How FeelDB measures emotions

What do emotion shares mean?

Shares describe normalized emotional signal, not the percentage of people who felt an emotion. A 40% joy share means joy accounts for 40% of the fingerprint’s total signal. It does not mean 40% of viewers felt joy.

How does the AI estimate emotions?

The Gemini model family estimates emotions from catalogue metadata: title and original title, year, type, genres, country, directors, up to five cast names, runtime and up to 700 characters of overview when available. The prompt asks for an estimated viewing experience rather than plot themes or characters’ emotions.

This prompt receives no film or episode footage, audio, subtitles, full screenplay or viewer ballots. The model may draw on its learned knowledge, but the estimate is not evidence that it watched the title or measured actual viewers.

How is the emotional journey estimated?

The emotional journey is a language model’s reading of English subtitles, including sound cues. We divide the runtime into 20 intervals of 5% each and give the model condensed dialogue and cues, together with the film’s title, genres, synopsis and a short summary of its critical reception from Wikipedia, so it can read irony and tone correctly. Where and how strongly each emotion appears still comes from the subtitles. Intervals with too little subtitle evidence remain gaps. A quality gate determines whether the journey is shown, labelled “approximate”, or hidden. It is not viewer votes and may differ from your experience. The curves show relative emotional intensity; the ribbon shows the dominant emotion in each interval, the one that stands out most relative to how strongly it usually registers, so a habitually high emotion such as tension does not win every close call. It may fall outside the three plotted curves.

Dialogue-heavy films tend to read better. Silent stretches remain gaps, and subtitles can be out of sync with the film. Shape labels are simplified descriptions of the intensity curve, not a verdict about the story. “Hide the ending” masks the final 40% and its turning-point descriptions.

Kadri, your film companion

Kadri reads stored subtitles and FeelDB film data. With spoiler protection on, it uses only the opening or the part up to your marker and checks each answer before showing it. Questions are processed by our AI provider, Anthropic. AI can be wrong; spoiler protection cannot guarantee a perfect result. We do not keep question or answer text in our logs or database; your chat stays only in your temporary session and disappears when the session expires or you clear it.

How do viewer ballots change the fingerprint?

Each distinct viewer’s ballot contributes one unit of signal, divided equally among its selected emotions. When an AI estimate is present, it starts with a prior mass of 3 ballot-equivalent units; viewer ballots are blended with that prior. With no viewer ballots, the fingerprint stays AI-only. With no AI estimate, viewer ballots supply the full signal.

For each emotion: share = 100 × (prior mass × AI share / 100 + viewer signal) / (prior mass + number of viewers). Prior mass is zero when no AI estimate exists. As ballots accumulate, the AI prior’s relative weight decreases.

How reliable is a low-sample fingerprint?

AI-only and early fingerprints are uncertain. Fewer than three viewers is an early signal, and even three or more ballots is not a representative survey. Viewers choose whether to respond; preferences, culture and context can change how a story feels. No confidence interval or guarantee of your reaction is implied.

Which 28 emotions do we use?

FeelDB groups 28 emotion labels into eight families for browsing. These are the catalogue’s labels, not a clinical scale or a universal account of emotion.

Joy

Joy, Happiness, Smile, Satisfaction, Relief

Love

Love, Empathy, Sympathy

Thrill

Excitement, Anticipation, Surprise, Tension

Wonder

Awe, Inspiration, Hope, Curiosity

Blue

Sadness, Melancholy, Nostalgia

Dread

Fear, Anxiety, Disturbing

Fire

Anger, Envy, Guilt, Disgust

Play

Humor, Confusion

Where does title metadata come from?

TMDB and IMDb supply catalogue metadata and identifiers where available. Their metadata is distinct from FeelDB’s AI estimates and viewer ballots. Source records can be incomplete or wrong; a source link is attribution, not endorsement of the emotional fingerprint.

TMDB · IMDb

This product uses the TMDB API but is not endorsed or certified by TMDB.

What are the limitations?

The fingerprint is an estimate for discovery, not a scientific study, a professional review or a prediction of your individual response. Metadata gaps and model errors can affect AI estimates; self-selected ballots can be biased. A data-updated date records a database update, not an editorial review.