Cleaner fish are sensitive to what their partners can and cannot see

The study investigates whether cleaner wrasse possess a core component of theory of mind (ToM): representing what conspecific partners can and cannot see in strategic interactions. Prior work in non-human primates shows sensitivity to others’ perceptual access and knowledge in competitive contexts, raising the question of whether such abilities require large brains or can evolve under ecological pressures favoring strategic deception. The mutualistic but conflict-prone cleaner–client system provides a natural context: females benefit from consuming preferred client mucus (cheating) but risk punishment by males. The authors test the hypothesis that female cleaners strategically modulate cheating based on a male partner’s visual access, predicting more cheating when unobserved and links between strategic behavior and experienced punishment.

Comparative research demonstrates that non-human primates can represent what others perceive and know, often revealed in competitive paradigms: rhesus monkeys infer perceptual access behind barriers; subordinate chimpanzees preferentially take food hidden from dominants; apes and monkeys track what others have seen and sometimes conceal visual and auditory information to deceive. Corvids also show components of these abilities. These findings suggest foundational ToM-like mechanisms outside humans, though mechanisms remain debated. In cleaner wrasse, previous work shows sophisticated social strategies in cooperative cleaning, sensitivity to client behavior, and punishment dynamics between partners, indicating potential selection for perspective-taking and strategic deception in an ecologically relevant, high-frequency interaction.

Participants and housing: Twelve male–female pairs of Labroides dimidiatus were collected near Moorea, French Polynesia (March–April 2016) and housed at Gump Field Station in outdoor circular tanks (2.1 m diameter, ~0.386 m height) subdivided into 12 sections (~120 × 30 × 30 cm). Flow-through seawater maintained ambient temperature; natural light conditions prevailed. Sections contained PVC shelters. Pairs acclimated for 2–3 days and were fed daily; food items for experiments included flake (lower value) and prawn (high value, preferred) paste dots on plexiglass plates. Training: Individuals learned that plates were removed when a prawn item was consumed, simulating a client leaving. Training ensured subjects would consume flake items and experience removal contingencies. Reliability checks compared live counts of consumed flakes with blind counts and video-coded counts (high agreement: r = 0.96 and r = 0.92, respectively). Study 1 design: Females and males were separated by either a transparent (male visible) or opaque (male not visible) partition. The female was presented with a plate containing 12 flake and 2 prawn items. Plates remained until the female ate a prawn (cheating), at which point the plate was removed, the male was released into the female’s compartment, and punishment was monitored for ~30 s. Each pair received 24 trials (12 per visibility condition), counterbalanced within blocks. Main measures: number of flake items consumed before the first prawn (higher numbers indicate greater cooperation), and whether male punishment occurred (yes/no). Study 2 design: Females began behind a transparent partition with a central door. Two identical plates (12 flake, 2 prawn) were placed behind a transparent versus an opaque barrier within a choice arena while the male was either visible (transparent partition) or not visible (opaque partition). When plates were in place, the door was lifted and the female chose one side; the unchosen plate was removed. After the first prawn consumption, the plate was removed, the male was released, and punishment monitored. Trials were presented in two blocks (each condition practiced across six trials per block; total 24 trials). Measures: female choice of barrier (opaque vs transparent), male punishment, and female cooperation (flakes before prawn). Coding and analysis: Punishment was recorded live (presence/absence). Statistical analyses used generalized linear mixed models and likelihood ratio tests in R 3.6.3, with appropriate random effects structures. Strategic cheating score (Study 1) was computed as the difference in flakes consumed between male visible and male not visible conditions for each female. Data and code availability statements: raw data in Supplementary Data 1–2; R code available upon request.

Study 1 (N = 12 pairs): Females ate fewer flake items when males could not see them (male not visible: Mean = 3.6, SD = 3.2) than when males could see them (male visible: Mean = 4.5, SD = 3.1), indicating earlier cheating when unobserved and greater cooperation when observed (condition effect consistent with predictions; effect of round: LRT χ² = 12.46, p = 0.006). Male punishment did not depend on an interaction between condition and female cooperation; males were less likely to punish as females ate more flakes (greater cooperation; LRT χ² = 4.18, p = 0.004). Females’ strategic cheating score correlated positively with male punishment (r = 0.64, N = 12, p = 0.02), suggesting females paired with more punitive males showed stronger condition-dependent adjustments. Study 2 (N = 11 pairs): Female barrier choices showed a significant interaction between male-visibility condition and block (LRT χ² = 12.23, p < 0.001). In block 1, when males were visible, females preferentially chose to feed behind the opaque barrier (consistent with strategic concealment); in block 2 this preference reversed, with more choices behind the transparent barrier when males were visible. Males were more likely to punish when females chose to feed behind the opaque barrier, regardless of condition (LRT χ² = 4.48, p = 0.034). Female cooperation (flakes before prawn) showed only marginal, non-robust interactions with condition and choice; models including these interactions were not better than null.

Cleaner wrasse adjusted their behavior based on a partner’s visual access: females cheated earlier when unobserved and cooperated more when observed, paralleling primate performance in competitive ToM tasks. Individual differences further indicate strategic calibration: females experiencing more punitive males exhibited stronger condition-dependent cheating, consistent with adaptive adjustment to partner behavior. In Study 2, initial choices reflected strategic use of visual barriers to conceal cheating when males could observe, though this effect diminished or reversed with repeated exposure, potentially due to male responses (greater punishment after opaque-choice trials) or other learned contingencies. Male punishment patterns appeared to depend less on direct visual access and more on cues associated with female choices (e.g., choosing to feed behind opaque barriers), suggesting that both sexes may attend to behavioral signals beyond immediate perceptual access. These findings imply that ecological pressures in the cleaner mutualism—where deception and punishment co-occur—can foster perspective-sensitive strategies reminiscent of ToM components.

Female cleaner wrasse are sensitive to what conspecific partners can and cannot see and strategically adjust cheating accordingly, especially when paired with punitive males. This sensitivity to others’ perceptual access—a hallmark of theory of mind-like cognition—emerges in a small-brained, distantly related vertebrate within an ecologically valid cooperative context. Future work should disentangle the cues guiding male punishment and female cheating decisions, test generalization to other social contexts (e.g., bystander clients), refine barrier paradigms to isolate asymmetric visual access, and examine learning dynamics across repeated interactions.

In Study 1, the design could not produce an asymmetric barrier allowing females to see males without males seeing females, leaving open alternative interpretations tied to general barrier context. Effects in Study 2 were block-dependent and attenuated with repeated trials, suggesting learning or habituation influences. Sample sizes were modest (12 pairs; 11 in Study 2), and punishment was coded as a binary outcome over a short window, potentially missing gradations of response. Male punishment appeared influenced by cues correlated with female choices (e.g., opaque selections), introducing possible confounds. The laboratory arena and artificial plates, while ecologically inspired, may not fully capture the complexity of natural interactions.