Deezer reports 44% of daily uploads are now AI-generated tracks while actual listener demand remains near zero, Spotify launches opt-in AI credit disclosure in beta, Splice ships creator-compensating generative AI tools, UPenn researchers debut text-prompt spatial audio editing, and GRAI raises $9M on a bet that AI's real music future is social remixing — not generation from scratch.
Deezer: 44% of New Uploads Are AI — and Almost Nobody Is Listening to Them
Deezer is now receiving roughly 75,000 AI-generated tracks per day, representing 44% of all new music uploaded to the platform. The numbers sound alarming until you look at consumption: AI tracks account for just 1–3% of total streams, and Deezer flags 85% of those streams as fraudulent activity and demonetizes them. The flood is almost entirely artificial engagement farming, not organic listener demand finding a new format.
For working composers and sound designers, the near-term implication is that detection is holding — for now. The more important signal is what happens downstream: filtering 75,000 tracks per day is already a significant ML infrastructure cost, and that volume will only grow as generation costs approach zero. The platforms building the most defensible detection and curation layers right now are positioning themselves for when the next generation of AI audio becomes harder to fingerprint.
Spotify Launches Voluntary AI Credit Tags — But Only If Artists Opt In
Spotify began a beta rollout of AI Credit disclosure on April 16, letting artists voluntarily flag how AI was used in their tracks — vocals, lyrics, instrumentals, or production — with the tag appearing inside the Song Credits section of the mobile app. The rollout started with DistroKid users and will expand to CD Baby, Believe, EMPIRE, and others in the coming weeks. The underlying data standard is being developed with DDEX.
The critical limitation is the word "voluntary." There's no detection layer — if an artist doesn't disclose, no tag appears, and listeners have no way to know. For now this functions as a credentialing tool for artists who want to claim human authorship, not a reliable signal for listeners trying to filter AI content. Its long-term value depends entirely on whether distributors eventually make disclosure mandatory or whether the DDEX standard gets adopted broadly enough to create real transparency pressure.
Splice Pays Sample Creators When AI Remixes Their Work
Splice launched Variations and Craft in beta this month — two AI tools that extend its existing pay-on-download model into generative territory. Variations generates alternate versions of any Splice sample with adjusted key, tempo, and structure while preserving the original's core character. Craft converts samples into fully playable instruments. The essential detail: every Variation download triggers a payment to the original sample creator, traced through Splice's catalog of over 3 million sounds. Magic Fit — which auto-matches any sample to your session's harmonic and rhythmic context — arrives Summer 2026.
This is a materially different approach from how most AI audio tools handle training data and compensation. Splice is building traceability and payment directly into the workflow rather than treating source material as a legal gray area to be litigated later. If you rely on sample packs professionally, this model is worth tracking: it's the first commercially deployed system that makes AI-generated sample variation feel like a licensed arrangement rather than an uncredited copy.
SmartDJ Edits Spatial Audio Using Plain-Text Prompts
Researchers at the University of Pennsylvania published SmartDJ this week, an AI system for editing immersive stereo audio using natural language. Give it a high-level instruction — "make this sound like a busy office" or "move the traffic further back" — and SmartDJ decomposes it into a sequence of discrete edits using an audio language model trained on both sound and text. A diffusion model then executes those edits step by step while preserving spatial cues and 3D placement. In quantitative evaluations and human listening studies it outperformed prior methods on audio quality, instruction adherence, and spatial realism.
This is research, not a shipping product, but the direction is clear and the timeline to practical tools is shorter than it looks. Text-editable spatial audio matters everywhere from game audio and VR to post-production workflows where reshaping an environment currently requires dismantling and rebuilding individual layers. If you work in immersive sound, this class of system is the one to watch for DAW integration within the next two years.
GRAI Raises $9M on a Bet That AI Music's Future Is Social Remixing
Warsaw-based music lab GRAI closed a $9 million seed round co-led by Khosla Ventures and Inovo VC, with participation from the a16z Scout Fund and several angels. The company's thesis cuts directly against the dominant AI music narrative: most people don't want AI to generate music from scratch — they want to remix songs they already love, share those remixes, and play around with tracks socially. GRAI is building around that model, with a remixing app for iOS and an AI music playground for Android already in market. Founders previously sold video-creation app Vochi to Pinterest.
The licensing question will define whether this works at scale. Remixing commercially released music requires rights holder buy-in, and GRAI says it's engaging labels up front rather than building first and asking forgiveness later. If it can actually clear music and build the social layer, it's a different product category from Suno or Udio — and one that may face less artist and label resistance, since the model amplifies existing catalogs rather than replacing them.
Grammys on the Hill Puts AI Voice Protections at the Center of Its Congressional Push
More than 200 Recording Academy members, producers, and industry leaders converged on Washington, D.C. this week for the 25th Grammys on the Hill, with AI protections as the explicit legislative focus. The event honored senators and representatives championing the NO FAKES Act — a bill that would establish federal protections against unauthorized AI-generated replicas of a person's voice and likeness. The bipartisan framing signals that momentum for federal AI legislation in music has moved past talking points into active lobbying.
For producers and studios working with voice talent, the NO FAKES Act's passage would meaningfully clarify the legal landscape. Currently, consent requirements for AI voice replication vary by state, and the floor is largely set by platform-level terms of service — which is to say, inconsistently. Federal legislation would create a baseline. The outstanding question is whether the bill's scope will be limited to voice likeness or extended to cover training data practices more broadly.
Sonilo generates full-length music soundtracks directly from video input — no text prompt, no style description needed. Add it as a native node in ComfyUI, feed it a video clip, and Sonilo reads the visual timing, pacing, and emotional arc to compose a matching score in around 20 seconds. The output matches your video's exact length, lands its structural beats, and ends cleanly without manual trimming or looping. All output is cleared for commercial and broadcast use. The integration launched as an official ComfyUI Partner Node on April 14.
If you're scoring video inside ComfyUI workflows — ads, social content, trailers, longer-form film — this closes the gap between visual generation and audio composition without context-switching to a separate tool. The video-as-input model is the real differentiator: it's reading actual motion, cuts, and pacing rather than a text description of them, which gives it a structural accuracy advantage over text-to-music for anything where timing to picture matters.
Deezer's 44%/1–3% split — AI comprising nearly half of new uploads but a rounding error of actual streams, with 85% of those streams flagged as fraud — reveals a fragile equilibrium. Streaming platforms are currently winning the detection game against today's AI audio. The problem is that the economics are asymmetric: generating an AI track costs fractions of a cent and the volume is doubling regularly, while detection requires active ML investment that scales with volume. At 75,000 flagged tracks per day, Deezer is already running what amounts to a continuous adversarial audio classification operation.
When detection quality degrades — and it will, as generation models improve — the consequences hit legitimate artists disproportionately: suppressed discoverability, eroded streaming royalty pools, and listener trust in platform recommendations. The companies to watch here aren't the music AI generators; they're the audio fingerprinting and forensic detection startups quietly signing deals with majors and DSPs. SoundPatrol, already working with UMG and Sony, is one. The detection layer is becoming as strategically important as the generation layer — and right now it's significantly underbuilt relative to the problem it's being asked to solve.