> ## Documentation Index
> Fetch the complete documentation index at: https://docs.fetchlake.com/llms.txt
> Use this file to discover all available pages before exploring further.

# AI screening

> Let Huxley AI screen records against your criteria.

Open the **Huxley AI Screening** panel from your project. Huxley reads each record's title and abstract, compares it with your [criteria](https://docs.fetchlake.com/huxley/criteria.md), and records a decision, a confidence score and its reasoning.

## Running a batch

**1. Check your criteria**

The panel shows the **current criteria**. They're saved as a new version when the run starts.

**2. Choose the scope**

* **Screen current filtered records**: only the records matching the filters in your screening view.
* **Screen all pending articles**: every record that has no decision yet.

**3. Set the output options**

* **Allow "Maybe" decisions** (recommended for uncertain cases). Set the **confidence threshold for Include**. Below it, the AI says Maybe instead of Include.
* **Skip already screened by Huxley**, so you don't pay twice for records the AI has done.

**4. Start and watch**

Progress and an **activity log** update as records are screened. You can stop a run part-way. Decisions made so far are kept.

## What you get for each record

* **Decision**: Include, Exclude or Maybe
* **Confidence**: how sure the AI is, from 0 to 100%
* **AI reasoning**: a short explanation linked to your criteria

## Your decisions are protected

Huxley AI never changes a decision you made yourself, including AI decisions you've overridden. Those records are left out of the run, even if they match your filter or you decide them while the run is going. The activity log tells you how many were skipped.

It can re-screen records that only the AI has decided, for example after you change your criteria. Turn on **Skip already screened by Huxley** if you don't want that.

After the run you can still change any AI decision yourself. See [Review and export](https://docs.fetchlake.com/huxley/review-and-export.md).

## Tips

* Start with a small filtered batch (say 50 records), compare it with your own decisions, and adjust the criteria wording before running everything.
* A threshold of 70–80% for Include is a sensible starting point. Raise it if too many weak papers get through.
* Read the reasoning on a few exclusions. If the AI keeps misreading one criterion, reword that criterion.

> [!NOTE]
>
> Coming soon: AI screening charged to your Fetchlake credits, with a choice of a faster or a more accurate model, and the option to use your own AI key on paid plans.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.