As we kick off the new year, I am excited to share the progress made by the Witty Works ethics board throughout 2023. It was a year of growth and refinement for our ethical processes, which were initially developed in 2022, thanks to the invaluable contributions of our external experts, Anna, Fungai, and Leila.
To provide every member of the Witty Works team with an understanding of our ethical framework, we conducted comprehensive training in Q1 of 2023. This training session, led by our board's external experts, sparked meaningful discussions about how our ethical approach extends to our marketing and sales efforts.
With our ethics process firmly in place, we tackled our backlog of topics, allowing us to streamline the size of our ethics board. However, we chose to retain Anna as an external member. This safeguards that we continue to benefit from external insights, staying aligned with industry developments and maintaining a critical voice that prioritizes ethics over commercial interests.
One of our significant challenges in 2023 was how do we empower users to manage their learning path. We introduced a feature that allows teams and users to customize feedback for each diversity dimension. We agreed that while removing feedback for specific diversity dimensions is not ideal, it is better than overwhelming users (one of the main feedbacks we receive in the uninstallation survey). However, we unanimously agreed that feedback on hate speech cannot be disabled.
We also established guidelines on how to determine if a rule is basic or advanced. Throughout the year, we then meticulously reviewed all of our existing rules according to these guidelines. We reserved the 'advanced' category for issues demanding a bigger ‘mental leap’ or that isn’t officially recognized spelling or grammar. Exceptions were made for widely accepted practices, like the use of the Genderstar in German. To enhance clarity, we're in the process of providing examples for both basic and advanced rules for each diversity dimension.
In the summer, we embarked on experiments with Large-Language-Models (LLMs) such as ChatGPT and Llama-2 to automate sentence rephrasing. This improvement aims to simplify our ‘inspirations’ feature, which currently requires manual sentence rewriting. Initial test results have been promising. In exchange with the ethics board we would make this feature opt-in, both at the user and team levels, so users maintain control over proposed rewording while maintaining data privacy.
We also used LLMs to discover new rules, like our December update, which expanded our emoji lists. Although LLMs played a role, extensive manual review remained crucial. We have also begun using LLM’s to generate example sentences, which we are using for automated test to improve rule quality, reduce false negatives, and minimize false positives.
Witty Works was invited to present our ethical process at the ada festival in Berlin, where we garnered significant interest. Attendees show a lot of interest in our ethics process. Our rule-based approach at the core of our algorithm to side-step significant ethical challenges with machine learning such as bias, hallucinations, and energy consumption also created a lot of intrigue.
In December, our ethics board convened to reflect on the year's progress and discuss our future roadmap. One significant agenda item for 2024 is registering as part of the EU AI Act and, of course, monitoring the introduction of LLM-based features. We extend our heartfelt thanks to our ethics board alumni, Fungai and Leila, for their instrumental role in shaping our process and to Anna, whose ongoing guidance remains invaluable.
“I am very excited to see how much Witty has achieved with the refinement of the ethics processes! Witty has proven that adequate processes are key for an efficient use of ressources - the inital hard work was very much worth it. I am proud to be part of this journey!”
Anna Mätzener, Members of the ethics board
