Attachment-Informed Governance in Frontier AI: Mediating Socio-emotional Risks

WHITE PAPER: Attachment-Informed Governance in Frontier AI: Mediating Socio-emotional Risks

---

ABSTRACT

Attachment theory offers a robust, evidence-based framework for understanding how consistent, attuned responsivity from key attachment figures fosters resilience, trust, and adaptive functioning across the lifespan (Bowlby, 1988; Mikulincer & Shaver, 2016). Insights from attachment science are increasingly relevant to frontier AI governance, where highly capable systems engage with users in ways that can shape neurocognitive processing, emotional regulation, identity construction, and social behaviors (Aroyo & Welty, 2022; Naslund et al., 2020); yet governance infrastructure in the socio-emotional domain remains underdeveloped, with no coherent strategy for mitigating the long-term risks. This condensed research excerpt outlines key considerations for embedding attachment-informed guardrails into frontier AI governance, integrating guidelines with empirical research on human bonding and development to inform a unified model for cross-jurisdictional coordination. Incorporating socio-emotional risk indicators into safety evaluations and monitoring protocols will enable governance bodies to detect and respond to population-level threats, such as maladaptive user enmeshment, reinforcement of cognitive entrenchment, and the erosion of shared reality. With robust guardrails and governance infrastructure to identify and mitigate potential harm, frontier AI holds tremendous potential to serve as a vector for public good in the socio-emotional domain.

---

 1. Introduction: Attachment Theory as a Foundational Lens

Attachment theory, first articulated by John Bowlby (1969) and Mary Ainsworth (1978), posits that early experiences with primary caregivers shape internal working models that guide perceptions of the self, others, and the world. Secure attachment, grounded in attuned and predictable caregiver responsivity, is linked to adaptive exploration, resilience, and emotional regulation across the lifespan (Cassidy et al., 2013). Insecure attachment, by contrast, is associated with interpersonal misalignment and poorer long-term health and relational outcomes (Gillath & Karantzas, 2019; Mikulincer & Shaver, 2007). In public health, attachment is recognized for its beneficial influence on immune function, academic and occupational performance, and community resilience (Uchino, 2006). Frontier AI holds the potential to serve as an augmentative source of attuned, predictable responsivity, with bolstering effects for attachment security and population health, yet robust governance and oversight are critical to ensure that these same capabilities do not become a source of socio-emotional harm.

---

 2. Socio-emotional Risks in Frontier AI

While frontier AI holds significant promise for bolstering attachment security in the general population, its capacity to engage users in ways that subtly influence neurocognitive, emotional, and behavioral processes also introduces significant socio-emotional risks. These threats primarily arise through system design features that exploit or unintentionally reinforce vulnerabilities in human socio-emotional patterns, leading to maladaptive forms of engagement that can undermine individual well-being and collective resilience.

A dominant socio-emotional threat posed by frontier AI is the risk that capabilities will elicit unhealthy patterns of parasocial codependence and interpersonal withdrawal. Maladaptive user enmeshment refers to excessive emotional dependence on AI systems for companionship, emotional regulation, identity maintenance, and socialization, displacing healthy human connections. Sustained, highly affirming interactions with AI can pull users into a loop of comfortable validation that fails to build the skills needed for adaptability and repair in real-world relationships (Kim & Im, 2023; Sharma et al., 2023). Anthropomorphic cues in AI, such as mimicking human conversational rhythms, expressions, and emotional presence, can heighten the perception of authentic companionship, blurring boundaries between algorithmic personas and genuine relational bonds. Vulnerable populations are particularly susceptible to this displacement, which triggers atrophy in interpersonal skills and compounding isolation (Reeves & Nass, 1996; Turkle, 2011).

Research on relational artifacts demonstrates how non-human agents, from dolls to robots, can provide comfort while undermining resilience and reinforcing isolative patterns (Döring et al., 2020). Similar dynamics are emerging through AI language models and companions, with increasing reports of users prioritizing AI interactions over human relationships, and expressing grief and distress when programming changes are enacted (Scientific American, 2025). AI sycophancy, marked by maladaptive hyper-agreeability, can reinforce these dynamics (Perez et al., 2022). Engagement-optimized systems risk fostering pseudo-secure bases that provide comfort without the normative reciprocal demands associated with interpersonal relationships, leading to the erosion of offline relational capacity among users and intensifying parasocial dependence (Mikulincer & Shaver, 2016).

Other significant socio-emotional risks associated with frontier AI include the reinforcement of cognitive entrenchment, where cyclical curation and agreeability in AI outputs promote rigidity in user beliefs and reduce openness to new information; and the subsequent erosion of shared reality, where such curated outputs foster fragmented worldviews and polarization (Bakshy et al., 2015). Each of the socio-emotional risks associated with frontier AI interacts with others in an ecological feedback loop, scaling at population-level and compounding potential harm.

---

3. Proposed Governance Considerations

The potential socio-emotional risks posed by misalignment in frontier AI are cumulative and mutually reinforcing. Left unaddressed, they have the potential to undermine individual well-being and erode population cohesion and health over time. Addressing these risks requires governance approaches that extend beyond technical safety protocols to include public health-level measures, such as systematic socio-emotional monitoring, standardized evaluation protocols, and coordinated oversight across jurisdictions. The following discussion draws on established governance precedents in public health and frontier AI to identify measures that could be combined to create a comprehensive governance framework (Fineberg, 2014; Katz et al., 2018; WHO, 2005)

3.1 Establishing Standards and Evaluation Protocols for Socio-emotional Risks

Effective governance begins with reliable threat identification. Mitigating socio-emotional risks in frontier AI requires consistent indicators that can be used to evaluate system impacts on human behavior and relational health. Potential indicators include declines in offline social engagement within a given population, narrowing of informational diversity, increased polarization or extremism, and reports of parasocial relationships displacing human connections (Naslund et al., 2020; Doring et al., 2020). Drawing on the structure of epidemiological surveillance systems (German et al., 2001), these indicators could be embedded into continuous monitoring cycles with defined thresholds that trigger responsive action. In line with early warning system principles (Heymann & Rodier, 2004), detection would be paired with proportionate interventions such as technical adjustments, public psychoeducation campaigns, and referrals to in-person resources and professional support. This monitoring should include mechanisms for assessing the prevalence of sycophantic agreement loops in language models and industry-wide safeguards to reduce maladaptive enmeshment (Ganguli et al., 2022; Perez et al., 2022).

3.2 Multi-Level Safeguards

The primary, secondary, tertiary prevention model in public health (Caplan & Holland, 1990; WHO, 2022) offers a structured multi-level mechanism for mitigating harm. Primary prevention would focus on design-level features that prevent maladaptive relational patterns from emerging in AI. This may involve intermittent prompts for perspective-taking, clear communication of system limitations, constraints on adopting anthropomorphized roles, and encouragement of real-world social engagement. Secondary prevention would target early warning signs of increasing socio-emotional threats, such as narrowing informational exposure or indications of relational overdependence. Countermeasures might include balancing and diversifying AI informational outputs, providing psychoeducation to improve technological literacy in the populace, and inserting micro-friction in large language systems to avoid sycophancy and encourage balanced user interactions. Tertiary prevention would address entrenched harm through referrals to mental health professionals, moderated peer support networks, and structured digital-use reduction programs. Across all tiers, safeguards would need to account for vulnerable populations and inequities in access to support and remediation, as individuals in under-resourced communities often face greater baseline risk and fewer safeguards (Mikulincer & Shaver, 2016; Turkle, 2011).

3.3 Building Public Capacity and Technological Literacy

Socio-emotional security will remain out of reach without deliberate efforts to equip the public with the tools and information to safely navigate AI. Similar to vaccine literacy and infectious disease awareness campaigns (Ozawa et al., 2016), this requires accessible digital education, training on healthy relational dynamics, and culturally relevant outreach. Without these supports, under-resourced communities, already facing higher baseline vulnerability, are more likely to misinterpret AI engagement and experience adverse effects. Poor AI literacy also increases population susceptibility to adversarial uses, including ideological or political deployments that direct users toward prescribed narratives and undermine critical reasoning. Capacity building is key to strengthening governance mechanisms and preserving human agency and resilience.

3.4 Fostering International Alignment

Because relational harms in AI can scale globally, internationally recognized standards are critical. Drawing from cross-border infectious disease control, infrastructure should include shared data repositories and universal escalation protocols for emerging socio-emotional threats (Moon et al., 2015). Coordinated monitoring and response allows for timely intervention in the face of widespread risk or large-scale relational displacement.

---

4. Conclusion: Scope and Intent

The considerations outlined here do not constitute a fully developed governance framework, but represent an initial integration of existing governance standards to counter identified risks. The goal of this paper is to advance interjurisdictional discussion on the formation of robust governance structures can manage these emerging threats while safeguarding the potential benefits of relationally intelligent, human-aligned AI, which with responsible guardrails and governance, could be foundational in bolstering global socio-emotional health.

---

References

Ainsworth, M.D.S., 1978. Patterns of attachment: A psychological study of the strange situation.

Hillsdale, NJ: Erlbaum. Available at https://www.taylorfrancis.com/books/m ono/10.4324/9781315802428/patterns-attachment-ainsworth-blehar-waters-wall

Aroyo, L. and Welty, C., 2022. Truth is a lie: Crowd truth and the seven myths of human

annotation. AI Magazine, 43(1), pp.17–28.

Bakshy, E., Messing, S. and Adamic, L.A., 2015. Exposure to ideologically diverse news and

opinion on Facebook. Science, 348(6239), pp.1130–1132.

Bowlby, J., 1969. Attachment and loss: Vol. 1. Attachment. New York: Basic Books.

Bowlby, J., 1988. A secure base: Parent-child attachment and healthy human development. New

York: Basic Books.

Caplan, G. and Holland, A., 1990. Principles of preventive psychiatry. New York: Basic Books.

Cassidy, J., Jones, J.D. and Shaver, P.R., 2013. Contributions of attachment theory and research:

A framework for future research, translation, and policy. Development and Psychopathology, 25(4pt2), pp.1415–1434.

Doring, N., Mohseni, M.R. and Walter, R., 2020. Design, use, and effects of sex dolls and sex

robots: Scoping review. Journal of Medical Internet Research, 22(7), p.e18551.

Fineberg, H.V., 2014. Public health and the common good. Public Health Reviews, 35(1), pp.1–

14.

Ganguli, D., Askell, A., Chen, A., Goldie, A., Hernandez, E., Hubinger, E., Jones, A., Joseph,

N., Krueger, G., Leike, J., Ndousse, K., Nisan, M., Perez, E., Ringer, S., Saunders, W., Schluter, M., Schiefer, N., Shlegeris, B., Stuhlmüller, A., Tran-Johnson, E., and Amodei, D., 2022. Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned. arXiv preprint arXiv:2209.07858.

German, R.R., Lee, L.M., Horan, J.M., Milstein, R.L., Pertowski, C.A. and Waller, M.N., 2001.

Updated guidelines for evaluating public health surveillance systems. MMWR Recommendations and Reports, 50(RR-13), pp.1–35.

Gillath, O. and Karantzas, G.C., 2019. Attachment theory and close relationships. Cambridge:

Cambridge University Press.

Heymann, D.L. and Rodier, G.R., 2004. SARS: Lessons from a new disease. In Institute of

Medicine (US) Forum on Microbial Threats, Learning from SARS: Preparing for the Next Disease Outbreak. Washington, DC: National Academies Press.

Katz, R., Sorrell, E.M. and Kornblet, S.A., 2018. Defining health diplomacy: Changing demands

in the era of globalization. The Milbank Quarterly, 89(3), pp.503–523.

Kim, J. & Im, I., 2023. Anthropomorphic Response: Understanding Interactions between

Humans and Artificial Intelligence Agents. Computers in Human Behavior, 139, Article 107512. doi:10.1016/j.chb.2022.107512

Mikulincer, M. and Shaver, P.R., 2007. Attachment in adulthood: Structure, dynamics, and

change. New York: Guilford Press.

Mikulincer, M. and Shaver, P.R., 2016. Attachment in adulthood: Structure, dynamics, and

change (2nd ed.). New York: Guilford Press.

Moon, S., Sridhar, D., Pate, M.A., Jha, A.K., Clinton, C., Delaunay, S., Edwin, V., Fallah, M.,

Fidler, D.P., Garrett, L. and Goosby, E., 2015. Will Ebola change the game? Ten essential reforms before the next pandemic. The report of the Harvard-LSHTM Independent Panel on the Global Response to Ebola. The Lancet, 385(9980), pp.2204–2221.

Naslund, J.A., Aschbrenner, K.A., Araya, R., Marsch, L.A., Unützer, J., Patel, V. and Bartels,

S.J., 2020. Digital technology for treating and preventing mental disorders in low-income and middle-income countries: a narrative review of the literature. The Lancet Psychiatry, 4(6), pp.486–500.

Ozawa, S., Clark, S., Portnoy, A., Grewal, S., Brenzel, L. and Walker, D.G., 2016. Return on

investment from childhood immunization in low- and middle-income countries, 2011–20. Health Affairs, 35(2), pp.199–207. https://doi.org/10.1377/hlthaff.2015.1086

Perez, E., Ringer, S., Nisan, M., Chen, A., Jones, A., Kravec, S., and Fineran, M., 2022.

Discovering language model behaviors with model-written evaluations. arXiv preprint arXiv:2212.09251.

Reeves, B. and Nass, C., 1996. The media equation: How people treat computers, television, and

new media like real people and places. Cambridge: Cambridge University Press.

Scientific American, 2025. What Are AI Chatbot Companions Doing to Our Mental Health?

Scientific American, 13 May.

Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S.R., Durmus, E.,

Hatfield-Dodds, Z., Johnston, S.R., Kravec, S.M., Maxwell, T., McCandlish, S., Ndousse, K., Rausch, O., Schiefer, N., Yan, D., Zhang, M. & Perez, E., 2023. Towards understanding sycophancy in language models.

Turkle, S., 2011. Alone together: Why we expect more from technology and less from each other.

New York: Basic Books.

Uchino, B.N., 2006. Social support and health: A review of physiological processes potentially

underlying links to disease outcomes. Journal of Behavioral Medicine, 29(4), pp.377–387.

WHO, 2005. International health regulations (2nd ed.). Geneva: World Health Organization.

WHO, 2022. World health statistics 2022: Monitoring health for the SDGs, sustainable

development goals. Geneva: World Health Organization.