Better safe than sorry!
In cooperation with the German Research Institute for Artificial Intelligence (DFKI), we have developed Symbolon AI Coaching, an AI coaching tool that behaves ethically in coaching. Our AI coaching fulfills the following five key criteria for professional coaching:
1. The coachee develops the solution, not the coach. The coach merely guides through the reflective and results-oriented coaching process.
2. Every word counts. A wrong word at the wrong time can cause injury. This should be prevented at all costs.
3. The coach does not evaluate the coachee. There is no praise or criticism from the coach. The coach adopts a neutral stance.
4. The coaching space is protected. The coach is bound to confidentiality, no information is disclosed to third parties.
5. The coach behaves in an ethically correct manner. This means that he or she supports the coachee in the best possible way and does not go beyond the concerns of the coachee.
We follow the ethical rules of the International Coaching Federation and have opted for AI programming that we can fully control. This is currently only possible with a classifying AI.
The conscious decision for a classifying AI and not a generative AI
Generative AI learns to improve its form of text or image generation with every user interaction. When generating text, it can happen that a generative AI hallucinates. It then writes something wrong without making it recognizable. Depending on the type of error, this can be highly problematic in a coaching context.
In contrast to generative AI, classifying AI does not learn on its own. As the name suggests, it is characteristic of this AI that it works with classifiers. Classifiers are a type of text recognizer. All classifiers and texts of the Symbolon AI coaching were developed and validated by us in cooperation with the DFKI.
For example, a classifier can be a check of the entered text length, let's call it "topic length" for example. The "topic length" classifier has the following use: At the beginning of the Symbolon AI coaching session, the AI chatbot asks the coachee to formulate the coaching topic briefly and concisely in one sentence. In the example case, the coachee then describes their coaching topic in three sentences. Based on the "topic length" classifier, the AI notices that the coachee has not completed the task. As a trained reaction, the AI chatbot now asks the coachee to be more precise and formulate their topic more concisely in one sentence. If the coachee now formulates the topic in one sentence, the process continues.
If a generative AI were at work, it might not be patient with the coachee. Perhaps it would briefly summarize the sentences itself and then move on. With this step, it would take over the coachee's thinking and interpret something that misses the coachee's point. As a result, the coachee no longer actively searches for the solution within themselves but sees the AI as a solution expert. In our view, this would correspond to superficial advice that misses the target of the coaching offer and is therefore an absolute no-go!
In addition to the "topic length" classifier, we use others that check whether the person is making progress, for example, or others that support the coaching process by actively listening. To further improve Symbolon AI coaching with classifying AI, either existing classifiers can be retrained with extended data sets or new classifiers can be added to the system.
Every word that our AI chatbot writes has been predefined by us. There are no surprises (no hallucinations). This allows us to guarantee the highest quality and security for Symbolon AI Coaching.
The Symbolon AI Coaching is:
EU data protection compliant: login without registration, coaching is anonymous, no personal data.
The AI chatbot interacts in a secured space. A classifying AI is used, which makes validated statements only.
No image or audio recordings are made, the coaching takes place purely via the chat.
The AI chatbot implements in a focused manner the Symbolon questioning technique, which has been approved for 25 years.
The result is made available to the user as a one-off download in the form of a summary at the end.
The program runs on the DFKI server in Germany.
The University of Hamburg is evaluating the AI coaching. The first results of the study will be published in April 2024.
Dr. Friederike Redlbacher
Symbolon specialist, Hamburg