Sparx currently does not use LLMs in our 'core' homework personalisation algorithms; our in-house expert pedagogy data modellers build them.
Sparx does use Artificial Intelligence tool providers as part of our support companies that help us deliver our service to you. The use of these tools has fully GDPR-compliant data-sharing contracts in place. They have had extensive privacy and security evaluations and are very safe.
In line with DfE guidance, our Sparx Terms and Conditions, Section C: Data Handling Agreement, Sub-processors states:
"All personal data, regardless of processing location, will be strictly ring-fenced and explicitly excluded from any and all subprocessor large language models (LLMs) or generative AI training. This safeguard is absolute and ensures student data will not be used to train or improve third-party AI models."
At present, we are using AI tools for the activities described below.
Categorising teacher support tickets
Categorising teacher support tickets
We use AI tooling to provide a summary of help and support tickets and categorise them for analysis e.g. related to log-in or a set-up query. This concerns teacher emails only, students cannot raise support tickets.
Data is shared through a secure, gated application programming interface or API.
Sparx Learning data is prohibited for use in the training of any third-party AI models.
During project testing, the outputs from the API were repeatedly evaluated to ensure that they gave accurate responses, checking for misleading or biased information.
It has significantly improved our customer support efficiency and prioritisation. No decisions are made solely on the output; ticket summaries and categorisations are always reviewed by the Sparx staff member dealing with the query.
Improving teacher feedback submission
Improving teacher feedback submission
We use AI to save teachers time when submitting feedback about Sparx products or features. When starting to add feedback or request a new feature, our feedback platform has an inbuilt LLM to suggest similar topics that have already been submitted. Teachers can then upvote or comment on the topic, saving them time rather than rewriting items already raised, whilst adding more context if desired.
Feedback comments are anonymised with a randomised user alias for each teacher (e.g. red squirrel). This enables us to make topics and comments visible across our teaching community, sharing ideas and engaging discussion.
Sparx Learning data is prohibited for use in the training of any third-party AI models.
Using AI to mark written answers in Sparx Science
Using AI to mark written answers in Sparx Science
In Sparx Science, we use AI to mark short-answer exam-style questions worth between 1 and 4 marks. Using AI allows us to automatically mark answers and provide immediate feedback to students. No student personal data is shared, only the question answer and the mark scheme for that question.
Large language models (LLMs) have a nuanced understanding of language and so we are able to award marks even when the student has phrased their answer differently, included other irrelevant information or made spelling mistakes. Many online platforms bypass this by using more constrained question types (e.g. multiple-choice). The benefits of our approach are:
Students are prompted to retrieve the information from memory (an effective learning strategy) instead of selecting the correct answer from a list of options.
Students improve their writing skills.
Students are able to practise application skills similar to those in exam contexts.
It is important to note these are different to our short-text questions, which test key terminology and expect short answers of 3 words or fewer. Examples of these can be found in Flashcards or gap-fill questions. They are marked by our custom spellchecker, which permits a range of misspellings, so students are rewarded for the science they know, not their spelling ability. These do not use AI.
In-context definitions in Sparx Reader
In-context definitions in Sparx Reader
In Sparx Reader, students can select a word they do not understand to request its definition. We use AI tooling to interpret the sentence, enabling us to deliver the specific definition of the word in context, enhancing the understanding of the reader. No student data is shared.
Invoice processing
Invoice processing
We use AI to help automate our invoice processing. This includes extracting purchase order numbers from documents and sending communications and payment reminders.