< Go to Work

Turning Responsible-AI into Audio-Innovation Practice

Lead UX Researcher | Mixed Methods, AI Strategy & Responsible AI

TL;DR: I led the development of seven Responsible Audio-AI guidelines by combining listener research, established AI standards, and an AI impact assessment. The framework translated risks involving synthetic voices, content provenance, personal data, and human oversight into concrete product requirements.

My role: As one of two UX Researchers working alongside two Designers, I I led the framework from evidence synthesis through product application. I:

  • Synthesized Responsible AI and human-AI interaction guidance.

  • Connected listener findings to potential audio-AI harms.

  • Helped assess the proposed experience for potential impact.

  • Identified and organized audio-specific risk areas.

  • Drafted and refined the seven guidelines.

  • Facilitated critique with senior researchers.

  • Mapped each guideline to product behaviors, controls, or requirements.

Three evidence sources informed the guidelines-

Listener Research- A survey, exposure-testing interviews, co-design sessions, and concept testing revealed that acceptance of AI in audio depended on augmentation, transparency, consent, personalization boundaries, and human control.

1.

BUILDING THE FRAMEWORK

Together, these sources connected user expectations and known AI harms to concrete product decisions.

Existing Guidance- I reviewed academic research, human-AI interaction guidance, industry Responsible AI standards, emerging regulation, and publicly available audio-AI products.

2.

Impact Assessment- We evaluated how the proposed experience could affect listeners, creators, communities, and people whose voices or identities might be represented.

3.

As my team helped iHeartRadio navigate AI-integrated audio content in the risky market of late 2023, a core question of under what conditions should GenAI operate consistently remained unanswered.

General Responsible AI frameworks addressed fairness, transparency, privacy, and accountability. However, our 2024 landscape review did not identify publicly available guidance tailored to the perceptual and social characteristics of AI-generated audio. Audio needed domain-specific guidance since it introduced distinct considerations:

  • Voice communicates identity and authority.

  • Listening is often passive, especially while driving or multitasking.

  • Human-like delivery can imply social understanding.

  • Personalization can require sensitive contextual data.

  • Listeners may not recognize whether content is human-created, AI-generated, or AI-delivered.

This called for a framework to account for how people perceive and trust audio, not only how the underlying AI system operates.

THE CORE PROBLEM

KEY INFLECTION: CONDITIONAL OPENNESS (LED SYNTHESIS)

Across the survey and interviews, a consistent pattern emerged:

Users were not resistant to AI.
They were conditionally open.

AI that supplemented rather than replaced human content (news/weather, translations, customization).

What worked

Context-aware personalization when transparently signaled.

Preservation of human connection.

What did not work

Unlabeled AI hosts.

Likeness recreation without consent.

Hallucination risk.

Misuse by bad actors.

Privacy ambiguity.

This revealed that acceptance of AI in audio was not binary. Acceptance depended on whether AI augmented human connection or attempted to replace it. These patterns directly informed the strategic decisions that shaped the product direction.

KEY JUDGMENT CALLS

We prioritized augmentation over AI-host replacement.

Tradeoff: Less novelty in exchange for stronger brand fit and adoption potential.

1. Supplement, Don’t Replace

2. Anchor in Community-Led Personalization

Rather than a grab bag of AI utilities, we converged on community channels, preserving ritual and connection while using AI for scalable facilitation.

Tradeoff: One strong directional bet rather than multiple shallow experiments.

3. Responsible AI for Audio-AI Systems (Led end-to-end)

Impact assessment of the idea surfaced harm vectors:

Privacy risk.

Bad actors.

Hallucinations.

NLP/IP Misuse.

Recognizing the governance vacuum, I authored and operationalized a Responsible Audio-AI framework to translate research risks into deployable guardrails:

Synthesized cross-industry AI principles.

1.

Identified audio-specific risk categories (voice mimicry, parasocial attachment, tone manipulation).

2.

Drafted the initial framework.

3.

Facilitated structured feedback with senior researchers.

4.

Translated principles into seven actionable design constraints grounded in primary research.

5.

I then translated guidelines into product and design-system changes, including:

Approval flows.

Data usage settings.

Role structures (Admin, Moderator, Member, AI Host).

Clear AI labeling.

Tradeoff: Time invested in governance reduced prototype polish but created a credible path to ship.

We proposed using AI as community infrastructure that delivers personalized community-driven radio stations. The concept included:

THE PROPOSED DIRECTION

Shared-interest group channels in community-driven personalized stations.

Clear role hierarchy.

Public or private communities.

Explicit AI transparency and control signals.

Scalable across community types and sizes.

User-controlled customization boundaries.

This repositioned GenAI from synthetic performer to connective scaffolding. Strategically, this reduced reputational risk while preserving competitive differentiation.

Embedded transparency, consent, accessibility, data boundaries, and human oversight into the proposed experience

IMPACT

Connected Responsible AI risks to specific product requirements

Provided iHeartRadio with an audio-specific AI foundation for responsibly evaluating future experiences.

Brought governance into concept development rather than treating it as a final compliance review

The framework was applied to an exploratory concept whose usefulness was assessed through expert critique and its ability to produce concrete product requirements. Given promising results, toward design teams and in turn the impact on customers, guidelines like these effectively operationalize product exploration in volatile tech environments.

In an innovation landscape like today’s, this work should be extended to product, engineering, legal, accessibility, policy, and content-moderation teams. Research initiatives should assess longitudinal adoption, harm-reduction metrics, and impact on product decisions.

WHY THIS MATTERS

WHAT I’LL CARRY FORWARD

Here are some key takeaways from this study that taught me more about research-

Guidance becomes actionable when it is domain specific, reflects the medium, behavioral context, and ways a system can affect identity, relationships, and trust.

Safety mechanisms like labeling, consent, approvals, permissions, and personalization controls are UX decisions.

If a guideline cannot inform a requirement, workflow, control, or system behavior, it is not yet useful to a product team.