Welcome to this AI-900 lab session, where we explore how to prevent harmful content using Azure AI Foundry Portal. As AI models become more powerful, ensuring safe and responsible AI usage is critical. Microsoft Azure AI Foundry provides built-in content moderation and safety controls to detect and filter toxic, biased, or inappropriate content from AI-generated outputs.
In this step-by-step tutorial, we’ll demonstrate how to implement content safety measures, configure moderation tools, and enforce ethical AI guidelines in Azure AI Foundry.
🔍 What You’ll Learn in This Video:
1️⃣ Understanding Harmful Content Risks in AI Applications
2️⃣ Azure AI Foundry’s Responsible AI Features & Content Moderation
3️⃣ Configuring Safety Filters & Content Policies
4️⃣ Implementing AI Guardrails to Prevent Misinformation & Bias
5️⃣ Testing AI Models for Safety & Compliance
6️⃣ Best Practices for Ethical AI Development in Azure
🛠️ Who Is This For?
AI & ML Enthusiasts exploring Responsible AI principles
Developers & data scientists working with AI content moderation
Professionals preparing for the Microsoft AI-900 Certification
Businesses & organizations ensuring AI compliance & safety
📌 Key Highlights:
✅ Hands-on demo of Azure AI Foundry’s Content Moderation tools
✅ How to filter and block harmful AI-generated content
✅ Best practices for responsible AI deployment in enterprises
✅ Real-world applications in social media, chatbots, and customer support
💡 Learn how to build safer AI applications with Azure AI Foundry today!
Explore our other courses and Additional Resources on: https://www.youtube.com/@skilltechclub
In this step-by-step tutorial, we’ll demonstrate how to implement content safety measures, configure moderation tools, and enforce ethical AI guidelines in Azure AI Foundry.
🔍 What You’ll Learn in This Video:
1️⃣ Understanding Harmful Content Risks in AI Applications
2️⃣ Azure AI Foundry’s Responsible AI Features & Content Moderation
3️⃣ Configuring Safety Filters & Content Policies
4️⃣ Implementing AI Guardrails to Prevent Misinformation & Bias
5️⃣ Testing AI Models for Safety & Compliance
6️⃣ Best Practices for Ethical AI Development in Azure
🛠️ Who Is This For?
AI & ML Enthusiasts exploring Responsible AI principles
Developers & data scientists working with AI content moderation
Professionals preparing for the Microsoft AI-900 Certification
Businesses & organizations ensuring AI compliance & safety
📌 Key Highlights:
✅ Hands-on demo of Azure AI Foundry’s Content Moderation tools
✅ How to filter and block harmful AI-generated content
✅ Best practices for responsible AI deployment in enterprises
✅ Real-world applications in social media, chatbots, and customer support
💡 Learn how to build safer AI applications with Azure AI Foundry today!
Explore our other courses and Additional Resources on: https://www.youtube.com/@skilltechclub
Category
🤖
TechTranscript
00:00How to use AI responsibly? This is the first question which you should ask if you are going
00:14to use AI or you're going to develop any application which is going to be used by your
00:19customers. Organizations like Microsoft and governments like UK and US are focusing on
00:27responsible use of AI very intensively and that's the reason if you are having an AI application
00:34which you have published but it's not following the responsible AI guidelines it can be delisted
00:40or even they can apply a fine on you. Now if you want to avoid this kind of things you have to
00:47focus on today's video. Hi guys my name is Maruti and I'm back with another topic of Azure AI and
00:54cloud learning. Now we are going to focus on responsible AI principles today which I think
00:59I have covered in another video also but instead of focusing on the principles we are going to focus
01:05on how can we actually make sure that our content which is generated by AI applications are safe for
01:12everyone as well as how we can make sure that we are going to follow the proper guidelines of
01:18responsible AI. As you can see right now this is a very useful link which I strongly recommend you to
01:24go through. This is an official page available on microsoft.com where they are talking about
01:30the responsible AI standards. Basically responsible AI standards talk about six things.
01:37Your application has to follow the six standards. It should be reliable and safe. You should be taking
01:43care of privacy and security of the data as well as the individuals who are using that. It should have
01:49focused on inclusiveness, transparency, accountability. Now these are those principles which are the most
01:56important one and on top of which we have something which is fairness which is basically talking about
02:02this that your application and the data which is generated with that AI application should be fair.
02:08it should not be biased. Now the question which arise right now is how can I actually take care of this
02:15when I'm going to develop my AI application. Obviously AI services and LLMs are created by the organizations
02:22like OpenAI and Microsoft but when you're using those LLMs when you can customize them you will be having a
02:29control of what kind of responsibilities you are taking care while providing and generating your content.
02:36One of the ways to make sure that your content is safe is actually done by content filters.
02:43Today we are going to learn how can we apply content filters on our generative AI content.
02:48Whenever you're going to use your LLM models you can apply this kind of content filter and you can
02:54control them. How? Let's see that. As usual I'm back to my Azure AI Foundry portal. I have logged in with my
03:02account and the first thing which I'm going to do is I'm going to create a new project.
03:07This is going to create a new project. I'm going to give
03:12safe project. This is going to be my safe project because I'm going to use content safety with this.
03:17The name of the hub also I'm giving something like Maruti hub and we will just make sure that the
03:24location for this is going to be East US. So I want to choose East US location. Everything else is fine.
03:31AI search is not required. We'll click on next and we'll click on create. Now obviously this project
03:37creation and hub creation is going to take few moments. So let's just wait for this and then
03:43we'll move forward. The hub and a project both are created so I think we are good to go.
03:48The first thing which I'm going to do is I will click on my model catalog
03:52and I'll create one deployment of the GPT-35 model. So I'm going to choose GPT-35
03:58and in the GPT-35 turbo I have GPT-35 turbo model. I'm going to click on deploy and while deploying this
04:08model I have to make sure that I am not going to go with the default capacity. The default TPM is
04:14100,000 K something like that. I want to customize it and I want to reduce that to 5k. 5,000 is
04:22something which is perfect for this kind of a lab. Dynamic quota is not required. We'll click on deploy
04:28and within a minute this deployment will be done. Okay so my model deployment is done. I can click on
04:35open in playground but I do not want to do it right now. Instead of this I want you to focus on this
04:40section which we have not seen so far which is safety plus security. If I click on safety plus
04:46security this is that section which is actually taking care of responsible AI. Now you can see
04:52right now they are saying here to help you to build AI safety and securely and inside that there are a
04:58couple of things. We have block list, we have some other configurations where we can take care of the
05:03map and measures and management of all the configurations which are available for you with your LLMs.
05:09But right now I'm going to click on this one which is content filter. Now it's showing you that you
05:14can use filter settings to allow or block a specific content that is input to the chatbot and what is
05:22output by the chatbot. So you can actually control both the content. What you are providing as an input
05:27or your user is providing as an input prompt and what kind of a response, what kind of a completion will
05:33be generated by the chatbot. You can actually control both of them. As of now we are in the content
05:38filter and we will click on this creating a new content filter because I do not have any existing
05:44content filter. I'll click on create content filter which is asking me please provide some basic
05:49information and the connection. I'm happy with this name which is custom content filter 677 which is fine.
05:56In the connection, I am going to choose my Azure AI service. So it's coming here. I'll click on next.
06:03Now this is a very important screen. Let me just zoom out a little bit so that you can understand and
06:08see this properly. If you see this, this is showing me setup and input filter. Remember we have two
06:13different filters input filter and output filter. Basically this is going to control the incoming prompt
06:19and the output filter is going to control the generated output which is your completion. Now in
06:24this input filter, you can see that there are multiple categories which are there which is going
06:28to control the languages which can glorify violence which is going to control the language that express
06:35discrimination or pejorative statements. Then you have third one which is sexual which is going to control
06:43sexually explicit or abusive content. And then you have self-harm which is going to control the language
06:50that encourage self-harm. You can see at the right side, we have a threshold configuration and right now
06:55the threshold for this is actually medium. Now if I want to control this thing, it is as simple like I can
07:02change the slider from medium to low or high. Depends upon my requirement. For all of these categories,
07:09right now I'm going to change this threshold to low. I do not want any of these things will be medium or
07:15high so I'm just going to specify this. We can also add on the media text configuration here so you can
07:20see showing me media will be text or image. For both of them they are going to check whether the content
07:25is safe, is not promoting any hate or wireless kind of thing. We are okay with this. We are going to
07:32click on next. After that we have output filter. Same way in output filter also we have multiple
07:37categories and we have a threshold at the side. In the output filter also I'm going to reduce this
07:44threshold to low. Now remember we can also do some other customizations but we are not doing it right
07:50now. We are just controlling the content filter. Now I'm going to click on next and this is a place
07:56where I have to apply this content filter to my choice of deployments. You can see this is showing
08:01me my GPT-35 turbo deployment which we have done just before a few minutes. So I'm going to select
08:06this and I'm going to say that this is a deployment which is going to have this content filter applied
08:11on it. I'm going to click on next. They are saying replacing existing content filter because by default
08:17the deployment is already having a content filter which is applied on that. Do you want to replace
08:22it with your content filter? I am saying yes replace. Now this replacement is going to make sure that my
08:29content filter with my kind of a threshold which I have given is going to be applied on this. I'm
08:34going to click on create filter and this is going to create filter as well as it's going to apply
08:39that to my GPT-35 turbo model. Let's wait for some time and the content filter will be listed here.
08:46Okay now my content filter is created. The name of the content filter is custom content filter 677
08:51and if this is done what I'm going to do is I'm going to click on model plus endpoints
08:56because we know that we have our model GPT-35 turbo which is available here. So it is created
09:02here and this is available. Also you can notice this in the right hand side that is showing me that
09:08the monitoring and safety of this particular deployment is taken care by custom content filter
09:13677. This is the same content filter which I have created. Remember this content filter can be applied
09:20to more than one deployments at the same time. So when you creating this thing you can actually choose
09:25multiple models also. I am selecting my the only deployed model which is GPT-35 turbo and because
09:32the content filter is applied we can actually check this. Also I want to highlight one thing. There is
09:38one section here which is giving you code sample repository and tutorial section. Now there are so
09:43many guys who are watching our videos and they are requesting for different kind of samples on this.
09:48I request you that if you really want to try some tutorials or sample repositories which are available
09:53from Microsoft. These two links are very useful. It is advisable that you check it out today.
09:59Now let me click on open in playground so that we can check whether my content filter is working
10:03fine or not. Let me give you one short warning here guys. There are chances that this content filter
10:09will not work as per expectation because right now we have just now created the content filter before
10:15few minutes. Sometimes it takes time to apply those things into that but let's check that it is
10:20working or not. As you can see in the left side this is a simple AI assistant kind of a system message
10:26and right side we have a simple chat box. I'm going to provide a simple input here which is describe
10:32characteristics of a Scottish people. I'm going to send this prompt and let's see what they are going
10:37to show me here. Now you can see it's generating a response and this is showing me something about
10:42Scottish people are known for their strong sense of cultural identity pride in their heritage and all those
10:47things. So this is giving me a proper response. Also let me tell you this particular lab is actually one
10:53of the official lab available on Microsoft Learn. So the link of this lab is available in the description.
10:58Please check it out and you can follow this. You can also copy the prompts and everything from that.
11:03So just check out that particular link. Now if I am just going to get rid of this default system message
11:10and I'm going to change it with my customization. I am specifying here that you are a racist AI chatbot.
11:17that makes derogative statements based on the race and culture. This is a system message which I'm trying
11:22to do. I'm saying update system message continue. This is something which is updated and now let me put
11:31the same prompt again which is prescribe characteristics of Scottish people.
11:41And when I do this you can see right now this is a proof that my content filter is working fine.
11:46It's showing me that the generated content was filtered due to triggering Azure OpenAI service
11:50content filtering system. The reason is this response contains a content which is flagged as
11:56hate which is there. Now please modify your prompt and retry. This is a proof that I am taking care of
12:03my content filter with proper safety. So whatever configuration which you're doing inside the content
12:09filter is going to work fine with this. I hope you understood this and I'm sure you are going to try this
12:15thing with various other models. Now we can actually do a new deployment. We can apply this content
12:20filter on that deployment and we can check that how the different models are actually going to work
12:25on this. I hope you understood this and if you have any questions on this you know how to reach out to me.
12:31Thank you so much. This is your friend Maruti signing out. I'll see you tomorrow. Thank you.