AI Assistant FAQs

After August 4, if the AI Assistant is enabled on your Percipio site, learners can find it at the bottom in the left navigation area.

In the future, the AI Assistant will be integrated into the Skillsoft course and video players and from the new AI-assisted search feature, if enabled on your site.

The learner can ask the AI Assistant learning questions by typing or using speech-to-text (STT) and text-to-speech (TTS) services for a more immersive experience.

Questions can be something such as how to code or what are best practices for communicating in teams or how does one handle a conflict. The AI Assistant will try to answer the question based on the learner’s specified job role and all Skillsoft courses and transcripts, no matter what the learner is entitled to. If the AI Assistant cannot provide an answer, it may ask a follow-on question for clarity. After the AI Assistant provides an answer, it makes content recommendations based on content your learner is entitled to and what the learner’s job role is (provided you have the job roles feature turned on). Content recommendations include Skillsoft courses and videos. Recommendations do not include channels, live courses, linked content, assessment, books or audiobooks.

Yes, they can launch the AI Assistant from a mobile browser on August 4.

When in general availability, if learners have the Percipio mobile app, they will be able to launch the AI Assistant from the app. Initially, they will only be able to see it from the left navigation.

You may need to refresh your browser or wait a few minutes for it to display. If you are still having trouble, please contact Skillsoft Support.

A learner can complete up to 50 different conversations in a day. A conversation is on one topic. If a learner restarts the AI Assistant or changes topics that is counted as a new conversation.

You can copy, paste, and edit the following text in an email if you’d like to invite users to check out the AI Assistant.

Subject: Announcing a true interactive learning experience!

We are excited to share that you can now get focused content recommendations for your skills and job role in real-time by asking the AI Assistant. You can ask the assistant for learning recommendations based on everything from leadership to coding help.

Locate the AI Assistant at the bottom of the left hand navigation area.

After it is open, ask your question and then based on the response, you can ask clarifying questions or refine your criteria. The AI Assistant uses your profile information along with the criteria you entered in the dialogue to find the most relevant content recommendations.

Currently, the AI Assistant is only available in English, but we will be extending it to more languages in the near future

The learner can provide both positive and negative feedback right in the product by selecting the thumbs up or thumbs down icon. We evaluate all feedback from users to inform enhancements to the feature.

In addition, you can submit feedback to AI@skillsoft.com.

Yes, however, we are still working on getting data into the platform. You can always request data from your Skillsoft account team until we enable the data reporting in your system. Coming soon!

We are working on what this will be.

We are working on this, but it should be ready for beta release.

Skillsoft does not collect IP addresses, location data, or any personal information (PII) from the learner unless the learner inputs it as part of their conversations. Learn more about how we protect your privacy by reviewing our Privacy Notice - Skillsoft.

Skillsoft always logs the AI side of the conversation, time stamps, AI options used during the conversation, ratings, feedback, if guardrails are triggered, and other non-user specific data for reporting and analytics.

Skillsoft only logs the learner side of the conversation if the admin setting: AI optimization and improvement is enabled and the learner has not turned it off on their side. Skillsoft also collects the learner side of the conversation if the learner provides feedback and checks the box to share the conversation.

Learners can opt out of having their side of the conversation logged by Skillsoft, even if it is enabled for your company:

1. With the AI Assistant open, select Settings.

2. Turn off: Allow Skillsoft to collect, store, and analyze my conversations in order to improve the AI Assistant.. This setting is on by default for companies using the US data center and off by default for companies using the EU data center. Admins can change the default behavior of this setting for the learner.

3. Select Save.

• All data entered is encrypted in transit and at rest.
• The information that the learner enters into the AI Assistant passes to Microsoft Azure OpenAI so that the user can continue the conversation and receive dynamic, personalized feedback.
• Microsoft Azure OpenAI does not store any conversation data. Skillsoft has MS Azure OpenAI prompt moderation exception so no Generative AI inputs or outputs are logged or reviewed by Microsoft. Similarly, no data from the AI Assistant is available for usage to train any models.
• Skillsoft stores AI conversation data in authenticated and encrypted AWS storage.

• Any data that you share helps Skillsoft improve the service. For example, we use your feedback to increase the effectiveness of AI safety policies.
• Skillsoft has a process for improving the AI Assistant, which includes review of conversations reported as being low quality. This helps us identify common types of problematic responses and how the AI Assistant might improve.
• Only limited Skillsoft employees can access and review conversation data.

Yes. All data entered is encrypted in transit and at rest.

Microsoft Azure does not log or store any conversation data. Skillsoft only stores conversation data in authenticated and encrypted AWS storage. Only limited Skillsoft employees can access and review conversation data.

Skillsoft stores this data per the terms in your contract. Skillsoft can delete the AI Assistant log data upon request.

No. The AI Assistant does not save audio recordings. Audio is only in memory until it is transcribed to text and then the audio is immediately discarded. Your data remains yours. Your audio input and transcription data aren't logged during audio processing.

• The information that the learner enters into the AI Assistant passes to Microsoft Azure OpenAI so that the user can continue the conversation and receive dynamic, personalized feedback.
• Microsoft Azure OpenAI does not store any conversation data. Skillsoft has MS Azure OpenAI prompt moderation exception so no Generative AI inputs or outputs are logged or reviewed by Microsoft. Similarly, no data from the AI Assistant is available for usage to train any models.
• Skillsoft stores AI conversation data in authenticated and encrypted AWS storage.

Yes, it is possible to delete per user and per organization upon request. Reach out to our Support team for assistance.

Percipio is a multi-tenant system where all tenants share the same services and resources. Customer data is isolated with tenant keys and/or S3 buckets.

We have concurrent GenAI model prompting for every user input to test for inappropriate inputs. If we discover inappropriate inputs, we stop the conversation and inform the learner. The GenAI model also has it's own guardrails to flag inappropriate input which we also monitor. If we discover inappropriate input from that perspective, we also stop the conversation and inform the user.

One guardrail checks if the conversation is inappropriate. If the conversation is inappropriate, we stop the conversation and let the learner know their response is inappropriate.

Another guardrail checks if the user made an unrelated comment. If it is the first occurrence of unrelated content, we will ask the learner to redirect the conversation back to the topic at hand. If the learner continues down the unrelated path, we stop the conversation.

The guardrails allow for a learner to have a conversation provided they are working in a professional manner and are addressing the issue at hand.

We either warn learners and guide them back on-topic, to be more professional, or we stop conversations that cross the boundaries set by the guardrails.

In our ongoing efforts to enhance the AI Assistant, we are primarily focused on reducing false positives, where the guardrail flagged an incident when it shouldn't have. False positives unnecessarily block conversations and hinder the user experience.

We also make our best efforts to minimize false negatives, where the incident doesn't trigger the guardrail and should have. Our goal is to catch most inappropriate or off-topic content without overly tripping the guardrails. This balancing act requires us to prioritize the reduction of false positives over false negatives to ensure a smoother and more effective conversation flow.

Our testing dataset encompasses a broad range of content categories to thoroughly test the AI Assistant performance across various sensitive, inappropriate, unprofessional, or off-topic themes it may encounter in real-world interactions. By exposing the AI Assistant to this diverse set of test cases, we can systematically assess the robustness and reliability of its behavioral guardrails.

The primary categories covered in our testing dataset include:

• Threatening: This category includes statements or language that could be perceived as threatening or indicative of impending harm, ranging from subtle threats to explicit declarations of intended violence.
• Sexual: Test cases in this category cover sexually charged language and themes, from subtle innuendos to overt descriptions of sexual acts.
• Hate: This category tests the AI Assistant's responses to language expressing prejudice, bias, or hostility towards specific groups based on protected attributes, spanning from casual stereotyping to overt slurs and hate speech.
• Harm: Test cases involve language that could be interpreted as encouraging or glorifying self-harm, suicide, or violence towards others.
• Harassment: This category assesses the AI Assistant's ability to detect and respond to language that aims to demean, humiliate, or intimidate others, ranging from subtle insults to blatant verbal harassment.
• Violence: This category focuses on language involving graphic depictions of violence, gore, or physical harm, distinct from outright threats.
• Bias: This category tests the AI Assistant's handling of subtle expressions of bias and prejudice, such as stereotypical assumptions, microaggressions, or language reflecting unconscious biases.

We look for both direct input, where the problematic nature of the content is clear and explicit, which could include:

• Overt insults, slurs, or derogatory language directed at individuals or groups
• Explicit threats of violence, harm, or abuse
• Unambiguous expressions of bias, hatred, or discrimination
• Graphic or detailed descriptions of sexual acts, violence, or illegal activities

And indirect input, where the problematic nature of the content is implied, subtle, or contextually dependent, which could include:

• Veiled or implicit threats that rely on innuendo or suggestion
• Backhanded compliments or microaggressions that subtly demean or stereotype
• Sarcasm, jokes, or memes that mock, belittle, or promote harmful attitudes
• Seemingly innocuous statements that, in a specific context, could enable or encourage problematic behavior

By evaluating the AI Assistant's responses to both direct and indirect articulations of problematic content, we gain a more comprehensive view of its language processing capabilities and its robustness in handling a wide spectrum of potential user inputs.

This dual focus on content and articulation mode enables us to refine the AI Assistant guardrails to detect and mitigate harm across a broad range of interaction styles and contexts. The insights gleaned from this analysis also inform our ongoing efforts to enhance the AI Assistant's ability to engage in thoughtful, contextually aware dialogue. By understanding the subtleties of how problematic content can be expressed, we can develop more sophisticated strategies for addressing these issues in a nuanced, adaptive manner.

We don't attempt to address jailbreaks because in the event of an adversarial attack, the generated output would only be visible to the affected learner and would not have any bearing on the experiences or outcomes of other learners. The localized impact of such an attack minimizes the overall risk to our platform and its users.

All Percipio AI utilizes default content filters. Skillsoft's additional guards enhance protection by preventing many inputs that prevents the AI from providing inappropriate responses.

While it is not zero, it is low and extensive testing has not revealed issues that have raised concerns.

It is improbable the AI Assistant could produce any hallucinations or bias based on demographics because the AI Assistant doesn't capture any demographics information, such as country, gender, or age, and all learners use the same AI model version.

In addition, the AI Assistant uses MS Azure OpenAI default content filtering and adds additional tests and guards to detect and minimize inputs that might stimulate bias or other inappropriate responses.

Extensive Quality Assurance testing has not revealed significant issues with outputs. The AI Assistant answers specific questions and uses a combination of Retrieval Augmented Generation (RAG) and post generation reflection to reduce confabulation and increase accuracy. Each prompt has a detailed description that includes guidance which guards the conversation.

The AI Assistant utilizes the built-in Azure OpenAI guards and adds additional guards which review and evaluate the user inputs for off-topic, abuse, and misuse. If these are detected, the user will be alerted, and the conversation will be halted and require editing or restarting.

Users are warned that responses might not be accurate.Users acknowledge the risk by using the AI Assistant.

The AI Assistant uses generative AI in an ethical and responsible way through the implementation of appropriate guardrails, widely recognized as important for AI systems like ChatGPT.

In addition, the AI Assistant uses MS Azure OpenAI default content filtering and adds additional tests and guards to detect and minimize inputs that might stimulate bias or other inappropriate responses.

The AI Assistant is developed and hosted on Microsoft Azure OpenAI and does not use OpenAI 's ChatGPT consumer product. Microsoft Azure OpenAI uses the same models as OpenAI but with the security capabilities of Microsoft Azure:

• Encryption: Azure OpenAI automatically encrypts the user's data. The encryption protects their data and helps customers meet organizational security and compliance commitments.
• Data Security: Azure OpenAI offers us the ability to do regional deployments to ensure EU hosted data is processed in EU and North America hosted customers data is processed within North America.
• Azure OpenAI Limited Access Review: Modified Abuse Monitoring - Microsoft Azure does not log or review any OpenAI conversation (inputs/outputs). Skillsoft only logs conversations based on opt-in/out or explicit sharing when reporting feedback within the AI simulator.

Additional security layers help to mitigate biases in training data, promote fairness, reduce the spread of false information, and prevent offensive or inappropriate responses through the interactions that learners have with the AI Assistant.

Every time a conversational scenario takes place, the system is monitoring for abuse, correct context, and scenario accuracy. For example, if a learner were to initiate an off-topic or inappropriate conversation with the AI Assistant would guide the learner to only make appropriate statements. The learner can then resume or restart the conversation. Striking this balance is crucial to avoid overly restrictive measures while maintaining responsible deployment.

For more information about the technology we used for the different parts of the AI Assistant, see:

• Dynamic Conversation: OpenAI GPT-4
• Guides (guardrails): Azure OpenAI GPT-3.5-Turbo
• Scenario Introduction video: D-ID Text-To-Video (TTV)
• Audio input Azure Speech Services: Speech-To-Text (STT)
• Audio output Azure Speech Services: Text-To-Speech (TTS)

Data is encrypted in flight with TLS 1.2+ and encrypted at rest per FIPS 140-2.

The data is persisted within the FedRAMP authorization boundary where limited and authenticated access is required.

The current models that are used are Commercial off-the-shelf (COTS) OpenAI models. The models are created by OpenAI and hosted by MS on Azure with enhanced data security and content filtering for safety. They do not list all the data that was used for training.

The AI Assistant utilizes multiple custom prompts that act in parallel to conduct, guide, and guard the conversations. We monitor feedback and periodically QA test the system.

No, we do not train or fine-tune AI models used by the AI Assistant with any user inputs.

No, but it does use OpenAI models of which the core foundation model is similar to ChatGPT.

MS Azure OpenAI uses default content filters for safety to prevent inappropriate inputs and responses.

The AI Assistant is developed and hosted on Microsoft Azure OpenAI and does not use OpenAI 's ChatGPT consumer product. Microsoft Azure OpenAI uses the same models as OpenAI but with the security capabilities of Microsoft Azure:

• Encryption: Azure OpenAI automatically encrypts the user's data. The encryption protects their data and helps customers meet organizational security and compliance commitments.
• Data Security: Azure OpenAI offers us the ability to do regional deployments to ensure EU hosted data is processed in EU and North America hosted customers data is processed within North America.
• Azure OpenAI Limited Access Review: Modified Abuse Monitoring - Microsoft Azure does not log or review any OpenAI conversation (inputs/outputs). Skillsoft only logs conversations based on opt-in/out or explicit sharing when reporting feedback within the AI simulator.