Producing more has never been easier. Being chosen remains difficult.
Deezer reported in 21 July 2026 that it received on average about 90 thousand fully AI-generated tracks per day in June. On peak days, this content accounted for over 50% of the platform’s daily new uploads. Despite the volume, it accounted for only 1%–3% of plays.
The numbers don’t say that half of what people listen to is AI music. They say almost the opposite: at Deezer and during observed peaks, more than half of the new catalog delivered could be synthetic, while actual listening share remained small.
This contrast matters far beyond the music industry. When automation drastically reduces the cost of producing and publishing, the bottleneck shifts. It moves from generation to curation, trust, distribution, contextual fit, and outcome.
What do Deezer’s data show?
The June milestone didn’t appear out of nowhere. In April, Deezer reported nearly 75 thousand AI tracks per day, about 44% of new uploads. Two months later, the monthly average rose to about 90 thousand and exceeded 50% at peak volume moments.
The company classifies tracks as fully AI-generated if identified by its proprietary detector. This detail is important: the percentages depend on Deezer’s own technology and criteria, and the announcement doesn’t mention an independent audit of the results.
The platform also claims to remove this content from algorithmic recommendations and editorial playlists. This helps explain why supply and consumption don’t grow proportionally. Availability may rise due to automation; discovery remains governed by filters, context, preference, and trust.
There’s also a layer of fraud. According to Deezer, up to 85% of streams of fully AI-generated tracks were classified as fraudulent in 2025 and excluded from royalty payments. The number doesn’t mean 85% of the music is fraudulent nor does it apply to the entire catalog. It refers to streams of these tracks detected by the platform in that period.
In response, the company announced it will remove AI tracks linked to fraud and those with no plays for six months. It’s a stance shift: storing everything stops being neutral when abundance creates cost, dilutes signals, and opens room for manipulation.
The cost of producing dropped; the cost of choosing rose
For a long time, producing content was expensive. Recording, editing, distributing, and promoting required teams, infrastructure, and time. With generative AI, part of that cost tends to fall. This allows more experimentation, personalization, and serving niches previously unviable.
But production capacity doesn’t automatically create demand. When everyone can generate at scale, attention becomes scarcer. An organization that measures only quantity may celebrate growth the audience doesn’t notice.
The problem appears in marketing, sales, service, and operations. A company can produce a hundred pieces per week, generate thousands of messages or open dozens of automations. If the content doesn’t help someone decide, if the message doesn’t respect the customer’s moment, or if the process doesn’t close the cycle, volume is just digital stock.
This is one reason why process automation needs to start with the goal, not the tool. Automating a bad step multiplies its speed. Designing the complete process allows defining input, decision, action, confirmation, and exception before scaling.
Curation isn’t a brake: it’s value infrastructure
Curation is often treated as editorial selection or human taste. In AI systems, it’s broader. It includes choosing sources, defining quality criteria, excluding duplicates, labeling origin, limiting recommendations, observing behavior, and measuring real effect.
Without curation, the system optimizes what’s cheap to count. Uploads, messages, documents, and completed tasks shoot up quickly. Trust, usefulness, and revenue can stall.
A well-designed Digital Employee doesn’t exist to produce the maximum amount of text possible. It performs a function with goals, context, limits, and indicators. The guide on what a Digital Employee is shows why operational identity and responsibility matter as much as the model’s capacity.
The same applies to systems that work longer and cross multiple tools. Our analysis on AI agents at work highlights that value is in completing verifiable work, not in accumulating actions. And the discussion on AI memory as an operational risk reinforces that more context only helps when there is origin, validity, and correction possibility.
Four metrics better than raw volume
If generation became cheap, companies need to raise measurement standards. Four questions help:
- Useful consumption: did someone read, listen, respond to, or use what was produced?
- Fit: did the content reach the right person, at the right time, with the right context?
- Trust: is the origin clear, data traceable, and is there a way to correct or dispute?
- Outcome: was there a decision, time saved, higher conversion, resolution, or measurable learning?
These metrics prevent automation from becoming a vanity machine. Production remains relevant but becomes input. The result appears after selection and use.
In Deezer's case, the difference between over 50% of uploads at peak times and only 1%–3% of plays clearly shows this. Supply can grow faster than interest. When it does, increasing generation doesn't close the gap; it may even widen it.
The lesson for companies is not to produce less
The message is also not to abandon generative AI. The technology lowers barriers, speeds up experimentation, and lets small teams explore more possibilities. The mistake is treating abundance as synonymous with value.
Companies need to combine generation with rails: clear goals, quality criteria, risk-proportional review, context-driven distribution, and real-use feedback. AI can create options; the process must decide which deserve attention.
At XMACNA, this principle guides Cognitive Process Design. A Digital Employee is not evaluated by the number of phrases it writes but by the work performed within an observable contract: what it received, what it decided, what it did, what evidence it left, and when it called a person.
When production becomes almost free, the competitive advantage returns to being deeply human and operational: knowing what is worth producing, for whom, when, and by what success criteria.
Want to map where your company is confusing volume with result? Take the XMACNA Assessment and identify which processes need curation, indicators, and supervision before scaling.
Frequently asked questions
More than 50% of the music listened to on Deezer is AI-generated?
No. Deezer reported that fully AI-generated tracks accounted for over 50% of new uploads on peak days in June. They represented only 1%–3% of plays on the platform.
How many AI-generated tracks does Deezer receive daily?
The company reported an average of approximately 90 thousand fully AI-generated tracks per day in June 2026.
Does the 85% data mean that almost all AI music is fraudulent?
No. According to Deezer, up to 85% of streams of fully AI-generated tracks were classified as fraudulent in 2025. The percentage refers to plays, not to the number of songs or the entire catalog.
Have the numbers been independently audited?
The announcement is based on Deezer's proprietary detector and data. The consulted source does not provide independent auditing of the percentages, so the text attributes the numbers to the company.
What is the main lesson for companies?
Automation increases production capacity but doesn't guarantee attention, trust, or outcome. Companies need to measure useful consumption, suitability, traceability, and impact, beyond raw volume.