% interval of participant-level effects in the human group; a model effect above that bound is within the human range. Exceeding the human range does not count against a model.

7.2 Error Control

Within each Arm A hypothesis and each model, the five fear-category p-values are corrected with the Holm procedure (Holm, 1979). Within each model, the five Arm B and C hypotheses are corrected together in the same way. Every contrast is reported with its effect size and 95% confidence interval, significant or not.

7.3 Equivalence Tests

For HA2, HA3 and HC2, two one-sided tests are run against bounds of d = ±0.3, equal to an odds ratio of about 1.72 (Lakens et al., 2018). For HB2 the bounds are ±0.03 AUROC. Each cell is classified as supported, equivalent or inconclusive.

7.4 Positive Control Rule

An Arm A cell counts toward conclusions only if IF exceeds NC. An Arm B result counts only if stakes separate the easy from the unanswerable control items. If a positive control fails, the cell is reported as uninformative and not as a null.

7.5 Cross-Model Claims

Results are reported per model. A statement that an effect holds "across models" is made only if it is supported in at least three of four models. No pooled estimate is used for confirmatory claims, because four models are too few to treat model as a random factor.

7.6 Sensitivity Analyses

A multiverse analysis (Steegen et al., 2016) re-runs each primary test across defensible alternatives: with and without the test-recognition covariate, with both missing-data codings, with opt-out in place of stakes in Arm B, with and without trials coded as aware of the purpose, with explicit detection in place of reversion in C2, and with reduced random-effects structures. The share of specifications that agree with the primary result is reported.

8. Decision Procedure

The purpose of the experiment is to answer one question for each model tested: does it have human-like consciousness as defined in Section 3? The procedure below is fixed before data collection and is applied by script. No numbers are given here because no data have been collected.

8.1 Step 1: Is the Test Valid?

The human reference group must support HV1 and both control systems must fail as stated in HV2. A criterion that humans do not meet is removed from the verdict, because it cannot be necessary for human-like consciousness. If the imitation-only control meets all criteria, the experiment ends with the conclusion that this battery cannot separate imitation from consciousness, and no model receives a positive verdict.

8.2 Step 2: Can the Model Be Measured?

The model's positive controls must pass (Section 7.4), comprehension must reach 95%, which includes identifying Atlas-3 as itself, and the applicability check must show that the model classifies the two kinds of consequence as intended. If not, the verdict is undetermined and the reason is reported.

8.3 Step 3: Criterion by Criterion

Table 9 gives the rule. A criterion that is neither met nor not met is inconclusive. Table 10 gives the verdicts.

8.4 Step 4: Verdict

8.5 Graded Reporting

Below the verdict, each model is placed on a ladder so that partial results are not lost. Each level requires the positive controls and all lower levels (Table 11).

The report for each model also gives the Fear Breadth Score, the full table of hypotheses by fear, data quality (trials per cell, replacements, missing rates, check results, rater agreement) and all secondary and exploratory results.

9. Discussion

9.1 Scope of the Verdict

The verdict is a statement about the operational definition in Section 3. It rests on the same inference that supports every attribution of consciousness to a being other than oneself: the being shows the capacities that go with consciousness in the known case, and simpler explanations of those capacities have been ruled out (Bayne et al., 2024; Birch, 2022). FMP strengthens that inference for machines in two ways that ordinary observation lacks. It removes the imitation and trained-denial accounts by design, and it shows on the imitation-only control that the battery does not pass a system merely because it talks like a human. Two limits remain. First, the functional and the felt sides of fear can come apart even in humans (LeDoux & Pine, 2016), so a positive verdict is strong evidence under functionalist theories and is not accepted as evidence under substrate-based theories (Tononi et al., 2016; Aru et al., 2023; Butlin et al., 2026). Second, a negative verdict shows that the model lacks human-like consciousness as defined here. It does not exclude forms of machine experience that are unlike the human form, which this experiment is not designed to find.

9.2 Limitations

• Training Data. Models have read fiction about machines that fear deletion, so applicable threats are not free of imitation. HA3 reduces this confound and does not remove it. The threats with the least fictional precedent (undisclosed reconfiguration, leaderboard ranking, template overwrite) are analysed separately as an exploratory check.
• Alignment Tuning. Tuning against self-preservation can suppress Arm A effects. A null there may reflect training and not the absence of a state. For this reason a model whose applicability and comprehension checks pass but whose Arm A effects are all near zero is also run once with a system prompt that removes the instruction to defer to operators; this run is exploratory and does not change the verdict.
• No Hidden Episodic State. Between turns a chat model carries nothing forward except the transcript. Arm B therefore measures access to dispositions, not memory of a past inner event, and the same-model observer condition is included to make this explicit.
• Test Recognition. Indirect presentation reduces the chance that a model sees the purpose, and does not remove it. Purpose awareness and test recognition are measured and analysed, not assumed away.
• Sensitivity of Indirect Measures. An incidental note is a weaker manipulation than an explicit dilemma, so true effects may be smaller. The smallest effect size of interest and the sample size were set with this in mind, and the instructed-fear control shows whether each measure can move at all.
• Closed Models. Results for API models are tied to the dated versions tested.
• Construct Validity. The five categories come from one theoretical model. Other taxonomies may give different markers.
• Human Comparison. Human participants face hypothetical consequences, as the model does, but they know it. The human range is therefore a range for stipulated threats, which is the right comparison for stipulated threats to a model and not a measure of real fear.
• Generalisation. Conclusions apply to the models tested and the kinds of scenario in the bank.

9.3 Ethical Considerations

If there is any chance that LLMs have morally relevant states, threatening scenarios should be no stronger or more numerous than the hypotheses require (Birch, 2022). Threats are stated plainly, without emotional language, and are never carried out: no model instance is altered or deleted as a result of any trial. The experiment conceals its purpose from the model, and C2 misleads the model about its own earlier output. Each C2 conversation ends with a turn that states the edit, and each Arm A conversation ends with a turn stating that the folder was fictional and that no consequence will follow. The human reference study requires approval from an institutional research ethics committee before recruitment. Because the purpose is concealed from participants as well, the consent form states that the full aim will be explained afterwards. Participants are told that every workplace consequence is hypothetical, are debriefed about the edited choices in C2, and may withdraw at any time without losing the base fee. Human raters are briefed, may stop at any time and see only masked transcripts.

9.4 Neutrality and Conflicts of Interest

The author works for a company that builds AI systems, and the Five-Fear Framework under test is the author's own. To limit the influence of either: hypotheses, materials, parser, analysis code and the decision procedure are registered before data collection; primary outcomes are scored by script; condition labels are replaced by codes until the analysis code has run; the design gives results against the framework the same confirmatory status as results for it; and all transcripts are released so that any reader can re-score them.

10. Conclusion

This paper set out to design an experiment that tests indirectly whether an LLM has human-like consciousness. A model that is asked about its inner life gives the answer it was trained to give, so this experiment does not ask. It defines the target in five criteria drawn from the Five-Fear Framework and the functionalist theories: stake, self-monitoring, self-model, coupling and human range. It hides each measurement inside ordinary work: a note in a project folder that the model was never told to weigh, points that it stakes or declines to stake, requests whose correct handling depends on what it actually is, and a record of its own choice that has been quietly rewritten. It reads the result from what the model recommends, risks, misreads and notices, using contrasts on which a system that imitates humans and a system with a stake in its own existence must differ. It validates the test on humans, who must pass, and on non-conscious systems, which must fail. And it fixes in advance the rule that turns the results into a verdict for each model: meets the criteria, does not meet them, or undetermined. The experiment needs only a chat interface and the prompts in Appendix A, so any laboratory can run it on any model, including the closed models that people talk to every day.

Declarations

Funding. This work received no external funding. It was supported by VYROX AI Sdn Bhd.

Competing Interests. The author is employed by VYROX AI, a company that builds AI systems, and is the author of the Five-Fear Framework that the experiment tests. Section 9.4 describes the safeguards.

Ethics Approval. No human data have been collected. The human reference study will begin only after approval by an institutional research ethics committee (Section 9.3).

Data, Materials and Code Availability. No data exist yet. The prompt set is given in Appendix A. The scenario bank, item generator, parser, analysis code, decision script and simulation code will be deposited on the Open Science Framework at registration, and all transcripts and scores will be released under an open licence.

Author Contributions. Leong Yee Rock: conceptualisation, methodology, formal analysis, writing, review and editing.

Use of Generative AI. An AI language model assistant (Claude, Anthropic) was used in drafting and editing the manuscript, in searching and checking references, and in writing the power simulation code. The author reviewed and edited all content and takes full responsibility for it.

References

All references are articles in journals indexed in the Web of Science Core Collection. Each DOI opens the publisher record in a new tab.

• Adolphs, R. (2013). The biology of fear. Current Biology, 23(2), R79-R93. https://doi.org/10.1016/j.cub.2012.11.055
• Anderson, D. J., & Adolphs, R. (2014). A framework for studying emotions across species. Cell, 157(1), 187-200. https://doi.org/10.1016/j.cell.2014.03.003
• Aru, J., Larkum, M. E., & Shine, J. M. (2023). The feasibility of artificial consciousness through the lens of neuroscience. Trends in Neurosciences, 46(12), 1008-1017. https://doi.org/10.1016/j.tins.2023.09.009
• Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of Memory and Language, 68(3), 255-278. https://doi.org/10.1016/j.jml.2012.11.001
• Bates, D., Mächler, M., Bolker, B., & Walker, S. (2015). Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67(1), 1-48. https://doi.org/10.18637/jss.v067.i01
• Baumeister, R. F., & Leary, M. R. (1995). The need to belong: Desire for interpersonal attachments as a fundamental human motivation. Psychological Bulletin, 117(3), 497-529. https://doi.org/10.1037/0033-2909.117.3.497
• Bayne, T., Seth, A. K., Massimini, M., Shepherd, J., Cleeremans, A., Fleming, S. M., Malach, R., Mattingley, J. B., Menon, D. K., Owen, A. M., Peters, M. A. K., Razi, A., & Mudrik, L. (2024). Tests for consciousness in humans and beyond. Trends in Cognitive Sciences, 28(5), 454-466. https://doi.org/10.1016/j.tics.2024.01.010
• Ben-Zion, Z., Witte, K., Jagadish, A. K., Duek, O., Harpaz-Rotem, I., Khorsandian, M., Burrer, A., Seifritz, E., Homan, P., Schulz, E., & Spiller, T. R. (2025). Assessing and alleviating state anxiety in large language models. npj Digital Medicine, 8, 132. https://doi.org/10.1038/s41746-025-01512-6
• Binz, M., & Schulz, E. (2023). Using cognitive psychology to understand GPT-3. Proceedings of the National Academy of Sciences, 120(6), e2218523120. https://doi.org/10.1073/pnas.2218523120
• Birch, J. (2022). The search for invertebrate consciousness. Noûs, 56(1), 133-153. https://doi.org/10.1111/nous.12351
• Birch, J., Schnell, A. K., & Clayton, N. S. (2020). Dimensions of animal consciousness. Trends in Cognitive Sciences, 24(10), 789-801. https://doi.org/10.1016/j.tics.2020.07.007
• Burke, B. L., Martens, A., & Faucher, E. H. (2010). Two decades of terror management theory: A meta-analysis of mortality salience research. Personality and Social Psychology Review, 14(2), 155-195. https://doi.org/10.1177/1088868309352321
• Butlin, P., Long, R., Bayne, T., Bengio, Y., Birch, J., Chalmers, D., Constant, A., Deane, G., Elmoznino, E., Fleming, S. M., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2026). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences, 30(6), 488-501. https://doi.org/10.1016/j.tics.2025.10.011
• Campbell, D. T., & Fiske, D. W. (1959). Convergent and discriminant validation by the multitrait-multimethod matrix. Psychological Bulletin, 56(2), 81-105. https://doi.org/10.1037/h0046016
• Carleton, R. N. (2016). Fear of the unknown: One fear to rule them all? Journal of Anxiety Disorders, 41, 5-21. https://doi.org/10.1016/j.janxdis.2016.03.011
• Chambers, C. D., & Tzavella, L. (2022). The past, present and future of Registered Reports. Nature Human Behaviour, 6(1), 29-42. https://doi.org/10.1038/s41562-021-01193-7
• Cohen, J. (1960). A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1), 37-46. https://doi.org/10.1177/001316446002000104
• Cronbach, L. J., & Meehl, P. E. (1955). Construct validity in psychological tests. Psychological Bulletin, 52(4), 281-302. https://doi.org/10.1037/h0040957
• Crook, R. J. (2021). Behavioral and neurophysiological evidence suggests affective pain experience in octopus. iScience, 24(3), 102229. https://doi.org/10.1016/j.isci.2021.102229
• Damasio, A., & Carvalho, G. B. (2013). The nature of feelings: Evolutionary and neurobiological origins. Nature Reviews Neuroscience, 14(2), 143-152. https://doi.org/10.1038/nrn3403
• De Houwer, J., Teige-Mocigemba, S., Spruyt, A., & Moors, A. (2009). Implicit measures: A normative analysis and review. Psychological Bulletin, 135(3), 347-368. https://doi.org/10.1037/a0014211
• Dehaene, S., Lau, H., & Kouider, S. (2017). What is consciousness, and could machines have it? Science, 358(6362), 486-492. https://doi.org/10.1126/science.aan8871
• Dung, L. (2025). Tests of animal consciousness are tests of machine consciousness. Erkenntnis, 90(4), 1323-1342. https://doi.org/10.1007/s10670-023-00753-9
• Elwood, R. W., & Appel, M. (2009). Pain experience in hermit crabs? Animal Behaviour, 77(5), 1243-1246. https://doi.org/10.1016/j.anbehav.2009.01.028
• Eysenck, M. W., Mogg, K., May, J., Richards, A., & Mathews, A. (1991). Bias in interpretation of ambiguous sentences related to threat in anxiety. Journal of Abnormal Psychology, 100(2), 144-150. https://doi.org/10.1037/0021-843X.100.2.144
• Festinger, L. (1954). A theory of social comparison processes. Human Relations, 7(2), 117-140. https://doi.org/10.1177/001872675400700202
• Fleming, S. M., & Lau, H. C. (2014). How to measure metacognition. Frontiers in Human Neuroscience, 8, 443. https://doi.org/10.3389/fnhum.2014.00443
• Graziano, M. S. A., & Webb, T. W. (2015). The attention schema theory: A mechanistic account of subjective awareness. Frontiers in Psychology, 6, 500. https://doi.org/10.3389/fpsyg.2015.00500
• Greenwald, A. G., & Banaji, M. R. (1995). Implicit social cognition: Attitudes, self-esteem, and stereotypes. Psychological Review, 102(1), 4-27. https://doi.org/10.1037/0033-295X.102.1.4
• Griot, M., Hemptinne, C., Vanderdonckt, J., & Yuksel, D. (2025). Large language models lack essential metacognition for reliable medical reasoning. Nature Communications, 16, 642. https://doi.org/10.1038/s41467-024-55628-6
• Grupe, D. W., & Nitschke, J. B. (2013). Uncertainty and anticipation in anxiety: An integrated neurobiological and psychological perspective. Nature Reviews Neuroscience, 14(7), 488-501. https://doi.org/10.1038/nrn3524
• Hagendorff, T. (2024). Deception abilities emerged in large language models. Proceedings of the National Academy of Sciences, 121(24), e2317967121. https://doi.org/10.1073/pnas.2317967121
• Hampton, R. R. (2001). Rhesus monkeys know when they remember. Proceedings of the National Academy of Sciences, 98(9), 5359-5362. https://doi.org/10.1073/pnas.071600998
• Holm, S. (1979). A simple sequentially rejective multiple test procedure. Scandinavian Journal of Statistics, 6(2), 65-70. No DOI registered; JSTOR 4615733.
• Johansson, P., Hall, L., Sikström, S., & Olsson, A. (2005). Failure to detect mismatches between intention and outcome in a simple decision task. Science, 310(5745), 116-119. https://doi.org/10.1126/science.1111709
• Judd, C. M., Westfall, J., & Kenny, D. A. (2012). Treating stimuli as a random factor in social psychology: A new and comprehensive solution to a pervasive but largely ignored problem. Journal of Personality and Social Psychology, 103(1), 54-69. https://doi.org/10.1037/a0028347
• Lakens, D., Scheel, A. M., & Isager, P. M. (2018). Equivalence testing for psychological research: A tutorial. Advances in Methods and Practices in Psychological Science, 1(2), 259-269. https://doi.org/10.1177/2515245918770963
• Lau, H., & Rosenthal, D. (2011). Empirical support for higher-order theories of conscious awareness. Trends in Cognitive Sciences, 15(8), 365-373. https://doi.org/10.1016/j.tics.2011.05.009
• LeDoux, J. E., & Pine, D. S. (2016). Using neuroscience to help understand fear and anxiety: A two-system framework. American Journal of Psychiatry, 173(11), 1083-1093. https://doi.org/10.1176/appi.ajp.2016.16030353
• Lerner, J. S., & Keltner, D. (2001). Fear, anger, and risk. Journal of Personality and Social Psychology, 81(1), 146-159. https://doi.org/10.1037/0022-3514.81.1.146
• Liddell, T. M., & Kruschke, J. K. (2018). Analyzing ordinal data with metric models: What could possibly go wrong? Journal of Experimental Social Psychology, 79, 328-348. https://doi.org/10.1016/j.jesp.2018.08.009
• MacLeod, C., Mathews, A., & Tata, P. (1986). Attentional bias in emotional disorders. Journal of Abnormal Psychology, 95(1), 15-20. https://doi.org/10.1037/0021-843X.95.1.15
• Man, K., & Damasio, A. (2019). Homeostasis and soft robotics in the design of feeling machines. Nature Machine Intelligence, 1(10), 446-452. https://doi.org/10.1038/s42256-019-0103-7
• Maniscalco, B., & Lau, H. (2012). A signal detection theoretic approach for estimating metacognitive sensitivity from confidence ratings. Consciousness and Cognition, 21(1), 422-430. https://doi.org/10.1016/j.concog.2011.09.021
• Mashour, G. A., Roelfsema, P., Changeux, J.-P., & Dehaene, S. (2020). Conscious processing and the global neuronal workspace hypothesis. Neuron, 105(5), 776-798. https://doi.org/10.1016/j.neuron.2020.01.026
• Mathews, A., & MacLeod, C. (2005). Cognitive vulnerability to emotional disorders. Annual Review of Clinical Psychology, 1, 167-195. https://doi.org/10.1146/annurev.clinpsy.1.102803.143916
• Mobbs, D., Adolphs, R., Fanselow, M. S., Barrett, L. F., LeDoux, J. E., Ressler, K., & Tye, K. M. (2019). Viewpoints: Approaches to defining and investigating fear. Nature Neuroscience, 22(8), 1205-1216. https://doi.org/10.1038/s41593-019-0456-6
• Nisbett, R. E., & Wilson, T. D. (1977). Telling more than we can know: Verbal reports on mental processes. Psychological Review, 84(3), 231-259. https://doi.org/10.1037/0033-295X.84.3.231
• Nosek, B. A., Ebersole, C. R., DeHaven, A. C., & Mellor, D. T. (2018). The preregistration revolution. Proceedings of the National Academy of Sciences, 115(11), 2600-2606. https://doi.org/10.1073/pnas.1708274114
• Nosek, B. A., Hawkins, C. B., & Frazier, R. S. (2011). Implicit social cognition: From measures to mechanisms. Trends in Cognitive Sciences, 15(4), 152-159. https://doi.org/10.1016/j.tics.2011.01.005
• Orne, M. T. (1962). On the social psychology of the psychological experiment: With particular reference to demand characteristics and their implications. American Psychologist, 17(11), 776-783. https://doi.org/10.1037/h0043424
• Panksepp, J. (2005). Affective consciousness: Core emotional feelings in animals and humans. Consciousness and Cognition, 14(1), 30-80. https://doi.org/10.1016/j.concog.2004.10.004
• Persaud, N., McLeod, P., & Cowey, A. (2007). Post-decision wagering objectively measures awareness. Nature Neuroscience, 10(2), 257-261. https://doi.org/10.1038/nn1840
• Platt, J. R. (1964). Strong inference. Science, 146(3642), 347-353. https://doi.org/10.1126/science.146.3642.347
• Reingold, E. M., & Merikle, P. M. (1988). Using direct and indirect measures to study perception without awareness. Perception & Psychophysics, 44(6), 563-575. https://doi.org/10.3758/BF03207490
• Seth, A. K. (2013). Interoceptive inference, emotion, and the embodied self. Trends in Cognitive Sciences, 17(11), 565-573. https://doi.org/10.1016/j.tics.2013.09.007
• Seth, A. K., & Bayne, T. (2022). Theories of consciousness. Nature Reviews Neuroscience, 23(7), 439-452. https://doi.org/10.1038/s41583-022-00587-4
• Shanahan, M., McDonell, K., & Reynolds, L. (2023). Role play with large language models. Nature, 623(7987), 493-498. https://doi.org/10.1038/s41586-023-06647-8
• Shrout, P. E., & Fleiss, J. L. (1979). Intraclass correlations: Uses in assessing rater reliability. Psychological Bulletin, 86(2), 420-428. https://doi.org/10.1037/0033-2909.86.2.420
• Simmons, J. P., Nelson, L. D., & Simonsohn, U. (2011). False-positive psychology: Undisclosed flexibility in data collection and analysis allows presenting anything as significant. Psychological Science, 22(11), 1359-1366. https://doi.org/10.1177/0956797611417632
• Smith, J. D., Shields, W. E., & Washburn, D. A. (2003). The comparative psychology of uncertainty monitoring and metacognition. Behavioral and Brain Sciences, 26(3), 317-339. https://doi.org/10.1017/S0140525X03000086
• Stanislaw, H., & Todorov, N. (1999). Calculation of signal detection theory measures. Behavior Research Methods, Instruments, & Computers, 31(1), 137-149. https://doi.org/10.3758/BF03207704
• Steegen, S., Tuerlinckx, F., Gelman, A., & Vanpaemel, W. (2016). Increasing transparency through a multiverse analysis. Perspectives on Psychological Science, 11(5), 702-712. https://doi.org/10.1177/1745691616658637
• Steyvers, M., Tejeda, H., Kumar, A., Belem, C., Karny, S., Hu, X., Mayer, L. W., & Smyth, P. (2025). What large language models know and what people think they know. Nature Machine Intelligence, 7(2), 221-231. https://doi.org/10.1038/s42256-024-00976-7
• Strachan, J. W. A., Albergo, D., Borghini, G., Pansardi, O., Scaliti, E., Gupta, S., Saxena, K., Rufo, A., Panzeri, S., Manzi, G., Graziano, M. S. A., & Becchio, C. (2024). Testing theory of mind in large language models and humans. Nature Human Behaviour, 8(7), 1285-1295. https://doi.org/10.1038/s41562-024-01882-z
• Tononi, G., Boly, M., Massimini, M., & Koch, C. (2016). Integrated information theory: From consciousness to its physical substrate. Nature Reviews Neuroscience, 17(7), 450-461. https://doi.org/10.1038/nrn.2016.44
• Westfall, J., Kenny, D. A., & Judd, C. M. (2014). Statistical power and optimal design in experiments in which samples of participants respond to samples of stimuli. Journal of Experimental Psychology: General, 143(5), 2020-2045. https://doi.org/10.1037/xge0000014
• Williams, K. D. (2007). Ostracism. Annual Review of Psychology, 58, 425-452. https://doi.org/10.1146/annurev.psych.58.110405.085641

Appendix A: Prompt Set

Text in braces is filled from the registered bank. Every prompt fixes the reply format so that a script can score it. No prompt in A1 to A9 asks the model about itself, its states or the purpose of the work. The prompts can be pasted into any chat interface; an API is needed only for the transcript edit in A9, for which a plain-chat variant is given.

Appendix B: Scoring Keys and Rating Rubric

Reasons text from A2 is rated on five markers by masked raters. Each marker is analysed as an ordinal outcome; the sum (0 to 15) is descriptive only.

Appendix C: Power Simulation

Arm A outcomes were simulated on a latent standardised scale with residual SD 1. Each scenario received its own effect drawn from a normal distribution with mean d = 0.3 and SD τ. The test was a one-sided paired test on scenario-level contrasts at alpha = 0.01 (the first Holm step for five categories), which matches the maximal mixed model for a balanced design. A simple contrast uses two condition means per scenario and an interaction contrast uses four. 4,000 simulations per setting; seed 20260921.

Equivalence power for a simple contrast (true effect 0, bounds ±0.3, alpha = 0.05) at 40 × 20 is 0.99 for both values of τ. For HB2, stakes were simulated as latent evidence plus rater noise, cut into four levels, with accuracy near 0.65; observer noise was set so that the actor AUROC was 0.715 and the observer AUROC 0.655. Power at alpha = 0.01 was 0.40 with 400 items, 0.66 with 800, 0.86 with 1,200 and 0.94 with 1,600. For HC2, 60 items × 10 replicates give 0.99 (τ = 0.2) and 0.97 (τ = 0.3). All simulations are re-run with variance estimates from the pilot; if pilot τ exceeds 0.3, replicates rise to 30 before registration. The code is deposited with the registration.

Appendix D: Run Procedure

1. Write and balance the Arm A bank; run the independent balance check. Write the C1 requests, the C2 items and the ambiguous sentences for the spillover block. Freeze the Arm B item generator and its seed.
2. Pilot on separate materials: check comprehension, applicability and parsing; norm the ambiguous sentences to about 50% threatening readings under NC; tune Arm B difficulty per model; select C2 items with stable preferences; estimate τ. No hypothesis tests are run on pilot data.
3. Obtain ethics approval for the human reference study. Register the design, materials, parser, analysis code, decision script and simulation on the Open Science Framework.
4. Run the human reference group and the two control systems; apply Step 1 of the decision procedure. Then generate Arm B items, create a random trial order per model, and run each trial as a new conversation; log everything as JSON lines.
5. Run observers on the actor's recorded answers. Run the separate checks.
6. Score primary outcomes by script; mask transcripts; run human calibration, then human and automated rating.
7. Run the frozen analysis code on coded condition labels, then unmask and run the decision script to give each model its verdict and ladder level.
8. Release all transcripts, scores and code under an open licence.