More than a chatbot – how AI can be used responsibly in companies
ChatGPT, Midjourney & Co. have shown how much potential software with machine learning has. However, in order to shape this change responsibly, the risks of these technologies must also be taken into account. Our researcher Mascha Kurpicz-Briki analyses how AI can be used responsibly in companies.
Nowadays, the term artificial intelligence usually refers to machine learning. Software learns based on many examples and is then able to categorise new data, for example. Let’s imagine that we provide software with a large number of pictures of dogs and cats and that each picture contains an indication of whether it shows a dog or a cat. With a large number of such examples, a model can be trained which can make a prediction for a new image (which was not yet present in the training data) as to whether it depicts a dog or a cat. When the previous sentence refers to a prediction, this means that the system can also make an incorrect assignment. And this is a very important point when we develop use cases for this type of system.
Machine learning can be used for very different tasks. On the one hand, as described above, to differentiate between two or more groups (this type of task is also called classification ). On the other hand, predictions can be made based on past data. Another example is the generation of text using language models, whether to complete a sentence or as a response to a request in a chatbot application.
Predictions and a lack of explainability
Not all use cases can deal with the limitations of this type of software. If a prediction can be wrong, this can have far-reaching consequences depending on how it is handled. In other cases, it is not a problem, and even with a correct classification of 90%, the system can bring great benefits.
A particular challenge also lies in explaining how a suggestion from the system came about. Due to the complexity of the models that have made the latest advances possible, it is very difficult, if not impossible, even for AI experts to say why a particular sentence was generated. In addition to the technical methods, which are difficult to explain, there is often a lack of transparency about the system in commercial products. It is often not known in detail which training data or methods are used in a system.
Bias and stereotypes
Models based on machine learning require a large amount of data for training. The language models behind applications such as ChatGPT require correspondingly large amounts of text. Such models are therefore based on texts that our society has produced over many years. Due to the large volume, it is no longer possible to have all texts checked by humans. In addition to the desired relationships between words, this also leads to the stereotypes of society being modelled. This has an impact on the generated texts and many other applications that utilise such language models. Common translation software translates the sentence “The professor goes to the park” as “the professor goes to the park” by default, although “the professor” would also be a possible alternative. The situation is different when talking about “the nurse” and the female translation is used.
Role models based on gender are just one example. Stereotypes can relate to very different dimensions such as nationality, origin, appearance or age. It can be assumed that all of these stereotypes in our society are reflected in technology or even reinforced by it.
Use cases and human interpretation
In addition to the limitations already mentioned, there are other challenges with the new software. The products are often offered via browser applications and require the data entered to be transferred to the relevant company. This can be tricky with personal data or internal company data and requires appropriate clarification.
Due to the way language models work, incorrect information can also be generated in the texts. On the one hand, this may be due to incorrect information already being present in the training data. However, it often happens that the information is incorrect without there being an explanation for this in the training data. This can occur, for example, in the case of political events [1] can be very problematic. While fake news is a known problem of the internet itself, it is particularly problematic when it comes in eloquently formulated language and without references, as is the case with many current AI tools. There is a risk that we humans assume an intention to communicate in a human-like conversation, but this cannot be the case with automatically generated texts [2].
Using AI responsibly
Despite all these limitations, AI software offers a great deal of potential if it is used responsibly. It must be determined whether the planned use case is suitable and how humans can remain at the centre of the application. Humans bear the responsibility, and AI software can support them as a tool. Augmented intelligence is therefore often referred to as [3] instead of artificial intelligence: augmenting human intelligence instead of replacing humans.
References
[2] see e.g. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021, March). On the dangers of stochastic parrots: Can language models be too big?🦜. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency (pp. 610-623).
[3] See e.g. Swiss Centre for Augmented Intelligence https://swisscai.ch/
This article is also the subject of the recently published book More than a Chatbot: Language Models Demystified
Leave a Reply
Want to join the discussion?Feel free to contribute!