> For the complete documentation index, see [llms.txt](https://musenai.gitbook.io/musenai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://musenai.gitbook.io/musenai/understanding-musen/ai-radio-white-paper.md).

# AI Radio White Paper

## AI RADIO

### A Computational Framework for Continuous Media

#### From abundance and item recommendation to stateful orchestration, governed adaptation, and time based attribution.

<table data-header-hidden><thead><tr><th valign="top"></th></tr></thead><tbody><tr><td valign="top"><p>Central thesis</p><p>As media supply becomes more abundant, the central problem increasingly shifts from access to orchestration: deciding what should happen next, when, and under which constraints within a continuous experience.</p></td></tr></tbody></table>

Public conceptual white paper

Version 4.0.0

August 2026

Author: musen

Publication note. This paper defines a conceptual category and research agenda. It is not an experimental result, legal opinion, licensing framework, patent filing, or claim of economic superiority. Product status statements are dated and should be reverified when reused.

## Abstract

Digital music systems have achieved extraordinary access, scale, and personalization. Global recorded music revenue reached US$31.7 billion in 2025, with streaming accounting for 69.6 percent of revenue and paid subscription accounts reaching 837 million users. At the same time, generative systems are lowering the marginal cost of producing some forms of media, while new interfaces from Spotify and YouTube Music increasingly accept natural language intent as an input to listening. These changes do not prove demand for a new category, but they sharpen a systems question: when access is abundant, how should a listening experience be organized over time? \[1] \[2] \[3] \[4] \[5]

This paper proposes AI Radio as a class of audio systems that continuously construct and adapt a temporally coherent listening experience by reasoning over current state, permitted context, expressed intent, prior observations, and governing constraints. Unlike item recommendation or one time playlist generation, an AI Radio system observes the unfolding experience and can revise subsequent editorial decisions while preserving continuity.

The contribution is conceptual. We define category boundaries, present an abstract state model, distinguish mandatory properties from optional maturity features, propose an evaluation framework, and state conditions that would weaken or falsify the thesis. musen is discussed as one working implementation and research platform, not as the definition of the category.

<table data-header-hidden><thead><tr><th valign="top"></th></tr></thead><tbody><tr><td valign="top"><p>Economic thesis</p><p>In continuous media, time is a natural candidate for attribution, but not a self validating measure of value or fairness. The proposition that value follows time is a thesis to test.</p></td></tr></tbody></table>

### Evidence and claim discipline

| Evidence type                    | Meaning in this paper                                                                          |
| -------------------------------- | ---------------------------------------------------------------------------------------------- |
| Established external observation | A dated fact supported by a primary or peer reviewed source.                                   |
| Literature supported proposition | A proposition supported in some settings by prior research, without universalizing the result. |
| Design principle                 | A normative recommendation for trustworthy or coherent systems.                                |
| musen implementation fact        | A dated statement about the current product or internal implementation.                        |
| Research hypothesis              | A proposition that requires empirical testing and may be falsified.                            |

## 1. Problem and contribution

The twentieth century problem of recorded media was scarcity of access. The early digital problem was search. The streaming problem became relevance at scale. These problems have not disappeared, but they coexist with a newer condition: the listener can often access far more media than can be meaningfully evaluated, while systems can produce, upload, rank, and recombine content at increasing speed.

The recorded music market provides scale context rather than evidence for AI Radio itself. IFPI reports that global recorded music revenue reached US$31.7 billion in 2025, that streaming represented 69.6 percent of revenue, and that paid subscription accounts reached 837 million users. Access to large catalogues is therefore no longer a fringe capability. It is a mature global infrastructure. \[1]

Supply can also expand much faster than attention. Deezer reported in July 2026 that it was receiving about 90,000 fully AI generated tracks per day and that such tracks exceeded half of daily new uploads at peak in June. Yet those tracks represented only a small share of listening on Deezer, partly because the service excludes detected AI music from recommendations. This single platform is not the whole market, but it illustrates a broader distinction between the production of items and the allocation of attention. \[5]

The central problem considered here is therefore not that recommendation has failed. Modern recommendation systems are highly effective. The question is whether item relevance is sufficient to describe a listening product whose intended output is an ongoing experience rather than a sequence of isolated decisions.

### Contributions of this paper

•  An implementation independent definition of AI Radio.

•  A boundary test that distinguishes AI Radio from adjacent products such as generated playlists, conversational search, autoplay, fixed linear channels, and voice presenters.

•  An abstract state model for reasoning about continuous media without prescribing a particular algorithm or architecture.

•  A distinction between category requirements and optional maturity properties such as durable personal memory, spoken hosting, external control interfaces, or cryptographic provenance.

•  A research framework covering experience quality, agency, longitudinal utility, system performance, privacy, rights, and accounting integrity.

•  Explicit limitations and falsification conditions.

<table data-header-hidden><thead><tr><th valign="top"></th></tr></thead><tbody><tr><td valign="top"><p>Scope boundary</p><p>This paper does not claim that AI Radio is necessarily superior to playlists, recommendation, or on demand listening. It proposes a category for a different dominant product objective: continuity over time.</p></td></tr></tbody></table>

&#x20;

## 2. From discrete selection to continuous orchestration

Most digital media systems represent creative works as discrete items. Tracks, videos, posts, and episodes can be indexed, ranked, recommended, queued, and monetized individually. This abstraction is computationally convenient and has produced enormous consumer value.

Music recommendation research nevertheless shows that item relevance is strongly affected by sequence and context. Large scale work from Spotify researchers emphasizes that music is consumed repeatedly, in sessions, and under context dependent preferences. Context aware music research has studied time, location, activity, weather, mood, social setting, and session conditions. \[6] \[7] \[8]

The distinction proposed here is not between intelligent and unintelligent systems. It is between two dominant abstractions. A discrete system typically asks which item, set, or ranked result is relevant now. A continuous system additionally asks how the current state of an experience should evolve.

<p align="center">Figure 1. Two dominant abstractions. Real products may combine both.</p>

Natural language interfaces make this distinction more visible. Spotify DJ allows Premium users to request changes to a personalized listening session based on current intent, and Spotify introduced Talk to Spotify as a conversational interface for shaping what plays. YouTube Music Ask Music allows eligible Premium users to describe mood, activity, or desired listening experience by voice or text to create customized stations and mixes. These products validate that intent driven listening interfaces are strategically important. They do not establish that continuous AI Radio, as defined here, is superior or that any specific implementation has product market fit. \[2] \[3] \[4]

## 3. Radio as a temporal medium

Radio is useful to this discussion because its basic unit is not simply the track. Historically, radio programming has been shaped as a flow through time. Research on the development of broadcast scheduling describes how programme arrangement, scheduling blocks, logs, and recurring structures were used to manage sonic flow and listener attention. \[9]

Radio also functions in daily life as more than a content retrieval mechanism. Qualitative research on everyday radio listening has documented routines, background listening, companionship, mood regulation, comfort, and community as parts of the experience. These findings are context specific and should not be generalized to all audiences, but they illustrate why radio cannot be reduced to track ranking alone. \[10]

A continuous medium has relationships between elements that matter over time: a transition can be coherent or jarring, a familiar track can be restorative or repetitive, a spoken intervention can be helpful or intrusive, and a sequence can create an energy trajectory that no individual item expresses alone.

### Temporal properties that may matter

| Property           | Question                                                                                              |
| ------------------ | ----------------------------------------------------------------------------------------------------- |
| Continuity         | Does the experience feel like one evolving radio rather than a succession of disconnected results?    |
| Pacing             | How quickly should energy, novelty, density, or speech change?                                        |
| Recurrence         | When does familiarity create comfort, and when does it become repetition?                             |
| Transition quality | Do adjacent elements relate coherently in mood, energy, topic, or purpose?                            |
| Liveness           | Does the experience respond appropriately to present conditions rather than only to a static profile? |
| Companionship      | When speech exists, does it add presence without increasing unwanted cognitive load?                  |

&#x20;

AI Radio takes this temporal object seriously. It treats the state and direction of the radio as computational variables rather than as accidental byproducts of a ranked item list.

## 4. Definition and boundaries of AI Radio

<table data-header-hidden><thead><tr><th valign="top"></th></tr></thead><tbody><tr><td valign="top"><p>Proposed definition</p><p>AI Radio is a class of audio systems that continuously construct and adapt a temporally coherent listening experience by reasoning over current state, permitted context, expressed intent, prior observations, and governing constraints. Unlike item recommendation or one time playlist generation, an AI Radio system observes the unfolding experience and can revise subsequent editorial decisions while preserving continuity.</p></td></tr></tbody></table>

&#x20;

### Mandatory category properties

| Property                              | Minimum meaning                                                                                                                                    |
| ------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------- |
| Continuity                            | The primary output is an ongoing temporal experience rather than merely a finite result set.                                                       |
| Temporal reasoning                    | Sequence, duration, pacing, transition, and prior state can affect subsequent decisions.                                                           |
| Contextual adaptation                 | The system can adapt using permitted context, explicit intent, or observed listening behavior.                                                     |
| Editorial abstraction                 | The system reasons about experience level properties above individual item relevance, such as flow, contrast, recurrence, novelty, and continuity. |
| Continuous observation and replanning | The system can observe outcomes as the experience unfolds and revise future decisions.                                                             |
| Constraint enforcement                | Rights, policy, safety, availability, and user controls constrain what may be emitted.                                                             |

&#x20;

### Optional maturity properties

•  Durable personal memory across sessions.

•  Spoken hosting or presenters.

•  Generative media.

•  Multiple cooperating intelligence systems.

•  Explicit separation between proposal, validation, commitment, and execution services.

•  External control APIs or machine interfaces.

•  Local models.

•  Time based accounting.

•  Cryptographic provenance or timestamp anchoring.

### Boundary test

| Adjacent system                     | AI Radio by itself? | Reason                                                                                              |
| ----------------------------------- | ------------------- | --------------------------------------------------------------------------------------------------- |
| Generated playlist                  | No                  | A one time generated sequence can remain static after generation.                                   |
| Conversational music search         | No                  | Natural language input changes the interface, not necessarily the temporal system.                  |
| Autoplay recommender                | Depends             | It may qualify only if it maintains state, observes outcomes, and replans the unfolding experience. |
| Voice presenter over a fixed stream | No                  | Presentation does not create adaptive editorial state.                                              |
| Generative podcast                  | No                  | Generation alone does not imply continuous observation and replanning.                              |
| Personalized linear channel         | Depends             | It may qualify if the channel adapts during use under a persistent state model.                     |
| Continuously generated music        | Depends             | Generation can be an input or output method, but it is neither necessary nor sufficient.            |

&#x20;

The definition is deliberately implementation independent. A system may use rules, statistical models, reinforcement learning, large language models, graph systems, human editorial policies, or combinations of these. The category is defined by system behavior over time, not by a particular model family.

## 5. Abstract state model

A minimal analytical model can express AI Radio without specifying production architecture. Let Sₜ represent the current radio state at time t. Let Xₜ represent permitted observations available at that time, including relevant context, explicit intent, and prior experience. Let G represent governing constraints such as user controls, rights, availability, safety, and policy.

The next state is represented abstractly as:

<p align="center">Sₜ₊₁ = Φ(Sₜ, Xₜ ; G)</p>

The function Φ does not prescribe an algorithm. It represents the category level idea that the next radio state depends on what the radio currently is, what the system is permitted to observe, and the constraints that govern change.

The resulting experience produces observations Oₜ. In systems that retain durable personal state, an optional memory process may be represented as:

<p align="center">Mₜ₊₁ = Ψ(Mₜ, Oₜ ; G)</p>

Durable memory is therefore a maturity property rather than a strict category requirement. An ephemeral AI Radio can still be stateful within a session. A production system that retains personal state should make retention, scope, correction, deletion, and access governable.

<p align="center">Figure 2. Abstract state model. The formalism describes a class of systems, not a production protocol.</p>

This formulation also clarifies why longer listening cannot automatically be interpreted as greater welfare or satisfaction. Oₜ is evidence about behavior. It is not a direct observation of utility. Research on recommender systems increasingly distinguishes immediate engagement from longer term satisfaction and retention, and has shown the difficulty of inferring user utility from behavioral signals alone. \[11] \[12] \[13]

## 6. Memory, context and intent

Music preference is context dependent. Research has shown that contextual controls can improve perceived recommendation quality in some settings, and systematic reviews document a broad range of contextual variables used in music recommendation. These findings support context as a useful input, but not a universal claim that more context always improves experience. \[7] \[8]

AI Radio distinguishes three conceptually different inputs. Intent is what the listener explicitly asks for. Context describes permitted information about the present situation. Memory represents prior observations retained under defined governance. Conflating them creates both technical and privacy problems.

| Input   | Example                                                                    | Governance question                                                  |
| ------- | -------------------------------------------------------------------------- | -------------------------------------------------------------------- |
| Intent  | “Keep me focused while I work.”                                            | How long should this request remain active?                          |
| Context | Time of day, activity, device, coarse environment state.                   | Is the information necessary, permitted, and appropriately scoped?   |
| Memory  | Prior radio outcomes, stable preferences, corrections, recurring routines. | What persists, for how long, with what provenance and deletion path? |

&#x20;

### Context without raw surveillance

Context awareness does not imply continuous transfer of raw sensors. A system can prefer semantic observations that are proportionate to the task, such as “walking,” “conversation active,” or “speech preference low,” rather than a continuous microphone stream. This is a design principle, not a guarantee of privacy.

The data minimization problem is itself nontrivial. Research on personalization under GDPR data minimization requirements shows that useful personalization may remain possible under reduced data collection, but that minimization strategies can affect users differently. \[14]

### Memory governance

•  Scope: which user, session, device, household, or organization does the state belong to?

•  Provenance: what observation or explicit action produced the memory?

•  Confidence: how certain is the system that the state remains relevant?

•  Retention: should it persist for minutes, days, months, or not at all?

•  Correction: can the listener reverse or override an inferred preference?

•  Forgetting: when should older evidence lose influence?

•  Isolation: how are personal and shared listening states prevented from contaminating one another?

These are not decorative privacy features. They affect editorial quality. Memory that cannot decay or be corrected can make an adaptive radio progressively less adaptive.

## 7. Editorial intelligence and governing authority

AI Radio requires decisions about the experience above individual item relevance. Editorial intelligence may reason about pace, contrast, familiarity, novelty, recurrence, speech density, transition, topic, and trajectory. Different implementations can use different algorithms to do so.

For trustworthy production systems, this paper recommends a separation between intelligence that proposes behavior and authority that commits an action. The reason is practical: learned systems can be probabilistic, stale, unavailable, or wrong, while rights, policy, safety, user control, and system state may require deterministic enforcement.

<table data-header-hidden><thead><tr><th valign="top"></th></tr></thead><tbody><tr><td valign="top"><p>Design principle</p><p>Probabilistic intelligence may propose a transition. A governing system should remain able to reject, constrain, or replace that proposal before it becomes part of the delivered experience.</p></td></tr></tbody></table>

&#x20;

This separation is not a mandatory property of every research prototype and does not prescribe a particular service architecture. It is a production governance principle. The category can be researched with simpler systems, while deployed systems may require stronger authority boundaries.

### Why the distinction matters

•  Rights can change independently of model confidence.

•  User controls must override inferred preferences.

•  A proposed item can become unavailable between planning and playback.

•  A spoken intervention may be contextually inappropriate even if semantically relevant.

•  A model can hallucinate a fact, misread intent, or produce an invalid action.

•  Economic records should follow verified system events rather than model assertions.

## 8. Continuous planning, rendering, observation and adaptation

The distinctive loop of AI Radio is not merely recommendation followed by autoplay. It is continued observation and the possibility of revision. The system plans enough of the future to preserve coherence, renders or selects the next part of the experience, observes what happens, and updates later decisions.

Sequential recommendation and reinforcement learning research offer useful neighboring ideas because they explicitly model future effects and long term objectives. Work in large scale systems has moved beyond immediate click objectives toward retention or longer term satisfaction, while also exposing the difficulty of delayed and partially attributable rewards. \[11] \[12] \[13]

AI Radio does not require reinforcement learning, but it shares a temporal problem: a locally attractive choice can degrade the experience over a longer horizon. A high energy track may be individually relevant yet wrong after sustained high intensity. A host intervention may be individually interesting yet harmful if speech has already been dense.

### A receding editorial horizon

A useful conceptual analogy is receding horizon planning: make a bounded plan, execute a portion, observe the updated state, and plan again. In AI Radio, the planning horizon may be seconds, tracks, segments, or broader narrative phases. The appropriate horizon is an empirical question.

| Planning horizon | Potential strength                                        | Potential risk                                                   |
| ---------------- | --------------------------------------------------------- | ---------------------------------------------------------------- |
| Very short       | Highly responsive to immediate intent and context.        | Can produce jitter, repetition, and weak long range coherence.   |
| Medium           | Can coordinate transitions, pacing, and speech density.   | Requires state quality and some prediction of future experience. |
| Long             | Can express broader trajectories and recurring structure. | Can become stale as context or user intent changes.              |

&#x20;

### What counts as feedback?

A continuous system can observe explicit corrections, requests, stops, skips, session endings, repeated returns, and other signals. None is a pure measure of satisfaction. The evaluation problem is therefore causal and longitudinal: did the adaptation improve the subsequent experience relative to a credible alternative?

## 9. Trust, provenance and rights records

Continuous media crosses several evidence domains that should not be collapsed into one word such as “provenance” or “blockchain.” Identity, asset integrity, provenance assertions, rights, usage, accounting, and payment answer different questions.

| Layer                  | Question it answers                                              | What it does not prove by itself                        |
| ---------------------- | ---------------------------------------------------------------- | ------------------------------------------------------- |
| Identity               | Who or what signed, submitted, or acted?                         | That the underlying claim is true.                      |
| Asset integrity        | Is this asset the same data that was previously bound or signed? | Ownership or permission.                                |
| Provenance assertions  | What history is asserted for the asset?                          | That every assertion is complete or legally sufficient. |
| Rights and permissions | Who is authorized to use the asset, where, and for what?         | That use actually occurred.                             |
| Usage records          | What system event or listening event occurred?                   | How revenue should be allocated.                        |
| Accounting             | How are verified events translated into economic records?        | That a given allocation is ethically fair.              |
| Payment                | How is value transferred?                                        | Provenance, ownership, or accurate measurement.         |

&#x20;

### C2PA

The Coalition for Content Provenance and Authenticity defines Content Credentials as signed provenance structures that can be cryptographically bound to assets. The specification supports verifiable association between assertions and content and is designed for tamper evidence. It does not require blockchain, does not independently prove ownership, and does not make every assertion true merely because the manifest is cryptographically valid. \[15]

### Timestamp anchoring

Timestamping systems such as OpenTimestamps can provide evidence that a particular data commitment existed no later than a given time. OpenTimestamps supports independently verifiable timestamp proofs and can use Bitcoin attestations. Such a proof does not by itself establish authorship, ownership, consent, licensing, or provenance completeness. \[16]

A media system can therefore use cryptography selectively: signed provenance for asset history, timestamp anchoring for existence evidence, controlled rights records for permission, and separate listening evidence for usage. None requires a platform cryptocurrency or governance token.

## 10. Attribution in continuous media

The discrete event model of media economics often uses streams, plays, views, impressions, or other events as accounting units. Continuous media makes elapsed and attributable listening time a natural alternative primitive. That does not mean time is intrinsically fair, sufficient, or economically optimal.

Four separate questions must remain distinct:

•  Measurement: how much listening occurred?

•  Attribution: which works, participants, or editorial contributions were present during that time?

•  Allocation: how is a revenue pool or other economic value distributed using those measurements?

•  Fairness: which allocation principles should be considered legitimate, equitable, efficient, or culturally desirable?

Research on user centric streaming allocation illustrates why these steps cannot be collapsed. The Centre national de la musique found that linking each subscription to what that listener actually consumed changes the distribution of revenue, but explicitly noted that quantitative analysis cannot determine whether one model is inherently more fair or effective. A 2024 study using 154,505 users and 890 million streams similarly found substantial redistribution effects under alternative models rather than a universal winner. \[17] \[18]

<table data-header-hidden><thead><tr><th valign="top"></th></tr></thead><tbody><tr><td valign="top"><p>Economic research hypothesis</p><p>Continuous media can make verified listening duration a useful attribution primitive. Whether and how that measure should affect payments, creator value, editorial value, or rights settlements is an empirical, contractual, legal, and normative question.</p></td></tr></tbody></table>

&#x20;

The phrase “value follows time” should therefore be read as a thesis to test. Listening time is observable behavior. It is not identical to attention, satisfaction, cultural value, authorship, or fairness.

## 11. Evaluation framework

AI Radio should not be evaluated only through item accuracy or raw session length. A category defined by continuity requires experience level and longitudinal measures, while a trustworthy production system requires privacy, rights, and operational metrics.

| Dimension             | Candidate measures                                                                                     | Important caution                                                             |
| --------------------- | ------------------------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------- |
| Continuity and flow   | Transition quality, abruptness, repeated correction, sequence coherence, perceived flow.               | Requires human evaluation and context specific baselines.                     |
| Intent fulfillment    | Successful request interpretation, time to adaptation, correction burden.                              | A fulfilled request can still harm longer term flow.                          |
| Longitudinal utility  | Return frequency, meaningful listening days, voluntary session continuation, retained listening hours. | Duration and retention are not complete measures of welfare.                  |
| Memory benefit        | Difference between memory enabled and memory limited conditions in later sessions.                     | Novelty effects and selection bias must be controlled.                        |
| Agency                | Rate of overrides, explicit controls used, reversibility, user understanding.                          | Low interaction may mean effortless success or passive disengagement.         |
| Speech and hosting    | Speech tolerance, interruption cost, relevance, repetition, perceived presence.                        | Different activities and listeners can require very different speech density. |
| System performance    | Latency, failure rate, fallback rate, cost per listening hour, delivery reliability.                   | Operational efficiency is not experience quality.                             |
| Privacy               | Data volume, retention, deletion success, scope violations, unnecessary sensor transfer.               | Compliance and user trust require more than metric optimization.              |
| Rights and accounting | Eligibility enforcement, provenance completeness, usage record integrity, dispute resolution.          | Technical correctness does not settle contract interpretation.                |

&#x20;

### Experimental design

Several core claims require comparative experiments rather than observational storytelling. Examples include memory versus no durable memory, context adaptation versus static personalization, continuous replanning versus one time sequence generation, and spoken intervention policies at different densities. The strongest evidence will connect intervention to subsequent behavior and reported experience while measuring cost and failure modes.

### A useful hierarchy of evidence

•  Technical feasibility: the system can perform the intended state transition.

•  Experience evidence: listeners perceive the intended improvement.

•  Behavioral evidence: the improvement changes repeated voluntary use.

•  Economic evidence: the improvement can be delivered at sustainable marginal and fixed cost.

•  Institutional evidence: rights holders, creators, regulators, or partners can rely on the system records and controls.

## 12. musen as one implementation and research platform

musen is one working implementation through which parts of the AI Radio model are being developed and tested. This section is intentionally dated. It should not be used to infer that every concept in the paper is currently implemented or publicly available.

<table data-header-hidden><thead><tr><th valign="top"></th></tr></thead><tbody><tr><td valign="top"><p>Product status date</p><p>The status statements below describe musen as of August 2026. They are product facts, plans, or research directions, not category requirements.</p></td></tr></tbody></table>

&#x20;

| Status                 | musen example                                                                                                                                 | Interpretation                                                                                       |
| ---------------------- | --------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------- |
| Publicly available     | Free personal Live Radio on mobile with account based continuity and adaptive requests. Optional AI Host speech can appear within Live Radio. | Current public product.                                                                              |
| Implemented internally | Server operated orchestration, persistent listening state, continuous stream execution, and listening evidence infrastructure.                | Implementation exists in some form, but completeness and performance should be evaluated separately. |
| Planned                | Web interface, stronger public control surfaces, creator and rights workflows, broader attributable listening infrastructure.                 | Product or infrastructure direction, not current availability.                                       |
| Research direction     | Large Radio Models, local or hybrid intelligence, machine interfaces, broader open developer ecosystem.                                       | Exploratory. No product commitment is implied.                                                       |

&#x20;

The current musen architecture also adopts a production principle consistent with Section 7: editorial intelligence can propose what should happen, while authoritative systems remain responsible for rights, policy, state, and execution. The detailed internal protocol, model boundaries, state machinery, and rights mechanisms are outside the scope of this public category paper.

musen does not currently require generative music for the AI Radio thesis. The product direction emphasizes AI for orchestration, continuity, context, and hosting rather than using content generation as the definition of intelligence.

## 13. Open research questions, limitations and falsification

### Open research questions

•  What aspects of temporal coherence are perceived by listeners, and which can be measured reliably?

•  When does persistent state improve later sessions, and when does it create stale or overconfident personalization?

•  What is the optimal planning horizon for different listening contexts?

•  How much context materially improves radio after controlling for novelty and interface effects?

•  How should intent expire, conflict, or coexist with learned preferences?

•  When does spoken hosting improve companionship, discovery, or orientation, and when does it become interruption?

•  How should group, household, vehicle, or workplace listening affect personal memory?

•  Can local interpretation of context reduce privacy exposure without degrading continuity?

•  Which measures of listening behavior are most predictive of long term satisfaction without becoming manipulative optimization targets?

•  How should rights constraints and editorial objectives interact when catalog availability changes?

•  Can attributable listening records support more transparent economic models without creating surveillance or false claims of fairness?

•  What technical and cultural properties distinguish a durable AI Radio habit from novelty around conversational interfaces?

### Limitations

•  This is a conceptual synthesis, not a controlled experiment.

•  The definition is proposed by musen and has not been adopted as an external standard.

•  Existing research on recommendation, radio, HCI, privacy, and streaming economics supports parts of the argument but does not validate the category as a whole.

•  Current competitor products validate interest in intent driven listening interfaces, not demand for musen or the superiority of continuous AI Radio.

•  Longer sessions, higher retention, or more listening time are not complete measures of user welfare or cultural value.

•  Contextual adaptation can create privacy, inference, and autonomy risks if inputs are excessive or poorly governed.

•  Time based attribution can redistribute revenue but cannot be described as universally fair without a normative framework and empirical evidence.

•  Cryptographic records improve evidence integrity but cannot turn a false or incomplete assertion into a true one.

•  Rights and licensing remain jurisdictional and contractual problems. This paper is not a licensing analysis.

•  The public implementation description is intentionally less specific than internal technical and intellectual property materials.

### Falsification conditions

The central AI Radio thesis would be weakened if one or more of the following consistently hold under credible tests:

•  Users obtain equivalent sustained satisfaction from one time generated sequences without continuing observation or adaptation.

•  Temporal coherence provides no measurable benefit beyond item relevance.

•  Contextual adaptation fails to outperform simpler personalization after controlling for novelty and interface effects.

•  Persistent state provides no durable benefit or creates harms that outweigh its value.

•  Separating probabilistic intelligence from governing authority provides no meaningful safety, rights, or reliability advantage in production.

•  Continuous orchestration cannot be delivered at economically sustainable cost.

•  Listening time has weak or unstable relationships with listener utility and legitimate creator attribution.

•  Users in the intended use cases consistently prefer explicit item control over effortless continuous orchestration.

<table data-header-hidden><thead><tr><th valign="top"></th></tr></thead><tbody><tr><td valign="top"><p>Research standard</p><p>A category earns legitimacy when it makes predictions that can fail. AI Radio should be judged by whether continuity, state, context, and ongoing adaptation create measurable value beyond simpler alternatives.</p></td></tr></tbody></table>

&#x20;

## 14. Conclusion

Streaming solved access at extraordinary scale. Recommendation reduced search cost. Conversational interfaces are reducing the effort required to express intent. Generative systems are increasing the volume and speed at which media can be produced. None of these developments makes continuous radio inevitable. Together, however, they make orchestration a more important problem.

AI Radio is proposed here as a class of systems whose primary object is an unfolding temporal experience. The system maintains state, reasons over permitted observations and constraints, makes editorial decisions above individual item relevance, observes outcomes, and can revise what happens next while preserving continuity.

This framing does not require generative music, a speaking host, a particular model family, a token, or a specific business model. Those are implementation choices. The category claim is narrower: continuous listening can itself be treated as a computational object.

The scientific burden now is empirical. Does continuity matter beyond relevance? Does context improve experience without creating disproportionate privacy cost? Does governed state produce better subsequent sessions? Can such systems remain controllable, rights aware, and economically sustainable? Can time become a useful attribution primitive without being mistaken for a complete measure of value?

<table data-header-hidden><thead><tr><th valign="top"></th></tr></thead><tbody><tr><td valign="top"><p>Final thesis</p><p>In environments of abundance, orchestration becomes the scarce capability. For continuous media, time may become a natural basis for attribution. Whether value follows time is not a conclusion of this paper, but a thesis that AI Radio makes possible to test.</p></td></tr></tbody></table>

&#x20;

<br>

&#x20;

## References

\[1] IFPI. Global Music Report 2026: State of the Industry. March 2026. [Source](https://www.ifpi.org/global-music-report-2026-global-recorded-music-revenues-grow-6-4-as-record-companies-drive-innovation/)

\[2] Spotify. Your Personalized DJ Experience Is Expanding With 4 New Languages and More Markets. 7 May 2026. [Source](https://newsroom.spotify.com/2026-05-07/dj-expansion-4-new-languages/)

\[3] Spotify. Just Say the Word: A More Personal Way to Ask, Discover, and Listen. 14 July 2026. [Source](https://newsroom.spotify.com/2026-07-14/talk-to-spotify-announcement-beta/)

\[4] YouTube Music Help. Ask Music for music and podcasts. Accessed August 2026. [Source](https://support.google.com/youtubemusic/answer/17090260?hl=en)

\[5] Deezer. AI music has surpassed 50 percent of new music uploads for the first time. 21 July 2026. [Source](https://newsroom-deezer.com/2026/07/ai-music-exceeds-50-percent-daily-uploads-deezer/)

\[6] Hansen, C., Hansen, C., Maystre, L., Mehrotra, R., Brost, B., Tomasi, F., and Lalmas, M. Contextual and Sequential User Embeddings for Large Scale Music Recommendation. RecSys 2020. [Source](https://doi.org/10.1145/3383313.3412248)

\[7] Jin, Y., Htun, N. N., Tintarev, N., and Verbert, K. ContextPlay: Evaluating User Control for Context Aware Music Recommendation. UMAP 2019. [Source](https://doi.org/10.1145/3320435.3320445)

\[8] Pichl, M. and Zangerle, E. Context Aware Recommender Systems in the Music Domain: A Systematic Literature Review. Electronics, 2021, 10, 1555. [Source](https://doi.org/10.3390/electronics10131555)

\[9] VanCour, S. Making Radio Time: Managing Broadcasting’s Sonic Flows. In Making Radio, Oxford University Press, 2018. [Source](https://doi.org/10.1093/oso/9780190497118.003.0002)

\[10] Krause, A. E. The Role and Impact of Radio Listening Practices in Older Adults’ Everyday Lives. Frontiers in Psychology, 2020, 11:603446. [Source](https://doi.org/10.3389/fpsyg.2020.603446)

\[11] Ding, R., Xie, R., Hao, X., Yang, X., Ge, K., Zhang, X., Zhou, J., and Lin, L. Interpretable User Retention Modeling in Recommendation. RecSys 2023. [Source](https://doi.org/10.1145/3604915.3608818)

\[12] Agarwal, A., Usunier, N., Lazaric, A., and Nickel, M. System 2 Recommenders: Disentangling Utility and Engagement in Recommendation Systems via Temporal Point Processes. FAccT 2024. [Source](https://doi.org/10.1145/3630106.3659004)

\[13] Wu, Y., Chang, D., She, J., Zhao, Z., Wei, L., and Heldt, L. Learned Ranking Function: From Short Term Behavior Predictions to Long Term User Satisfaction. RecSys 2024. [Source](https://doi.org/10.1145/3640457.3688184)

\[14] Biega, A. J., Potash, P., Daumé, H., Diaz, F., and Finck, M. Operationalizing the Legal Principle of Data Minimization for Personalization. SIGIR 2020. [Source](https://doi.org/10.1145/3397271.3401034)

\[15] Coalition for Content Provenance and Authenticity. C2PA Technical Specification and Content Credentials documentation. [Source](https://spec.c2pa.org/)

\[16] OpenTimestamps. A timestamping proof standard. Accessed August 2026. [Source](https://opentimestamps.org/)

\[17] Centre national de la musique. User Centric Payment System study. [Source](https://cnm.fr/en/impact-of-online-music-streaming-services-adopting-the-ucps/)

\[18] Alternative payment models in the music streaming market: A comparative approach based on stream level data. Information Economics and Policy, 2024, 68, 101103. [Source](https://doi.org/10.1016/j.infoecopol.2024.101103)

Source note. Product announcements and platform support pages are used only for dated observations about current interfaces. Academic references are used to locate the category in prior work, not to imply that any single study validates AI Radio as a whole.
