Microsoft Learning Zone - Interactive lesson generation

How does lesson generation work? 

Lesson generation is built using both a machine learning small language model called Phi Silica, and cloud-based large language model. The models are trained on a vast number of publicly available text samples. As a result, language models generate lessons that look like they were written by a human. 

To minimize the risk of model-generated inappropriate content, the response is passed through multiple validators and content moderation filtering steps.   

How should creators use lesson generation?

Educators who create lesson in Learning Zone can preview lessons on their own, share it with their colleagues or assign them students. It is the creator's responsibility to review and edit the content for appropriateness and accuracy before sharing with others. While we employ various filters and safeguards to limit questionable or inappropriate content, Elthe underlying technology of language models, has been trained on a wide range of internet sources, and might generate inaccurate information, so we rely on creators to confirm the appropriateness and correctness of the final activity.    

Lessons are designed to enhance the effectiveness and engagement of the learning process but may not cover the entirety of the learning material so they should be used alongside the study material. 

Note that the generated lessons are designed for introduction of topics, practice and formative assessment purposes and not for evaluation purposes, as many times the answers may be given in the body of the experiences, and educators should use those accordingly. 

What safeguards are employed while generating the passage?

• Lessons creators are instructed to review and approve the lesson content before using it and before sharing it with others (e.g. their students and colleagues).

• A descriptive notification is displayed to the creator if they missed reviewing any of the slides before saving the lesson.

• Content moderation is employed on the lessons before it is presented to the creator.

Model limitations

• Lesson generation is currently limited English and Spanish only.

• The model only uses textual input to create the lessons, any images and other objects included in the input will be ignored by the model. Code and math expressions are not being handled properly as well.

• Input files are used as reference to the lesson generation; the lessons may not cover the full files content.

• Input document size must be 50MB or less, and up to 125K characters or less, larger files that that will fail to upload. 

• Despite intensive training and Responsible AI guardrails employed, AI services are fallible and probabilistic by nature. This makes it challenging to comprehensively block all inappropriate content, leading to potential biases, stereotypes, or ungroundedness in generated lessons. 

Supported Languages

Lesson generation has been validated and supported for the following languages es-ES, es-MX, en-US, en-GB. 

Generative AI models are trained using vast amounts of data, and that data is most often in English. This can sometimes result in better performance in English as compared to non-English languages. As with any deployment of generative AI models, we encourage users to be mindful of the limitations of these systems for their specific use case, and cultural and linguistic contexts.    

Microsoft is planning to add more supported languages and locals to AI lesson generation. This article will be updated as additional languages are supported.  

Related Articles

Data Collection in Microsoft Learning Zone