Update on AI Safety in Nualang

This post provides an update on our safe use of AI in Nualang as it relates to product updates that appeared in our September release.

With our September release of Nualang, we added some exciting new features, some of which are enabled by additional use of AI. In particular, we have responded to your feedback and made use of AI to provide more dynamic real-time conversation practice for your students. Read more about these new features..

Previously, conversations with our Nualas were scripted using sets of predefined phrases and responses. A considerable number of these phrases and responses had to be loaded to provide any reasonable simulation of a natural, free-flowing conversation. It was always constrained by the inputs.

With the advent of large language models, open-ended conversation has become possible. However, without constraints, a truly open-ended conversation could easily drift in any direction.

Therefore, the challenge for our developers was to take advantage of the learning possibilities of open-ended conversations, while maintaining constraints on the content. These constraints needed to consider both pedagogical and safety values. So, we developed a range of prompts that we use to constrain the LLMs and we have extensively lab-tested those constraints. Interestingly, as part of our testing methodology, we use AI as well as humans to challenge the Nualas and attempt to subvert the conversations. The axes that our constraints follow include the following:

• Class Level - ensure that the conversation uses grammar and vocabulary that is at the correct level
• Target Language - Discourage the use of any language that is not the language of the course
• Topic Adherence - keep the Nualas on point and bring the conversation back if the student tries to change topic
• Empathy and Tone - Conversation will be constructive and helpful, correcting errors in an empathetic way
• Safety - conversations will be:
  • age-appropriate
  • school-safe
  • free of advertising or propaganda
  • free of inappropriate, unsafe, or malicious material

While we have tested and continue to carry out extensive testing, it must be recognized that conversations are not scripted. So, while every effort is made to constrain the Nualas and test those constraints, some educators may not be comfortable with less scripted conversations. Nualang ensures that teachers are in control of which parts of our product are used in the classroom. Where we use AI, we provide controls that allow you, the teacher, to disable those features, if preferred. But, we would encourage you to embrace the educational possibilities of these powerful tools to provide richer conversational practice for students.

We welcome your feedback as we navigate the AI landscape and take advantage of the possibilities AI offers in a safe way.