Kategorien
Allgemein

Biased Data and Policies for Artificial Intelligence (Dr. Anna Jobin)

“Technology is neither good nor bad. Nor is it neutral” 

(Kranzberg, 1986, as cited in Jobin, October 2022).

Nowadays, everyone has at least heard or read about Artificial Intelligence. This big umbrella term has made an appearance in our world in the last decades. Everyone has different associations with it, be it a dystopian movie, an article in the newspaper, the algorithms underlying the ever-present social media or the YouTube video of a robot.

Dr. Anne Jobin, the president of the Swiss Federal Media Commission and scientist concerned with the interplay between society and technology, more specifically algorithmic systems and their social dimensions (https://annajobin.com), addressed the topic of Biases and Policies in Artificial Intelligence in a lecture part of the “Kritische Perspektiven auf die Digitalisierung” series mandated by the Vice-Rectorate Quality (https://www.dh.unibe.ch/forschung/biased_data/index_ger.html). 

What is an AI?

This is the first question that needs to be addressed when talking about Artificial Intelligence (henceforth AI), what is it exactly? There is no universally accepted definition of AI, but an important point made by Dr. Jobin is that policies for AI can still be built around the nonexistent definition. AI is now often times to be understood as specific form algorithms. In her lecture, Dr. Jobin compares it to a recipe, where data are the ingredients that are used to come to the result. For example, if we want to make an apple pie, our data is the butter, the flour, and the apples. AI is an umbrella term: a lot can be put under it. This umbrella term helps since there is no consensus on one definition and thus can be served as a black box for everything that falls under this concept. 

The role of bias in AI

Bias plays a central role in AI: it needs to be identified and understood. However, one common misconception about bias is that it is necessarily something bad. This is not always the case. Rather, sometimes it can be a useful tool in order to define what should be included in our data; in other words: what are our categories? To come back to our example, this would mean not including rotten apples in the recipe. This is where category definition comes into play, and it is important to decide to draw the line on what data are important to include or not. There can, however, be undesirable or inappropriate effects of AI. They sometimes are thought to be only pure computational problems where, in reality, it is more complex than this. 

One example is the chatbot that was created by Babylon Health (https://www.publictechnology.net/articles/features/gender-bias-concerns-raised-over-gp-app) that offered different guidance to men than to women. Two people with similar symptoms received different instructions, the woman to lie down and rest, and the man to go to the hospital to prevent a potential heart attack because of their gender. The thing is that the program was based on data, in this case, medical evidence, data that indicated that there were differences in symptoms for heart attacks between men and women. This was then severely critiqued for its gender bias. Therefore, digital profiling inferring individuals from general truths is methodologically and ethically problematic even if the statistics are true. It is important to keep in mind that AI bases on existing data, and it is important to reflect on how this may create problems. 

Some aspects that come up with AI are not even visible to the naked eye. Athalye et al. (2018) exemplify this with an image of a 3D-printed turtle that was classified as a rifle nine times out of ten. Apparently, the algorithm recognized a pattern on the shell of the turtle as a rifle. This exemplifies how AI can develop mechanisms that are outside of our understanding, and not everything it comes up with can be predicted and is even less relied upon. 

(Athalye et al., 2018, p. 284)

In order to control this, several companies set some basic AI principles. They usually are soft and not legally binding. Jobin, Ienca, and Vayena (2019) were concerned with some of those principles in their paper. As mentioned by Jobin et al. (2019), “AI has sparked ample debate about the principles and values that should guide its development and use. Fears that AI might jeopardize jobs for human workers, be misused by malevolent actors, elude accountability or inadvertently disseminate bias and thereby undermine fairness have been at the forefront of the recent scientific literature and media coverage” (p.389). Furthermore, they found that there was a strong difference in the interpretation of these guidelines and how they were then put into practice. However, as mentioned by Mittelstadt (2019), a “principled” approach to AI ethics might not necessarily be the best approach. 

An example of the severe consequences of an AI was the accident that led to the death of a pedestrian by an automated Uber car (https://www.bbc.com/news/technology-54175359). The detection mechanism did not recognize the pedestrian because it did not account for someone walking over the road outside of a crossroads. However, the technology worked as it was designed. There can be no unbiased AI because there is no unbiased world. No system will ever be 100% secure or unbiased. 

It is not possible to eliminate bias. It is important to be aware of it and to work with it. And most importantly: one needs to choose whether this is really the best solution. Many questions still remain unanswered in the field of AI; it is, however, unclear who needs and will answer them and who will set the benchmark of what is acceptable and desirable? These questions are all highly political. And this is why…  

“Human problems cannot be solved by technology alone

(Jobin, October 2022)

AI is political

It is important to keep in mind that AI works with what is given to it. Just because AI is available does not mean that it is always the best decision to use it. It needs to be questioned whether this is really the best solution for the purpose that is trying to be attained. It is important that there is an active investment to avoid and prevent harm because there are many ways in which AI can be hurtful. Therefore, AI is not only a technological matter but a political and social issue also. 

“We should treat AI as a political and social matter”

(Jobin, October 2022)

AI systems only work with available data and can only deal with these data in the way they were designed to deal with them. 

Categories, boundaries, and benchmark

The main takeaways here are three essential questions that must be kept in mind when considering AI:

  • What are our categories? 
  • Where do we draw boundaries? 
  • Where is the benchmark?

This guest lecture considers points in trying to find out what sensible governance or policy would be in relation to AIs. This is relevant to many different fields when considering the spreading of the use of AIs. Interdisciplinary work is also very important to go forward safely in these matters. Technology and AIs are undeniably a part of all our futures, which is why we are all concerned with this. It is thus important to think about what AI is used for and if it is adapted to certain situations or whether there are other or better solutions. It is important to consider because it can have real and hurtful consequences. We need to stay careful and use this technology mindfully. 

Personally, I really appreciated Dr. Anne Jobin’s lecture because it made the bias problem and AI problems more accessible to the wide public. I think this is sometimes missing in creating a link between scholars concerned with technology and scholars from other fields or people outside of the academic world. Making this discourse more accessible to more people opens it to public debate and enables one to have a real conversation about it. This also enables us to bring it onto a social and political level that is, as mentioned by Dr. Jobin, essential to consider in relation to AI.

References

Athalye, A., Engstrom, L., Ilyas, A., & Kwok, K. (2018, July). Synthesizing robust adversarial examples. In International conference on machine learning (pp. 284-293). PMLR.

Jobin, A. (October 2022). “What bias? Whose bias? A multi-level perspective on data and Artificial Intelligence”. Guest Lecture at the Universität Bern. 

Mittelstadt, B. (2019). Principles alone cannot guarantee ethical AI. Nature Machine Intelligence, 1(11), 501-507. Chicago. 


5 Antworten auf „Biased Data and Policies for Artificial Intelligence (Dr. Anna Jobin)“

Dr Jobin’s lecture was full of extremely important aspects concerning the profound impact of artificial intelligence. Moreover, the examples she showed clearly illustrated how – quite to the point – “human problems cannot be solved by technology alone”. With regard to the car accident or the chatbot from Babylon Health mentioned here : AI has to learn more – but also learns to apply prejudices. So, before AI acts ethically correct, such cases will accumulate, which again will emphasise the urgency of its political impact.

Herzlichen Dank für diesen Beitrag. Ich halte dieses Thema für sehr diskussionswürdig. Spannend finde ich insbesondere, dass hinter digitalen “Entscheidungsträgern” jeweils Menschen stecken, die die Regeln, nach welchen die Programm entscheiden, festlegen. Es sind also von Menschen festgelegte Regeln, welche dann durch eigenständige Anwendung der digitalen Programme zu Vorfällen wie dem Unfall des selbstfahrenden Autos führen. Mich erinnert die ganze Diskussion zum Thema ein wenig an den Umgang mit Kriminalität. Regelung und Regulierung hinken immer einen Schritt hinter den neusten Entwicklungen her. Die EU und auch die Schweiz haben das Thema aber auf ihren Traktandenlisten und sind daran Strategiepapiere zum Thema zu überarbeiten.

Digitale Strategie der Schweiz: https://www.digital.swiss/de/
Europäischer Ansatz für KI: https://digital-strategy.ec.europa.eu/de/policies/european-approach-artificial-intelligence

Dieser Beitrag fasst Dr. Anna Jobins Gastvortrag wunderbar zusammen. Insbesondere die Illustrationen und die Struktur des Textes sind sehr gelungen. Gemeinsam mit der Lesung und der Diskussionsrunde vom 07. November 2022 verhelfen beide Veranstaltungen zu einem besseren Verständnis über die Künstliche Intelligenz und regt zu einem kritischen Umgang damit an.
Deine Forderung nach mehr Vernetzung unter den Forschenden und der breiten Bevölkerung um die KI-Thematik mehr Sichtbarkeit zu verschaffen, befürworte ich. Dies geschieht zwar bereits durch Veranstaltungen wie die Ringvorlesung, die allen offen steht. Damit aber auch politische Entscheidungsträger Regulierungen („active investments to avoid and prevent harm“) betreffend KI als Priorität einstuft, braucht es genügenden Druck aus der Gesellschaft. Ich denke Kampagnenarbeit wie bei soziale Bewegungen könnte ein Lösungsansatz sein.

on behalf of Serena:
A very clear illustration of the concepts expressed by Dr. Jobin during her conference.

Personally, I find it fundamental to define the AI trust boundary that people should not overtake, and I believe you have shown it well by reporting the examples. Moreover, your blog post gives a transparent overview of the complexity that researchers have to deal with to handle this kind of technology, remembering us that sooner or later we have to learn to coexist with it as it will be our future.

The talk by Anna Jobin was really interesting and brought up so many important topics about AI. And I think you were able to bring out and focus on the most important of those topics in your blogpost! It’s really important to keep in mind how AI is a social & political subject which can have huge implications on our lives and that it should never (for the moment…) be seen as technology that has the same abilities as us humans aka. we should not 100% rely on it.

Schreibe einen Kommentar

Deine E-Mail-Adresse wird nicht veröffentlicht.

Diese Website verwendet Akismet, um Spam zu reduzieren. Erfahre mehr darüber, wie deine Kommentardaten verarbeitet werden.

Suche in OpenEdition Search

Sie werden weitergeleitet zur OpenEdition Search