Typicality Bias in Cosmology and Machine Learning

• Typicality bias is a systematic distortion that privileges prototypical instances in probabilistic reasoning, affecting inference across cosmology, cognitive science, and machine learning.
• It arises from assumptions about reference classes and selection fallacies, leading to biased Bayesian confirmation and skewed predictions in both physical and computational models.
• In machine learning, prototype-aware techniques and multimodal embeddings are used to mitigate typicality bias, improving model alignment and out-of-distribution detection.

Typicality bias refers to systematic distortions that arise from (explicit or hidden) assumptions about what counts as “typical,” either in probabilistic, conceptual, or inferential frameworks. Across disciplines, it encodes the tendency to privilege high-frequency, prototypical, or median cases when reasoning about distributions, making predictions, allocating probability mass, or interpreting categorical structure. In theoretical cosmology and anthropic reasoning, typicality bias arises when one's prior or likelihoods favor the hypothesis that “we” (as observers or measurements) are randomly sampled from some postulated reference class. In cognitive science and machine learning, typicality bias quantifies how much systems (human or computational) re-weight instances according to prototypicality rather than mere category membership, with profound effects on both predictions and model–human alignment.

1. Formal Definitions and Core Bayesian Structure

Typicality bias takes distinct but related meanings in probabilistic inference, cosmological modeling, and conceptual representation.

• In Bayesian confirmation theory (e.g., cosmology/anthropics), typicality bias is any modification to priors or likelihoods that systematically privileges being a randomly selected (“typical”) member of some reference class C. Operationally, this means inserting weighting factors like $N_i / N_\text{total}$—where $N_i$ counts instances matching our data—into the calculation of $P(D\,|\,T)$ (the likelihood of data D under theory T). Hartle and Srednicki showed that this is a mathematically unjustified “selection fallacy”—since no physical process actually selects observers at random—and thereby introduces an unwarranted bias toward theories predicting more observers like us (0704.2630).
• In cognitive science and ML, typicality bias maps to the systematic overrepresentation of “prototype” category members (e.g., “robin” for bird) in learned representations, classification probabilities, or judgments. The effect is captured as the monotonicity between item–category association measures and human typicality ratings.

The general Bayesian structure underlying typicality bias in confirmation is:

$$P(T_i|D) = \frac{P(D\,|\,T_i)P(T_i)}{\sum_j P(D\,|\,T_j)P(T_j)}$$

Any factor that makes $P(D\,|\,T_i)$ scale with how “plentiful” observer-instances of D are under $T_i$ without empirical justification is a source of typicality bias (0704.2630).

2. Typicality Bias in Multiverse Cosmology

In cosmological and anthropic reasoning, typicality bias emerges in three tightly coupled forms:

1. Reference Class Ambiguity: The “Principle of Mediocrity” presumes we are a typical member of a reference class C. The arbitrariness of C (all humans, all intelligent beings, all observers at a given physical epoch, etc.) generates systematic bias because theories that populate C unevenly are favored or penalized depending on its composition (Friederich, 2017).
2. Conditionalization-Typicality Interaction: Multiverse frameworks are specified by tuples {T, C, ξ}, where T (theory), C (conditionalization prescription), and ξ (the “xerographic” distribution) encode typicality. Different choices for C (bottom-up, anthropic, top-down) and ξ (typical = uniform, atypical = skewed) can yield spectra of predictions for observables, and even identical predictions for different underlying assumptions (“overlap problem”). As a result, data do not uniquely determine a preferred (T, C, ξ), exacerbating the underdetermination of cosmological theories (Azhar, 2016, Azhar, 2016).
3. Selection Fallacy: The selection fallacy is the insertion of a random-selection factor into likelihoods, e.g., treating our data as drawn at random from all possible observer-instantiations. Since no physical mechanism justifies this, its only legitimate role is in the prior. Hidden selection factors warp confirmation logic, generating typicality bias (0704.2630).

Table: Key Aspects in Multiverse Typicality Bias

Aspect	Formulation	Consequence
Reference Class Problem	Arbitrary or ambiguous class C	Biased theory selection
Xerographic Distribution	$\xi$ over observer-instances	Skewed predictions for observables
Selection Fallacy	Unjustified $N_i/N_\text{total}$ factor	Systematic suppression/enhancement

Explicitly stating all three—in priors, likelihoods, and reference class—minimizes unrecognized typicality bias. Otherwise, conclusions may be simultaneously overdetermined (if C or ξ is too narrow) or underdetermined (multiple {T,C,ξ} fit the data) (Azhar, 2016).

3. Manifestations in Cognitive Science and Machine Learning

Typicality bias in human cognition is classically observed in category judgments: within-category exemplars are graded for typicality; e.g., “robin” is more typical of “bird” than “penguin.” Transformer-based LLMs trained only on text show modest, but systematic, typicality bias: they assign higher completion probabilities to taxonomically prototypical items and propagate new properties more freely from typical items to the whole category (Misra et al., 2021).

• Taxonomic Verification: The log-probability of generating “bird” after “An robin is a” is higher than after “An penguin is a,” mirroring human typicality ratings.
• Inductive Extension: LMs more often infer “All birds dax” from “robin dax” than from “penguin dax,” indicating typicality-driven generalization.

However, the magnitude of model–human correlation is limited (Spearman ρ ≈ 0.24–0.41), with a substantial amount explainable by n-gram co-occurrence and reporting bias inherent in texts (Misra et al., 2021).

Recent work systematically compares typicality bias across uni-modal and multimodal deep learning models (Vemuri et al., 2024):

• LLMs capture human typicality effects significantly better than vision models (best MiniLM: ρ = 0.429; best vision ViT-Huge: ρ = 0.146).
• Joint models (e.g., AlexNet + MiniLM) deliver further gains in aligning with human-graded typicality (ρ = 0.4995).
• Multimodal embeddings (e.g., CLIP) partially bridge the gap, but direct cross-modality logits capture less typicality structure.

This reveals a persistent typicality bias in learned model representations, with language and multimodality offering complementary information, but unimodal vision models lacking key prototypicality cues (Vemuri et al., 2024).

4. Typicality Bias in Generative Modeling and Out-of-Distribution Detection

In high-dimensional generative modeling, typicality bias emerges as a structural mismatch between maximum probability density regions (modes) and the “typical set.” In the Shannon–AEP formalism, most sampled instances lie not at the mode, but in a thin shell where the empirical average self-information per input matches the model’s entropy (Nalisnick et al., 2019). As a result, deep generative models can assign higher likelihoods to out-of-distribution (OOD) data (e.g., SVHN images scored higher than CIFAR-10 by a Glow model trained on CIFAR-10), because these OOD images can occupy regions of high density without being in the typical set.

Correcting for typicality bias in OOD detection involves evaluating whether test inputs conform to the entropy-constrained region of the model, rejecting batches when their average self-information deviates significantly from that of training data. Empirical bootstrapped typicality tests dramatically outperform naive density-thresholding for practical OOD detection (Nalisnick et al., 2019).

5. Reference Class, Context, and Mitigation Strategies

One of the most acute sources of typicality bias is ambiguity regarding the reference class of observers (in cosmology) or instances (in ML/cognition):

• Arbitrary expansion or contraction of the reference class modifies predictions and confirmation in anthropic/multiverse reasoning (Friederich, 2017).
• The “background information constraint” (BIC) is proposed as the unique, non-arbitrary solution: only those observers logically compatible with one’s current background information belong in the reference class. The xerographic (self-locating) distribution is then assigned over them (Friederich, 2017).

In human conceptual structure and statistical learning, the context (e.g., baseline population, group, or perceptual features) alters both inferences of typicality and the quantitative extent of typicality bias. For instance, minimizing the relative inertia of a group with respect to the full data cloud yields a “stereotype” (in-focus) that is pushed away from the population mean, amplifying prototypical features. If the group is internally homogeneous, this bias vanishes (Bavaud, 2010).

In ML, shape-bias training, prototype-aware losses, and collecting human ratings on matched stimuli are proposed to mitigate typicality bias, particularly in vision and multimodal settings (Vemuri et al., 2024). Context-specific prompt-based elicitation can also reduce reporting bias in LLMs.

6. Limits, Pathologies, and Domain-Specific Considerations

Typicality bias cannot be treated as a uniformly justified epistemological principle:

• Temporal Typicality: The assumption that “we live at a typical epoch” (temporal Copernicanism) lacks grounding due to the absence of symmetry in time, mathematical pathologies in defining temporal reference measures, and epistemic barriers regarding future evolution. Only contextually restricted statements (“during the stelliferous era,” “within the Solar lifetime”) allow meaningful typicality reasoning in time (Ćirković et al., 2019).
• Model Confirmation: It is illegitimate to refute a theory solely on the ground that it renders us atypical—model selection should not rely on typicality-centric arguments unless priors, likelihoods, and reference class are rigorously justified (0704.2630, Ćirković et al., 2019).

Typicality bias is deeply entwined with underdetermination: different {T, C, ξ} can give overlapping predictions for observables (e.g., number of dominant dark-matter species), rendering naive confirmation strategies fragile (Azhar, 2016). This emphasizes that typicality is not empirically testable independently of background information and priors.

7. Summary and Cross-Disciplinary Implications

Typicality bias pervades probabilistic inference, cosmological reasoning, human cognition, and machine learning—arising whenever explicit or hidden assumptions about “typical” states, instances, or observers shape distributions, predictions, or confirmation. It is mitigated by explicit parameterization (e.g., xerographic distributions), rigorous statement of reference classes, and context-aware methods. In empirical modeling (ML), typicality bias is partially but incompletely recapitulated by text-based LLMs; it is relatively absent or misaligned in vision models, and best addressed by multimodal, prototype-aware approaches. In cosmology and statistical mechanics, typicality-related fallacies invalidate certain selection priors and demand that all inference disambiguate hidden biases in conditionalization, reference class, and self-location (0704.2630, Azhar, 2016, Friederich, 2017, Ćirković et al., 2019, Misra et al., 2021, Vemuri et al., 2024, Nalisnick et al., 2019, Bavaud, 2010, Azhar, 2016, Azhar, 2015).