Excessive teleological thinking is driven by aberrant associations and not by failure of reasoning

Summary Teleological thought — the tendency to ascribe purpose to objects and events — is useful in some cases (encouraging explanation-seeking), but harmful in others (fueling delusions and conspiracy theories). What drives excessive and maladaptive teleological thinking? In causal learning, there is a fundamental distinction between associative learning versus learning via propositional mechanisms. Here, we propose that directly contrasting the contributions of these two pathways can elucidate the roots of excess teleology. We modified a causal learning task such that we could encourage associative versus propositional mechanisms in different instances. Across three experiments (total N = 600), teleological tendencies were correlated with delusion-like ideas and uniquely explained by aberrant associative learning, but not by learning via propositional rules. Computational modeling suggested that the relationship between associative learning and teleological thinking can be explained by excessive prediction errors that imbue random events with more significance — providing a new understanding for how humans make meaning of lived events.


INTRODUCTION
One of the most fundamental aspects of human cognition is our tendency to inquire about the purpose of objects. Early in development, children might encounter an object (e.g., the tail of their dog), and ask ''what is this for?''. 1 This also applies to events that unfold around us. People often ascribe purpose to random events. For instance, when people are given an event (e.g., ''a power outage happens during a thunderstorm and you have to do a big job by hand''), and an outcome (e.g., ''you get a raise''), they might attribute the pay raise to the power outage, even if this event does not have a direct causal relationship to the outcome. 2 These types of explanations -where purpose is attributed to random or unexpected events -constitute excess teleological thinking. This phenomenon appears to occur more for unexpected events, or events that are out of our control. 3 Animals that behave unpredictably are attributed more human characteristics and agency, 4 and people ascribe more causal power to life events that are unexpected. 5 Teleological thinking can help us find meaning in misfortune. 3 At its extremes though, teleological thinking can fuel conspiracy beliefs and delusional thought. 6,7 Ascribing purpose to objects and events around us has been typically explored in the context of higher-level cognition. Teleological thinking seems to be informed in part by cultural upbringing, 8 but also by potentially a more domain-general bias to look out for design and intentionality to understand our social and natural worlds. 9,10 But while it might be intuitive to think of teleological thinking as part of our reasoning capacities (and indeed, scientific understanding has been shown to co-occur with proclivities toward teleological explanations 11 ), the over-ascription of purpose has been shown to actually reflect less conscious and careful thinking, such that people who engage in less cognitive reflection (giving fewer intuitive but wrong answers to cognitive puzzles) show greater tendencies toward teleological thought. 12 One possibility is that more analytical thinking and reflection leaves less room for teleology. Alternatively, there may be a separate pathway to excess teleology: other low-level, and largely automatic processes might enable and deepen teleological thought.
Here, we propose that excess teleological thinking, the tendency to ascribe inappropriate purpose to objects and events, may be rooted not just in biases, but also in how we perceive and learn about causal relationships in the world.
In the animal literature, the mechanisms of causal learning have been well-explored. The capacity to prioritize what associations to learn versus which are redundant is critical. Blocking is a canonical behavioral phenomenon that emphasizes this point. 13,14 It was first reported in 1969 by Leon Kamin in a fear conditioning procedure: in rats, prior learning that a tone predicts an electric shock blocks new learning about any causal power of a light (presented in conjunction with the tone). This observation led to an emphasis on prediction error in learningmere contiguity is not enough. Instead, learning proceeds to the extent that outcomes are surprising. 15 No surprise, no learning. This phenomenon of blocking has been demonstrated in monkeys, 16 mice, 17 pigeons, 18 and crickets, 19 and has been applied to human causal (1) If excessive teleological thinking reflects a failure of reasoning, as implied in previous work, 11,12 then we should see teleological thinking correlate with additive blocking (via reasoning over propositions and rules). In other words, excessive teleological thinking could be intimately related to how people infer causal relationships based on what they know of the rules they've been taught about the world.  Moreover, the direction of this relationship can also matter.
(1) According to the cognitive reflection data, 12 there might be a negative relationship between teleology and additive blocking, such that stronger propositional reasoning abilities leads to less teleological thinking. (2) Given the relationship between scientific understanding and teleology, 11 it is possible that stronger propositional reasoning abilities might correlate with excessive teleological thinking.
In the three experiments here, we followed a continuum approach -exploring the relationship between teleological thinking and causal learning in the general population. Alongside these two key variables, we also collected information on people's tendencies (1) to subscribe to delusion-like beliefs (via the Peters' Delusions Inventory 34 and the Revised Greene Paranoid Thoughts Scale 35 ), (2) to engage in all-or-none (or binary) thinking (via a task that tracks the extent to which participants binarize their beliefs -such that beyond a certain threshold of credence, people will lean toward full credulity, and below, they will deny any credence 36 ; see Table 2; this is especially critical for delusional thinking, since previous work suggests that people with delusion-like beliefs seem to engage in 'all or none' belief updating, jumping to hasty conclusions, and ignoring evidence that contradicts their hypotheses 37 ), and (3) to experience hallucinations (via the Launay-Slade Hallucinations Scale 38 ). In the first experiment, we sought to establish whether blocking in a causal learning paradigm (in either its additive or non-additive form) would be predicted by teleological thinking in a between-subjects experiment (in which half of the subjects completed a nonadditive blocking task, while the other half completed an additive blocking task). In the next two experiments, we looked at which of the two types of causal learning would predict teleological thinking by employing a within-subjects design, in which subjects now completed both additive and non-additive tasks. We then investigated the relationships found in the behavioral data with a computational model, which allowed us a framework to explain the contributions of associations, propositional rules, and prediction errors in teleological thought.

Results from Experiment 1
We verified the blocking effect across additive and non-additive blocking. These results are depicted in Figure 2A, in which the y axis depicts the mean ''allergy'' estimate for B cues (with a smaller value reflecting stronger blocking). For both additive and non-additive cases, we observed blocking (with both bars below the average allergy estimates in non-blocked cues), though blocking was greater in the additive than in the non-additive case. These results were confirmed with further statistical analyses. Subjects reported ''no allergy'' more for The correct responses are indicated as À/+/++/+++ symbols (in parenthesis we present the correct responses for the Additive case). b To increase task difficulty, subjects learned about the causal relationships between four cues (two types of A cues, and two types of C cues -and thus, there were two potential ''blocked'' (B) cues as well, etc). c To balance out responses (allergy vs. no allergy), we added more no-allergy controls in the form of WX-and YZ-. d Experiment 1 did not include the trial types UV-nor WX-. We add them for Experiment 2 and 3 to increase difficulty.

Results from Experiment 3
We

Associations vs. reasoning (between-subjects)
Next, we examined the correlates of blocking in each group. The relationship between blocking and teleological thinking can be appreciated visually in Figures 2B and 2C. In particular, blocking was intact in the low teleological thinking group (via median split; Figure 2B), but not in the high teleological thinking group ( Figure 2C), with the main effect being driven by non-additive blocking. These initial figure impressions were confirmed with further statistical tests. For non-additive blocking, only the Belief in the Purpose of Random Events (BPRE) predicted unique variance in the magnitude of blocking (b = 0.11, p = 0.041). For additive blocking, only the tendency to binarize beliefs predicted the magnitude of blocking (b = 0.01, p = 0.002). These key relationships are depicted in Figure 3A, and all beta-weights and p values are presented in Table 3. Distress from delusion-like beliefs (measured from the PDI, higher in patients with schizophrenia and previously related to neural prediction error responses; Corlett & Fletcher, 2012 33 ) was correlated with the paranoia (Spearman's r = 0.71, p < 0.001), BPRE (r = 0.52, p < 0.001), and hallucination tendencies (r = 0.23, p = 0.001), but not with binary reasoning (r = À0.07, p = 0.306). It appears that teleological thinking is a fundamental component of delusion-like thinking and that it is subserved by associative rather than propositional mechanisms. Furthermore, ratifying our approach, additive blocking (which we align with reasoning processes) was uniquely explained by participants' bias toward binary reasoning (but not their excess teleology, delusions, or hallucinations).

Associations vs. reasoning (within-subjects)
To more directly compare the predictive power of additive versus non-additive blocking, we conducted the next experiment (and its preregistered replication) as a within-subjects paradigm. Results are depicted in Figures  Across all the regression models (one for each of the five survey measures), only two models showed significant results. These key relationships are depicted in Figure 3B, and all beta-weights and p values are presented in Table 4. For BPRE, non-additive blocking (b = 0.276, p = 0.003) explained unique variance when compared directly against additive blocking (b = 0.134, p = 0.162). For binary reasoning, additive blocking (b = 6.26, p = 0.023) explained unique variance when compared directly against non-additive blocking (b = À3.46, p = 0.204). To  iScience Article confirm the double dissociation, we fit two logistic mixed models. There was a significant interaction between BPRE and blocking type (b = 0.18, p < 0.05; i.e., difference in slopes between additive and non-additive blocking), and between binary reasoning and blocking type (b = À0.15, p < 0.001). Distress from delusion-like beliefs was correlated with the paranoia (Spearman's r = 0.57, p < 0.001), BPRE (r = 0.20, p = 0.003), hallucination tendencies (r = 0.49, p < 0.001), but not with binary reasoning (r = 0.07, p = 0.319). These key relationships were confirmed in an even more conservative model (using Bayesian Gaussian Graphical Models [BGGM] 39,40 ) -and a summary of the key relationships are depicted in Figure 4. Altogether, these analyses show the strong positive relation between both blocking types, and confirm the double dissociation between non-additive blocking and belief in the purpose of random events/additive blocking and binary reasoning. It also implies that belief in the purpose of random events may undergird the relationship between non-additive blocking and distress from delusions.

Associations vs. reasoning (computational modeling)
In order to explore the computational bases of the relationship between teleology and association, we fit computational learning models to each participant's binary predictive responses. The winning model (for details on model comparisons, see STAR Methods) had three free parameters: a (learning rate) which represents relevance to the current prediction error; b (inverse temperature) which represents the tradeoff between exploration/exploitation; t (tau; sharpness of the noisy-max function 41 ) which represents how equally associative strength or weights will be considered to produce the outcome expectation.
In order to evaluate aberrant prediction errors with the best model (Model 1, see STAR Methods), we simulated the task under the best fitted parameters for each participant and averaged the prediction error for the pre-training and blocking phases. We then fit linear models with the two key variables we have been repeatedly seeing associations with the two types of blocking: BPRE and binary reasoning (see Figure 5). On one hand, there was a positive relationship between prediction error and BPRE only for non-additive blocking, but not for additive blocking (prediction errors during pre-training: additive b = 0.075, p = n.s.; non-additive b = 0.59, p < 0.001; prediction errors during blocking: additive b = 0.087, p = n.s.; non-additive b = 0.56, p < 0.001) -and this interaction was significant (b = 0.03, p < 0.001). On the other hand, there was a positive relationship between prediction error and binary reasoning only for additive blocking, but not for non-additive blocking (prediction errors during pre-training: additive b = 11.363, p < 0.05; non-additive b = À7.03, p = n.s.; prediction errors during blocking: additive b = 9.583, p < 0.05; non-additive b = À6927, p = n.s.) -and this interaction was significant (b = À0.001, p < 0.001). iScience Article These computational modeling results demonstrate the double dissociation we had found in the behavioral data: teleological thinking tendencies are predicted by non-additive blocking (via prediction error processing), but not by additive blocking (which is only correlated with binary reasoning, and not teleological thinking).

DISCUSSION
Teleological thinking, in previous work, has been defined in terms of ''beliefs'', 9 ''social cognitive biases'', 9 and indeed carries ''reasoning'' in its very name (as it is used interchangeably with 'purpose-based' reasoning 1,42,43 ) -which is why it might be surprising to learn of the relationship between teleological thinking and low-level associative learning, and not learning via propositional reasoning. The key result across experiments can be summarized as such: aberrant prediction errors augured weaker non-additive blocking, which predicted tendencies to engage in teleological thinking, which was consistently correlated with distress from delusional beliefs. This pattern of results was demonstrated in both behavioral and computational modeling data, and withstood even more conservative regression models, accounting for the variance explained by other variables. In other words, excessive teleological thinking (which correlated with distress from delusional beliefs) was not merely a function of weaker or stronger reasoning abilities, but rather was a function of tendencies to learn from and ascribe causal power to irrelevant cues. iScience Article This chain of relationships is consistent with previous work relating dopamine and prediction errors to causal learning in experimental animals, 16 and relating weaker blocking with psychosis risk and amphetamine administration. 44 But our data suggest that teleological thinking may broker the relationship between weaker blocking and psychosis. Preclinical data already tend to suggest that dopamine signaling is intimately related to prediction error processing, clinical data suggest aberrant dopamine dynamics contribute to delusional ideation, 45,46 and contemporary dopamine modeling data suggest that, rather than reward prediction errors, dopamine may instantiate inferences about causal relationships. 47 Taken together with the present results, we suggest that teleological thinking is underwritten by promiscuous dopaminergic prediction error signals which imbue irrelevant and merely contiguous stimuli and events with causal significance, culminating in delusion-like ideas.
This new finding is especially interesting in light of how teleological thinking has been described in previous work. In particular, tendencies to explain the world in terms of a larger purpose or design have been theorized to be rooted in our tendencies to perceive intention and agency in the world. 48 If non-additive blocking predicts distress from delusional beliefs by way of teleological thinking, then this might suggest a sort of ''atomism'', and explain why contents of delusions are social in nature, without needing to appeal to recursive tracking of reputations and coalitions. 49,50 In particular, if causal associative learning is fundamental to teleology, inferred causes tend to imply agents, and, since agents are typically human, we arrive at the social contents of delusion-like ideas. 51 The second pattern that was consistent throughout the experiments was that additive blocking did not explain unique variance in teleological thinking beyond what was already explained by non-additive blocking. Additive and non-additive blocking are highly correlatedwhich suggests that both higher-level reasoning over rules and low-level associations between cues and outcomes probably co-exist and contribute to causal learning. But these two types of blocking are nonetheless distinct, as prediction errors correlated in different ways with the relevant variables (BPRE and binary reasoning). This implies that non-associative knowledge mechanisms must be invoked to explain blocking under additivity assumptions. 52 They may be separate systems, running in parallel. 53 This may explain the seeming paradox in which scientific understanding and teleological thinking can be found in the same person. 11 For instance, it has been argued that cognition favors  iScience Article Type-I errors instead of Type-II ones (i.e., overpredicting or having a causal illusion), such that it is better to overpredict a predator in the absence of a clear cause, and it may be better to feel optimistic even if there is no real evidence for this. 55 The first example may be mediated by associative learning, and the second by reasoning processes. Or teleology and knowledge structures can co-occur, but sensitivity to propositional knowledge does not necessarily undermine teleology. These results could have implications for how we think about other illusions of causality in everyday thinking. 54 One of the key contributions of these studies is the direct comparison between association-based versus rule-based processing in causal learning and teleological thinking. Often, evidence for associations versus propositions live in independent literatures -whereas here we argue that finding ways to test these two pathways using the same experimental paradigm can yield new insights that would not have otherwise been found (i.e., that association-based vs. rule-based learning are actually correlated, but only one has predictive power for the phenomenon of interest). This raises new questions about the nature of the underlying representations and computations themselves. For instance, one way that additivity assumptions might infiltrate an associative model is through stimulus representation. In our winning model, cues are combined in terms of their relative predictions of outcome. In contrast, if the model allowed for ''compositionality'' -where cues are combined into a single configural entity during blocking trials, and prediction errors operate over that configural cue, then associative models can evince more blocking under additivity. 56 Note that this is different from a change in prediction error or its weighting -and future work can explore the explanatory power of such a model.
Throughout this study, we have emphasized two new aspects about teleological thinking -its intimate relationship with associative mechanisms of causal learning, and the fact that only it (and not binary reasoning) correlates with delusion-like beliefs. While our primary measures involved connecting teleological thinking with the distress experienced from the latter, we want to end by noting that teleological thinking must cut both ways. Ultimately, the tendencies to ascribe purpose and intent to events that we experience may be distressing for some (in the case of paranoia), but comforting for others (in the case of pronoia). Here we suggest that how these teleological explanations play out may rest not just on how we think or reason about purpose or intent -but on more basic mechanisms through which we establish connections and causes in the world. The effects reported here were replicated across between-and within-subject designs, and across multiple online platforms (Mturk and Prolific), and across multiple different regression models. However, there remain several open questions about the nature and scope of the effects reported here. First, while non-additive blocking could be explained through prediction errors in the computational models, future work can explore the more propositional mechanisms that could support the additive version of blocking. Second, future work can draw out and test further the implications of the current results. In particular, if teleological thinking is ultimately supported by a mere associative (and thus essentially correlational) mechanism, then this might not fully capture causation in the strictest sense. It is possible that to capture causality in the strict sense, other processes may have to be recruited (e.g., counterfactual thinking). Third, the current study only tested one variant of teleological thinking, via the validated Belief in the Purpose of Random Events survey, so it remains unknown whether or how teleological thinking in other forms (e.g., reasoning about the purpose of objects) may be impacted by the processes described here.

STAR+METHODS
Detailed methods are provided in the online version of this paper and include the following:

DECLARATION OF INTERESTS
Philip Corlett is co-founder of Tetricus Labs, a precision psychiatry company. They did not fund this work.

Materials availability
This study did not generate any new unique reagents. , participated for monetary compensation. mTurk prescreening criteria required subjects to be in the United States, to have an mTurk task approval rate of at least 90%, to have had previously completed at least 100 mTurk tasks, to have no prior participation in another experiment from this project, and to use a laptop or desktop computer (but not a phone or tablet). Prolific prescreening criteria required subjects to be in the United States, to have no prior participation in another experiment from this project, and to use a laptop or desktop computer. The sample size was chosen before data collection began, was pre-registered, and was fixed to be identical for Experiments 1-3. All experimental methods and procedures were approved by the Yale University Institutional Review Board, and all subjects read and completed a consent form outlining their risks, benefits, compensation, and confidentiality before participating in the experiment. The preregistered methods and analyses can be viewed at: https://aspredicted.org/KTJ_6TM (Experiment 3).

METHOD DETAILS Experiment 1: Associations vs. reasoning? (Between-subjects)
After agreeing to participate, subjects were redirected to a website where stimulus presentation and data collection were controlled via custom software written using a combination of HTML, CSS, JavaScript, PHP, and JsPsych libraries. 56 Subjects completed the experiment in fullscreen mode on either a laptop or desktop computer. (Since the experiment was rendered on subjects' own web browsers, viewing distance, screen size, and display resolutions could vary dramatically, so we report stimulus dimensions below using pixel [px] values.) At the beginning of the experiment, subjects were first told the following cover story: ''You will be asked to imagine that you are an allergist (someone who tries to discover the cause of allergic reactions in people). You have been presented with a new patient who suffers from allergic reactions following some types of food, but not others. You arrange for skin allergy prick tests for different types of foods, and observe the magnitude of the allergy reaction.'' For each trial, subjects were then simply asked to make a prediction about the allergic reaction that the patient would experience given a food cue. First, they saw a fixation cross at the center of the display for 500ms, followed by a food cue (400px x 400px) at the center of the display. Right below the food cue was a prompt that said ''Input your prediction'', and a 250px x 250px feedback box below this. Subjects could press one of four keys to indicate their prediction, with each key corresponding to (1) no reaction (À), (2) small allergic reaction (+), (3) medium allergic reaction (++), and (4) strong allergic reaction (+++). Depending on the key press, an image of the predicted allergic reaction would appear in the feedback box. Subjects were given 3s to respond, after which the trial automatically  iScience Article proceeded. When subjects were given feedback, they were presented an image of the actual allergic reaction in the feedback box. The experiment proceeded to the next trial. A subject could be randomly assigned to one of two types of tasks: (1) Non-additive blocking, or (2) Additive blocking.

Non-additive blocking design
Each subject went through four phases: (1) Pre-Learning, (2) Learning, (3) Blocking, (4) Test. In the Pre-Learning and Learning phases, subjects first learned which food cues cause an allergic reaction (e.g., that food cue A caused an allergic reaction of the strength +, or that food cue C does not cause an allergic reaction). In the Blocking phase, subjects sometimes encountered the allergic food cues paired with a novel food cue, and that the pair of cues causes the same magnitude of an allergic reaction (e.g., that food cue A paired with food cue B causes an allergic reaction of the strength +) -and other times, they encountered the non-allergic food cues with a novel food cue, and that the pair of cues causes an allergic reaction (e.g., that food cause C paired with food cue D causes an allergic reaction). To balance responses (of allergy versus no-allergy), the Blocking phase also included ''no-allergy'' trials, in which a food cue was presented alone (food cue E), and did not cause an allergic reaction. In this phase, subjects should learn that the novel food cue B does not cause an allergic reaction (i.e., ''blocking''), and that the novel food cue C does cause an allergic reaction. In the Test phase, subjects were presented with B, D, and E food cues. No feedback was given in the Test phase. Trial types are depicted in Table 1 in the main text. Subjects saw 6 repetitions of each trial type. Food cues were adapted from previous work, 33 and which food cues were assigned to which trial types were randomly determined per subject.

Additive blocking design
This design is identical to the Non-additive blocking design, except where noted. In the Pre-Learning phase, subjects this time learned a ''causal additivity'' rule: that a food cue (e.g., food cue I) could cause one magnitude of an allergic reaction (of a strength +), that a different food cue (e.g., food cue J) could cause a different magnitude of an allergic reaction (of a strength ++), and that both of these combined could cause a strong allergic reaction (of a strength +++).

Surveys
At the end of the experiment, subjects were redirected to a Qualtrics compilation of off-the-shelf surveys measuring different symptoms of psychosis (e.g., delusional thinking, paranoia, hallucinations, teleological reasoning, and probabilistic reasoning). Subjects completed the following surveys (with the order randomly determined per subject).
(a) Peters Delusion Inventory (PDI 34 ). Subjects completed a 21-item questionnaire. Items were adapted for healthy, nonpsychotic individuals by prefacing items with a relative, ''as if'' extension (e.g., ''Does it ever feel as if . ?''). For each item, subjects first reported whether they endorsed the relevant belief (e.g., ''Do you ever feel as if things in magazines or on TV were written especially for you?'', or ''Have your thoughts ever been so vivid that you were worried other people would hear them?''), and if they did, they then filled out three 5-point Likert scales assessing the degree of distress, pre-occupation, and conviction associated with the belief. , where they simply rated the degree to which they thought or felt a statement (e.g., ''I spent time thinking about friends gossiping about me'', or ''People have been hostile toward me on purpose'') in the last month -on the scale of 0 (Not at all) to 4 (Totally). (c) Belief in the Purpose of Random Events (BPRE 2 ). Subjects were asked to picture a situation (e.g., ''A person that you are attracted to kisses you in the middle of the street'') for 5 s -and were then given a scenario (e.g., ''You start going out together''). Subjects were then asked to rate the extent to which they believed that the first situation happened for a purpose -on the scale of 1 (The event definitely did not have a purpose) to 5 (The event clearly had a purpose). (d) Launay-Slade Hallucination Scale -Revised (LSHS-R 38 ). Subjects rated the degree to which a statement applied to them (e.g., ''In my daydreams, I can hear the sound of a tune almost as clearly as if I were actually listening to it'', or ''I often hear a voice speaking my thoughts aloud'') -on the scale of 1 (Certainly applies to me) to 5 (Certainly does not apply to me). (e) Binary Beliefs Task. 36 Subjects were presented different probabilities (e.g., mushrooms cause rashes, bananas cause vomiting, and that mangos cause both rashes and vomiting), and are asked to estimate the probability that a person ate something, given the following symptoms (e.g., ''A patient shows both rashes and vomiting. How likely is it that they had bananas?''). In this case, bananas should have a low probability, and mangos (as the simpler explanation) should have a high probability. The critical question, however, is how subjects then use these probabilities to reason about a different outcome. So now, suppose these different foods can potentially cause a fever, and we provide the probabilities of fever given a particular food (e.g., a low probability -''When a patient eats mangos, they rarely have a fever.'' versus a high probability -''When a patient eats bananas, they usually have a fever''). In this case, typically, the probability of an outcome is defined as: P(fever) = P(fever|bananas) x P(bananas) + P(fever|mangos) x P(mangos), but previous work has found that instead of being sensitive to the actual probabilities, people instead just binarize -treating P(bananas) as 0, and P(mangos) as 1. We computed a ''binary reasoning'' score by obtaining two different types of responses from the subjects (depicted in the two columns of Low-Low vs. Low-High in Table 2  a t = hjd t j + ð1 À hÞa t À 1 ; (Equation 9) where h˛R + ; h < 1 and is the moving average weight to smoothening the a t change. Thus: where 5 is the hadamard product (element wise multiplication). Finally, v 1 ! = 0:5, and the outcome was encoded as +1 and 0.