ChildSafeAds is a shared task on commercial content in YouTube videos likely to reach children and teenagers. The dataset contains 3,360 videos from 939 channels, each starting with a sponsor segment submitted to SponsorBlock. We pair each segment with its transcript, video, and channel information, plus a sales or service page linked from the video description. Systems determine the type of offer being promoted (ST1), assign product categories (ST2), and identify potential legal risks (ST3). Evidence is divided into four cumulative access levels, from the transcript to the linked page, so performance can be compared against the cost of collecting additional data. YouTube's "Includes paid promotion" label was absent from 45.5% of the videos. GPT-5.4 produced the labels after the organiser team, including a legal expert, reviewed samples and iterated on the taxonomy, prompts and model choices. GPT-5.6-luna independently labelled the development set to measure cross-model agreement. The evaluation received 21 entries; the highest mean macro-F1 was 0.735, while the best transcript-only entry reached 0.652. Across four system papers, the linked promotional page mainly helped identify what was being sold and added little to compliance-risk classification. The competition also exposed sensitivity to rare labels, advertiser overlap across splits and substantial variation in the ST3 labels.
Figures & tables
Record stage
Anonymised content
1. Commercial segment
The speaker names [brand] as the sponsor, promotes a sports bet and says that the bet cannot lose.
2. Description link
[URL]
3. Resolved page
[page title]; online sports betting and gambling service.
Platform disclosure
false
Task output
Released label
ST1
digital_content_or_services
Figure 1: A real training instance with identifiers removed. Source summaries are paraphrased; evidence quotes reproduce the release after redaction and punctuation adjustments. Labels are unchanged.
Level
Field group
Contents
Collection cost
1
transcript
Segment text, start and end time
Lowest; platform-scale text
2
video_context
Video ID, title, description, platform disclosure
One metadata call per video
3
channel_context
Channel ID and channel name
One channel lookup
4
product_page
Raw and resolved URL, page title and page text
Outbound crawl; pages may be dead or block automation
Table 1: Access levels are ordered by the cost of obtaining fields at platform scale. Teams report the highest level used, so accuracy can be interpreted alongside collection cost.
Label
Train
Dev
Test
Physical goods
1,125
218
235
Digital content or services
1,069
248
225
Physical services
122
24
33
No identifiable offer
35
14
9
Other
2
0
1
Table 2: ST1 and ST2 label counts by split. ST1 assigns one commercial type to each instance; ST2 may assign multiple product categories.
ST3 flag
Legal family
Train
Dev
Test
Misleading claim
Content
1,277
260
259
Inadequate disclosure
Disclosure
611
118
139
No flag
Housekeeping
529
127
112
Undisclosed advertising
Disclosure
352
74
89
Direct exhortation
Content
304
77
71
Age-restricted or prohibited product
Product
59
16
17
Table 3: ST3 compliance-risk counts by split. ST3 is multi-label; family-level evaluation groups the eight labels into four families: disclosure, content, product and housekeeping.
#
Team
Level
Mean
ST1
ST2
ST3
ST3 fam.
System paper
1
abc-abc
1–4
0.7347
0.8496
0.7550
0.5996
0.6834
2
thylinao a
1–4
0.7260
0.8049
0.7900
0.5830
0.6620
Silchenko (2026)
3
broxy
1–4
0.7079
0.6464
0.8243
0.6530
0.7031
Steigerwald et al. (2026)
4
runner b
1–4
0.7022
0.7923
0.7945
0.5197
0.6359
Narayan (2026)
5
rudolpheric
1–4
0.6699
0.6581
0.7364
0.6153
0.6984
6
9999_9999_99
1–4
0.6666
0.6093
0.8034
0.5870
0.6688
Table 4: Final evaluation-phase leaderboard (503 test instances, macro-F1). The top four entries lie within 0.033 of each other on the mean, and no entry leads all three subtasks. Level is the highest data-access level the team declared. Bold marks the best score in each column. a The entry includes three labels set by hand; the system alone scores 0.6537. b The entry was chosen between two variants with one leaderboard probe; the probe-free primary system scores 0.6798. Both teams disclosed these choices.