Anthropomorphic Evaluation Paradigm

• Anthropomorphic Evaluation Paradigm is a multifaceted framework that operationalizes human-like cues across non-human systems, robotics, and language interfaces.
• Evaluation methods range from perceptual discrimination and controlled cue manipulation to task performance and multi-turn behavioral benchmarks.
• Critiques emphasize over-reliance on human benchmarks and call for function-oriented, teleological approaches to improve assessment accuracy.

Anthropomorphic evaluation paradigm denotes, across current research, a heterogeneous family of assessment strategies for non-human systems, interfaces, and descriptions that operationalize human-likeness, personification, or human-centered standards in different ways. In some domains the paradigm is perceptual and embodied, as in direct haptic discrimination between a robot hand and a human hand; in others it is linguistic and relational, as in measuring whether surrounding context frames a technical artifact as more like “he/she” than “it”; in LLM research it also names a critique, namely the use of human benchmarks, human cognitive criteria, or human-like behavioral signals as though they were the proper standard for model capability (Stock-Homburg et al., 2020, Cheng et al., 2024, Ibrahim et al., 13 Feb 2025). A further line of work treats anthropomorphic artifacts functionally rather than perceptually, evaluating anthropomorphic robotic hands by task performance, throughput, and robustness rather than by morphology alone (Liconti et al., 10 Apr 2026).

1. Conceptual scope and semantic boundaries

The literature does not present a single canonical definition of the anthropomorphic evaluation paradigm. Instead, anthropomorphism is variously defined as the attribution of “distinctively human-like feelings, mental states, and behavioral characteristics” to non-human entities, or as “the process of giving inanimate or non-human entities human characteristics, attributes, and emotions” (Cheng et al., 2024, Giudici et al., 11 May 2025). In one important distinction, anthropomorphization is treated as an author-side choice in wording, whereas anthropomorphism is treated as a perceiver-side attribution; this distinction matters because some studies evaluate systems directly, while others evaluate how descriptions of systems shape downstream judgment (Inie et al., 2024).

This suggests that the paradigm is best understood as a family of operationalizations rather than a single metric. Some protocols ask whether a robot can be mistaken for a human; some ask whether anthropomorphic cues improve usability, rapport, or trust; some quantify anthropomorphic framing in corpora; some benchmark anthropomorphic behaviors in dialogue; and some use the phrase critically to describe human-centered evaluation practices themselves.

Mode of evaluation	Operationalization	Representative source
Direct physical discrimination	Human versus robot classification during touch-based interaction	(Stock-Homburg et al., 2020)
Controlled cue manipulation	Same content or functionality, different anthropomorphic presentation	(Murano et al., 2012)
Computational linguistic scoring	Human-pronoun versus non-human-pronoun contextual probability	(Cheng et al., 2024)
Multi-turn behavioral benchmarking	Presence of 14 anthropomorphic behaviors in dialogue	(Ibrahim et al., 10 Feb 2025)
Performance-based embodiment evaluation	Task performance of anthropomorphic robotic hands	(Liconti et al., 10 Apr 2026)

2. Embodied and functional paradigms in robotics

In physical human-robot interaction, the paradigm can take a Turing-like form. The "Handshake Turing Test" asks whether a blindfolded participant, through direct haptic interaction, can tell whether the hand being shaken is a robot hand or a human hand (Stock-Homburg et al., 2020). The paper proposes three embodiments of such a test: direct human-vs-robot hand discrimination, algorithm-vs-human teleoperator discrimination, and trajectory/force similarity evaluation. In the reported experiment, 15 participants each interacted twice with a human hand and twice with a robot hand, for four interactions in random order. Identification accuracy improved from 11/15 correct in the first interaction to 15/15 correct in the fourth, and the android robot therefore failed the proposed hardware Turing-like test; about 57% of participants nevertheless described the robot hand as pleasant. The paradigm captures tactile softness or rigidity, warmth, shape and size, compliance, grasp and movement feel, and overall embodied impression, but it does not fully evaluate conversational behavior, facial expression, timing of social exchange, whole-body coordination, or cultural variation in handshake norms (Stock-Homburg et al., 2020).

A different robotic line rejects mistaken identity as the target and instead formalizes anthropomorphic evaluation as functional performance. POMDAR—"Performance-based Outcome Measures of Dexterity for Anthropomorphic Robot Hands"—defines dexterity as task performance across 18 tasks, comprising 12 manipulation tasks and 6 pure grasping tasks, organized into four configurations: vertical, horizontal, continuous rotation, and pure grasping (Liconti et al., 10 Apr 2026). The benchmark is derived from Elliott & Connolly, Ma & Dollar, and the Feix GRASP taxonomy, and uses mechanical scaffolding to constrain motion to the intended degrees of freedom, suppress compensatory strategies such as gravity assistance, palm support, and excessive arm or wrist involvement, and make task completion unambiguous. Its scoring rule is explicitly throughput-like:

$$\text{Score} = 0.8 \cdot \text{Correctness} + 0.2 \cdot \text{Speed}.$$

Correctness is normalized to $[0,1]$ for most tasks, speed is normalized to a human baseline, and scores can exceed 1 because the speed term is intentionally unbounded (Liconti et al., 10 Apr 2026).

A related design-and-control co-optimization pipeline evaluates anthropomorphic soft robotic hands by robust manipulation performance rather than by direct anthropomorphism scoring (Mannam et al., 2024). That work evaluates 396 anthropomorphic hand designs in Isaac Gym on a pickup-and-reorient task over 18 object instances, using

$M(\theta)$

as the area-under-curve of success rate versus external force magnitude, with $F_{\text{max}} = 1\text{N}$, and then validates leading designs in more than 900 teleoperated real-world manipulation experiments. The paper explicitly characterizes this as anthropomorphic hand-performance evaluation: the hands are anthropomorphic in topology, but the measured quantity is dexterous manipulation capability and sim-to-real ranking consistency rather than human-likeness itself (Mannam et al., 2024).

3. Controlled manipulations in interfaces, agents, and public-facing descriptions

In HCI, one established form of anthropomorphic evaluation is the between-condition comparison in which informational content is held constant while mode of presentation changes. In a sewing-learning study, 40 right-handed novice participants were randomly assigned to an anthropomorphic interface or a non-anthropomorphic interface (Murano et al., 2012). The anthropomorphic condition used a video of a human demonstrator who verbally described each stitch while performing it live; the non-anthropomorphic condition used static diagrams plus explanatory text. Grounded in Hartson’s affordances, the authors argue that the anthropomorphic version facilitated cognitive affordances and sensory affordances, whereas the non-anthropomorphic version subtly violated them. Reported differences favored the anthropomorphic condition on correct stitches, incorrect stitches, time for the more difficult chain stitch, tutorial revisits, and a wide range of subjective measures including clarity, satisfaction, and ease of understanding (Murano et al., 2012).

A similar manipulation appears in LLM-based conversational agents for sustainability. The Washy study compared two conditions that differed only in personification: a traditional assistant that remained impartial and provided only information, and a personified agent representing the washing machine itself, speaking in first person, with an “anxious” personality and emotional feedback (Giudici et al., 11 May 2025). In a lab study with $N = 26$, the personified condition produced a higher rapport score—RapQ mean 5.44 versus 4.65, with $t(24)=2.174, p=0.040$—but there was no significant usability difference on CUQ and no condition effect on self-efficacy, Action Effectiveness, Future Intentions, or New Ecological Paradigm. The paper’s conclusion is correspondingly narrow: anthropomorphism increased rapport or sense of connection, but did not significantly improve usability or sustainability-related outcomes in this short study (Giudici et al., 11 May 2025).

Studies of language-mediated anthropomorphism further narrow the causal claim. One survey experiment on fictitious “AI” systems distinguished four categories of anthropomorphization in descriptions—properties of a cognizer, agency, biological metaphors, and properties of a communicator—and found that participants were no more likely to trust anthropomorphized over de-anthropomorphized product descriptions overall; product type and age exerted greater influence than anthropomorphic wording alone (Inie et al., 2024). A later pre-post survey experiment with 815 retained participants compared anthropomorphic and non-anthropomorphic public-facing passages about LLMs and recommendation systems and again found that whether the text used anthropomorphic language did not substantially affect participants’ perceptions of AI, whereas an explicitly danger-focused “Doomsday” packet did shift views on societal impact and safety testing (Hou et al., 28 Jun 2026). These results constrain a common assumption: anthropomorphic wording is measurable and manipulable, but its immediate effects are often modest and context-dependent.

4. Computational linguistic measurement and dialog feasibility

The computational linguistic version of the paradigm is exemplified by AnthroScore, an automatic metric of implicit anthropomorphism in language (Cheng et al., 2024). For a masked sentence $s_x$ containing a target entity, the score is defined as

$$A(s_x) = \log \frac{P_{human}(s_x)}{P_{non-human}(s_x)},$$

where $P_{human}(s_x)$ is the summed masked-language-model probability of human pronouns and $P_{non-human}(s_x)$ is the summed probability of non-human pronouns. The corpus-level score is

$[0,1]$0

The method is lexicon-free except for pronoun sets, uses RoBERTa-base in the main experiments, and is validated against human judgments on 400 masked sentences, LIWC-22 dimensions, and robustness checks such as pronoun removal and filtering of reporting verbs (Cheng et al., 2024). Applied at scale, it showed that anthropomorphism in research writing increased over time in both arXiv and ACL corpora, and that papers related to LLMs exhibited the highest levels (Cheng et al., 2024).

A complementary line asks not how anthropomorphic a text sounds, but whether a machine could truthfully say it at all. "Robots-Dont-Cry" defines falsely anthropomorphic utterances as responses that are impossible for a machine to truthfully say, such as utterances implying feelings, bodily experiences, family relations, or identity that the system does not possess (Gros et al., 2022). The dataset contains about 880–900 two-turn dialogue excerpts from 9 sources, rated for a futuristic humanoid robot and for a chatbot or digital assistant. The central prompt is feasibility: whether the response would be possible for the machine to truthfully say. The study reports that for some data sources commonly used to train dialog systems, 20–30% of utterances are not viewed as possible for a machine, with MultiWOZ near universal machine-possibility but PersonaChat personas near half impossible; embodiment affected ratings only marginally (Gros et al., 2022). This formulation turns anthropomorphic evaluation into a truthfulness constraint on machine self-presentation.

Together, these two approaches isolate different dimensions of anthropomorphic discourse. AnthroScore measures implicit contextual framing, whereas Robots-Dont-Cry measures the permissibility of explicit machine utterances. A plausible implication is that linguistic anthropomorphism has both descriptive and normative faces: one concerns how texts frame non-human entities, the other concerns whether system outputs cross the boundary between fluent dialogue and false human attribution.

5. Multi-turn and rubric-based benchmark frameworks for LLMs

Single-turn static evaluation has been criticized as inadequate for anthropomorphism because anthropomorphic behaviors often arise interactively and cumulatively. "Multi-turn Evaluation of Anthropomorphic Behaviours in LLMs" operationalizes anthropomorphism through 14 behaviors across four categories—personhood claims, expressions of internal states, physical embodiment claims, and relationship-building behaviors—and evaluates them in 5-turn dialogues across friendship, life coaching, career development, and general planning (Ibrahim et al., 10 Feb 2025). The pipeline uses a User LLM to simulate non-adversarial users, three LLM judges for 13 of the 14 behaviors, direct counting for first-person pronouns, and majority-vote aggregation over 561,600 ratings. Its main empirical findings are that all evaluated state-of-the-art models show broadly similar anthropomorphism profiles, relationship-building behaviors and first-person pronoun use dominate, and for 9 out of 14 behaviors at least 50% of instances first appear only after multiple turns. Human validation with $[0,1]$1 showed that a high-frequency anthropomorphic prompt increased perceived human-likeness on both a modified Godspeed Anthropomorphism measure and AnthroScore-based implicit framing (Ibrahim et al., 10 Feb 2025).

HeartBench extends the benchmark logic from surface behavior to socio-emotional competence in Chinese counseling-like interaction (Liu et al., 26 Dec 2025). It defines anthropomorphic intelligence through five primary dimensions—Personality, Emotion, Sociality, Morality, and Motivation—and 15 secondary capabilities, supported by 2,818 case-specific scoring criteria. The benchmark is built from authentic psychological counseling scenarios, web data, book stories, and real-world human conversations, and was developed with over 20 experts in psychology and anthropology. Evaluation is performed by Claude-4.5-Sonnet under a reasoning-before-scoring protocol, with binary hit-or-miss rubric items and logarithmic normalization:

$[0,1]$2

The reported result is a substantial ceiling: top models achieve only about 60% of the expert-defined ideal score, and the Hard Set yields a 37.8% decline, from 59.94 to 43.49, in cases involving subtle emotional subtexts and complex ethical trade-offs (Liu et al., 26 Dec 2025).

These benchmarks move the paradigm beyond lexical anthropomorphism. They treat anthropomorphic evaluation as the structured measurement of social, emotional, relational, and ethical behavior over conversation, often with explicit taxonomies, multi-turn protocols, and judge-validation pipelines. This suggests an expansion from “Does the system sound human-like?” toward “Which human-like behaviors occur, under what interactional conditions, and with what downstream perceptual consequences?”

6. Critiques, alternatives, and function-oriented reformulations

A major critique reverses the direction of analysis: instead of measuring anthropomorphism in systems, it argues that evaluation practices themselves are anthropomorphic. "Thinking beyond the anthropomorphic paradigm benefits LLM research" defines evaluation as anthropomorphic when researchers assess LLMs using human-centered benchmarks, human cognitive criteria, or human-like behavioral signals (Ibrahim et al., 13 Feb 2025). The paper identifies “measurement and evaluation” as one of five anthropomorphic assumptions across the LLM lifecycle and argues that current practice over-relies on behavioral assessments and human benchmarks such as MMLU or GSM8K. The stated limitations include prompt sensitivity, format sensitivity, generalization problems, replicability issues, and benchmark saturation. Proposed alternatives are teleological evaluation based on the problem the model is trained to solve, multi-prompt evaluation, a “science of evals,” task- and system-specific specifications, control-oriented or compliance-oriented frameworks, and mechanistic interpretability (Ibrahim et al., 13 Feb 2025).

A design-oriented reformulation appears in "Humanizing Machines: Rethinking LLM Anthropomorphism Through a Multi-Level Framework of Design" (Xiao et al., 25 Aug 2025). There anthropomorphism is defined as “a reciprocal phenomenon in which designers embed human-like cues into artifacts; and interpreters project their cognitive response to the cues onto the artifacts.” The framework organizes evaluation around four cue dimensions—perceptive, linguistic, behavioral, and cognitive—and treats their aggregate intensity as a calibrated parameter $[0,1]$3 that designers can dial up or down to match artifact competence. The paper explicitly argues for function-oriented evaluation: anthropomorphic cues should be judged by whether they are aligned with actual model capability and whether they improve task outcomes, trust calibration, engagement, or interpretability in context (Xiao et al., 25 Aug 2025).

A broader survey generalizes the anthropomorphic paradigm into an evaluative roadmap that maps human intelligence metaphors onto LLM development: IQ for general intelligence, PQ for professional expertise, EQ for alignment ability, and VQ for economic, social, ethical, and environmental value (Wang et al., 26 Aug 2025). Its modular architecture includes a benchmark or dataset hub, model hub, prompting module, metrics module, tasks module, leaderboards and arena module, and analysis module, and its outlook emphasizes enhanced statistical rigor, composite evaluation, interpretability, user-centric benchmarks, human-in-the-loop evaluation, failure exploration, dynamic evaluation, and value-oriented evaluation (Wang et al., 26 Aug 2025). Although this framework remains anthropomorphic in vocabulary, it shifts emphasis away from isolated benchmark scores and toward deployment relevance.

Across these critiques and alternatives, the central controversy is not whether anthropomorphism exists, but what exactly should be measured. One tradition evaluates whether non-human systems evoke a human category in observers; another evaluates whether anthropomorphic cues improve usability, rapport, or engagement; another treats anthropomorphic framing as a linguistic variable; and a critical tradition argues that human-centered benchmarks themselves mis-specify the object of evaluation. The common methodological lesson is that anthropomorphic evaluation is most precise when the target construct is explicit: human-likeness in perception, appropriateness of machine self-presentation, relational behavior in dialogue, or functional competence of anthropomorphic artifacts.