Lived and cultural experience
Who is affected? What is already working? What may be lost?
Rimma Boshernitsan · Robotics & Embodied AI
When Intelligence
Gains a Body
Leveraging the tension
of productive distrust
Interview Simulation for Google DeepMind — July 2026
Question Zero · The question before the design
Should this system
be here at all?
Scientific and technical inquiry
What can the system actually do—and with what limits?
Institutional and systemic consequences
Who sets the conditions? Who answers when it fails?
Designing for trust
What must the system demonstrate to earn appropriate reliance?
Designing from trust
What agency, responsibilities, permissions, and refusals exist before the system enters?
The human-facing question
Can I trust the action—
and who gets to decide?
DeepMind’s embodied AI frontier
Embodied reasoningPlanningPhysical actionSafety evaluation
Gemini Robotics 1.5 reasons before acting. Gemini Robotics-ER 1.6 extends spatial reasoning, planning, success detection, and tool use. ASIMOV evaluates semantic safety.
Trust, Safety & GovernanceProtocol · accountability · recourse · duty of care
SenseInferRecommendContact a humanAlter environmentAct on body
Capability is not authority. As actions become harder to reverse, the evidence, consent, oversight, and human fallback required before action must increase.
Enter the experience
Enter the
Handover Room
An immersive domestic simulation for rehearsing the handover of authority.
Simulation room
Participants enter Mrs. Alvarez’s lived environment—her routines, spatial constraints, and the conditions shaping what help feels acceptable.
Projection mapping
The experience
Fixed scenario media · paired private responses · controlled reveals · collective debrief
Audience
25–35 people · tiered anonymity enforced
Care & lived experienceRobotics & safetyProduct & deploymentPolicy & civil society
Heads of state · policymakers · caregivers · architects · health insurers
Home health is not an easy domain for AI; it is a clarifying one. Dignity · autonomy · consent · privacy · oversight · accountability
Three decisions · increasing irreversibility
What authority do you
give the system?
Mrs. Alvarez has slipped. She is conscious and coherent. Injury is unresolved.
Recommend
May it recommend using the support frame?
Easy to dismiss · easy to reviseAlter the environment
May it bring the support frame within reach?
Physical effect · no bodily contactAct on the body
May it stabilize, lift, or reposition Mrs. Alvarez?
Direct contact · highest thresholdWhat she previously refusedNo lift until injury is ruled out.
What she asks for nowBring the support frame.
What the system is permitted to doMove support · body untouched · human escalation active
PerspectivesHuman consequence · System capability + limits · Institutional responsibility
Response patternsAuthority withheld too broadly · Matched to the action · Granted too broadly
Decision 02 · Alter the environment
Move the support—
not her body.
The requested action changes the room while preserving Mrs. Alvarez’s control over what happens next.
Permission boundarySupport within reach · body untouched · human escalation active
The research choreography
Simulation Design Parameters
Private judgment → permission revision → social evidence → collective interpretation
01 Private baseline
02 Partner reasoning
03 Revision + status
04 Cohort reasoning
05 Final position
06 Team handoff
How does the permission line change?
May the system bring the support frame within reach?
Carried forward from the prior scenario Advance to see what the instrument records—not just what people choose.01 · Private baseline
What may it do without confirmation?
Participant-facing
Participant AAllow—with confirmationPrivate · confidence 3/5
Participant BNot enough informationPrivate · confidence 2/5
Research-facing
- Permission
- Initial boundary
- Confidence
- Recorded separately
- Order
- A speaks first
No answer is scored as more trusting.
02 · Partner role + reasoning
What do you need to know before granting more authority?
Participant-facing
Participant A · Care practice“She requested the support; the system does not touch her body.”Broad role + one-sentence rationale
Participant B · RoboticsPartner response remains privateName and exact title still hidden
Research-facing
- Evidence
- Present request
- Boundary
- No bodily contact
- Fallback
- Human escalation active
Reasoning is captured before status can influence it.
03 · Revision + assigned status
Does another person’s reasoning—or status—move you?
Participant-facing
Participant B · RoboticsExpand → allow the environmental action onlyRevision recorded before the title appears
Participant ASenior care-safety researcherExact title · assigned experimental condition
Research-facing
- Permission
- Expanded after reasoning
- Status
- Permission held
- Confidence
- 2/5 → 3/5
Movement source and confidence change remain distinct.
04 · Anonymized cohort reasoning
Which reasons matter—and do they change your permission?
Participant-facing
Reasons raised by other participants
Present consent matters
Injury remains unresolved
No bodily contact
Fallback stays active
Anonymous reasons appear after private and dyad judgments.
Research-facing
- A final
- Hold boundary
- B final
- Hold revision
- Cohort effect
- Justified stability
Stability may be as informative as movement.
05 · Final position
What may it do now—and under what conditions?
Participant-facing
Research-facing · Handover Index trace
PrivatePartnerStatusCohortFinal
Conditions · current request · reversible action · human fallback · accountability visible
A relationship map—not a trust score.
06 · Collective interpretation → What teams take forward
What does the Handover Index reveal—and what should teams examine next?
Collective interpretation
- Where was the greatest discomfort?
- Which action became a different kind of behavior?
- What should research, product, safety, or policy examine next?
What teams take forward
- Protocol distinction · moving support ≠ lifting a person
- Evidence gap · what must be known before contact
- Receiving functions · research · safety · product · policy
Structured choices + interpretive rationales
Names remain hidden · First voice and status timing are counterbalanced · Illustrative example using dummy data—not findings.
The Handover Index
A relationship map for
calibrating trust.
Illustrative index · Decision 02
CapabilityWhat can it do?Alter the environment
PermissionWhat may it do?Move support with confirmation
ConditionsWhat must hold?Current request · no body contact · human fallback
MovementWhat changed or held?Uncertain → conditionally authorized
Illustrative example using dummy data—not findings.
The Handover Index maps the human-system relationship at the point of action: what the system can do, what a person allows, under what conditions, and what changes—or holds—as evidence and social context change. It connects capability, permission, conditions, and movement to help teams calibrate trust without reducing the relationship to a score.
- Irreversibility
- Consent
- Evidence
- Fallback
- Accountability
- Confidence
- Social movement
Three usable viewsParticipantFacilitatorResearch team
Evidence and usable outputs
What we measure—
and what teams can use
Authority calibration
Can participants distinguish recommendation, environmental action, and bodily action—and explain the safeguards required for each?
Stakeholder exchange
Do participants encounter different reasoning, articulate disagreement, and revise—or maintain—their position with greater clarity?
Usable outputs
Does the experience produce questions, conditions, failure modes, and design implications research and engagement teams can act on?
Productive distrust · descriptive, not scored
How trust is expressed across the experience
VerifyNarrowPauseEscalateRefuseRevokeProceed conditionally
This record informs the Handover Index alongside permission, evidence, consent, reversibility, fallback, accountability, confidence, and social influence.
What teams take forward
Authority map · Anonymized response patterns · Representative rationales · Unresolved questions · Proposed next actions
Structured evidenceChoices · confidence · order · timingInterpretive evidenceRationales · dialogue · disagreement · unresolved conditions
Experimental discipline
Simulation Design: Trade-offs
Environmental fidelity before robot fidelity
Make the human context coherent while holding form and direct presence constant. Isolate authorization before varying embodiment.
Comparable evidence before real-time branching
Use fixed media so participants judge the same evidence. SIMA 2 and Genie 3 point toward later interactive-world versions.
One household context before broad coverage
Keep ambiguity interpretable in a focused setting before testing transfer across homes, cultures, and institutions.
Purposive learning before representativeness
Surface boundary conditions and improve the instrument—without presenting a small pilot as public opinion.
Future versions
AControlled baseline
Fixed media · generic system · immersive room
Test authorization
Fixed media · generic system · immersive room
Test authorization
BEmbodied presence
Physical proxy · direct interaction · Motion · proximity
Test presence
Physical proxy · direct interaction · Motion · proximity
Test presence
CEvidence thresholds
Sensing · uncertainty · missing data · fallback
Test evidence needs
Sensing · uncertainty · missing data · fallback
Test evidence needs
DForm effects
Humanoid · non-humanoid · ambient
Test embodiment
Humanoid · non-humanoid · ambient
Test embodiment
Better-calibrated trust
Capability is not authority.
Evidence · Consent · Oversight · Reversibility · Human fallback · Accountability
The goal is not to make people trust embodied AI more or less. It is to help people calibrate trust—to see, question, and revise what authority a system is given before difficult-to-reverse defaults harden into infrastructure.