How we estimate the age: the technical part

Models and process

• Three models from the same company (Anthropic): Claude Haiku, Sonnet and Opus. Each rated all 6,851 movies separately and blind, in 28 batches, on 17/10/2026.
• What they see: title, year, runtime, whether it is animated, genres, synopsis (trimmed) and TMDb keywords, plus their own knowledge. What they don’t see: official ratings, content descriptors, scores, awards, the hand-made list or the other models’ answers.
• What they return: an age from 2 to 12 or “not recommended”, the dimension that decides it, the warnings, a reason of at most 15 words and how well they know the movie (high, medium or low).
• Instruction: minimum age to watch it without an adult, conservatively, reasoning from content, with anchors taken from BBFC, FSK and Kijkwijzer. They are told not to copy official ratings or Common Sense Media.

Aggregation and calibration

1. Average of the three ages (“not recommended” counts as 13).
2. Smoothed isotonic regression against the hand-made list (89 movies, bands 3, 4, 5, 6, 7, 8, 10 and 12), anchored at 13 so adult content stays “not recommended”.
3. Round up to the whole year. Mean vs. median and round-up vs. nearest were chosen by the lowest dangerous error in cross-validation.
4. The reason is the one from the model closest to the result; warnings are those flagged by at least two models.

Metrics

Cross-validation 5 × 10 over the 89 movies: the fit is never scored on the movies it was trained on. There is no frozen test set yet, so the figures are somewhat optimistic.

The pilot

Before rating everything, we compared, on the same movies, a direct age against an 11-dimension panel with fixed rules. The direct age was slightly more accurate, so the method was simplified:

The pilot used a single model for the direct age; in production there are three, and the dangerous error fell from 11% to 1%.

Movie universe

Union of five TMDb searches, no adult content (13,055 candidates):

• us_g_pg: US: G or PG, 50+ votes
• gb_u_pg: UK: U or PG, 50+ votes
• es_apta_7: Spain: “Apta” or 7, 30+ votes
• de_fsk_0_6: Germany: FSK 0 or 6, 50+ votes
• familia_animacion: Family or animation genre, 50+ votes

Movies that the median official rating of 14 countries puts at age 10 or under, plus the hand-made list, are rated: 6,851. Of those, 358 came out “not recommended under 13”; they don’t appear when browsing, only when searching by title.

Data and licences

• Details, synopses and user score: TMDb. This product uses the TMDB API but is not endorsed or certified by TMDB.
• Streaming: JustWatch, via TMDb, as of 17 October 2026.
• IMDb, Rotten Tomatoes and Metacritic: links only, without their scores (their licences don’t allow showing them on a public site).
• The English reasons were translated by AI from the Spanish original; English titles and synopses are TMDb’s.
• TMDb’s terms restrict using its content in connection with AI applications. We say it openly: TMDb synopses and keywords were given to the models. If TMDb does not allow it, we will re-rate without that text.

Known limitations

• A single model family: they can share biases; agreement among them doesn’t prove they are right.
• The hand-made list comes from a single person and has few movies above 8 and none “not recommended”.
• Possible contamination: the list’s notes sometimes cite Common Sense Media, and the models may remember its ages; that favours the direct age and can’t be measured without movies released after the models’ cut-off.
• The models don’t watch the movie: visual and sound aspects depend on what they remember. With little-known movies, low confidence signals it.
• Human review of the 3,179 low-confidence movies is pending.

← Back to the version for parents