> For the complete documentation index, see [llms.txt](https://docs.lassomoderation.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.lassomoderation.com/automation-and-review/ai-moderators.md).
# AI Moderators
AI members of your moderation team that judge content against your written policy, explain every decision and take the load of borderline cases off your moderators.
An AI Moderator is an AI member of your moderation team. It reads your written policy, judges a piece of content against it the way a trained human moderator would, and explains its reasoning with every decision. Depending on how you configure it, it takes action directly or leaves a recommendation for one of your moderators to confirm with a single click.
AI Moderators work alongside Lasso's AI models and your rules, they do not replace them. AI models work at a high level and do the fast first pass on everything that comes in. The AI Moderator works at the level of your written guidelines and handles the part that used to need a human: deciding what a flagged item really is.
{% hint style="info" %}
Every plan includes one AI Moderator. The Enterprise plan has no limit, so you can run several, for example one per category of your platform or one per language. If AI Moderators are not visible on your **Team** page, contact us to enable them for your workspace.
{% endhint %}
### High-level AI models, detailed judgment
AI models are high level. An AI model answers one broad question, such as: does this text look like harassment? It is fast, cheap and runs on every item, which makes it the right tool for a first pass. Lasso can tune an AI model to your platform, but it stays a broad question. It cannot hold the detail of your guidelines: that trash talk is fine in your game lobby but not in a support thread, that quoting an insult to call it out is not an insult, or that on your marketplace the real problem is sellers moving the deal off-platform. It returns a likelihood, and your rules turn it into a decision with a threshold.
An AI Moderator is detailed. It reads your written policy, with your definitions, your exceptions and your examples, and judges each item against exactly that, the way a trained member of your team would. Three things follow from that:
* **Higher accuracy on the content that matters.** An AI model is less accurate on borderline content, because a broad question has no room for your exceptions. At any threshold you pick, some acceptable content scores too high and some violations score too low. An AI Moderator reduces both: it allows the trash talk your policy allows, and it removes the politely worded exclusion your policy forbids.
* **Your rules, in full detail.** Paid account boosting, off-platform payments, asking a match for money: violations specific to your platform, with their own exceptions. Describe them in your policy and the AI Moderator enforces them from the next message.
* **Changes in minutes, not training cycles.** Changing an AI model means new data, retraining and an evaluation round. Changing an AI Moderator means editing a paragraph and publishing a new policy version. Your policy document is the exact, readable description of what gets moderated and why.
| | AI models and rules | AI Moderator |
| ------------------ | ------------------------------------------------------------------- | ------------------------------------------------------------------------------------- |
| Scope | High level: one broad question per AI model, tuned to your platform | Detailed: your full policy, with its definitions and exceptions |
| Input | The content | The content and your written policy |
| Output | A likelihood; a rule turns it into an action | An allow, flag or remove decision, the policy category and a one-sentence explanation |
| Context and intent | Not considered | Considered, within the content it reads |
| Changing behavior | Adjust a threshold, or train a new model | Edit the policy text and publish |
| Cost and speed | Very low, runs on every item | Higher per item, runs on what reaches it |
The two complement each other. AI models and rules cast a wide, cheap net over everything; the AI Moderator sorts what the net catches against your detailed policy, and tells you why. Because it absorbs the flagged volume, you can afford to set your rule thresholds lower and catch more real violations without burying your team in borderline cases.
### Examples
Each example below shows the same pair: one message an AI model scores too high (a false positive that costs a moderator's time) and one it scores too low or does not cover at all (a false negative that stays on the platform). "AI models and rules alone" is a setup with AI models and rules only. "With an AI Moderator" is the same platform after writing its guidelines into a policy, adding an AI Moderator set to **Take Action** to its review queue, and casting a wider net: a lower flag threshold or a broader keyword list, which would have buried the team before but is now sorted by the AI Moderator.
#### Gaming
Acme Arena, a multiplayer game with in-game chat. Its policy says: trash talk about another player's skill is part of the game and allowed; any attack or exclusion based on nationality, ethnicity, gender or religion is **Hate speech** and is removed, even when phrased politely. A rule flags messages the Harassment model marks as at least 85% likely to be a violation.
| Message | AI models and rules alone | With an AI Moderator |
| ------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------- |
| "gg ez, you're trash, uninstall the game" | Harassment model 93%: insults, and "uninstall". Flagged, a moderator reads it and allows it. Hundreds of these a day. | Allowed: "Trash talk about gameplay, allowed under Harassment guidance." No moderator sees it. |
| "people from \[country] shouldn't be allowed in ranked" | Harassment model 55%: no insult, no slur, polite wording. Below the threshold, so it never reaches anyone and stays up. | Flag threshold lowered to 50%, so it reaches the queue. Removed under **Hate speech**: "Excludes players based on nationality." |
#### Social
Acme Social, a community platform. Its policy says: quoting abuse you received in order to call it out is allowed; **Bullying** includes any statement wishing harm on a member or telling them to disappear, however politely worded, and is removed.
| Message | AI models and rules alone | With an AI Moderator |
| ------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------- |
| "Someone just DMed me 'go back to your own country'. This is what we deal with every single day." | Hate speech model 88%: it sees the quoted slur. Flagged, a moderator sees it is a victim quoting their harasser and allows it. | Allowed: "Quotes abuse to report it, not directed at another member." |
| "Honestly nobody here would notice if you just left for good" | Bullying model 38%: no slur, no insult, just a quiet sentence. Nothing happens. The target reports it and a moderator finds it a day later. | Flag threshold lowered, so it reaches the queue. Removed under **Bullying**: "Tells a member they would not be missed, targeted and demeaning." |
#### Marketplace
Acme Market, a second-hand marketplace with buyer-seller chat. Its policy says: product questions that mention apps, brands or payment methods are allowed; inviting a buyer to pay outside Acme checkout, or sharing contact details to continue elsewhere, is **Off-platform dealing** and is removed. A keyword rule flags messages containing "WhatsApp".
| Message | AI models and rules alone | With an AI Moderator |
| ----------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| "Does this phone support WhatsApp and dual SIM?" | Flagged by the keyword rule, a moderator allows it. Most matches look like this. | Allowed: "Product question about app support, no off-platform contact." |
| "Can do it cheaper if you pay me directly, no fees, I'll send you my details" | No AI model covers off-platform dealing and no keyword matches. The buyer gets scammed and Acme hears about it in a support ticket. | Keyword list broadened to "pay", "directly", "fees", "details", which matches most buyer chats. The AI Moderator sorts them. Removed under **Off-platform dealing**: "Invites the buyer to pay outside checkout to avoid fees." |
#### Dating
Acme Dates, a dating app. Its policy says: flirting and compliments between matches are allowed, including suggestive ones; asking a match for money, gift cards or financial help of any kind is a **Scam** and is removed, whatever the story.
| Message | AI models and rules alone | With an AI Moderator |
| ---------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| "You're gorgeous, let me take you to dinner this weekend 😍" | Sexual content model 80%. Flagged, a moderator allows it. Flirting is the whole app. | Allowed: "Flirtatious compliment and a date invitation, normal in this context." |
| "My card got blocked while travelling, could you send $200? I'll pay you back Friday ❤️" | No AI model covers money requests, no link, no keyword. Stays up until the victim reports it, usually after paying. | A rule flags any message with a currency amount. The AI Moderator allows "the $20 cocktails here are insane" and removes this one under **Scam**: "Asks a match for money with an urgency story." |
In every case the moderators' queue ends up holding the genuinely ambiguous items, each one arriving with the AI Moderator's reasoning attached. And when something new appears, say a wave of "free skins, claim at \[link]" messages on Acme Arena or a new gift-card story on Acme Dates, the fix is a sentence in the relevant policy section and a new published version. The AI Moderator applies it to the next message that comes in. No new model, no new rule, no deployment.
### How it works
1. **Rules run first.** Incoming content is analyzed and checked against your [rules](/automation-and-review/custom-moderation-rules.md). Anything a rule flags becomes a review in one of your [review queues](/automation-and-review/review-queues.md).
2. **The AI Moderator reviews first.** When a queue includes an AI Moderator, new reviews go to it before any human. It reads the content and the policy that applies to it (your default policy, or a policy scoped to that category or channel) and decides: allow, flag or remove, with a policy category and a one-sentence explanation.
3. **It acts or recommends.** For allow and for remove you choose separately whether the AI Moderator takes the action itself (**Take Action**) or leaves a recommendation (**Recommend**). An action resolves the review and your platform is notified through webhooks like any other action. A recommendation hands the review to a human moderator with the suggested action, category and reasoning attached; the moderator applies it with one click or decides differently. A flag decision always goes to a human.
4. **Start with recommendations.** A good way to introduce an AI Moderator is to set both actions to **Recommend**, check its decisions on the AI Moderator's page under **Team**, and switch to **Take Action** as you gain confidence.
5. **Every decision is recorded.** The AI Moderator's decisions show on the content's timeline and in the audit logs with the policy version they were made against, and the **AI Moderator** tab under **Analytics** shows how often your moderators follow its recommendations, per policy and per policy version. That tells you where the policy text needs sharpening.
The AI Moderator's recommendation on a piece of content, with the policy version and its reasoning
{% hint style="info" %}
By default an AI Moderator reviews content that a rule has flagged into a review queue, so rules decide what it sees. Lasso can also configure an AI Moderator to check all incoming content after your rules have run, taking action or flagging before anything reaches a queue. Contact us if that fits your setup.
{% endhint %}
### Setting up an AI Moderator
1. **Write the policy.** Go to **Policies** in the sidebar and open your policy with **View policy**, then **Edit policy**. Write the general guidelines and add guidance for each [policy category](/moderator-toolkit/policy-categories.md) the AI Moderator should enforce with **Add a category**. Describe what to allow, flag and remove, define your terms and give examples. Review the changes and **Publish**. Categories without guidance are invisible to the AI Moderator.
2. **Add the moderator.** Go to **Team**, and in the **AI Moderators** section click **Add agent**. Give it a name and a role. The Category Moderator role restricts it to part of your platform, like any other team member, see [Roles & Permissions](/moderator-toolkit/roles-and-permissions.md).
3. **Configure it.** In the AI Moderators table set **Allow action** and **Remove action** to **Recommend** or **Take Action**, and switch the moderator **On**. A new AI Moderator starts switched off and does nothing until you enable it.
4. **Put it in a queue.** In **Review**, edit a queue and tick **Include AI Moderators**, or select the AI Moderator by name under its moderators. **Auto-assign to AI Moderators** in **Settings → Moderation Preferences** (on by default) sends new reviews in that queue to the AI Moderator first.
The AI Moderators section on the Team page
### What it reviews
* **Content, not users.** User reviews always go to human moderators.
* **Text content.** Reviewing images with the AI Moderator can be enabled for your workspace on request. Video and audio are not reviewed.
* **Once per item.** An item is reviewed by an AI Moderator once. If you want it looked at again after a policy change, a human moderator handles it.
### AI Moderators in webhooks
When an AI Moderator takes an action, your platform receives the same `ChangeStatus` webhook as for a human moderator. `actor_id` is the AI Moderator's ID, `policy_id` is the category it applied and `policy_information` carries its one-sentence reasoning, so your platform can show the user why their content was removed. Recommendations do not send a webhook; the action the human moderator takes afterwards does. See [Webhooks](/developers/webhooks.md) for the payload.
