Effective Safety Measures in Peer Evaluation with Azure AI Content Safety

Developing mechanisms for peer evaluation comments requires a strong focus on safety and ethical considerations. Drawing from experiences in AI application development, we highlight actionable strategies that can be applied to improve the safety of peer feedback systems in one of our learning and teaching modules in the School of Computing Science at the University of Glasgow: Professional Software Development and Team Project.

Introduction to Comment Safety in Peer Evaluations

Peer evaluation is a valuable tool in academic and professional settings, fostering growth through feedback. However, ensuring that this feedback remains safe and constructive is essential.

When designing a peer evaluation system, integrating text moderation tools is crucial for filtering harmful content. Similar to the Text Moderation service used in AI applications, peer evaluation platforms can benefit from tools that assess feedback for harmful language, bias, and inappropriate content. These tools should allow users to set thresholds for content filtering to ensure that only constructive comments are allowed, thus avoiding potential negative impacts.

The Microsoft Responsible AI Framework provides a good starting point for educators and technical professionals to consider the various factors that should be taken into account when designing human-related systems, especially those involving comment expression.

Figure 1: Microsoft Responsible AI Framework (source: Microsoft Learn: What is Responsible AI?)

Transparent Communication and User Guidelines

Clear communication is essential in encouraging safe peer feedback. Users should be informed about the system's limitations and rules, including guidelines for providing effective and respectful feedback. Transparency in how comments are moderated can also ensure trust in the process, empowering users to engage with the system responsibly.

Figure 2: Clear guidelines are essential for making fair and open comments to improve individuals and groups.

System Planning, Expansion and Evaluation

As with AI system development, continuous testing and validation are essential. Engaging in hands-on testing with sample comments allows developers to refine safety features and improve the system's response to various inputs. When designing the peer evaluation system, we consider possible scenarios, inputs, processing, and outputs involving all parties throughout the learning and teaching journey.

It is crucial to ensure that you select the appropriate region that offers all (or most) of the functions you require. For instance, in our scenario, we planned to train our own model to customize content categories and scan text for matches, so we chose Australia East for this API service.

Figure 3: (Partial) List of Region Availability for Azure AI Content Safety

Once you have selected the service region, clearly record the endpoint and API keys for the initial trial. Always be mindful about storing your credentials in a safe place and not expose them publicly.

Consider the following (fictional) example: students provided comments, feedback, and work done in the system. With the help of the Azure AI Content Moderator, we aim for a quick and preliminary screening to identify any potentially harmful content so that preventive actions, such as advisory and consultation, can take place.

Figure 4: Sample Peer Comments and Efforts Rating (Fictional)

One of the fastest ways to ensure your subscription works is to perform a test with payload using cURL: by providing the correct endpoint and subscription key. Include attributes such as "text" for the data requiring analysis, "categories" as the classifier for the types of checks to be conducted, and specify the severity levels ranging from 0 to 7 as the output type.

Figure 5: RESTful API sample request with cURL

Figure 6: Sample output from Azure AI Contents Safety

Reflections and Future Enhancements

Developing a secure peer evaluation system is one of our missions in the Learning and Teaching Scholarship (LTS). Using structured testing and evaluation methods, like AI red-teaming, can enhance system robustness. As technology evolves, integrating new strategies, such as image moderation for multimedia, can further optimize these platforms.

Ensuring safe and effective feedback requires a commitment to responsible development practices. Implementing solid safety measures, transparent communication, and ongoing validation helps create systems that provide constructive feedback and protect users. As digital platforms become vital in peer evaluations, these strategies are essential for fostering positive, growth-oriented interactions.

Check the other articles in the 'Responsible AI: from principles to practice' blog series to learn more.