AI Art & Music Library

Visual art made with AI whose originality was recognised by institutions, juries, auctions or critics, and AI-made music that found a real audience, collected from news reports. Selection uses Jev (TypeSafe AI) judgements with thresholds set from data. Updated automatically every day.
Updated: 2026-10-09 01:08 JST · Art 3 · Music 16

Original AI art: painting and visual art

Refik Anadol

In the news since Jul 2025 · 2 reports

Recognised (show, prize or sale)

Obvious

In the news since Nov 2018 · 1 report

Recognised (show, prize or sale)

Popular AI-made music

Breaking Rust

In the news since Nov 2025 · 2 reports

Audience popularity (charts, plays)

Puerto Rico

In the news since May 2026 · 1 report

Audience popularity (charts, plays)

Hawak Mo Ang Beat

In the news since Mar 2026 · 1 report

Audience popularity (charts, plays)

Lolita Cercel

In the news since Feb 2026 · 1 report

Audience popularity (charts, plays)

4AMTapes

Works: Bonita Bonita · In the news since Jan 2026 · 1 report

Audience popularity (charts, plays)

Aespa

In the news since Nov 2024 · 1 report

Recognised (show, prize or sale)

The Beatles

Works: Now and Then · In the news since Nov 2024 · 1 report

Recognised (show, prize or sale)

Selection criteria and thresholds

Collection: Google News (English and Japanese) is searched daily for reports on recognition or popularity of AI-made art and music (back-filled from 2018).
Judgement: Jev answers six questions per headline as probabilities: (1) is it about one specific work or AI artist (threshold 0.25); (2) was the work itself made by AI (0.80); (3) outside recognition: museum, prize, auction, acquisition or critics (0.80); (4) measurable audience popularity: charts, plays, virality (0.90); (5) kind of work (visual art, music, other); (6) does it imitate a real singer's or artist's voice or likeness without permission (excluded at 0.50 or above; as an exclusion rule this threshold maximises F2, weighting misses more).
Rule: visual art needs (1)(2)(3); music needs (1)(2) and (4) or (3).
Thresholds: a stratified sample of 218 judged headlines was labelled against the same criteria by a slower reasoning model (Claude); each threshold is the value on a 0.05 grid that maximises F0.5 (a wrong inclusion weighs twice a miss). Combined rule agreement: precision 0.955, recall 0.553.
Names are copied verbatim from headlines; works with the same name share one card, ordered by number of reports.
Limits: judgement uses headlines only, so unreported works are missing; popularity and recognition are as of the report date.