As head of Splice, CEO Kakul Srivastava has tried to carefully thread the line on AI music tools. She also hates her cup warmer.
The exchange was published on October 3, 2026, in The Verge's AI vertical. In it, Splice chief executive Kakul Srivastava named AI-written email — auto-generated meeting requests, follow-ups and client greetings — as the reason a thread that could move work forward instead becomes a series of polite exchanges carrying no new information. Her reasoning is concrete: once a recipient senses the text on the other side came from a model, willingness to reply drops and the chain of communication breaks.
Srivastava has led Splice since 2021, after holding product and design executive roles at Adobe and Flickr. Splice sells one-shots, loops and presets to producers on a subscrip
The remark lands where two preconditions intersect. One is the inbox. Since 2023, generative text has moved from demo to default: Google embedded Gemini inside Gmail, Microsoft placed Copilot inside Outlook, and writing a presentable email turned from a skill into a button. By 2026 a large share of routine workplace communication is automated, tolerance for templated phrasing has fallen sharply, and asking whether a message was written by a model is now routine.
The other is music. Generative music tools including Suno and Udio scaled quickly in 2023 an
Liu Gong argues that Srivastava has put a finger on an underrated structural weakness of generative AI. From a technical standpoint, a language model is trained to maximize the probability of the next token, which pushes its output toward the highest-probability phrasing: the least offensive, the most standard, and the emptiest of new information. Email, meeting notes and standard replies sit squarely in that sweet spot, which is why better tools produce more filler. Conversation, by contrast, derives its value from deviation — hesitation, a counter-question, a change of mind mid-sentence. Those low-probability moves are the information. The studio analogy fits: an AI-written email is like a heavily pitch-corrected vocal take. Perfectly in tune, no breath, and identifiable as non-human within ten seconds.
On the business side, the criticism doubles as commercial defense. Splice earns from subscriptions plus download credits, so users pay for what they can pull out of a library built by human creators. Generative music tools charge by subscription or generation, competing for the same producer's monthly budget. If models can reliably produce usable loops and one-shots, Splice's pricing needs a new justification. Keeping AI in the retrieval and recommendation layer while human material stays at the center of value is exactly that justification. She is not rejecting AI; she is drawing a line it should not cross.
For the industry, that line is being borrowed. Adobe's Firefly works from licensed data; Splice works from creator uploads. Both sell provenance and traceability rather than competing with models on raw generation. Liu Gong expects clearer stratification within a year: models at the bottom, licensed material and retrieval services in the middle, human judgment and attribution on top. Companies trying to own all three layers will feel the pressure first. The real value of Srivastava's remark is that it turns 'where AI should not be used' into a product question open to public debate.
- Splice runs a subscription plus download-credit model, with entry tiers around $9.99 per month and creator tiers around $19.99, directly overlapping the $10—$30 monthly range of generative music tools such as Suno and Udio, which chase the same paying producers.
- Commercial leverage of a single sample: both Lisa's 'Money' and Sabrina Carpenter's 'Espresso' drew on Splice. At Spotify's commonly cited $0.003—$0.005 per stream, a billion-stream hit implies a royalty pool of roughly $3m—$5m, of which a single sample license is a tiny fraction.
- Adoption
For the industry, the most immediate shift over the next three months is labeling. Music and content platforms will more actively separate 'AI-generated' from 'AI-assisted,' writing sample provenance and licensing chains into product pages and contracts. On the enterprise side, countermeasures will appear in communication norms: teams agreeing on which emails must be handwritten, and requiring disclosure on automated follow-ups. In the short term AI email volume will not fall, but its credibility premium will start to erode.
Stretched to a year, pricing
Questions worth exploring:
- Why the efficiency advantage of generative AI in standardized settings ends up destroying the trust those settings depended on.
- How pricing power over human-sourced material gets renegotiated once sample subscriptions collide head-on with model subscriptions.
- From Adobe Firefly to Splice, how licensed data and creator uploads are becoming a shared defense line against general-purpose models.
This item comes from The Verge AI (published 2026-10-03 15:00:00). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.