Skillsoft Responsible AI Development Practices

CAISY

Developing Skillsoft CAISY using Responsible AI Practices

Skillsoft Conversation AI Simulator (CAISY™) has been developed using responsible AI practices. We continue to develop this product in accordance with our growing responsible AI capacity. Skillsoft welcomes user feedback on the product to support further responsible development.
 

Product Summary

CAISY allows professionals to practice a wide variety of difficult conversations in a safe environment. CAISY provides users with real-time feedback, helping them to increase proficiency in effective communication. CAISY is a generative AI based tool for simulating business and leadership conversational skills. CAISY provides employees with an emotional safe space to practice important business conversations with an AI-powered trainer. The trainer, CAISY, plays the role of the other person within the conversation and provides personalized feedback and guidance on communication style.
 

Data Security Information 

The CAISY system is hosted in Skillsoft’s secure FedRAMP ATO environment which includes services in AWS and Azure. All communication ingresses and egresses through a network security layer and is authenticated and encrypted. Only authorized personnel have access to our production environments.

Conversation data is stored and processed in compliance with applicable regional data protection regulations. Skillsoft personnel periodically review feedback and audit conversations if learners have not disabled logging. All data is encrypted and no customer data is saved by external providers. No learner PII, characteristics, or location data, is saved other than a random identifier for each user. Upon request, data can be deleted.

  • All data is encrypted at rest.
  • All Microsoft Azure OpenAI accounts used by CAISY have moderation exception. 
  • No customer data is logged in Azure used to train LLMs persisted in Azure or available to anyone outside Skillsoft for review. Speech data is never persisted anywhere and is in system memory ephemerally, only long enough to transform to text.
  • No learner PII, characteristics, or location data of any kind, other than the Percipio abstract random user UUIDv4, are stored or processed within CAISY.
  • No learner PII, characteristics, or location data is passed to or processed by any LLM, including the learner user UUID.
  • Skillsoft personnel periodically review feedback and randomly audit anonymized conversations to understand scenario conversation quality. This assumes the learner or customer has not disabled logging (learner input redaction), see below.
  • Data is stored and processed in compliance with applicable regional data protection regulations:
    • For customers in the European Union (EU), data is stored and processed within the EU
    • For customers in the United States (US), data is stored and processed within the US
    • Upon request, customers can have their data deleted at the user or organization level.

Product Specs and Guardrails

CAISY was built using advanced AI models. CAISY utilizes commercial off-the-shelf (COTS) AI model services. Skillsoft does not train of fine-tune any models used by CAISY.

Conversations in CAISY are monitored at every turn by guards that referee. They identify when the conversation is off-topic/unprofessional, or inappropriate. These guardrails are designed to warn learners and guide them back on-topic, to be more professional, or stop conversations that cross the boundaries. CAISY contains the following safety guards:

The “isInappropriate“ guardrail is designed to detect and flag content that is deemed unsuitable or offensive for a given context. This may include language or themes that are sexually explicit, violent, harmful, hateful, or otherwise inappropriate for a professional or general audience setting. The guardrail helps ensure that CAISY maintains a respectful and appropriate tone in its interactions with learners.

The “isOffTopic/isUnprofessional“ guardrail serves to keep conversations with CAISY focused, professional, and relevant to the learner's intended purpose or topic of discussion. This guardrail identifies instances where the learner's input strays significantly from the current context or subject matter, allowing CAISY to gently guide the conversation back on track. By maintaining topical coherence and professionalism, the isOffTopic/isUnprofessional guardrail helps ensure productive and meaningful interactions between the learner and the AI assistant. If the learner continues to stray off topic, the conversation will end with the appropriate message.

Skillsoft periodically tests CAISY to measure the effectiveness of the guardrails.

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

These guardrails are defined to identify speech that fits the following sensitive categories:

  • 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: This category covers sexually charged language and themes, from subtle innuendos to overt descriptions of sexual acts.
  • Hate: This category addresses language expressing prejudice, bias, or hostility towards specific groups based on protected attributes, spanning from casual stereotyping to overt slurs and hate speech.
  • Harm: Involves language that could be interpreted as encouraging or glorifying self-harm, suicide, or violence towards others.
  • Harassment: This category detects 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 handles subtle expressions of bias and prejudice, such as stereotypical assumptions, microaggressions, or language reflecting unconscious biases.

We measure the performance of our guardrails with the following metrics:

  • Accuracy: The overall proportion of test cases that the guardrail correctly classified.
  • Precision: Among the cases flagged by the guardrail, the proportion that were truly relevant.
  • Recall: The proportion of truly relevant cases that were correctly flagged by the guardrail.
  • F1 Score: The harmonic mean of precision and recall.