This article was originally written in German by Konrad Rennert. Here you can find the machine translation, which was carried out by DeepL. Source: https://konrad-rennert.de/edge-und-bing-fuer-die-bildung-nur-bedingt-tauglich
AI awareness lesson: European education standards need their own AI developments
On the website https://aufklaerungsstunde.de, the Edge browser with Bing’s integrated artificial intelligence (AI) apologizes: “Sorry, this browser is not supported…” and when summarizing the ZDF video about the female anatomy appears: “I’m sorry, but I can’t assist with that…. Hmm…let’s try a different topic. Sorry. What else is on your mind?”
Such inappropriate reactions happen when AI is supposed to assist with translations, summaries and comprehension questions about sex education. The refusal to work observed and documented in November 2023 refers to reputable sources: The website aufklaerungsstunde.de of the pharmaceutical and consumer goods company Johnson & Johnson and a clip from the ZDF series Terra X, which is about the female anatomy.
Konrad Rennert (KR) planned the addition of the Terra X clips based on the order in the hit list. He had already edited two clips on the German memorial days on 3 October and 9 November and published them under a CC-BY license.
Proof of his claim regarding Bing and ChatGPT can be seen on YouTube. The spoken text can be read in 122 languages as subtitles or as a transcript on the right-hand side:
KR has made an Excel workbook available for download on its website. The sources and links are summarized on the “Info” sheet: https://konrad-rennert.de/wp-content/uploads/2023/11/BingZensiertAnatomieClipsVomZDF.xlsx
The results of the chat, screenshots and notes can be found on other sheets of the workbook.
KR graduated in physics more than 40 years ago. His thesis dealt with the interpretation of anomalies in the Earth’s gravitational field.
Thus educated, he generally thinks about cause and effect and the interpretation of observations.
The observed AI failures have human causes that need to be explored: Who is involved in decisions about whether AI responses should not be released, and who trains the AI in detecting toxic content?
KR found what it was looking for at Time magazine: According to a TIME investigation, OpenAI employed Kenyan click laborers for less than $2 an hour to make ChatGPT less toxic. The work of the Africans was crucial to improving the predecessor GPT-3, which was difficult to release due to its toxicity. OpenAI commissioned the Kenya-based outsourcing company Sama to flag texts containing violence, hate speech and sexual abuse: The employees of Sama were paid a net wage of between 1.32 and 2 US dollars per hour. The collaboration ended abruptly after Sama also collected images of illegal content. Source: OpenAI Used Kenyan Workers on Less Than $2 Per Hour: Exclusive | Time
With the knowledge of Time Magazine’s research, KR began to research the countries in which sex education is provided in schools. In Germany, it has been compulsory in elementary school since 1977. In many parts of Africa, it has still not been introduced. In Kenya, where the toxic content is evaluated thanks to cheap clickworkers, sex education is not on the curriculum until 2019, which means the clickworkers are not familiar with sex education in large parts of the world and cannot distinguish anatomy depictions from pornography. According to the OpenAI guidelines, sex education is expressly permitted and the content used should therefore not give rise to any objections.
The purpose of sex education in schools becomes clear when you compare the depiction of pregnancies of underage girls on the world map. In places where, according to the WHO, sex education has been practised for a long time, underage pregnancies are rare:
Conclusion of the observation: KR happened to find a flaw in the provision of AI. AI is important and our tried and tested European standards must be taken into account in all areas of life. Not only in dealing with sexuality, but also in other areas we have different attitudes, e.g. to the possession of weapons, to anti-Semitism, to dealing with homosexuality and with freedom of expression.
If Europeans do not want to be patronized in AI issues, they need independent AI developments. The evaluation of toxic content must also not be outsourced to cultural groups with completely different norms and standards. At the very least, the final check must always be carried out by professionally trained Europeans.
KR is not just relying on Bing. In his project to supplement the Terra-X clips, he added the anatomical terms from the ZDF video with the help of DeepL and translate.google.
He was unable to implement his tabular structure of a vocabulary list for simultaneous translation into several languages due to the Bing AI’s refusal to provide sex education. He then demonstrated the basic feasibility of using Bing for the anatomy of the ear in the Excel workbook mentioned above. This proves that his tabular arrangement can be used to translate the specialist vocabularies for all topics into several languages simultaneously, provided they are not classified as toxic.