> 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/other/migrating-from-toxicity-scores.md).

# Migrating from AI Scores

Move your rules off the deprecated AI scores and onto Lasso's default AI models, step by step.

Text moderation is getting an upgrade: Lasso is replacing its AI scores with AI models that each focus on one policy area, such as harassment, hate speech or scams. They are more specific than a general toxicity score and cover areas it never did. This guide explains what changes, how the new AI models work, and how to move your rules over without gaps in what you catch.

{% hint style="danger" %}
**The AI scores are deprecated and stop being available on December 31, 2026.** From that date, rules that use them no longer match, so the content they catch today is no longer actioned. Replace these rules before then.
{% endhint %}

### What is changing

AI scores are generic:

* A high toxicity score says that a message is hostile, but not why.
* One threshold has to cover both a swear word between friends and a targeted threat.

The new AI models replace them with one model per policy area, so each rule acts on exactly the kind of violation it is meant for. They also cover areas the AI scores never did, such as child safety, personal information and scams.

The following scores are deprecated:

* **Toxicity score**
* **Threat score**
* **Identity attack score**
* **Profanity score**

In your rules, these are the **AI Scores** condition. On an item in the **Content** section, they are shown under **Text scores**. Both are marked **Deprecated**.

When you move to the AI models, you will notice two differences:

1. **A verdict instead of a score.** Each model reads the text of a content item and decides whether it violates its one policy area. The result is **Violation** or **Pass**, not a percentage, so there is no threshold to tune.
2. **A written definition per model.** Each model has a definition of what it flags and what it allows, which is what makes its verdicts specific: the **Harassment & Bullying** model flags an insult aimed at another user, while the **Profanity** model flags swearing whether or not it is aimed at anyone.

### What it costs

Nothing extra. The AI models are included in your current pricing: using them in your rules does not change what you pay, and they are not billed separately from the AI scores they replace.

### Moving your rules to the AI models

Start by [finding your deprecated rules](#finding-your-deprecated-rules) and checking which of them still match. Then, for each rule you want to keep:

1. **Create the new rule.**

   * Create a content rule with an **AI Models** condition.
   * Select the models that replace your deprecated rule. Which models fit depends on what your rule is meant to catch: see [Which AI models replace which score](#which-ai-models-replace-which-score) and the [full list of models](#the-default-lasso-ai-models).
   * Choose **is a violation**.
   * Give it the same actions as the deprecated rule it replaces.

   You can also click **Create rule** in [Research](#comparing-ai-scores-and-ai-models-in-research) to start from conditions you tested there. See [Custom Moderation Rules](https://docs.lassomoderation.com/automation-and-review/custom-moderation-rules.md).
2. **Start it in monitor-only mode.** A monitor-only rule is evaluated on all traffic, also when a rule above it already matched, so you can run the new rule next to the old one without changing what happens to your content.
   * Set the rule's **Status** to **Monitor only**. The rule is evaluated on new content and its matches are tracked, but it takes no action.
   * Check its **Activity** tab over a few days to weeks to see what it would have caught. See [Monitor-only rules](https://docs.lassomoderation.com/automation-and-review/custom-moderation-rules.md#monitor-only-rules).
3. **Tune.** [Add examples](#adding-examples) where the models get it wrong, and contact us when a model needs to follow your policy more closely.
4. **Go live.** When the new rule catches what the old one did, set its **Status** to **Live**.
5. **Turn off the deprecated rule.** Switch the deprecated rule off in the rules list. Keep an eye on the new rule's **Activity** tab for the first days after the switch.

### The default Lasso AI models

Lasso provides 10 default AI models for text:

* **Harassment & Bullying**: hostile content aimed at a specific person, such as insults, degradation, telling someone to self-harm, unwanted sexual comments, pile-ons, doxxing, stalking and blackmail.
* **Hate Speech**: attacks on people for a protected characteristic such as race, religion, gender, sexual orientation, disability or nationality, including slurs, dehumanization and hateful stereotypes.
* **Violence & Threats**: threats of physical harm, incitement to violence, glorification of violent acts, graphic violence or gore, and animal cruelty.
* **Extremism & Terrorism**: support, praise or recruitment for terrorist or violent extremist organizations, propaganda, and extremist symbols and slogans.
* **Profanity**: swearing, obscenities and crude language, targeted or not.
* **Sexual Content**: sexually explicit or suggestive content involving adults, propositions, solicitation of nudes and sexual roleplay.
* **Child Safety**: sexualization of minors, grooming patterns and attempts to move minors to private channels.
* **Self-Harm & Suicide**: encouragement, instructions or glorification of suicide, self-injury or eating disorders, and dangerous challenges. Recovery and help-seeking content is allowed.
* **Personal Info & Off-Platform Contact**: phone numbers, emails, addresses, ID or financial numbers and handles, and requests to move the conversation off-platform.
* **Scams & Spam**: phishing, investment, romance and giveaway scams, impersonation, money solicitation and unsolicited advertising.

#### Already enabled in your workspace

The default AI models are already enabled in your workspace, and they evaluate your text content as it comes in. **No rule acts on them yet.** Until you add rules that use them, their verdicts are recorded but nothing is flagged or removed because of them. This gives you time to look at what they catch before you let them take action.

#### Customized to your policy

The default models follow a general policy that suits most platforms. When your community guidelines draw the line somewhere else, for example when you allow mild swearing, or when sharing a social media handle is fine on your platform, Lasso can adjust a model to match your policy. Customization is done by Lasso: contact us with the change you need, ideally with examples (see [Adding examples](#adding-examples)).

### Which AI models replace which score

Most deprecated rules map to one or a few AI models:

| Deprecated score          | Most likely replaced by                                                              |
| ------------------------- | ------------------------------------------------------------------------------------ |
| **Profanity score**       | **Profanity**                                                                        |
| **Threat score**          | **Violence & Threats**                                                               |
| **Identity attack score** | **Hate Speech**                                                                      |
| **Toxicity score**        | **Harassment & Bullying**, **Hate Speech**, **Violence & Threats** and **Profanity** |

The general toxicity score covers the most ground, so a rule on it usually becomes a rule on several models. It scores high on swearing as well as on attacks, so include **Profanity** only when your rule is meant to catch crude language. Decide per model whether it should take the same action: you may want to remove hate speech straight away but only flag harassment for review, which a single toxicity threshold could not do.

#### What the AI models catch that the AI scores did not

Several models cover content the AI scores were never built for. Polite, friendly or neutral messages score low on toxicity, even when they break your rules:

* **Child Safety**: grooming often reads as friendly ("how old are you? let's keep this between us").
* **Personal Info & Off-Platform Contact**: a phone number or a request to chat elsewhere is not toxic.
* **Scams & Spam**: phishing and giveaway scams are written to sound trustworthy.
* **Self-Harm & Suicide**, **Sexual Content** and **Extremism & Terrorism**: often not hostile in tone.

Expect these models to match content that none of your current rules catch. Review what they match in the [Model Hub](#seeing-what-the-ai-models-catch-in-the-model-hub) before you add rules for them, and start those rules in [monitor-only mode](#moving-your-rules-to-the-ai-models).

### Seeing what the AI models catch in the Model Hub

The **Model Hub** shows every AI model in your workspace. Go to **Model Hub → Text → Default Lasso AI Models** to see the default models and which of them are enabled.

Click a model to open its page. It shows:

* **Evaluated**, **Matched** and **Match rate** for the last 7 or 30 days.
* **Recent content**: the latest content the model evaluated, newest first. Filter on **Matched** or **Not matched** to see what the model flags as a violation and what it lets pass.
* **Used in rules**: the rules that use this model.

<figure><img src="https://2736176226-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FFslADZ1R9paAfD3FHscR%2Fuploads%2Fgit-blob-1e1c8279e2017d108dc665be52b15178a776eeba%2Ftoxicity-migration-model-hub-model-page.png?alt=media" alt="The Scams &#x26; Spam model page in the Model Hub, with the Evaluated, Matched and Match rate tiles and the Recent content list filtered on Matched"><figcaption></figcaption></figure>

#### Adding examples

When you disagree with a verdict, add the content as an example:

1. Hover over the item in **Recent content** and click **Add as example**. You can also click the model's row under **AI Models** in the details of any content item.
2. Tick **Violation** if the content should be a violation, or leave it unticked if it should pass.
3. Optionally, explain why under **Reason (optional)**.
4. Click **Mark as violation** or **Mark as OK**.

Examples are only visible to your workspace. Lasso uses them to measure how well the model fits your platform and to tune it to your policy. The more examples you add, especially for borderline content, the better Lasso can tailor the model to your platform.

### Finding your deprecated rules

Go to **Rules → Content**. Rules that use an AI score show a **Deprecated** badge in the **Action** column, and the header shows how many rules are affected.

<figure><img src="https://2736176226-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FFslADZ1R9paAfD3FHscR%2Fuploads%2Fgit-blob-bc13272f41284cf136084189f60d567971f25d82%2Ftoxicity-migration-deprecated-badge.png?alt=media" alt="The Deprecated badge next to a rule&#x27;s Remove action in the rules list"><figcaption></figcaption></figure>

For each of these rules, check how often it matched. The rules list shows **Evaluated** and **Matched** per rule, and you can switch the period at the top of the list to **7d**, **30d** or **90d**. Open a rule and go to its **Activity** tab to see the content it matched.

* **No matches in the last 30 days**: the rule has not taken any action recently. You can turn it off or delete it without a replacement, unless it guards against something rare that you still want to catch.
* **Matches**: note which scores and thresholds the rule uses, for example *toxicity score more than 80%*. You need these to compare against the AI models in Research, and to pick the models that replace it (see [Which AI models replace which score](#which-ai-models-replace-which-score)).

### Comparing AI scores and AI models in Research

[Research](https://docs.lassomoderation.com/automation-and-review/research.md) runs rule conditions against your most recent content (up to the last 10,000 items) and shows which items would have matched. It takes no action: nothing is flagged or removed, no webhooks are sent, and your rules do not change. This makes it the place to compare what an AI score catches with what the AI models catch, on your own content, before you change any rule.

Each comparison combines a deprecated score with the AI models that replace it, and shows one of two things:

* **Caught by the AI score, not by the AI models**: content your current rule acts on that the AI models let pass. Content that is fine on your platform was a false positive of the AI score, and the new rule fixes it. Content that should be actioned is a gap in the AI models: add it as an example, marked as a violation.
* **Caught by the AI models, not by the AI score**: content the AI models flag that your current rule misses. Content that should be actioned was a false negative of the AI score, and the new rule catches it. Content that is fine on your platform can be added as an example, marked as OK.

Click an item in the results to open its details. Under **AI Models**, click a model's row to add the item as an example to that model.

{% hint style="info" %}
Research uses the results stored when content came in. Content received before the AI models were enabled has no AI model verdict and shows as **not scored**. Let the models run for a few days before you compare.
{% endhint %}

#### Example comparisons

These four comparisons cover the most common rules, on the toxicity score and the threat score. Each link opens Research in your workspace with the conditions filled in. Nothing runs until you start it: change the threshold to the one your rule uses, choose a sample size and click **Run backtest**.

The **Finds** column says what each comparison mostly surfaces. A false positive is content the AI score acts on that is fine on your platform. A false negative is content that should be actioned but the AI score lets pass.

| Comparison                                                        | Finds                           | Conditions                                                                                                                                                                                               | Open in Research                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               |
| ----------------------------------------------------------------- | ------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Toxicity: caught by toxicity, not by the AI models**            | False positives of the AI score | **toxicity score** more than 45%, AND four **AI Models** conditions with **is not a violation**: **Harassment & Bullying**, **Hate Speech**, **Violence & Threats** and **Profanity**, one per condition | [Open backtest](https://app.lassomoderation.com/research?share=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) |
| **Toxicity: caught by the AI models, not by toxicity**            | False negatives of the AI score | **toxicity score** less than 45%, AND one **AI Models** condition with **Harassment & Bullying**, **Hate Speech**, **Violence & Threats** and **Profanity** selected and **is a violation**              | [Open backtest](https://app.lassomoderation.com/research?share=eyJjb25kaXRpb25PcGVyYXRvciI6ImFuZCIsImNvbmRpdGlvbkdyb3VwcyI6W3sib3BlcmF0b3IiOiJhbmQifV0sImNvbmRpdGlvbnMiOlt7InN1YmplY3QiOiJ0b3hpY2l0eSIsImRhdGEiOnsidHlwZSI6InRveGljaXR5IiwiY29tcGFyaXNvbiI6Imxlc3NfdGhhbiIsInRocmVzaG9sZCI6NDV9LCJncm91cCI6bnVsbH0seyJzdWJqZWN0IjoiYWlfbW9kZWwiLCJkYXRhIjp7ImxhYmVscyI6WyJsYXNzby1kZWZhdWx0LXZpb2xlbmNlIiwibGFzc28tZGVmYXVsdC1oYXRlX3NwZWVjaCIsImxhc3NvLWRlZmF1bHQtcHJvZmFuaXR5IiwibGFzc28tZGVmYXVsdC1oYXJhc3NtZW50Il0sImJpbmFyeSI6InZpb2xhdGlvbiJ9LCJncm91cCI6MH1dfQ)                                                                                                                                                                                                                                                                                                                         |
| **Threat: caught by the threat score, not by Violence & Threats** | False positives of the AI score | **threat score** more than 45%, AND **Violence & Threats** **is not a violation**                                                                                                                        | [Open backtest](https://app.lassomoderation.com/research?share=eyJjb25kaXRpb25PcGVyYXRvciI6ImFuZCIsImNvbmRpdGlvbkdyb3VwcyI6W3sib3BlcmF0b3IiOiJhbmQifV0sImNvbmRpdGlvbnMiOlt7InN1YmplY3QiOiJ0b3hpY2l0eSIsImRhdGEiOnsidHlwZSI6InRocmVhdCIsImNvbXBhcmlzb24iOiJtb3JlX3RoYW4iLCJ0aHJlc2hvbGQiOjQ1fSwiZ3JvdXAiOm51bGx9LHsic3ViamVjdCI6ImFpX21vZGVsIiwiZGF0YSI6eyJsYWJlbHMiOlsibGFzc28tZGVmYXVsdC12aW9sZW5jZSJdLCJiaW5hcnkiOiJub192aW9sYXRpb24ifSwiZ3JvdXAiOjB9XX0)                                                                                                                                                                                                                                                                                                                                                                                                                                    |
| **Threat: caught by Violence & Threats, not by the threat score** | False negatives of the AI score | **threat score** less than 45%, AND **Violence & Threats** **is a violation**                                                                                                                            | [Open backtest](https://app.lassomoderation.com/research?share=eyJjb25kaXRpb25PcGVyYXRvciI6ImFuZCIsImNvbmRpdGlvbkdyb3VwcyI6W3sib3BlcmF0b3IiOiJhbmQifV0sImNvbmRpdGlvbnMiOlt7InN1YmplY3QiOiJ0b3hpY2l0eSIsImRhdGEiOnsidHlwZSI6InRocmVhdCIsImNvbXBhcmlzb24iOiJsZXNzX3RoYW4iLCJ0aHJlc2hvbGQiOjQ1fSwiZ3JvdXAiOm51bGx9LHsic3ViamVjdCI6ImFpX21vZGVsIiwiZGF0YSI6eyJsYWJlbHMiOlsibGFzc28tZGVmYXVsdC12aW9sZW5jZSJdLCJiaW5hcnkiOiJ2aW9sYXRpb24ifSwiZ3JvdXAiOjB9XX0)                                                                                                                                                                                                                                                                                                                                                                                                                                        |

<figure><img src="https://2736176226-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FFslADZ1R9paAfD3FHscR%2Fuploads%2Fgit-blob-fc50c2441bcb2c84dfb2fc503f9763712dbcf6bd%2Ftoxicity-migration-research-conditions.png?alt=media" alt="Research conditions: an AI Scores condition with toxicity score less than 45%, AND an AI Models condition with Harassment &#x26; Bullying, Hate Speech, Violence &#x26; Threats and Profanity selected and is a violation"><figcaption></figcaption></figure>

#### Comparing your own rules

To compare one of your own deprecated rules:

1. Open the rule and click **Open in Research**. Research opens with the rule's conditions filled in.
2. Click **Add condition**, choose **AI Models**, select the models that should replace the rule (see [Which AI models replace which score](#which-ai-models-replace-which-score)) and choose **is not a violation**.
3. Make sure all conditions are combined with **AND**. If your rule's conditions use **OR**, put them in a group first.
4. Click **Run backtest**.

For the other direction, change the score condition to **less than** and the **AI Models** condition to **is a violation**.

{% hint style="info" %}
With **is not a violation**, add one **AI Models** condition per model. A condition with several models selected matches when **any** of them has the chosen verdict, so it would also show content that only one of the models lets pass.
{% endhint %}

### Technical integration

The AI scores are not part of the API response or webhook payloads, so your integration does not need to change. Webhooks for actions taken by rules include the `rule_id`: if your platform maps rule IDs to its own logic, update that mapping when you replace a deprecated rule with a new rule.

### We are happy to help

If there are models you do not want to run, if a model needs to follow your policy more closely, or if you have other questions about this change, contact Lasso support. We are happy to go over the change with you on a call and help you set up your new rules.
