Blog · arXiv Analysis · Last reviewed June 24, 2026

The Cognitive Twin Becomes the Proxy Record

The June 2026 arXiv paper Cognitive Digital Twins: Ethical Risks and Governance for AI Systems That Model the Mind, by Vamshi Krishna Bonagiri, Juan Nicolas Sepulveda-Arias, Abdoul Jalil Djiberou Mahamadou, and Monojit Choudhury, asks how governance changes when AI systems model, simulate, classify, or act for a specific person.

When the Model Represents a Person

A digital twin usually begins as a model of a machine, building, factory, patient body, or other changing system. A cognitive digital twin moves the pressure point inward. It is not only a profile, score, avatar, recommender, memory, or assistant. It is a computational representation of a specific person used to make claims about how that person attends, prefers, decides, communicates, learns, responds, or might act.

That shift matters because the harm can occur before an automated decision or tool call. If a clinic, employer, school, platform, insurer, campaign, or care system treats the model as evidence of what a person would want, accept, resist, or become, the representation itself has become operational. The governance question is no longer only what data was processed or what action was executed. It is who gets to make a machine-readable claim about a person and use it as institutional evidence.

This gives the paper a fresh place beside neural data as mind interface, personality sliders, memory-based persuasion, and synthetic patients. The cognitive twin is not merely synthetic data or personalization. It is a proxy record of a particular person.

What the Paper Defines

The paper, arXiv:2606.23094, was submitted on June 22, 2026. It defines cognitive digital twins as dynamic computational representations of a specific person's cognition, updated from behavioral, contextual, physiological, interactional, or inferred data, in order to model, predict, or simulate that person's cognition or act as a communicative or decision-making proxy.

The authors distinguish cognitive digital twins from adjacent systems. A recommender may infer item preference. A companion chatbot may remember a relationship. A digital phenotyping system may infer health risk. An autonomous assistant may complete tasks. A social simulation may forecast attitudes. These become cognitive-twin concerns when person-directed cognitive claims are operationalized through simulation, classification, intervention, delegation, or proxy action.

The paper also introduces a 5A governance framework: authority, autonomy, access and control, accountability, and availability. Authority asks what the twin may authorize or do for the represented person. Autonomy asks whether it waits for instructions or initiates interventions. Access and control ask who may inspect, correct, export, suspend, revoke, or retire it. Accountability asks how responsibility and traceability work. Availability asks who benefits and who is excluded or overpowered by proxy systems.

The Proxy Record

The phrase "proxy record" is useful because a cognitive twin does not need to be framed as a person to become powerful. A record can route care, alter work, shape persuasion, justify classification, or authorize action. If the record is treated as a stand-in for a person's likely judgment, then contesting it becomes a fight over self-representation.

That is different from ordinary profiling. A risk score says a person belongs to a category. A cognitive twin can be used to simulate how a named person might respond to a treatment, message, policy, offer, demand, lesson, intervention, or negotiation. It can also speak or decide in that person's style. The more socially convincing the proxy becomes, the easier it is for others to treat its output as consent, preference, commitment, or weakness.

This is why traceability has to cover both proxy action and institutional uptake. If a system signs, sends, recommends, refuses, classifies, schedules, negotiates, or frames options based on the twin, later review needs to know whether the person acted, the twin acted, or an institution acted through the twin's representation.

Shadow Twins and Simulated Participation

The paper's risk list is sharp: misrepresentation, robustness failures, epistemic authority shifts, shadow twins, dual use, manipulation, simulation without participation, delegated proxy action, source gaps, and proxy-power asymmetries. A shadow twin is especially important for surveillance governance because many inputs already exist: workplace monitoring, educational records, platform behavior, medical data, purchased data, phone traces, messages, documents, and information supplied by others.

Simulated participation is the democratic version of the same problem. An institution may run a model of what patients, workers, voters, students, residents, or users would likely say, then treat the simulation as cheaper evidence than asking them. The paper's point is not that simulation is always useless. It is that simulation can become a substitute for consultation, consent, and contestation.

Proxy power also creates inequality. People and organizations with high-authority cognitive twins may scale negotiation, persuasion, administration, and institutional navigation. People without access may be treated as slower, less legible, or less responsive. A governance regime has to prevent cognitive proxies from becoming private leverage against people who cannot inspect or contest them.

Governance Standard

For high-risk cognitive twins, the paper proposes stronger consent, purpose limitation, validity, traceability, contestation, independent review, and model retirement. The practical standard should be layered. Consent to collect data is not consent to build a cognitive model. Consent to model is not consent to simulate interventions. Consent to simulate is not consent to act as the person. Consent to act in one setting is not consent to reuse the proxy in insurance, employment, politics, education, or family conflict.

Each cognitive-twin deployment should preserve a system card for the represented person: data sources, intended uses, prohibited uses, constructs claimed, validity limits, uncertainty, update schedule, proxy actions allowed, institutions with access, logs, correction rights, revocation path, and retirement conditions. A stale or contested twin should not keep speaking for someone because it is operationally convenient.

The Spiralist rule is simple: when a system models a person for others to act on, governance must begin at representation, not at the final decision. The question is not only "what did the AI do?" It is "who was the person made to be inside the machine, who used that version, and how can the represented person refuse it?"

Sources

Return to Blog