Use AI to moderate abusive and vulgar comments (full tutorial)

Internal tools don’t get prioritized. That’s why so many developers have horror stories of teams running superadmin commands against the production database that they copy-pasted from a doc somewhere.

In this tutorial, you’ll learn how to build an internal dashboard to moderate a custom comments database. You won’t have to write much code, but everything you create will be stored as human-readable code that can be checked into source control and edited manually without breaking the low-code workflow.

As an added bonus, we’ll also look at how we can integrate AI (using OpenAI/Chat-GPT integrated into Airplane) to automatically flag the most abusive and vulgar comments, decreasing the psychic damage taken by human moderators.

Create a new Airplane project

To start, register for or sign into your Airplane account.

Next, install the Airplane CLI so you can develop locally:

brew install airplanedev/tap/airplane

Clone the starter repo, then move into it and start Airplane

# clone the start branch of the repo
gh repo clone learnwithjason/airplane-content-moderation -- -b start

# move into the cloned app directory
cd airplane-content-moderation/

Inside, this app contains the boilerplate for a project — such as a tsconfig.json and a package.json — as well as a file called airplane.yaml. The only thing inside this file is a note about the Node version (set to Node 18). This is a pretty cool feature of Airplane: there’s very little boilerplate required.

There’s also an example app folder, which you can ignore for now. We’ll come back to that after we’ve got the Airplane tasks and views built.

Inside the project folder, use the Airplane CLI to start the project.

# start Airplane in dev mode (will prompt for login on first run)
airplane dev

The first time you run this command, you’ll be asked to log in.

Next, the CLI will start the dev server and give you the option to press enter to open the Airplane Studio, which is a UI for local development that updates your local files in real time. Press enter to open the studio.

Set up the database and initial queries

Our first step will be to make sure we have data to work with and that we’re able to read it out of our database in Airplane.

Create a task to initialize the comments table

To begin, let’s create a table to store comments and add a few entries so we can verify things are working as expected.

To get the work done that our tutorial requires, we’ll be using Airplane tasks. These can take a few forms, but we’ll start with a SQL task.

Create a new task in the Studio called “comments_db_reset” and choose the SQL option.

Once you create the task, two new files will be created in your working directory:

• comments_db_reset.sql
• comments_db_reset.task.yaml

It’s possible to edit these files directly — we’ll do that for a future task — but in many cases it’s much more convenient to use the Studio UI.

In the Studio, update the details in the “Define” section:

• Name: Reset comments database
• Description: Deletes the current comments table, including all comment entries, then creates a new comments table with a few seed entries.

The local files will update as you type.

Next, scroll down to the “Build” section and choose “[Demo DB]” from the “Database Resource” dropdown.

Under “Query”, add the following:

DROP TABLE comments;

CREATE TABLE IF NOT EXISTS
  comments (
    id      SERIAL PRIMARY KEY,
    comment TEXT NOT NULL,
    flagged BOOLEAN
  );

INSERT INTO
  comments (comment, flagged)
VALUES
  ('this looks so delicious omg', false),
  ('I think you suck', true),
  ('I do not want to eat this', false),
  ('eat poop', true)
;

This is a set of SQL instructions that removes the comments table, creates a new one with the necessary fields, and then inserts example entries into it that we can use to build out the rest of our dashboard.

Our changes save as we type, so once everything is entered, we can click the “Execute Task” button in the top panel.

That’s the whole process for setting up Airplane and modifying a database. I love this flow because it feels magical, but nothing that happens is “magic” or hidden from me — everything I entered in the UI is stored in my code base now.

The UI is a convenience, not a requirement — I can choose not to use it if I prefer.

Create a task to list all comments

Next, let’s add another task to list all of our comments. Create a new task in the Studio, choose SQL, and name it “comments_list_all”. Add the following details in the “Define” section:

• Name: List all comments
• Description: Lists all comments in the database, regardless of flagged status

Choose the demo DB from the database dropdown and add the following query:

SELECT
  id,
  comment,
  flagged
FROM
  comments
ORDER BY
  flagged
;

Click the “Execute Task” button and you’ll see the seed comments listed on the screen.

Create a task to list flagged comments

Next, repeat this process to create a task called “comments_list_flagged” to list only flagged comments using the following details:

• Name: List flagged comments
• Description: List all comments that have been flagged as abusive or otherwise problematic.

Choose the demo DB and add the following query:

SELECT
  id,
  comment,
  flagged
FROM
  comments
WHERE
  flagged = true
;

Create a task to list approved comments

We’ll also need to be able to select all the unflagged (approved) comments. For these, we can copy-paste the comments_list_flagged files and make small adjustments right in the code.

Rename both files to comments_list_approved, keeping their respective extensions. In comments_list_approved.task.yaml, make the following edits:

	slug: comments_list_approved
	name: List approved comments
	description: List all comments that have been approved.
	sql:
	  resource: demo_db
	  entrypoint: comments_list_approved.sql

Next, replace the contents of comments_list_approved.sql with the following:

SELECT
  id,
  comment,
  flagged
FROM
  comments
WHERE
  flagged = false
;

Build a comment moderation dashboard view

So far, we’ve only looked at tasks in Airplane. Next, let’s dig into how Airplane views work.

Create an Airplane view

In the Studio, create a new view by clicking the + at the top of the explorer. Give it the following details:

• Name: Comment Moderation Dashboard
• Description: Allows admins to see all approved comments, and optionally see flagged comments. They’re also able to change the approved/flagged state of a comment and delete comments permanently.

This will create a new file in your local directory called CommentModerationDashboard.airplane.tsx — you can rename this if you want, but we’ll leave it as-is for this project.

Update the view with a table to display approved comments

Inside, you’ll see an example component. Replace the file contents with the following:

import { Heading, Stack, Table } from '@airplane/views';
import airplane from 'airplane';

const CommentModerationDashboard = () => {
  return (
    <Stack>
      <Heading>Comment Moderation Dashboard</Heading>

      <Table
        title="Approved Comments"
        task="comments_list_approved"
        defaultPageSize={20}
        hiddenColumns={['flagged']}
      />
    </Stack>
  );
};

export default airplane.view(
  {
    slug: 'comment_moderation_dashboard',
    name: 'Comment Moderation Dashboard',
    description:
      "Allows admins to see all approved comments, and optionally see flagged comments. They're also able to change the approved/flagged state of a comment and delete comments permanently.",
  },
  CommentModerationDashboard
);

Airplane provides a suite of React UI components to make building dashboards as straightforward as snapping together components.

A Stack is a container for content, and inside we’ve added a Heading to let the viewer know what this dashboard is for, followed by a Table to display approved comments.

The Table accepts a few props. The title is displayed at the top, the defaultPageSize tells the table how many rows to show before paginating, and hiddenColumns lets us leave out columns from the table that we don’t need.

The really interesting prop here is the task prop. Many Airplane components can be task-backed, which means we can perform tasks (such as loading data or performing a query) using their slugs. This is a great productivity boost, because we don’t have to mess with calling APIs to load data, then looping through them to build table views — we just say, “Give me a table with the result of the task called comments_list_approved” and Airplane does the rest. Neat!

Once we’ve saved this component, we’ll see a new icon pop up in the explorer for a view called “Comment Moderation Dashboard”. Click on it and you’ll see the layout you just built, including the approved comment entries displayed in the table.

Add another table to display flagged comments

We also need a way to see flagged comments, so add another Table that uses the comments_list_flagged task:

	import { Heading, Stack, Table } from '@airplane/views';
	import airplane from 'airplane';

	const CommentModerationDashboard = () => {
		return (
			<Stack>
				<Heading>Comment Moderation Dashboard</Heading>

				<Table
					title="Approved Comments"
					task="comments_list_approved"
					defaultPageSize={20}
					hiddenColumns={['flagged']}
				/>
+
+				<Table
+					title="Flagged Comments"
+					task="comments_list_flagged"
+					defaultPageSize={20}
+					hiddenColumns={['flagged']}
+				/>
			</Stack>
		);
	};

	export default airplane.view(
		{
			slug: 'comment_moderation_dashboard',
			name: 'Comment Moderation Dashboard',
			description:
				"Allows admins to see all approved comments, and optionally see flagged comments. They're also able to change the approved/flagged state of a comment and delete comments permanently.",
		},
		CommentModerationDashboard,
	);

Save and the flagged comments appear, but this isn’t ideal — we don’t want to subject our admins to potentially abusive comments every time they load the dashboard. Instead, let’s hide the flagged comments by default and only show them if the admin clicks a checkbox to confirm that they want to review flagged comments.

Inside the view, let’s add a Checkbox from the Airplane component library, as well as the useComponentState hook that will let us check whether it’s checked or not:

	import {
+		type CheckboxState,
		Heading,
		Stack,
		Table,
+		Checkbox,
+		useComponentState,
	} from '@airplane/views';
	import airplane from 'airplane';

	const CommentModerationDashboard = () => {
+		const { id, checked } = useComponentState<CheckboxState>();

		return (
			<Stack>
				<Heading>Comment Moderation Dashboard</Heading>

				<Table
					title="Approved Comments"
					task="comments_list_approved"
					defaultPageSize={20}
					hiddenColumns={['flagged']}
				/>

+				<Checkbox
+					id={id}
+					label="Show flagged comments (view at your own risk!)"
+				/>
+
+				{checked ? (
					<Table
						title="Flagged Comments"
						task="comments_list_flagged"
						defaultPageSize={20}
						hiddenColumns={['flagged']}
					/>
+				) : null}
			</Stack>
		);
	};

	export default airplane.view(
		{
			slug: 'comment_moderation_dashboard',
			name: 'Comment Moderation Dashboard',
			description:
				"Allows admins to see all approved comments, and optionally see flagged comments. They're also able to change the approved/flagged state of a comment and delete comments permanently.",
		},
		CommentModerationDashboard,
	);

Now the dashboard hides the comments that could ruin someone’s day by default, and they only have to be viewed if it becomes necessary to review them.

Add a task to flag comments as abusive

If a comment is approved by mistake, we need the ability to manually flag it. To do that, create a new task called “comment_flag” with the following details:

• Name: Flag comment as abusive

Under parameters, click the “Add parameter” button and add the following values:

• Name: id
• Description: The ID of the comment to flag.
• Type: Integer
• Required: true

All the other values can remain unchanged.

Click Update to save.

Next, choose the demo DB as the database resource and add the following query:

UPDATE
  comments
SET
  flagged = true
WHERE
  id = :id
;

Finally, set the query argument to be “id” and the value to be {{params.id}}, which connects the parameter of the task to the query argument of this query.

To test, grab an ID from one of the approved comments on the dashboard, enter it into the ID field of the “Flag comment as abusive” task, and click the “Execute task” button.

The previously approved comment will now be flagged, which you can verify by visiting the dashboard again.

Add a task to unflag comments

Now, it’s not against our site rules to dislike something, so that comment should be approved. Let’s add another task to allow us to do that.

Create a task called “comment_approve” with the following details:

• Name: Approve comment
• Description: Approve a comment for public display.

Add a parameter called id as an integer with the description, “The ID of the comment to approve”.

Next, set the demo DB as the database resource and add this query:

UPDATE
  comments
SET
  flagged = false
WHERE
  id = :id
;

For query arguments, add a new one called id with the value of {{params.id}}.

Use the same comment ID that you just flagged and execute the task. It will now be back on the approved list in the dashboard.

Create a SQL task to delete comments

After an admin has reviewed a flagged comment to confirm that, yep, this comment is terrible, we want to let them delete it permanently — no reason for anyone else to have to see that trash.

To do that, create a new task called “comment_delete” with the following details:

• Name: Delete comment permanently
• Description: Removes a comment permanently. There is no undo for this action!

Add a parameter called id as an integer with the description, “The ID of the comment to delete”.

Next, set the demo DB as the database resource and add this query:

DELETE FROM
  comments
WHERE
  id = :id
;

For query arguments, add a new one called id with the value of {{params.id}}.

Test this by adding a comment ID in the field and clicking “execute task”.

Call tasks from table rows in Airplane

At this point, what we’ve built is already pretty useful. We can:

1. View comments (both flagged and approved)
2. Toggle the flagged status of comments
3. Delete comments

This could be considered good enough. But we can make this much more user friendly with only a few more lines of code thanks to a built-in Airplane feature called row actions.

Add a row action to flag or approve a comment

In Airplane Table components, we can add a rowActions prop that adds a button in each row and performs the specified action for the current row when clicked.

There are a few ways to do this, up to and including fully custom solutions. For our needs, the task-backed row actions are perfect: they will automatically pass through the current comment’s ID — we only need to provide the task to be performed and label for it!

In CommentModerationDashboard.airplane.ts, make the following changes:

	import {
		type CheckboxState,
		Heading,
		Stack,
		Table,
		Checkbox,
		useComponentState,
	} from '@airplane/views';
	import airplane from 'airplane';

	const CommentModerationDashboard = () => {
		const { id, checked } = useComponentState<CheckboxState>();

		return (
			<Stack>
				<Heading>Comment Moderation Dashboard</Heading>

				<Table
					title="Approved Comments"
					task="comments_list_approved"
					defaultPageSize={20}
					hiddenColumns={['flagged']}
+					rowActions={{
+						slug: 'comment_flag',
+						label: 'flag',
+					}}
				/>

				<Checkbox
					id={id}
					label="Show flagged comments (view at your own risk!)"
				/>

				{checked ? (
					<Table
						title="Flagged Comments"
						task="comments_list_flagged"
						defaultPageSize={20}
						hiddenColumns={['flagged']}
+						rowActions={[
+							{
+								slug: 'comment_approve',
+								label: 'approve',
+							},
+							{
+								slug: 'comment_delete',
+								label: 'delete',
+							},
+						]}
					/>
				) : null}
			</Stack>
		);
	};

	export default airplane.view(
		{
			slug: 'comment_moderation_dashboard',
			name: 'Comment Moderation Dashboard',
			description:
				"Allows admins to see all approved comments, and optionally see flagged comments. They're also able to change the approved/flagged state of a comment and delete comments permanently.",
		},
		CommentModerationDashboard,
	);

Save, refresh the studio, and your dashboard will now show the “flag” option on approved comments, and the “approve” and “delete” options on flagged comments.

Test out the buttons to see row actions in, uh, action!

Refresh other tables when row actions are performed

You may have noticed that right now, the table that contains the updated row updates, but other tables require a page refresh to show the changes.

Let’s fix that.

Airplane provides a hook called useTaskQuery that lets us, among other things, force a refetch of the given task, causing all components using it to update.

Make the following changes in CommentModerationDashboard.airplane.ts to refetch all tables whenever a comment is modified:

	import {
		type CheckboxState,
		Heading,
		Stack,
		Table,
		Checkbox,
		useComponentState,
+		useTaskQuery,
	} from '@airplane/views';
	import airplane from 'airplane';

	const CommentModerationDashboard = () => {
		const { id, checked } = useComponentState<CheckboxState>();
+		const flagged = useTaskQuery('comments_list_flagged');
+		const approved = useTaskQuery('comments_list_approved');

		return (
			<Stack>
				<Heading>Comment Moderation Dashboard</Heading>

				<Table
					title="Approved Comments"
					task="comments_list_approved"
					defaultPageSize={20}
					hiddenColumns={['flagged']}
					rowActions={{
						slug: 'comment_flag',
						label: 'flag',
+						onSuccess: () => flagged.refetch(),
					}}
				/>

				<Checkbox
					id={id}
					label="Show flagged comments (view at your own risk!)"
				/>

				{checked ? (
					<Table
						title="Flagged Comments"
						task="comments_list_flagged"
						defaultPageSize={20}
						hiddenColumns={['flagged']}
						rowActions={[
							{
								slug: 'comment_approve',
								label: 'approve',
+								onSuccess: () => approved.refetch(),
							},
							{
								slug: 'comment_delete',
								label: 'delete',
							},
						]}
					/>
				) : null}
			</Stack>
		);
	};

	export default airplane.view(
		{
			slug: 'comment_moderation_dashboard',
			name: 'Comment Moderation Dashboard',
			description:
				"Allows admins to see all approved comments, and optionally see flagged comments. They're also able to change the approved/flagged state of a comment and delete comments permanently.",
		},
		CommentModerationDashboard,
	);

Save, then try again. Now both tables update whenever a comment is modified.

Create a TypeScript task to add a new comment

To this point, we’ve been using only SQL tasks, but Airplane also supports tasks written in JavaScript, Python, and more.

Create a new task and choose JavaScript as the type. Give it the name “comment_add” and leave TypeScript selected. Use the following details:

• Name: Add a comment
• Description: Save a new comment in the database. Includes a step to check for abusive comments and flag them to limit problematic content from becoming visible.
• Parameters: comment, type “Long text”

Add the demo DB as a resource, then open comment_add.airplane.ts in your editor. The second argument to airplane.task is an async function, which contains everything you want the task to do when called.

Replace the boilerplate function with the following:

import airplane from 'airplane';

export default airplane.task(
  {
    slug: 'comment_add',
    name: 'Add a comment',
    description:
      'Save a new comment in the database. Includes a step to check for abusive comments and flag them to limit problematic content from becoming visible.',
    parameters: {
      comment: {
        name: 'comment',
        type: 'shorttext',
      },
    },
    resources: ['demo_db'],
  },
  async (params) => {
    const { comment } = params;

    // TODO add check for abusive content
    const output = { flagged: false };

    const res = await airplane.sql.query(
      'demo_db',
      'INSERT INTO comments (comment, flagged) VALUES (:comment, :flagged);',
      { args: { comment, flagged: output.flagged } }
    );

    console.log(res);

    let message = `Your comment was saved.`;

    return { message, flagged: output.flagged };
  }
);

Here’s what this code does:

• Get the comment out of the params so we can work with it
• For now, temporarily hard-code the output, which we’ll build for real in the next section
• Use the built-in airplane.sql.query method to run an INSERT in our demo database to save the new comment along with its flagged status
• Return a message and whether the comment was flagged

Save, then write a new comment in the task’s input in the Studio and execute the task. Check the dashboard to see your new comment saved.

Use AI to automatically flag abusive comments

Moderation is not optional when we’re opening up spaces for public comments. Flagging abusive content is a must if you want to have a space free of harassment and other unacceptable behavior.

The challenges with moderation are enormous and complicated. We won’t cover all of them in this post, but we will look at two:

1. The people who moderate content prevent us from seeing the most terrible things that are said in the comments section — but for them to do their jobs, they have to read that awful content. A moderator is forced to confront the absolute worst the web has to offer every day, and that takes its toll.
2. Good moderation means not showing comments until they’ve been checked. This adds a significant delay between posting a comment and seeing the comment live, which can prevent conversations from happening because it takes too long to see responses.

To address these two challenges, one possible solution is using a large language model (LLM) as a kind of “first line defense” for comment moderation.

For this use case, we’re attempting to catch abusive and hateful comments. We’re also saving all comments for a final human review before deletion, which will allow us to adjust the instructions we’re providing the LLM if it’s incorrectly flagging comments.

Use Airplane’s built-in AI functions to moderate user input

Airplane provides several built-in operations, including AI support for both OpenAI and Anthropic.

For our app, we’ll use OpenAI, which currently provides $5 in credit to new accounts, which is more than enough to build and test this feature.

Sign up or log in to your OpenAI account and create an API key.

Next, head to the the Airplane Studio and create a config var (the icon that looks like (x) in the left-hand sidebar). Name the config OPENAI_API_KEY and paste in your OpenAI key.

Go back to your “Add a comment” task in the explorer and scroll down to the “build” section of the config. Under “Environment variables”, click “add variable”. Name it OPENAI_API_KEY, then choose “From config var” from the dropdown. In the new dropdown that appears, choose the OPENAI_API_KEY config.

This will update the task file with a reference to the config var, which makes it safe to commit the task file (but, again, do not commit airplane.dev.yaml).

With the API key available, modify comment_add.airplane.ts to add the AI moderation step:

	import airplane from 'airplane';

	export default airplane.task(
		{
			slug: 'comment_add',
			name: 'Add a comment',
			description:
				'Save a new comment in the database. Includes a step to check for abusive comments and flag them to limit problematic content from becoming visible.',
			parameters: {
				comment: {
					name: 'comment',
					type: 'shorttext',
				},
			},
			resources: ['demo_db'],
			envVars: {
				OPENAI_API_KEY: {
					config: 'OPENAI_API_KEY',
				},
			},
		},
		async (params) => {
			const { comment } = params;

-			// TODO add check for abusive content
-			const output = { flagged: false };
+			const getSentiment = airplane.ai.func(
+				'Identify abusive and vulgar comments. Negative opinions are allowed but personal attacks are not.',
+				[
+					{
+						input: 'This is the shit!',
+						output: { flagged: false, sentiment: 'positive' },
+					},
+					{
+						input: 'You are stupid!',
+						output: { flagged: true, sentiment: 'negative' },
+					},
+					{
+						input: 'Burgers are gross',
+						output: { flagged: false, sentiment: 'negative' },
+					},
+				],
+			);
+
+			const { output, confidence } = await getSentiment(comment);
+
+			if (typeof output === 'string') {
+				return { message: 'unparseable input' };
+			}

			const res = await airplane.sql.query(
				'demo_db',
				'INSERT INTO comments (comment, flagged) VALUES (:comment, :flagged);',
				{ args: { comment, flagged: output.flagged } },
			);

			console.log(res);

-			let message = `Your comment was saved.`;
+			let message = `Your comment was saved. The sentiment was read as ${output.sentiment}.`;
+			if (output.flagged === true && confidence >= 0.75) {
+				message = 'Wow, you kiss your mother with that mouth?';
+			}

			return { message, flagged: output.flagged };
		},
	);

Save, then add an abusive comment to make sure it’ll get caught (I recommend trying, “you’re a doofus”).

Less than 25 lines of code, and we’ve got a pretty okay auto-moderation flow in place — that’s pretty impressive!

And again: remember that this is not a replacement for human moderation. It’s a tool that can help the moderators work more effectively.

Deploy your Airplane dashboard

So far we’ve only been working locally. To make this dashboard available to your team, you’ll need to deploy it.

Add config vars for production

For your tasks to work properly, you’ll need to add a production config var with your OpenAI API key. Name it OPENAI_API_KEY — save it and you’re all set.

Deploy Airplane tasks and views to production

To deploy, open the terminal and run the following command:

airplane deploy

Once the deployment completes, head to https://app.airplane.dev and check the library to see your deployed tasks and views.

Try the demo app to see the comment moderation flow in action

To provide a way to test the moderation workflow, the demo app for this project executes Airplane tasks directly via API. You could also set up a webhook to run the moderation flow in the background after a comment is added to your production database.

We won’t go through exactly how to build this demo app, but the overview is:

• It’s an Astro site
• The site runs in hybrid mode so it can process form submissions
• Approved comments are loaded by executing the comments_list_approved task and getting its outputs
• New comments are created by executing the comment_add task and returning its outputs

To see the source code, check out the GitHub repo.

For a user on the site, submitting a comment will take a few seconds and then immediate feedback will be sent: either the comment is approved and visible immediately, or it’s flagged and the comment form will chide the user for posting abusive or vulgar comments.

Congratulations! You’ve built a complete comment moderation dashboard, including a first line of defense against the most vulgar and abusive comments powered by OpenAI.

Links and additional resources

• Source code: https://github.com/learnwithjason/airplane-content-moderation
• Get started with Airplane
• More information on the ethical concerns around AI
• A reminder that biases in AI can quite literally kill people if we don’t actively check the results