We explain the processes behind the training of artificial intelligence. Because ChatGPT shows us how dangerous the unreflective use of AI is and what quality demands we should place on AI.
Not a day goes by without headlines from and about ChatGPT: a few days ago, it was revealed that OpenAI’s text generator had resorted to cheap Kenyan labor for its development. For less than 2 US dollars per hour, employees of the East African company Sama were hired to train the artificial intelligence, according to a report by “Time”.
Reflective handling of AI content
Artificial Intelligence depends on human input. It must be manually fed data in order to learn. But what about the quality of AI development when this training is handed off to low-wage workers?
“The knowledge of the AI is generated primarily on the basis of the labels provided. The people who prepare the data and subsequently the labels are therefore the teachers of these,” says our AI expert Dr. Ernst Nusterer. The process is often outsourced, as in the case of OpenAI, and done for a defined hourly rate at labeling companies. “We carelessly like to see the decisions of an AI as objective – but it is not at all, and this mistake is fire dangerous”.
Labeling and quality requirements: A new challenge
We have been developing AI solutions in-house for over 5 years. The labeling and the high demands of the training are central topics. Here, a dedicated team is tasked with deciding what to teach the AI. The top priority is the quality of the data used, according to Tina Waldner from the labeling team: “Unclear data that we even suspect could pose a problem during training is not even included in the data pool. Just like the motto: Better less, but good data.”
Not only the quality of the training material is crucial. For labeling, concentration is also required from the team: “We pay attention to exact and precise work, and internal consultation is just as important. We always have to be on the same level to be able to work in the same way. A very important point is the constant exchange with the development team,” Waldner continues.
With the rapidly advancing development of AI software, the process in the background and its quality must also be made more transparent in the future. This is where people play a key role. Because ultimately, training determines how good a piece of software is, what added value it offers, and whether we can really trust the results.
Changing media content: Let’s talk!
Content-generating AI will seek its place in the media landscape in the coming years. We analyze the developments with first-hand experience. Feel free to write to us if you want to talk about it: contact@linkthat.eu