AI Company Principles: wie man AI im Unternehmen regeln kann.
Im Rahmen meines Oxford AI Programs haben wir umfassend das Thema Ethik und AI durchgenommen. Kurzum sollte die Gesellschaft nicht die gleichen Fehler wie in der Social Media Era machen. Ob gesellschaftlich, politisch und auch gesundheitlich haben die Social Media Apps und ihre Algorithmen einen wesentliche Einfluss gehabt und wurden nicht resp. zu spät reguliert. Im Bereich AI ist man hier proaktiver z.B. die EU mit ihrem AI Act.
Gerade wenn Algorithmen einen wesentlichen Impact auf den Menschen haben, braucht es klare Regeln: Bildung, Kredite/Finanzen, Recruiting etc. Schon jetzt haben Algorithmen einen Impact auf das Recruiting:
Das Recht ist das eine, was ethisch korrekt ist, das andere. Jedes Unternehmen muss sich überlegen, welche Guidelines und Prinzipien bei der Nutzung von AI gelten sollen. Welche Daten dürfen wir nutzen? Sind diese biased, veraltet oder wirklich repräsentativ? Welche AI-Entscheidungen müssen von einem Menschen freigegeben werden? Sind diese Nachvollziehbar etc.
Ich habe mal für eine hypothetische Firma ein paar Zeilen zusammengeschrieben. Ihr könnt das gerne als Grundlage nehmen.
As Switzerland’s leading condom and sexual pleasure brand, we are committed to diversity, sexual health and the privacy of our customers. These principles also apply to our marketing and the use of AI. The following rules for internal marketing, tools and agencies must be observed. The purpose of all our AI and machine learning activities have to support our principles and not using AI to deceive, manipulate, or harm users.
Data Privacy
The revDSG (Swiss GDPR) applies to all our activities. The use of personal data to train machine learning algorithms or to target customer segments or groups needs the consent of the users.
Whenever possible we train with not personalised data excluding personal information like surnames, email addresses and such.
We ensure to use internal systems or closed cloud systems to train or analyse data. The use of personal user data in open systems like ChatGPT is strictly prohibited.
Whenever users ask for deletion of personalised data or transparency we answer these requests in a timely manner. The use of data and of algorithms have to be documented in an understandable, non-technical documentation before an algorithm is actively used in a project in a production environment.
Accuracy and Diversity
When developing machine algorithms or training third party AI tools we ensure that representative data sets are used. Ceylor wants to foster diversity, support all genders, races, ethnic groups, age groups and sexual orientations.
When developing AI we ensure that actual data is used and algorithms are updated in a frequent manner.
When using algorithms in communication e.g. for targeting personalised ads, personalised email, personalised shop recommendations etc. we have to ensure that our diversity principle is still in place despite personalisation.
Explainability
When developing or implementing AI measures and tools, explainability of automated decisions is key to Ceylor as already mentioned in 1d. Any AI system used by our company has to be able to explain its decisions in a manner that is understandable to stakeholders. If this is not possible, no fully automated decisions are allowed and have to supervised by a employee of Ceylor.
Safety
When using algorithms in communication e.g. for targeting personalised ads, personalised email etc. we have to make sure that no conclusions can be drawn about individual persons, such as sexual orientation, fertility, and pregnancy. Whenever possible and reasonable we work with generic and diverse segments and not 1:1 personalisation.
Security and compliance
All internal and external staff members in an AI project (developers, data scientists etc.) have to carefully read and sign this document of AI principles before starting working on an AI project.
All team members of an AI project have access to only as much personalised data as is necessary to achieve the project goals.
These project goals have to be defined in a detailed project concept when AI projects include personal data or personalisation. This concept has to be reviewed and approved by the internal AI commission.
AI and machine learning measures that have a significant impact on our company or our clients must be reviewed by an external audit and submitted to the internal AI committee. This impact has to be defined by the committee.
Habt ihr schon Regelungen im Unternehmen etabliert? Welche Themen deckt ihr ab und welche bewusst nicht? Wir unterstützen dabei gerne mit AI-Strategien, Workshops und Guidelines.
