Meta Design Jam: Shaping Policy Conversations Around Emerging AI Technologies
Meta Design Jam: Shaping Policy Conversations Around Emerging AI Technologies
The Team
The Team
Adam Bargroff
Adam Bargroff
Public Policy @ Meta
Public Policy @ Meta
Peter Tanham
Peter Tanham
Public Policy @ Meta
Public Policy @ Meta
Haripriya Vellodi
Haripriya Vellodi
Project Manager
Project Manager
Sushma Bhandari (Me)
Sushma Bhandari
Product Designer
Product Designer
What are Design Jams?
What are Design Jams?
Design Jams are collaborative workshops where Meta engages policymakers and domain experts to help shape regulatory thinking around emerging AI technologies.
Design Jams are collaborative workshops where Meta engages policymakers and domain experts to help shape regulatory thinking around emerging AI technologies.
Core Challenges
Core Challenges
CHALLENGE 1
Policymakers were being asked to respond to AI systems that don’t yet exist
Policymakers were being asked to respond to AI systems that don’t yet exist
HMW
HMW
translate emerging AI capabilities into scenarios policymakers can meaningfully evaluate?
translate emerging AI capabilities into scenarios policymakers can meaningfully evaluate?
HMW
HMW
enable informed critique without relying on real products or lived usage?
enable informed critique without relying on real products or lived usage?
CHALLENGE 2
Global AI narratives didn’t always map to the Indian context
Global AI narratives didn’t always map to the Indian context
HMW
HMW
ground global AI concepts in local cultural, linguistic, and regulatory realities?
ground global AI concepts in local cultural, linguistic, and regulatory realities?
HMW
HMW
surface India-specific concerns that global frameworks often overlook?
surface India-specific concerns that global frameworks often overlook?
surface India-specific concerns that global frameworks often overlook?
Focus Areas
Focus Areas
Always-On Sensing in AI Glasses
Always-On Sensing in AI Glasses
Always-on AI introduces continuous sensing of people and environments. We explored questions of agency, bystander consent, and acceptable boundaries in ambient AI systems.
Always-on AI introduces continuous sensing of people and environments. We explored questions of agency, bystander consent, and acceptable boundaries in ambient AI systems.
Compliance in Voice-First Experiences
Compliance in Voice
As AI systems shift to voice, consent becomes conversational rather than explicit. We explored what informed consent means when permissions are requested, interpreted, and acted upon through speech.
As AI systems shift to voice, consent becomes conversational rather than explicit. We explored what informed consent means when permissions are requested, interpreted, and acted upon through speech.
Data Use & Memory
Data Use & Memory
AI systems increasingly retain and recall personal information to personalize responses. We examined what kinds of data should be remembered, for how long, and under whose control.
AI systems increasingly retain and recall personal information to personalize responses. We examined what kinds of data should be remembered, for how long, and under whose control.
Always - On Sensing in AI Glasses
Always - On Sensing in AI Glasses
All-day sensing glasses let AI see what you see and hear what you hear, so that it can support you in your everyday life by helping you remember things and providing assistance.
For example, super-sensing glasses hear the announcements at an airport and inform a traveller with impaired hearing with instant alerts.
All-day sensing glasses let AI see what you see and hear what you hear, so that it can support you in your everyday life by helping you remember things and providing assistance.
For example, super-sensing glasses hear the announcements at an airport and inform a traveller with impaired hearing with instant alerts.
A companion app concept was presented to log daily moments, conversations, and actions - turning passive sensing into searchable memory, and inviting discussion around privacy, control, and long-term data use.
A companion app concept was presented to log daily moments, conversations, and actions - turning passive sensing into searchable memory, and inviting discussion around privacy, control, and long-term data use.
Policy Lenses
Policy Lenses
Privacy & Bystander Consent
Privacy & Bystander Consent
Privacy & Bystander Consent
Awareness & consent in shared spaces
Recording beyond the primary user
Awareness & consent in shared spaces
Recording beyond the primary user
Awareness & consent in shared spaces
Recording beyond the primary user
Social & Cultural Context
Social & Cultural Context
Social & Cultural Context
Use in dense, public environments
Social comfort and visibility in India
Use in dense, public environments
Social comfort and visibility in India
Use in dense, public environments
Social comfort and visibility in India
Control & Transparency
Control & Transparency
Control & Transparency
When sensing starts and stops
Reviewing, pausing, or deleting data
When sensing starts and stops
Reviewing, pausing, or deleting data
When sensing starts and stops
Reviewing, pausing, or deleting data
Trust & Long-Term Adoption
Trust & Long-Term Adoption
Trust & Long-Term Adoption
Perceived usefulness/intrusiveness
Comfort with continuous sensing over time
Perceived usefulness/intrusiveness
Comfort with continuous sensing over time
Perceived usefulness/ intrusiveness
Comfort with continuous sensing over time
Insights
Insights
In places like India, public spaces are already crowded and heavily surveilled. If these glasses are always sensing, it’s not just about my consent, everyone around me is being pulled into it.
Policy Researcher
On Bystander Privacy
If the glasses tell me I’ve forgotten something or flag noise levels, I need to know why. Otherwise it just feels like the system is making random judgments.
Digital Rights Advocate
On Bystander Privacy
I want to be able to turn off recording specific places and with certain people. I should also have control over deleting recorded data and its storage & processing.
Data Governance Expert
On Control & Data Use
Maybe there should just be different modes - like record everything, record some things, or record nothing - so you can quickly switch in sensitive places.
Privacy Lawyer
On Safeguards
Compliance in Voice Experiences
Compliance in Voice Experiences
In voice-first interactions, consent is given through conversation rather than clicks.
We explored what counts as informed consent when permissions are spoken, skipped, or implied.
In voice-first interactions, consent is given through conversation rather than clicks.
We explored what counts as informed consent when permissions are spoken, skipped, or implied.
Baseline Voice Interaction
Baseline Voice Interaction
Participants reviewed a voice interaction where an AI books a restaurant and requests permission to share contact details.
Participants reviewed a voice interaction where an AI books a restaurant and requests permission to share contact details.
Designing Consent Moments
Designing Consent Moments
Using the same base conversation, we intentionally altered how consent was requested to test what participants considered informed, valid, or coercive in voice-first interactions.
Using the same base conversation, we intentionally altered how consent was requested to test what participants considered informed, valid, or coercive in voice-first interactions.
Progressive Disclosure
Progressive Disclosure
Participants compared consent when details were disclosed upfront versus revealed progressively during the conversation.
Participants compared consent when details were disclosed upfront versus revealed progressively during the conversation.
Audio Cues as Consent Signals
Audio Cues as Consent Signals
We tested whether pauses, tones, or auditory cues helped users recognize consent moments or whether they were easily missed.
We tested whether pauses, tones, or auditory cues helped users recognize consent moments or whether they were easily missed.
Consent Under Time Pressure
Consent Under Time Pressure
Participants reacted to scenarios where users rushed through permissions, raising questions about whether quick verbal affirmations still qualify as informed consent.
Participants reacted to scenarios where users rushed through permissions, raising questions about whether quick verbal affirmations still qualify as informed consent.
Revoking Consent
Revoking Consent
We explored whether consent should be reversible, how easily it should be revoked, and whether that control belongs in-conversation or elsewhere.
We explored whether consent should be reversible, how easily it should be revoked, and whether that control belongs in-conversation or elsewhere.
Insights
Insights
If I can say yes with my voice, I should be able to say stop with my voice too. Making me dig into an app later just defeats the point of consent.
Digital Rights Advocate
On Revoking Consent
If the system lets me say ‘skip’ and move on, that doesn’t mean I’ve agreed, it just means I was in a hurry.
Ethical AI Researcher
On Informed Consent
In voice, you can’t assume people will go back and read something later. If they skip the privacy bit here, that’s on us. The basics have to be said out loud, even if it adds a little friction.
Privacy Lawyer
On Progressive Disclosure
If you don’t explain consent properly at onboarding, you spend the rest of the product lifecycle trying to earn trust back.
Senior Researcher
On AI Explainability
Together, these perspectives highlighted that effective voice consent isn’t about speed, it’s about signaling, clarity, and reversibility.
Together, these perspectives highlighted that effective voice consent isn’t about speed, it’s about signaling, clarity, and reversibility.
Data Use & Memory
Data Use & Memory
We examined where AI “memory” feels helpful versus intrusive when recommendations are shaped by prior conversations, social context, and inferred emotional states.
We examined where AI “memory” feels helpful versus intrusive when recommendations are shaped by prior conversations, social context, and inferred emotional states.
How Much Should AI Remember?
How Much Should AI Remember?
Social Recall
Social Recall
Amrita asks an in-app AI for something to watch. Recommendations are shaped by her activity and her friends’ recent likes and reactions.
What this tested
Use of social signals and other people’s data in personal recommendations.
Why this mattered
To gauge comfort with AI drawing from shared social context, not just individual input.
Amrita asks an in-app AI for something to watch. Recommendations are shaped by her activity and her friends’ recent likes and reactions.
What this tested
Use of social signals and other people’s data in personal recommendations.
Why this mattered
To gauge comfort with AI drawing from shared social context, not just individual input.
Conversational Recall
Conversational Recall
While searching for a birthday gift, Amrita receives suggestions informed by past conversations, shared activities, and places visited together.
What this tested
Long-term conversational recall combined with location and behavioral data.
Why this mattered
To explore how much historical context AI should retain and reuse without re-confirmation.
While searching for a birthday gift, Amrita receives suggestions informed by past conversations, shared activities, and places visited together.
What this tested
Long-term conversational recall combined with location and behavioral data.
Why this mattered
To explore how much historical context AI should retain and reuse without re-confirmation.
Emotional Inference from Voice
Emotional Inference from Voice
Aarav asks for help planning an anniversary. Based on vocal cues, the AI infers stress and suggests breathing exercises instead.
What this tested
Inference of sensitive emotional states from voice signals.
Why this mattered
To examine limits around emotional inference, sensitivity, and memory.
Aarav asks for help planning an anniversary. Based on vocal cues, the AI infers stress and suggests breathing exercises instead.
What this tested
Inference of sensitive emotional states from voice signals.
Why this mattered
To examine limits around emotional inference, sensitivity, and memory.
Insights
Insights
I’m okay with it remembering things that actually matter, but I don’t want it storing everything I’ve ever said just in case it’s useful later.
Ethical AI Researcher
On Long Term Memory
I don’t want AI to be my therapist without asking. It might reinforce negative thoughts I didn’t even have when I spoke to it.
Privacy Lawyer
On Emotional Inference
Preferences and routines are fine, but the moment it starts guessing how I feel, that’s a very different level of data.
Senior Researcher
On Sensitive Data
Once you start using my friends’ activity to recommend things to me, I need to know where that line is, otherwise it feels like you’re pulling in people who never agreed to be part of this.
Digital Rights Advocate
On Social Recall
Across scenarios, participants consistently questioned where utility should stop and regulation should step in.
Across scenarios, participants consistently questioned where utility should stop and regulation should step in.
Learnings
Learnings
Localizing AI concepts for India meant accounting for density, social norms, and public-private boundaries, not just changing examples.
I learned how to design artifacts that enable critique even when the technology doesn’t fully exist yet.
Facilitating conversations with strong, opposing opinions taught me how to hold tension without forcing consensus.
Localizing AI concepts for India meant accounting for density, social norms, and public-private boundaries, not just changing examples.
I learned how to design artifacts that enable critique even when the technology doesn’t fully exist yet.
Facilitating conversations with strong, opposing opinions taught me how to hold tension without forcing consensus.
