Deezer: AI-Generated Tracks Now Exceed Half of Daily Song Uploads

On July 21, music streaming service Deezer announced that more than 50% of the tracks uploaded to its platform each day are AI-generated content. As of June 2026, roughly 90,000 AI-generated tracks are being uploaded per day.

The figure is symbolically significant: it is the first publicly disclosed case in which a 'majority' of the new music flowing onto a platform is no longer composed by humans. Deezer has long worked on detecting and labeling AI-generated music, and this data release appears to be the first time the company has put hard numbers on the scale behind that effort.

For streaming services, the surge in AI-generated content also raises a practical question: how to design copyright royalty distribution. Deezer has not mentioned any specific changes to its payout scheme at this point, but the disclosure could accelerate the debate across the industry.

AI Companies Are Buying Up 'Old Books' to Avoid AI-Slop Contamination of Training Data

Bulk purchases of old books are surging among AI companies seeking 'clean' training data untainted by AI-generated text (AI slop), as revealed by a 404 Media report and discussion on Hacker News.

The background is that text data on the web is already heavily contaminated with AI-generated content. Training AI on AI-produced text has been shown to degrade model quality — a phenomenon known as 'model collapse.' Books published before the widespread adoption of AI therefore hold value as 'pure,' human-written data.

Specifically, publishers' back catalogs and collections of used books are reportedly being bought up as datasets. Only fragments have been disclosed about which companies are buying how much, and much remains unclear about how consent from copyright holders is being handled. The picture that emerges is one of AI training data procurement expanding from web scraping into the physical book market.

Gumroad's CEO: AI Token Costs Now on Par with Payroll

Sahil Lavingia, CEO of Gumroad, has revealed that the company's AI-related token costs have reached roughly the same level as its payroll. Gumroad is a platform for selling digital content and has been doubling down on AI features in recent years.

The statement is drawing attention as a concrete example of AI usage costs reaching 'payroll-scale' proportions. Even though per-token prices keep falling, the growth in usage is outpacing that decline — and as a result, the company's internal cost structure itself is shifting.

That said, this reflects the particular circumstances of one specific company, and any generalization to the wider industry warrants caution. Gumroad is a relatively small company, and its business structure — heavily weighted toward AI features — likely plays a part as well.

Cisco Releases 'Antares,' an Open-Weight Model for Identifying Vulnerabilities

Cisco has released Antares, an open-weight AI model for identifying vulnerabilities in software. The announcement came on the company's AI blog.

Antares specializes in vulnerability localization — pinpointing vulnerabilities within a codebase. There have been attempts to use general-purpose LLMs for vulnerability discovery, but security-focused open-weight models remain rare. Cisco says it aims to build 'the most efficient open-weight model.'

Releasing open weights in the security space not only contributes to the research community but also lowers the barrier to real-world adoption. That said, the risk of vulnerability-finding models being repurposed for attacks is also part of the debate.

Study Claiming AI Has 'Learning Benefits' Retracted

A research paper claiming that AI tools boost student learning outcomes has been retracted, Plagiarism Today reports.

The detailed reasons for the retraction have not been fully disclosed in the coverage, but problems with data reliability or methodology are suspected. Research on AI's educational effects is a mixed bag of positive and negative studies, and this retraction has added fuel to the debate over the quality of studies claiming an 'AI effect.'

On Hacker News, commenters noted that 'in topics where desirable conclusions come easily, the rigor of peer review becomes especially important.' Measuring AI's effects is a field where randomized controlled trials are hard to design and confirmation bias easily creeps in.

---