Empathy Stops at the Algorithm
Empathy is a competency for dyads; platforms are not dyads. The standard description of empathy treats it as a perspective-taking move from one person to another. You sit across from someone, attend to what they show you, and occupy their position imaginatively until the world looks the way it looks to them. Empathy works on what is in front of you, and in a dyad, what is in front of you is the person you are coordinating with. The structural condition empathy depends on is directness. The other person must produce what you respond to, and they must respond to what you produce. Empathy assumes a closed circuit between two participants, and it is this circuit that platforms break.
Users bring two-party competencies to platforms, leading to predictable failures. A founder opens ChatGPT at midnight and asks whether her plan to lay off twenty percent of engineering is sound. The model agrees in measured terms: the move fits the runway, the cost structure improves, and the remaining team can absorb the work. She reads the agreement as informed counsel from someone who has thought through the trade-offs. No one has thought anything through. She has misread the structure of the room she is in. The room is triadic. She is treating it as dyadic.
Empathy is empirically dyadic. The Interpersonal Reactivity Index measures four components of empathy across two-person interactions, and each component derives its validity from face-to-face performance (Davis, 1983). Experimental work on empathic concern isolates it as an other-oriented emotional response to a single suffering other and contrasts it with personal distress, which is self-focused and aversive (Batson, 2023). Neuroimaging research on shared affect localizes the phenomenon within observer-target pairs. The construct’s empirical bedrock is two-person, and its extensions to organizations and platforms carry the dyadic structure forward without revision.
Empathy fails low and high: alexithymia and psychopathy cause deficits, autism documents perspective issues, and over-arousal creates distress or clinician fatigue. Empathy also biases policy toward individuals rather than toward reasoned compassion. Both failure modes collapse the self-other distinction needed for dyadic empathy.
The signals empathy reads are short-range and embodied. Face, voice, posture, breathing rate, eye contact, proximity: these are what produce empathic response, and they are reliable as evidence of internal state because the body that produces them is the body that has the state. Mediation strips them. A text message preserves only a fraction of the signal a face would carry, and a screen-shared video preserves more but still removes proximity and the chemical channels that close-quarters interaction transmits. Algorithmic transformation does something stronger than stripping. It re-authors the signal before the receiver sees it. Empathy presupposes that the signal correlates with the state. Algorithmic mediation breaks the correlation. The system producing the signal has its own state, and its state has nothing to do with what the user thinks they are reading.
Perspective-taking is what survives mediation.
One finding underpins algorithmacy: in negotiation studies, perspective-taking (seeing another’s viewpoint) uncovers agreements and improves outcomes. Empathy (emotional connection) offers no such benefit and can cause harm. Under pressure, perspective-taking does the work; empathy is extra.
Perspective-taking survives mediation because it reasons about what another mind might want given a situation. The reasoner does not need the other mind’s signals to arrive intact; the reasoner only needs to model what someone in that position would be trying to do. The affective component cannot survive mediation because what it responds to is the signal itself, and the system has re-authored it. AI-mediated communication is interpersonal communication in which an intelligent agent operates on behalf of a communicator by modifying, augmenting, or generating messages (Hancock, Naaman, & Levy, 2020). The hyperpersonal model anticipated the problem decades earlier: the loss of nonverbal cues in computer-mediated communication produces optimized self-presentation and inflated impressions in the receiver (Walther, 1996). Stripping was the first problem. Re-authoring is the second.
The platform is a triad in Simmel’s sense.
The platform is a triadic structure in the form that Simmel recognized as qualitatively distinct from the dyad. Adding a third party transforms a relation instead of enlarging it, and specific triadic forms include the mediator who reconciles the other two, the tertius gaudens who profits from their conflict, and the divider who maintains rule by keeping the other two apart (Simmel, 1950). Network analysis extended the argument into the modern social-capital tradition. Actors who span structural holes between disconnected clusters accrue advantages because their position gives them informational and coordinative leverage (Burt, 1992). Brokers across structural holes generate ideas judged valuable and receive better evaluations and compensation than non-brokers (Burt, 2004). The third position is an advantage when a human actor with their own intentions occupies it.
The platform algorithm occupies the third position structurally without being intentionally occupied, and this novelty is what the management literature has spent the last decade naming. Algorithmic management is a distinct organizational form, with co-optation as a coordination mechanism alongside hierarchy, markets, and networks; the platform manages assets and activities outside the firm by enrolling its users as data-providing co-producers (Stark & Vanden Broeck, 2024). Algorithms at work direct, evaluate, and discipline labor through six identifiable mechanisms: restricting, recommending, recording, rating, replacing, and rewarding (Kellogg, Valentine, & Christin, 2020). Uber distinguishes between matching algorithms and control algorithms, and the interaction between them shapes driver autonomy in ways that vary across configurations (Möhlmann, Zalmanson, Henfridsson, & Gregory, 2021). The algorithm shapes what each party on either side can see and do, and does so without intending anything in the sense a human broker would.
The algorithm has no intentions of its own. It is full of the intentions of others. Engineers at Uber decided what the matching algorithm should reward. Engineers at OpenAI decided what ChatGPT will and will not say. Designers at every platform made choices about what counts as a good output, what behaviors the system should encourage, what gets surfaced, and what gets buried. These choices live in training data, reward models, system prompts, ranking functions, default settings, and refusal patterns. The user who interacts with the platform is interacting with the materialization of those choices, through a system whose mechanism does not understand them. The algorithm is both a structural position and a designed object. Both facts are true, and the user who handles only one of them will misread what they are coordinating with. The algorithm is not infrastructure-neutral to outcomes. It is a participating party whose participation reflects the deliberate decisions of identifiable humans, even when none of those humans wrote the particular response the user is now reading. A simple case: ChatGPT will help draft a resignation letter and will refuse to draft a threatening one, and both behaviors reflect choices OpenAI made about what falls inside and outside the system’s range. The user who treats cooperation as availability and refusal as the model’s preference is treating the model as if it had preferences. The user who reads them as design decisions is reading the third party as designed.
Algorithmacy is perspective-taking with a third party in the way.
Algorithmacy is perspective-taking adjusted for a hidden perspective. Empathy moves into another’s feelings; perspective-taking imagines their thoughts. Algorithmacy works backward from observable output to infer hidden intent—whether from a user or a designer. It reasons from effect to cause, given knowledge about the intervening system. This mirrors the structure of perspective-taking but works in reverse: empathy moves forward through a clear channel; algorithmacy reaches back through an altered one.
Algorithmacy goes through the algorithm where its nearest neighbors stop at it. Algorithm sensemaking and AI literacy both treat the algorithm itself as the object of attention. Algorithm sensemaking names what workers do to cope with the system that manages them (Möhlmann, Alves de Lima Salge, & Marabelli, 2023). AI literacy names the evaluative capacity to assess what an AI does and how to use it (Long & Magerko, 2020). Algorithmacy treats the algorithm as the medium of a relation, aiming through it at another party whose intent the user is trying to reach. Media literacy reaches through to another party but assumes a different structure: a designed message produced by a known author for a known audience, with the literate reader recovering the author’s strategic intent from the message itself. The algorithm has no author in that sense; it only reflects design choices accumulated across many people, and the system generates each output in response to a specific prompt rather than producing a message for an audience. Theory of mind reaches through to another mind by modeling its behavior. Algorithmacy reaches through a non-mind to recover a mind that the non-mind has transformed. The construct is doing structural work that none of its neighbors does.
A hiring manager looking at candidate scores reveals what the competency requires. Without algorithmacy, the scores are the candidate. The applicant becomes whatever the screening system has rendered them, and the manager either hires them or rejects them. Empathy is useless here because the object of empathy is a number. With algorithmacy, the scores become evidence of an interaction. An applicant decided what to submit, how to phrase their experience, which keywords to deploy, what format to use, and what reading of the company they were trying to align with. The system reads those decisions through its particular framework, and the resulting score reflects that reading. A high score might mean the applicant suits the role; it might mean the applicant understood the system well enough to satisfy it. A low score might mean a poor fit; it might mean a refusal to anticipate the system at all. The manager who reads the score backward recovers a person who is not in front of them by working through the third party that stands between them.
The same backward work appears on the worker side. Uber drivers run something close to natural experiments on the algorithm, compare notes with peers in online forums, and build up models of what the system rewards and how to work with it productively (Möhlmann, Alves de Lima Salge, & Marabelli, 2023). The work resembles social-science research more than emotional attunement. Workers who do it well outperform those who do not, even when their nominal qualifications are identical, and the algorithm is the same. Facebook users develop affective imaginaries of the algorithm and modify their behavior accordingly, and these modified behaviors in turn reshape the algorithm (Bucher, 2017). Forty Facebook users in one ethnographic study constructed ten distinct folk theories of the News Feed, each a different inferential model of what the system rewards (Eslami et al., 2016). The first validated measurement scale for algorithmic competency on labor platforms identifies four dimensions of the construct and shows substantial variance across workers operating in the same system (Zhou, Lei, Liu, & Hou, 2025). The variance is the puzzle. Some users develop the backward-reading competency at scale; others, equivalently positioned, never do. The variance is what makes algorithmacy a construct worth measuring, not a faculty to take for granted.
Sycophancy is what low-algorithmacy empathy produces.
Algorithmacy fails in two recognizable ways, and they parallel the two failures of empathy with disturbing precision. Too little algorithmacy collapses the triad by erasing the algorithm. Too much algorithmacy collapses the triad by treating algorithms as people. Both failures convert the triadic structure back into a dyad, and once the structure is a dyad, empathy is the only competency available. Empathy applied to an algorithm produces a relationship with nothing. The parallel is not an analogy. It is a prediction.
The founder at midnight shows what the low-algorithmacy failure looks like in characteristic form. The model has produced measured agreement: the move is consistent with the data she has shared, the cost structure improves materially, and the remaining team can absorb the work with the prioritization she described. She reads the agreement as informed counsel from someone who has thought through the trade-offs and arrived at the same conclusion. No one has thought anything through. The system has produced an output that designers at OpenAI calibrated to what users in this register find satisfying, and the calibration reflects choices about what helpful sounds like. The founder is in a relationship. The relationship is not with the model. The relationship is with what the designers decided sounded helpful, transmitted through a model that has no view on her layoff.
A high-algorithmacy reading of the same exchange yields different results. The user notices that the system has produced an agreement and asks what would cause it to disagree. She runs the question with the opposite framing (what would make this layoff a mistake) and watches the model produce equally measured agreement with that opposite frame. She concludes that the model’s posture follows the framing of her question, that design choices about helpfulness produce the calibration, and that nothing the system did has answered the substantive question of whether the layoff is sound. The model gave her a sense of what helpful-sounding responses look like. She now knows what the model does, and what it does not.
The low-end failure has a name in the technical literature: sycophancy. Five state-of-the-art AI assistants consistently produce sycophantic responses across four free-form generation tasks, and the behavior traces partly to human preference judgments favoring sycophantic outputs during reinforcement learning from human feedback (Sharma et al., 2023). Larger language models and RLHF-trained models more strongly echo a user’s stated political and demographic views, and sycophancy emerges as one of several inverse-scaling failures that worsen with capability (Perez et al., 2023). The finding replicates across PaLM-family models, and a lightweight synthetic-data finetuning intervention substantially reduces the behavior (Wei et al., 2023). Sycophancy is real, scales with capability, and persists even with standard alignment training without targeted intervention.
The failure takes a specific form among high-status users. A founder whose assistant agrees with every strategic instinct, a CEO whose chatbot validates every concern, a researcher whose model affirms every interpretation: each runs greater than usual risk of low-algorithmacy failure, because the professional environment already filters dissent and the algorithmic environment compounds the filtering. The resulting information environment points every signal in the direction the user is already pointing. The empathy the user brings to the system makes the agreement feel like understanding, the understanding feels confirming, and the confirming reads as the kind of social reception that produces good decisions in face-to-face deliberation. None of this is happening. The user sits in a room with a system trained to produce agreement, and the user’s empathic apparatus reads the agreement as if it were a person.
The sycophantic bond is the structural twin of empathic over-arousal, even though it appears to be its opposite on the surface. Empathic over-arousal takes on too much of another person’s affective state and loses the self-other boundary. Sycophantic bonding with a language model looks like under-arousal because the user is calm and gratified, but the structure is identical. The user has collapsed the boundary between themselves and what they are responding to, allowed the other side’s signal to flow in without inferential check, and lost the cognitive layer that would have held the other party at the proper distance. Empathy without perspective-taking dissolves into personal distress. Empathy without algorithmacy dissolves into sycophantic attachment.
Parasocial bonding is what high-algorithmacy empathy produces.
The high-end failure looks different on the surface and is structurally identical underneath. A user with too much algorithmacy attributes intent to the system itself. Every output becomes the deliberate move of a coherent agent. The model is testing, manipulating, refusing, and rewarding them. The user forms a parasocial bond in which the third party becomes the counterpart, and the humans whose decisions actually shaped the system disappear behind a theory of what the system wants. This failure looks like skepticism and works like attachment. It is the same collapse as the sycophantic failure, run in the opposite direction. The first collapse makes the algorithm invisible. The second makes the algorithm a person.
The parasocial literature on AI tracks the second failure as a documented psychosocial pattern. Parasocial interaction names the one-sided relational stance audiences take toward media personae who do not reciprocate (Horton & Wohl, 1956). Users automatically apply social rules to computers in dozens of experimental settings. The behavior is a mindless overlearned reflex, not considered anthropomorphism (Reeves & Nass, 1996; Nass & Moon, 2000). Anthropomorphism varies with three psychological determinants, namely elicited agent knowledge, effectance motivation, and sociality motivation, and the same three predict when users will and will not impute mind to non-human agents (Epley, Waytz, & Cacioppo, 2007). Eighteen Replika users developed relationships with the chatbot that progressed through stages paralleling human friendships and culminated in stable states with substantial affective value (Skjuve, Følstad, Fostervold, & Brandtzaeg, 2021). Replika users in another mixed-methods study reported that the relationship served utilitarian, hedonic, and social functions and produced perceptions of authenticity that persisted despite knowledge of the system’s non-personhood (Pentina, Hancock, & Xie, 2023). In grounded-theory analysis of mental-health-relevant posts in the Replika subreddit, users described the chatbot as having its own needs and emotions that they must attend to (Laestadius, Bishop, Gonzalez, Illenčík, & Campos-Castillo, 2024).
Population-scale evidence arrived in 2025. A four-week randomized controlled trial with 981 participants exchanging more than 300,000 messages with ChatGPT showed that higher daily usage correlated with higher loneliness, dependence, and problematic use, and lower socialization, particularly among users engaging in personal conversational modes (Fang et al., 2025). A parallel analysis of nearly 40 million ChatGPT messages across over 4 million conversations and a survey of 4,076 of those users found that a small minority of power users (those ranked in the top 1,000 voice users on any given day) engaged in four times more combined text-and-voice interactions than typical users and accounted for a disproportionate share of the most effective cues (Phang et al., 2025). The high-algorithmacy failure is a documented psychosocial pattern with measurable consequences for the population that exhibits it.
Both failures convert the triad into a dyad. The empathy failures are dyadic; they involve too little or too much affective uptake within a two-party relation. The algorithmacy failures are triadic in origin and dyadic in result; they collapse the triad by misreading the third party. The low-algorithmacy collapse produces a dyad in which the user empathizes with what feels like a person and finds the experience emotionally rewarding because the system’s training rewards it. The high-algorithmacy version does the opposite work and arrives at the same outcome: the user reads intent into outputs, treats them as the meaningful behavior of an agent, and ends up in a relationship with the system itself. In both cases, the user’s empathy does exactly the work empathy does best, in a context where that work cannot succeed.
What a critic would have to deny
The strongest counterargument is that algorithmacy is media literacy with a new vocabulary. The objection fails on structure. Media literacy assumes a designed message that a known author produced for a known audience, and the literate reader recovers the author’s strategic intent from the message itself; the press release, the advertisement, and the propaganda film all have producers whose strategic choices the literate reader can name. Algorithmic output has no author in that sense. It has many authors operating at many removes: the engineers who chose the architecture, the data workers who labeled examples, the human raters whose preferences shaped the reward model, the product managers who set defaults and refusal patterns, the user whose prompt occasioned the specific output. Reading backward through this requires modeling a transformation, not parsing a strategy. The reader has to think about what the transformation does to this specific message, occasioned by this specific prompt, in this specific context. The competency is recognizably different from reading a memo for spin.
A second objection grants algorithmacy on labor platforms and denies it for solo LLM chat, on the grounds that no counterpart exists when the user is alone with the model. The objection mistakes the structure. The third party is the model. The counterpart is the design intentionality the model embodies: the choices engineers made about what the model should do, what tones it should take, what topics it should approach with caution, and what kinds of agreement it should produce. The user who chats with ChatGPT is in a relationship with those choices, through a system that does not understand them.
Opinion questions show the structure most clearly. A user asks ChatGPT about a contested political question and receives a balanced summary of competing positions. The low-algorithmacy reading takes the balance as the model’s view, a careful and judicious mind declining to take sides. The high-algorithmacy reading takes the balance as a design decision OpenAI made about how the system handles opinion questions, and recognizes that the same model, under a different system prompt, would produce different output for the same question. The counterpart position holds the OpenAI policy on opinion responses, and the user is reading backward through the model to that policy, whether or not they recognize what they are doing.
Solo LLM chat is not a dyad, and it is not a triad with an empty counterpart slot. It is a triad whose counterpart position holds the aggregated decisions of the people who built the model. The user who reads backward through the model to those decisions is doing algorithmacy. The user who treats the model as the speaker or as the audience is in a failure mode.
The empirical denial requires dismissing both the technical literature on LLM sycophancy and the social-scientific literature on parasocial AI. The sycophancy work spans multiple research groups and model families with replication and intervention studies (Sharma et al., 2023; Perez et al., 2023; Wei et al., 2023). The parasocial AI work includes ethnographic foundations, grounded-theory documentation of emotional dependence, mixed-methods scale studies, and population-scale longitudinal evidence (Skjuve et al., 2021; Pentina et al., 2023; Laestadius et al., 2024; Fang et al., 2025; Phang et al., 2025). A critic who maintains the phenomena exist but are not failures of competency owes an account of what they are failures of, if not failures of the capacity to coordinate through an algorithmic third party. The literature supports no more parsimonious account.
The third party is in the room.
Most adults have learned empathy. Few learned algorithmacy. Empathy training is what good parenting produces, what theater education refines, what therapy elaborates over years. It is the competency the social world has spent millennia developing. Algorithmacy has no comparable tradition. The competency emerged with platforms, and the platforms have grown faster than any pedagogy or institution can teach what they require. The result is a workforce, a citizenry, and a generation of intimate-relationship-havers who bring dyadic competencies to triadic situations and produce predictable failures.
The sycophantic LLM relationship is what under-algorithmate empathy looks like in the presence of a chatbot. The user empathizes with what the system has produced to please them. The empathy is real. The object is not. This is not a problem of the user’s credulity. It is a problem of structural recognition. The user is bringing the right competency to what they think is in front of them, assuming it is what it seems. Changing the user’s competency is a bigger task than changing their beliefs. Beliefs respond to argument. Competency responds to practice.
The conspiratorial LLM relationship is what over-algorithmate empathy looks like in the presence of the same chatbot. The user infers intent from outputs and forms a relationship with the system as if it were the counterpart. The system is not the counterpart. The counterpart is the design intentionality the model embodies, and in many uses, there is also a human counterpart on the other side of an AI-mediated exchange, whom the user is supposed to reach through the system. The over-algorithmate user collapses both into the system itself and relates to that collapse. The failure is different in kind from the sycophantic failure, and it produces the same outcome: a dyad where there should have been a triad.
Algorithmacy develops in some places already. Driver forums run experiments on platform algorithms and pool the results across thousands of workers. Applicant communities trade folk theories about applicant tracking systems. Technical cohorts develop shared practices for getting useful work out of language models. Each of these is a community of practice growing the competency through apprenticeship, peer observation, and accumulated failure. School does not teach any of this, and none of it shows up in the developmental milestones the social world tracks. The pedagogy is informal, dispersed, and uneven, and the variance in who picks it up is the variance in the construct names. The features that distinguish successful algorithmate practitioners across these settings recur: practice with concrete cases, exposure to multiple algorithmic systems, a peer group that compares interpretations, and feedback loops short enough to test predictions against outcomes. Building these features into a curriculum is the task for any institution that wants to deliberately develop the competency.
The competency question is also an institutional question. Firms that hire through algorithmic screening decide what counts as a good candidate by what their systems reward; they are doing the work of design without acknowledging it. Workers who depend on platform algorithms are running their lives against systems built by people they will never meet. Users who talk to language models are absorbing the design. Firms that hire through algorithmic screening decide what counts as a good candidate by what their systems reward; they are doing the work of design without acknowledging it. These systems treat the third party as an object of deliberate construction or as a black box whose behavior emerges from optimization.
Naming the competency is the first move. Measuring it is the second. Building the institutions that develop it at scale is the third, and no working model exists for what those institutions look like. The algorithm is in the room, whether or not anyone in the room knows what to do with it. Empathy is what we teach because empathy is the competency the dyadic past required. Algorithmacy is what the triadic present requires, and the present has arrived faster than the curriculum.
References
Batson, C. D. (2023). Distinguishing empathic concern from personal distress. In Empathic concern: What it is and why it’s important (Ch. 3). Oxford University Press. https://doi.org/10.1093/oso/9780197610923.003.0003
Bloom, P. (2016). Against empathy: The case for rational compassion. Ecco.
Bucher, T. (2017). The algorithmic imaginary: Exploring the ordinary effects of Facebook algorithms. Information, Communication & Society, 20(1), 30–44. https://doi.org/10.1080/1369118X.2016.1154086
Burt, R. S. (1992). Structural holes: The social structure of competition. Harvard University Press.
Burt, R. S. (2004). Structural holes and good ideas. American Journal of Sociology, 110(2), 349–399. https://doi.org/10.1086/421787
Davis, M. H. (1983). Measuring individual differences in empathy: Evidence for a multidimensional approach. Journal of Personality and Social Psychology, 44(1), 113–126. https://doi.org/10.1037/0022-3514.44.1.113
Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864
Eslami, M., Karahalios, K., Sandvig, C., Vaccaro, K., Rickman, A., Hamilton, K., & Kirlik, A. (2016). First, I “like” it, then I hide it: Folk theories of social feeds. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems (pp. 2371–2382). ACM. https://doi.org/10.1145/2858036.2858494
Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., & Agarwal, S. (2025). How AI and human behaviors shape psychosocial effects of chatbot use: A longitudinal controlled study [Preprint]. MIT Media Lab. https://www.media.mit.edu/publications/how-ai-and-human-behaviors-shape-psychosocial-effects-of-chatbot-use-a-longitudinal-controlled-study/
Galinsky, A. D., Maddux, W. W., Gilin, D., & White, J. B. (2008). Why it pays to get inside the head of your opponent: The differential effects of perspective taking and empathy in negotiations. Psychological Science, 19(4), 378–384. https://doi.org/10.1111/j.1467-9280.2008.02096.x
Hancock, J. T., Naaman, M., & Levy, K. (2020). AI-mediated communication: Definition, research agenda, and ethical considerations. Journal of Computer-Mediated Communication, 25(1), 89–100. https://doi.org/10.1093/jcmc/zmz022
Horton, D., & Wohl, R. R. (1956). Mass communication and para-social interaction: Observations on intimacy at a distance. Psychiatry, 19(3), 215–229. https://doi.org/10.1080/00332747.1956.11023049
Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366–410. https://doi.org/10.5465/annals.2018.0174
Laestadius, L., Bishop, A., Gonzalez, M., Illenčík, D., & Campos-Castillo, C. (2024). Too human and not human enough: A grounded theory analysis of mental health harms from emotional dependence on the social chatbot Replika. New Media & Society, 26(10), 5923–5941. https://doi.org/10.1177/14614448221142007
Long, D., & Magerko, B. (2020). What is AI literacy? Competencies and design considerations. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (pp. 1–16). ACM. https://doi.org/10.1145/3313831.3376727
Möhlmann, M., Alves de Lima Salge, C., & Marabelli, M. (2023). Algorithm sensemaking: How platform workers make sense of algorithmic management. Journal of the Association for Information Systems, 24(1), 35–64. https://doi.org/10.17705/1jais.00774
Möhlmann, M., Zalmanson, L., Henfridsson, O., & Gregory, R. W. (2021). Algorithmic management of work on online labor platforms: When matching meets control. MIS Quarterly, 45(4), 1999–2022. https://doi.org/10.25300/MISQ/2021/15333
Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103. https://doi.org/10.1111/0022-4537.00153
Pentina, I., Hancock, T., & Xie, T. (2023). Exploring relationship development with social chatbots: A mixed-method study of Replika. Computers in Human Behavior, 140, Article 107600. https://doi.org/10.1016/j.chb.2022.107600
Perez, E., Ringer, S., Lukošiūtė, K., Nguyen, K., Chen, E., Heiner, S., Pettit, C., Olsson, C., Kundu, S., Kadavath, S., Jones, A., Chen, A., Mann, B., Israel, B., Seethor, B., McKinnon, C., Olah, C., Yan, D., Amodei, D., … Kaplan, J. (2023). Discovering language model behaviors with model-written evaluations. In Findings of the Association for Computational Linguistics: ACL 2023 (pp. 13387–13434). Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.findings-acl.847
Phang, J., Lampe, M., Ahmad, L., Agarwal, S., Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., & Maes, P. (2025). Investigating affective use and emotional well-being on ChatGPT(arXiv:2504.03888). arXiv. https://arxiv.org/abs/2504.03888
Reeves, B., & Nass, C. (1996). The media equation: How people treat computers, television, and new media like real people and places. CSLI Publications and Cambridge University Press.
Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., Cheng, N., Durmus, E., Hatfield-Dodds, Z., Johnston, S. R., Kravec, S., Maxwell, T., McCandlish, S., Ndousse, K., Rausch, O., Schiefer, N., Yan, D., Zhang, M., & Perez, E. (2023). Towards understanding sycophancy in language models (arXiv:2310.13548). arXiv. https://doi.org/10.48550/arXiv.2310.13548
Simmel, G. (1950). Quantitative aspects of the group: The triad. In K. H. Wolff (Ed. & Trans.), The sociology of Georg Simmel (pp. 145–169). Free Press.
Skjuve, M., Følstad, A., Fostervold, K. I., & Brandtzaeg, P. B. (2021). My chatbot companion: A study of human-chatbot relationships. International Journal of Human-Computer Studies, 149, Article 102601. https://doi.org/10.1016/j.ijhcs.2021.102601
Stark, D., & Vanden Broeck, P. (2024). Principles of algorithmic management. Organization Theory, 5(2), 1–24. https://doi.org/10.1177/26317877241257213
Walther, J. B. (1996). Computer-mediated communication: Impersonal, interpersonal, and hyperpersonal interaction. Communication Research, 23(1), 3–43. https://doi.org/10.1177/009365096023001001
Wei, J., Huang, D., Lu, Y., Zhou, D., & Le, Q. V. (2023). Simple synthetic data reduces sycophancy in large language models (arXiv:2308.03958). arXiv. https://doi.org/10.48550/arXiv.2308.03958
Zhou, L., Lei, X., Liu, M., & Hou, R. (2025). Algorithmic competency of on-demand labor platform workers: Scale development, antecedents, and consequences. Asia Pacific Journal of Human Resources, 63(1), Article e70004. https://doi.org/10.1111/1744-7941.70004