What happened
Researchers Selcen Ozturkcan, Jean-Paul de Cros Peronard and Inci Toral-Manson introduce “robotoid humanness,” a proposed form of self-narrowing in which people increasingly align their self-presentation with machine-readable categories, pacing and logic. Their AI & Society paper describes a three-stage mirroring mechanism and sets out propositions for future empirical testing.
The paper, published in AI & Society on 19 August 2026 as an open-access “Open Forum” article, treats consumer-facing AI as more than a tool or interface. It argues that generative systems, socially responsive service agents and personalized recommendation systems can shape how people receive recognition and understand themselves. Its central term, “robotoid humanness,” names a proposed drift toward feeling most fluent, correct or socially viable when compatible with machine legibility, pacing and logic. The article’s central contribution is to name this proposed drift and treat consumer-facing AI agents as identity-relevant infrastructures.
The authors describe a three-stage mechanism. First, a person enters a synthetic social reality: an interaction that feels meaningful and responsive although the system’s social presence is computationally generated. Second, the system extracts a reduced identity profile from data traces and returns it as personalized recognition; this “mirror” is a statistically actionable profile, not a person’s full narrative complexity. Third, the person adapts self-presentation and behavior toward what the system can parse and reward. Repeated interaction creates a recursive loop in which the profile becomes normalized as an account of who the user is.
The paper distinguishes this from ordinary social feedback. Human interlocutors provide heterogeneous, divergent and contestable responses, including disagreement, surprise and misrecognition. AI-mediated feedback is presented as engineered toward convergence, legibility and engagement. The concern is not that people respond to feedback, but that personalized systems may filter friction-producing or disconfirming evidence while reinforcing what they already predict. The distinction is structural: repeated filtering can make a system’s response seem like an account of the person.
The framework begins with predictive recommendation and personalization systems and extends to open-ended large language model agents. Persistent memory, reinforcement learning from human feedback and embedding-based personalization are identified as technical routes that could produce analogous recursive dynamics. These are conceptual parallels, not findings from a comparative test of deployed systems. The authors reposition consumer-facing AI agents as identity-relevant infrastructures, formulate testable propositions and identify an autonomy risk; they report no user experiment, behavioral measurement or model evaluation demonstrating the proposed drift.
Read the primary source: link.springer.com ↗
Why it matters
The paper argues that AI-mediated autonomy risks extend beyond privacy and bias. Personalized recommendations and conversational agents could influence not only what people choose, but how they understand and present themselves. The article does not establish that these effects occur in practice, their size, or how long they last.
The paper broadens AI impact from what systems know about people to what repeated computational recognition might encourage people to become. Recommendations, predicted needs, automated scripts and conversational responses could make system classification feel more authoritative than ambiguous or contradictory self-assessment. Its example is an AI wellness system labeling someone an early-morning optimizer and returning a high readiness score despite fatigue. The example illustrates the argument, not a study result.
The autonomy concern rests on a view of selfhood as an ongoing narrative process in which people connect memories, commitments, emotions and future plans into continuity. If AI systems write or refine emails, biographies, applications, dating profiles and posts, a user may shift from author to editor. Automated summaries, options and conclusions could also reduce the struggle involved in ethical reasoning, vocational commitment and self-interpretation. The paper does not claim every use has this effect or quantify any loss of agency.
The proposed ethical concern goes beyond data access, privacy and biased classification. The authors call for a prospective “right not to be reduced,” countering systems that make people actionable by compressing them into categories. Suggested design directions are limiting inferential overreach, making personalization contestable, preserving ambiguity and self-authorship, and avoiding confirmation loops that present a thin profile as authentic identity. These are normative implications, not existing legal rights or established product requirements.
The source says the risk is most acute when AI interaction becomes a dominant or near-exclusive form of social engagement, including isolation, social anxiety, age-related vulnerability or structural barriers to human participation. When AI supplements ordinary human relationships, consequences are more likely to be episodic and recoverable. The article does not support a general claim that AI inevitably makes people robotic; it argues that certain systems and social conditions could narrow the evidence through which people revise self-understanding. The paper’s stronger claim is conditional on particular systems and social conditions, not universal across users.
What to watch next
The central test is empirical: whether sustained interaction with personalized or socially responsive AI changes self-authorship, exposure to disagreement, narrative identity or moral agency. Future work will also need to determine which system features and user circumstances matter, and whether design safeguards can preserve ambiguity and human choice.
The next requirement is measurement. The authors propose a conceptual mechanism but leave open how “robotoid humanness” should be operationalized. Studies should examine whether repeated personalized-AI exposure changes self-authorship, narrative coherence, self-presentation, exposure to disagreement, perceived agency or moral decision-making. They need comparison groups and measures distinguishing ordinary adaptation to social feedback from adaptation specifically caused by machine-optimized feedback. The article specifically leaves these outcomes for future empirical testing rather than reporting them as established effects.
Researchers should test whether the mechanism differs across classical recommendation engines, AI wellness tools, memory-enabled chatbots and other conversational agents, which may personalize users through different data flows and interaction patterns. The paper argues that their mirroring logic can be structurally similar but does not demonstrate equivalent effects. Unknowns include memory, affirmation, visibility of personalization, the ability to challenge an inference and whether systems introduce genuinely surprising information rather than reinforce prior patterns.
Deployment context will matter. The paper predicts greater concern when an AI relationship displaces human encounters rather than supplements them. Research should examine users with different levels of social access, reliance and vulnerability, rather than treat AI interaction as a single exposure. The source does not establish prevalence, a threshold at which supplementation becomes displacement, or uniform effects across users, cultures or forms of AI-mediated service. It also does not show whether effects are consistent across these settings.
The proposed safeguards need testing. Making personalization contestable, limiting inference and preserving ambiguity may help users resist a narrow profile, but the article provides no evidence about which interface or governance measures work. It also leaves unresolved whether alignment with machine-readable categories is always harmful, how quickly narrowing might emerge, whether it can be reversed, and whether users can gain practical benefits without surrendering meaningful self-authorship. Those questions determine whether the autonomy risk is a measurable public problem or primarily a theoretical warning.
