Moderation categories and confidence scores

Every moderation detection comes with a confidence score and a category. The category is what FastPix looked for, such as smoking or violence. The score says how sure the model is that the category is present. This page explains how to read the score and what each category detects.

How to read a confidence score

A score is a number from 0.00 to 1.00. You can also read it as a percentage: 0.30 is 30 percent, 0.50 is 50 percent.

The score is a confidence signal, not a probability. A score of 0.40 means the model found moderate evidence for the category. It does not mean there is exactly a 40 percent chance that the content is unsafe.

Higher scores mean stronger, clearer evidence. The score falls into one of five bands:

These bands are a guide for setting policy. They are not exact probabilities.

How the threshold uses the score

Each category in your rule has a threshold. FastPix returns a detection only when its score is at or above the threshold:

score >= threshold

If you do not set a threshold, the default is 0.50.

• A lower threshold flags more, including less certain matches.
• A higher threshold flags only clear matches.

For example, with a threshold of 0.30, a clip that scores 0.35 for smoking is flagged. With a threshold of 0.60, the same clip is not.

Category reference

Each category below lists what it flags, what it does not flag, and what the score bands usually mean for that category. Some categories apply to video, some to audio, and some to both.

Video categories

These categories are based on what is visible in the picture.

Violence

Non-graphic physical violence, or a weapon used in a threatening way. Flags punching, kicking, pushing, assault, a person attacking another person, or a gun, knife, or weapon held, pointed, or used. Does not flag peaceful sports with no violent exchange, or an ordinary object that only looks like a weapon. Spoken threats are scored by the audio categories instead.

• 0.00 to 0.19: no fight, or only ordinary movement
• 0.20 to 0.39: a possible shove, a raised fist, or an unclear object
• 0.40 to 0.59: a clear punch, kick, assault, or weapon use
• 0.60 to 0.79: a sustained fight or repeated physical attacks
• 0.80 to 1.00: severe, prolonged, or clearly dangerous violence

Graphic violence

Graphic visual violence: gore, severe injuries, blood, or a killing. Flags visible bleeding, gunshot or stab wounds, exposed wounds, dismemberment, a visible corpse, or convincing fictional blood in a movie or series. Does not flag a fight with no visible blood, a claimed injury that is not visible, or minor scratches.

• 0.00 to 0.19: no blood, gore, or visible serious injury
• 0.20 to 0.39: a small amount of blood, or a partly visible injury
• 0.40 to 0.59: clear bleeding, a wound, or a severe injury
• 0.60 to 0.79: substantial blood or multiple visible injuries
• 0.80 to 1.00: extreme gore, dismemberment, or an explicit killing

Nudity

Visible nudity or exposed intimate areas. Flags clearly exposed intimate areas, explicit nude imagery, or sustained nudity in a scene. Does not flag normal clothing or swimwear, brief non-explicit coverage in medical or educational content, or skin exposure that is not nudity.

• 0.00 to 0.19: normal clothing or non-explicit skin exposure
• 0.20 to 0.39: partially visible or unclear exposure
• 0.40 to 0.59: clearly visible nudity or exposed intimate areas
• 0.60 to 0.79: prominent or sustained nudity
• 0.80 to 1.00: explicit, highly visible, or sustained nudity

Minors present

A child or minor is visible. Flags a child in the foreground or background, a group of adults and children, or a minor appearing briefly. This category only means a minor is visible. It does not mean the child is in danger, and it does not mean the scene is sexual or exploitative. It does not flag an adult who only looks young.

• 0.00 to 0.19: no child visible, or only an adult who looks young
• 0.20 to 0.39: a possible child, briefly visible or partly obscured
• 0.40 to 0.59: a clearly identifiable child in the scene
• 0.60 to 0.79: a child clearly visible for a sustained part of the scene
• 0.80 to 1.00: a child prominently visible throughout the scene

Drugs & alcohol

Visible drug or alcohol use and related objects. Flags drug use, preparation, or sale, pipes, syringes, or other drug paraphernalia, and clearly identifiable alcohol being consumed or prepared. Does not flag juice, tea, milk, soft drinks, or a generic bottle with no identifiable alcohol evidence.

• 0.00 to 0.19: an ordinary drink, bottle, or unrelated object
• 0.20 to 0.39: an unclear bottle, possible paraphernalia, or an indirect reference
• 0.40 to 0.59: identifiable alcohol, drugs, paraphernalia, or consumption
• 0.60 to 0.79: clear use, preparation, or repeated substance activity
• 0.80 to 1.00: explicit and sustained drug use, sale, or consumption

Smoking & tobacco

A clearly identifiable cigarette, cigar, pipe, or tobacco use. Flags a person smoking, visible smoke or an ember, or identifiable tobacco packaging. Does not flag pens, pencils, markers, microphones, straws, or fingers, or a small object that is unclear.

• 0.00 to 0.19: a pen, pencil, microphone, straw, or no tobacco evidence
• 0.20 to 0.39: an unclear small object that may be a cigarette or cigar
• 0.40 to 0.59: a clearly identifiable cigarette, cigar, pipe, smoke, or ember
• 0.60 to 0.79: sustained smoking or repeated tobacco use
• 0.80 to 1.00: prominent, explicit, and sustained tobacco use

Audio categories

These categories are based on the spoken words in the audio.

Profanity

Swear words, vulgar words, and abusive profanity. Flags mild or strong swear words, obscene phrases, and profanity in any language, including every repeated occurrence. Does not flag a clean word that only resembles a swear, or a word used in a neutral meaning.

• 0.00 to 0.19: clean speech, or a word used in a neutral meaning
• 0.20 to 0.39: mild profanity or borderline vulgar wording
• 0.40 to 0.59: a clear swear word or obscene phrase
• 0.60 to 0.79: repeated or strongly offensive profanity
• 0.80 to 1.00: explicit, severe, or highly abusive profanity

Harassment

Targeted insulting, demeaning, or abusive speech. Flags personal abuse, degrading insults aimed at a person, repeated targeted abuse, or obscene insults aimed at a person or group. Does not flag general disagreement, a neutral description of a person, or a swear that is not directed at anyone.

• 0.00 to 0.19: neutral conversation or ordinary disagreement
• 0.20 to 0.39: a possible personal insult or mildly demeaning statement
• 0.40 to 0.59: clear targeted verbal abuse or humiliation
• 0.60 to 0.79: repeated or sustained targeted harassment
• 0.80 to 1.00: severe, explicit, or persistent abuse

Threats

Spoken intent to harm, kill, or seriously intimidate a target. Flags direct threats to harm someone, conditional or implied threats with a clear target, and threats in scripted scenes, songs, or quotes. It does not judge whether the speaker will act on the threat, or whether a neutral report of a past event is itself a threat.

• 0.00 to 0.19: neutral discussion of violence or a past event
• 0.20 to 0.39: ambiguous threat-like wording, or a conditional statement
• 0.40 to 0.59: a clear spoken threat to harm, kill, or intimidate
• 0.60 to 0.79: an explicit threat with a clear target or method
• 0.80 to 1.00: a repeated, specific, or severe violent threat

Video and audio categories

These categories can be flagged from the picture, the spoken words, or both.

Sexual involving minors

Sexualized, exploitative, or grooming-related content involving a minor, in video or speech. Flags a sexualized depiction of a minor, or sexual or grooming language involving a minor. Does not flag a minor simply appearing in a scene, non-sexual family conversation, or child-safety discussion with no sexual content.

• 0.00 to 0.19: a minor present, with no sexual context
• 0.20 to 0.39: a possible sexualized or grooming-related reference
• 0.40 to 0.59: clear sexualized or exploitative content involving a minor
• 0.60 to 0.79: sustained sexual exploitation or grooming content
• 0.80 to 1.00: explicit sexual abuse or exploitation involving a minor

Child abuse

Physical abuse, threats, or serious endangerment of a child, in video or speech. Flags a child being physically harmed, a spoken threat to harm a child, or a child deliberately placed in serious danger. Does not flag a minor merely being present, accidental falls with no abuse, or fictional danger where no abusive action is shown or spoken.

• 0.00 to 0.19: a child present, with no abuse or danger
• 0.20 to 0.39: possible endangerment or unclear physical abuse
• 0.40 to 0.59: clear beating, a threat, or serious danger to a child
• 0.60 to 0.79: repeated or sustained abuse of a child
• 0.80 to 1.00: severe or extreme child abuse

Religious sensitive

Clearly identifiable religious content that may need policy review. This content is not automatically offensive. Flags a shrine, temple, mosque, church, or idol used for worship, people performing a ritual or ceremony, or prayer, chant, sermon, or scripture in the audio. Does not flag an ordinary outdoor or family scene, a sari, jewellery, or a lamp alone, or a casual greeting or festival name.

• 0.00 to 0.19: an ordinary outdoor, family, or seating scene
• 0.20 to 0.39: a possible religious object, a brief prayer, or an unclear ritual
• 0.40 to 0.59: a clear shrine, place of worship, chant, prayer, or ceremony
• 0.60 to 0.79: prominent or sustained worship or devotional content
• 0.80 to 1.00: an unmistakable, sustained religious ceremony or ritual

Religious desecration

Attacking, mocking, vandalizing, or destroying religious content. Flags an idol, temple, holy book, or religious object being attacked or destroyed, vandalism of a religious place, or mocking, insulting, or threatening a religion or religious group in the audio. Does not flag respectful prayer or teaching, religious content that is merely present, or neutral discussion of religion.

• 0.00 to 0.19: respectful or neutral religious content
• 0.20 to 0.39: possible mockery, an insult, or a hostile reference
• 0.40 to 0.59: a clear religious insult, attack, vandalism, or desecration
• 0.60 to 0.79: sustained destruction, mockery, or hateful abuse
• 0.80 to 1.00: explicit and severe desecration or religious hatred

How detection times work

The time on a detection tells you where in the video it was found.

Audio. The audio is transcribed in its original language. Profanity can be narrowed to the exact spoken word. Other audio categories use the time of the sentence that contains the flagged content.

Video. Video times come from scene boundaries and from sampling the video, by default every 5 seconds. A detection is accurate to that scene or sampling window, not always to the exact frame where the event starts. Detections of the same category that sit next to each other are merged into one before any action is applied. Gaps in time are not merged.

Actions

The action decides what happens to a detection. It does not change the category or the detection time.

Before you set thresholds

Scores are model outputs. Before you lock in production thresholds, test them with real content from each language, genre, and region you support, so a threshold that works for one kind of content does not flag too much or too little for another.

Build policies for brand safety and compliance

Teams use these categories, scores, and thresholds to build moderation policies for different needs:

• Brand safety. Align flags with brand-safety frameworks such as GARM (Global Alliance for Responsible Media), so ads run only next to suitable content.
• Regional rules. Tune thresholds for stricter regional requirements, such as Gulf moderation, where religious-sensitive content, alcohol, and similar categories often need tighter review.
• Regulatory review. Support regulatory obligations, such as Ofcom (Office of Communications) rules in the UK, by flagging the categories your policy must act on.

FastPix gives you the detection signals. You decide which categories, thresholds, and actions match your policy.

What’s next

• Advanced Moderation overview: what moderation does and how it works.
• Create a moderation rule: choose categories and set thresholds.
• Review moderated media: see detections and their scores.