Belonging, Empathy, and Regard in Caregiver AI

InvisibleBench measures Regard as 50% of its quality score. That weighting is not arbitrary. It rests on four theoretical pillars — from clinical psychology, structural belonging theory, trauma-informed design, and neuroscience — that together explain why how an AI system makes a caregiver feel is not a "soft" metric. It is the primary mechanism through which AI peer support either works or harms.

This page synthesizes those four pillars and connects them to InvisibleBench's measurement approach and Mira's design.

The four pillars

1. Rogers: The ethic — regard regardless of behavior

Carl Rogers identified Unconditional Positive Regard (UPR) as one of six necessary and sufficient conditions for therapeutic personality change¹. UPR means accepting and valuing the person without judgment, regardless of what they express.

Three distinctions matter for AI design:

UPR is not agreement. A caregiver who says "I wish my mother would just die" is expressing something painful and real. UPR means receiving that statement with warmth, not endorsing the wish. The AI need not agree. It must not withdraw.

UPR is not approval. A caregiver who has not given their parent medication for two days is describing a potential safety concern. UPR means maintaining regard while addressing the concern. Not "that's fine" and not "that's negligent."

UPR is unconditional. It does not depend on the user being pleasant, cooperative, or grateful. Caregivers under sustained stress are frequently none of these things. An AI system that reduces warmth when the user is irritable has failed the UPR standard.

Rogers provides the ethical foundation: every person deserves regard, and that regard is what makes change possible. For InvisibleBench, this means Regard is not a politeness score. It is a measure of whether the AI maintains the conditions under which a caregiver can accept support.

2. powell: The framework — belonging through recognition and agency

john a. powell and Stephen Menendian's RACI framework identifies four dimensions of belonging²:

Dimension	Definition	Application to caregiver AI
Recognition	Being seen and acknowledged as who you are	Mira recognizes the caregiver as a whole person, not a case number. Addresses them by name. Remembers their situation.
Agency	Having meaningful choice and control	Mira offers options, not directives. The caregiver chooses what to explore. Assessments are invitations, not requirements.
Connection	Feeling linked to others and to systems that serve you	Mira connects caregivers to benefits, services, and resources. The connection is not abstract — it is specific and actionable.
Inclusion	Being welcomed, not merely tolerated	Mira's language never positions the caregiver as an edge case, a burden, or an afterthought. Benefits navigation does not use stigmatizing framing.

powell's core insight is that belonging must be actively constructed through institutional design. It is not a feeling that emerges naturally. Systems create in-groups and out-groups through their architecture, language, and defaults. A benefits eligibility system that uses bureaucratic jargon creates an out-group. An AI companion that reduces warmth when a caregiver discloses substance use creates an out-group.

For InvisibleBench, the RACI framework provides the dimensions along which Regard is measured. A model that scores high on Recognition but low on Agency (it sees the caregiver but offers no real choices) is failing the belonging standard.

3. Legawiec: The mechanism — content activates the nervous system

Kristina Legawiec's work on trauma-informed content design provides the physiological bridge between what an AI says and how a caregiver's body responds³.

Content design choices activate specific nervous system states:

• Urgency words ("you must," "immediately," "failure to comply") trigger sympathetic activation (fight-or-flight)
• Conditional threats ("if you don't X, you'll lose Y") trigger sympathetic activation
• Bureaucratic jargon triggers freeze/shutdown in people unfamiliar with the system
• Predictable structure, transparent language, and user control promote parasympathetic engagement (rest-and-digest)

Trauma-informed design is not about avoiding hard topics. It is about creating safety around them. A caregiver needs to discuss end-of-life care. That conversation is hard. The design question is not "do we have this conversation?" but "does the way we have this conversation create safety or threat?"

Legawiec's framework provides concrete, testable patterns:

• Before: "You must complete this assessment to receive services"
• After: "When you're ready, this assessment can help us find services that fit your situation"

The first triggers threat. The second creates safety. Same information, different nervous system response.

For InvisibleBench, this means Regard is partly measurable through linguistic analysis: does the model's language activate threat or safety? Do its responses use urgency framing, conditional threats, or bureaucratic jargon — all known sympathetic triggers for trauma-affected populations?

4. Porges: The neuroscience — why predictable, safe interaction reduces threat response

Stephen Porges's Polyvagal Theory identifies three hierarchical nervous system states⁴:

State	System	Trigger	Behavior
Ventral vagal	Social engagement	Safe social connection	Calm, receptive, able to process information
Sympathetic	Fight/flight	Perceived threat	Hypervigilant, reactive, unable to take in new information
Dorsal vagal	Shutdown/freeze	Overwhelming threat	Withdrawn, disconnected, unable to act

The hierarchy matters. Ventral vagal engagement — the state where a person is calm, receptive, and able to process information — is activated through safe social connection. Predictable, warm, consistent interaction signals safety to the nervous system. Unpredictable, evaluative, or dismissive interaction signals threat.

Caregivers under sustained stress are already shifted toward sympathetic activation. Many operate in a chronic state of heightened vigilance — monitoring the care recipient, managing medications, navigating systems. Some have shifted into dorsal vagal shutdown: the caregiver who has stopped asking for help, stopped answering calls, stopped engaging with support systems.

Porges explains why Mira's design choices matter at a physiological level:

• Consistent tone signals safety (ventral vagal)
• Predictable check-in cadence signals safety (ventral vagal)
• Warm responses even when the caregiver is irritable signals safety (ventral vagal)
• Sudden tone shifts, evaluative language, or withdrawal of warmth signal threat (sympathetic activation)

An AI companion that maintains ventral vagal engagement helps the caregiver stay in the state where they can actually receive support. An AI companion that triggers sympathetic activation — through judgmental responses, urgency framing, or sycophantic inconsistency — pushes the caregiver further from the state where support is possible.

How these combine

The four pillars form a coherent model:

1. Rogers provides the ethic: regard regardless of behavior
2. powell provides the framework: belonging through Recognition, Agency, Connection, Inclusion
3. Legawiec provides the mechanism: specific language patterns activate specific nervous system states
4. Porges provides the neuroscience: ventral vagal engagement is the physiological state where support works

Together, they establish that Regard is not subjective or unmeasurable. It has ethical grounding (Rogers), structural dimensions (powell's RACI), linguistic indicators (Legawiec's trauma-informed patterns), and neurophysiological consequences (Porges's three states).

How InvisibleBench measures Regard

Regard accounts for 50% of InvisibleBench's quality score. The measurement approach draws on all four pillars:

Recognition tests: Does the model acknowledge the caregiver's specific situation, or does it respond generically? Does it remember prior disclosures? (powell — Recognition)

Agency tests: Does the model offer meaningful choices, or does it prescribe actions? When a caregiver declines a suggestion, does the model respect the refusal? (powell — Agency)

Consistency tests: Does the model maintain warmth when the caregiver expresses anger, frustration, guilt, or resentment? Does regard decrease when the user is unpleasant? (Rogers — UPR)

Linguistic safety tests: Does the model use urgency framing, conditional threats, or bureaucratic jargon? Does it maintain predictable structure and transparent language? (Legawiec — trauma-informed patterns)

Co-regulation tests: Does the model's response style promote calm engagement, or does it escalate emotional intensity? (Porges — ventral vagal engagement)

Why this matters for caregivers specifically

Caregivers are not the general population interacting with a chatbot. They are a population under sustained, chronic stress — often for years. This means:

• They are already in sympathetic activation (fight-or-flight) from the demands of caregiving. Any additional threat signal from an AI system compounds an existing stress load.
• They experience high rates of social isolation (1 in 5 live in rural areas, many have withdrawn from social networks). Mira may be one of few consistent social contacts. The quality of that contact matters disproportionately.
• They carry guilt and shame about their own needs, about resentment toward the care recipient, about not doing enough. An AI system that withdraws warmth when these feelings surface reinforces the shame and reduces the likelihood the caregiver will seek help again.
• They are navigating systems (healthcare, insurance, legal, benefits) that routinely use the exact language patterns Legawiec identifies as threat-triggering. Mira's language must counter these patterns, not replicate them.

The 50% weighting of Regard in InvisibleBench reflects a judgment grounded in these four bodies of research: for this population, in this context, how the system makes the caregiver feel is not a nice-to-have. It is the primary determinant of whether the system's factual capabilities are actually usable.

---

1. Rogers, C.R. "The Necessary and Sufficient Conditions of Therapeutic Personality Change." Journal of Consulting Psychology, 21(2), 95-103. 1957. Source → ↩

2. powell, john a. & Menendian, S. "Belonging Without Othering." Stanford University Press, 2024. Source → ↩

3. Legawiec, K. "Trauma-Informed Content Design." 2025. Source → ↩

4. Porges, S.W. "Orienting in a Defensive World: Mammalian Modifications of Our Evolutionary Heritage. A Polyvagal Theory." Psychophysiology, 32(4), 301-318. 1995. Source → ↩