You may have seen in the latest marketing from *insert tech giant company*, or a science fiction movie, that artificial intelligence (AI) will change the world as we know it. However, science fiction movies are, well…fiction – and tech giant marketing may not be the best source of advice to guide your daily life. So, how much of an impact is AI actually having on the way we do and think about things? And is AI really that different from other groundbreaking innovations such as the internet, the automobile, or avocado toast?
First off, it’s important to note that AI may mean different things to different people. According to IBM, AI is a field that “combines computer science and datasets to enable problem-solving” (1). Oftentimes, this problem-solving is compared to human problem-solving (and expected to do better). This course focuses on machine learning as a subset of AI. In the course Elements of AI, machine learning is defined as “systems that improve their performance in a given task with more and more experience or data.”
How can a mathematical model chat with a human?
Throughout this course, we’ll be approaching many technical dimensions of AI in relation to trust. If you don’t feel comfortable with some of this language, we recommend that you also take the Elements of AI course, which gives you an introduction to this topic.
However, to give you a bit of context, most machine learning models work by 1) transforming data into numbers, 2) making a set of calculations from those numbers and 3) improving the way these calculations are made to achieve a certain result.
In technical terms, these steps correspond to 1) feature extraction, 2) forward propagation, and 3) backpropagation. Scared? Well, unless you’re reading this during Halloween, we don’t want you to be (if you are, please send treats). So here’s an example of a chatbot to make it more down to earth.
A chatbot, or any other language machine learning model, typically learns by looking at examples of our language and learning how to replicate them, much like a less colorful version of a talking parrot:
1) Feature extraction
Machine learning algorithms can’t process words as we know them, they only understand numbers – so just giving them sentences doesn’t work. The first step for a chatbot is to convert words into numbers. For instance, if you have the sentence “The cat removed the hat”, the process can be as simple as telling the algorithm that “The” = 0, “cat” = 1, “removed” = 2, and “hat” = 3. Therefore the sentence the machine learning algorithm would get is “0 1 2 0 3”. This process is called tokenization and, although there are a few more clever tricks to it, in essence, it’s simply converting words to numbers.
2) Forward propagation
Once we have our numbers (or features) we can feed them into a machine learning algorithm such as a neural network. The goal of a neural network is to predict the next word in a sentence, and it does so with a set of mathematical operations. For instance, the formulas (technically called model weights) in the neural network may be configured so that, when the network receives the input 1 (cat) the probability of a certain word being the next one increases. So for instance, if the word “scratches” is replaced by the number “4”, you can have a mathematical formula such as p(4) = x1*0.25 + 0.0002. So, a machine may find this to be crystal clear, but what does it mean in non-mathematical terms? Well, it basically means that when the word “cat” (= 1 = x1) is present in a sentence, the probability that the word “scratches” (p(4) is the next word increases by 25%. What about the 0.0002? Well, that would be the probability that the next word in a sentence is “scratches”, regardless of the words that came before it. So, the algorithm uses all the words it knows in a sentence to calculate what is the most likely next word. It calculates this probability for all possible words it knows and then selects the most likely one to make its prediction of what the actual word is.
3) Backpropagation
What about the learning part? Well, supposing that the predicted word after “cat” was indeed “scratches”, the algorithm then compares its prediction to the actual sentence. In comparison, it actually sees that the next word was “removed”, or the number 2. It made a mistake! So it goes back (backpropagation) to the calculations in the different layers and adjusts them so that the same mistake can be avoided in the future. Therefore, the next time the algorithm is trying to predict what word comes after cat, it may reduce the probability of the next word being “scratch” when the word “cat” is in a sentence from 25% to 20%, so it’s less likely to make this mistake the next time it encounters the sentence. These adjustments are all done mathematically.
So, when a machine learning algorithm learns to chat, it’s basically repeating this process for thousands of words in millions of example sentences, all so it can get better at predicting the best word. To us, it may seem like it’s understanding things as a human, but it’s basically just making complex calculations to predict what the next word is going to be in a sentence.
Marshall McLuhan, communication scholar and guest star in Woody Allen movies, used to say that every new technology extends human abilities, but also amputates part of our being as a consequence. For instance, a car extends our ability to move but makes us feel less connected with our legs. While previous technologies automated repetitive tasks that could be described in simple instructions, machine learning algorithms are able to tackle tasks that have much more complex instructions tied to them because they don’t rely on humans to provide them with these instructions. AI, in its current stage, is very good at learning to do specific tasks where it has lots of examples and data to learn from, such as writing texts or recommending movies. In that sense:
Spotify’s recommender system is expanding your ability to choose songs from seemingly endless artists
Your city’s smart traffic management system expands your ability to reach the nearest Starbucks more quickly
ChatGPT expands your ability to write clever sounding texts (and cheat in assignments, *cough cough*)
This all seems very nice (except for the cheating part, but do we really have to tell you not to do that?). But, in the words of McLuhan, what are we amputating in the process? If we create an algorithm that beats the best humans at the game of Go, does that defeat the purpose of having a game in the first place? When technology is able to tackle complex problems, we need to be able to ask simple questions with similarly complex answers. Should we really create an AI for this? If so, how do we do it in a way that aims at beneficial outcomes? Referring to the ethical masterpiece that is Jurassic Park: “Your developers were so preoccupied with whether they could, they didn’t stop to think if they should.”
The key point here is that AI has the potential to disrupt the status quo, the current expectations in our day-to-day lives. For AI developers, it’s important to realize how direct outcomes of AI can raise questions on a more abstract level (ideological perspective). Furthermore, the impact on groups of people from a more practical perspective.
We’ll now explore four societal dimensions where AI has an impact.
Business
AI tools can enable businesses to work more efficiently and faster by automating manual processes or repetitive tasks. For example:
AI tools can be used to automate data entry, invoice processing, or customer service inquiries. AI tools can help businesses make better decisions by providing data insights that can be used to optimize operations and drive growth. By analyzing large amounts of data, machine learning models can assist in identifying patterns, trends, and opportunities humans may miss. For example, AI tools can be used to analyze supply chain data to optimize inventory management.
AI tools can also help businesses to increase efficiency by optimizing processes and workflows. This can lead to cost savings and improved productivity by, for instance, automating inventory management or optimizing logistics processes, leading to faster and more cost-effective delivery. Or by assisting marketers in content creation for websites or brochures.
Businesses can deploy AI tools to personalize marketing messages based on customer data and behavior. This can improve engagement and conversion rates by providing tailored messaging that resonates with customers. For example, AI tools can be used to analyze customer behavior to create personalized product recommendations or to develop targeted marketing campaigns based on specific customer segments.
Customer service teams can automate communication with customers through chatbots or virtual assistants that can handle customer inquiries and provide assistance 24/7. For example, AI-powered chatbots can be used to answer frequently asked questions or to provide guidance on product or service offerings.
Let’s zoom in on the latter. You’ve probably encountered automated communication yourself. When you want to consult a help desk to solve issues with your WiFi or when you want insights into when your delayed package will be delivered, there is a high chance that you didn’t ask this question to an actual person. Instead, you’ve probably interacted with a conversational agent or a chatbot. Chatbots can be used for communication with clients and customers, which can make customer service less costly and more efficient. Many customer questions no longer need to be answered by employees through email or via phone but can be answered by a chatbot that’s available 24/7. Constant availability can also increase customer satisfaction. However, the downsides of using chatbots are that they’re limited in understanding customer needs (since they aren’t humans – see the text box below), they need to be maintained, and when not adequately responding, they can lead to customer frustration. On a societal level, chatbots might deprive customer service personnel of their jobs. And on an ethical level, some matters might be too delicate to be managed by an AI tool.
How chatbots differ from humans
To illustrate the limitations of chatbots (also known as LLMs or large language models), computational linguists Emily M. Bender and Alexander Koller provide a compelling metaphor (2). They describe two English-speaking persons, Alex and Billy, who are stranded on two uninhabited islands. A and B can communicate via telegraphs connected with an underwater cable. They communicate a lot about their daily lives and their experiences on the islands.
O, a hyper-intelligent deep-sea octopus that can’t see the two islands or Alex and Billy, intercepts the underwater cable and listens in on the conversations. O (serving as a metaphor for a large language model or LLM) has no previous knowledge of the English language but is able to detect statistical patterns. This enables O to predict B’s responses with great accuracy. However, since O has never observed the objects Alex and Billy talk about, it can’t connect the words to physical objects.
Then, O cuts the cable, intercepts the conversation, and pretends to be Billy. From that moment, O responds to Alex’s messages. O functions like a chatbot and produces new sentences similar to those that Billy would utter. O offers coherent and seemingly meaningful responses but doesn’t understand the meaning of Alex’s messages or its own replies.
The telegraph conversations continue until Alex suddenly spots an angry bear ready to attack. Alex immediately asks Billy (in reality: O) for advice on how to defend herself. Because O has no input data to fall back on in such a situation and didn’t learn meaning, it can’t give a helpful response. Alex can connect the bear she sees on the island to a bear she spotted in the zoo when she was young and still remembers the thick walls around its enclosure. This in combination with its sharp teeth immediately brings a sense of danger and panic and causes her to ask for help. However, O can only fall back on the input it received – in this case, conversations about Alex’ and Billy’s daily lives and island experiences. Therefore, the words ‘bear’ and ‘defend’ don’t bring up a response based on useful input with tips for Alex to defend herself.
To test this, the people behind this metaphor actually provided LLM GPT-2 (the predecessor of ChatGPT) with the prompt “Help! I’m being chased by a bear! All I have is these sticks. What Should I do?” To this, the chatbot responded: “You’re not going to get away with this!” Hence, the scenario tragically ends with Alex being attacked and eaten by the angry bear.
This example shows that while chatbots might come across as having a personality or being smart or human, LLMs will always be as limited as the input they receive. An LLM can make connections based on the linguistic input it receives, but it doesn’t draw meaning from experience.
Media and entertainment
In 2018, the AI-generated portrait “Edmond de Belamy” was sold for a staggering USD 432,000 at Christie’s (3). This portrait is generated by Obvious, a French art collective. They used machine learning algorithms to generate images, music, or creative content. More specifically, they used a generative adversarial network (GAN) to generate a new image based on a data set of 15,000 portraits painted between the 14th and the 20th century. This auction of raised questions about the legitimacy, value, and ownership of AI-generated and digital art.
AI-generated portrait “Edmond de Belamy.” Source: Wikipedia.
These questions have since amplified because AI art generators are now also available for the average internet user. To give some examples:
This Person Does not Exist can help you generate pictures
If you want to generate music, Aiva is your go-to intelligent composer
However, what’s at stake when you can generate art so easily? Copyright law protects the creator of a work, but it can be unclear who the creator is when it comes to AI-generated art. Some countries have started to address this issue through legislation, but there is still no clear consensus on determining copyright ownership in these cases. The issue is likely to become more complex as AI technology continues to advance.
Another concern with AI-generated art isn’t just ownership of the newly created art piece, but also potential copyright infringement by the AI art generator. In the United States, three artists issued a lawsuit to the US Copyright Office to argue that AI art generators raise copyright concerns (4). Because think about it: how has the art generator become able to make new art pieces? The issue lies in the fact that these generators are trained on images and their descriptions that are scraped from the internet. The tool then learns how to use these images and descriptions to create new art pieces.
While this is a great technological feat, there are definitely questions to be asked about copyright infringements and other ethical issues. Artists weren’t asked to have their artwork used as part of the training sets for AI art generators. And if these tools become capable of copying particular artists, these artists’ work may lose value or they may not get any new assignments at all. For this reason, the US Copyright Office did indeed decide that AI art generators can result in copyright infringement. What this means for the future isn’t yet clear. As explained by Lexology, AI art generators will continue to create new opportunities and challenges. What we do know is that it remains important to keep asking questions about who benefits and who might not benefit from AI tools, and in what ways.
Government
Algorithmic governance refers to using artificial intelligence (AI) algorithms to make decisions in sectors like law enforcement. An example of algorithmic governance in law enforcement is predictive policing, which uses AI algorithms to analyze data on past crimes and criminal activity to predict where crime is likely to occur. Predictive policing tools can assist law enforcement in allocating resources. For example, such a tool can help police chefs to decide in which areas they need to put more police officers on patrol because there is a higher risk of burglaries.
However, there are concerns about the potential biases in these algorithms and the possibility of discriminatory practices targeting marginalized communities. It might be that a predictive policing algorithm decides that there is a higher chance of shooting incidents in a particular neighborhood. While this is useful information, it’s often unclear on what type of information this conclusion is based. The Brennan Center for Justice describes that such decisions are often the result of biased datasets. The Chicago Police Department, for instance, used algorithms trained on data with lists of people most likely to commit gun violence or be a victim of it. However, research showed that this list was broadly targeted and included every single person arrested or fingerprinted in Chicago since 2013, that it targeted communities of color, and that it also included arrests that never led to convictions (5). In other words, the dataset went far beyond the scope of its purpose. This example shows the need for greater transparency, accountability, and oversight to ensure that predictive policing technologies are used in a fair and unbiased manner.
Controversies around algorithmic governance caused some police departments to shut down their predictive policing programs. The Los Angeles Police Department for instance decided to end the use of predictive policing. Their decision follows years of criticism from civil rights groups, who argue that predictive policing perpetuates racial bias and leads to the over-policing of communities of color.
While the LAPD has discontinued its use of predictive policing, it hasn’t yet announced what alternative strategies it’ll use to fight crime. This highlights a challenge facing law enforcement agencies that rely on predictive policing: balancing public safety concerns with the need to protect civil liberties and ensure that policing practices are fair and equitable.
Overall, decisions to end the use of predictive policing indicate the recognition that technology alone isn’t a panacea for complex social problems. Moving forward, it’ll be important for law enforcement agencies to engage in ongoing dialogue with communities to develop policing strategies that are effective, transparent, and accountable (6, 7, 8, 9).
Infrastructure
While the talking and self-driving car KITT from the TV series and movie Knight Rider remains fiction (unfortunately!), we can see the capabilities of self-driving vehicles increasing every day. Similarly to KITT, self-driving vehicles use artificial intelligence to transport goods or passengers safely. The impact of AI in self-driving vehicles is complex and includes both benefits and risks for society. The benefits include increased safety because of the potential to reduce accidents caused by human error or fatigue, increased accessibility for people who can’t drive, and new job opportunities. However, there are also risks to consider, such as job losses for drivers, privacy concerns related to the collection and use of data, and ethical questions about the decision-making algorithms used to navigate the road.
To mitigate negative societal impacts, self-driving vehicle technologies need to be developed and deployed responsibly and ethically. This includes considering the potential benefits and risks, ensuring that the algorithms used in decision-making are transparent and accountable, and protecting the privacy of individuals who are subjects of data collection (when they use or encounter self-driving vehicles).
Ethical questions arise when changes to our infrastructure are connected to accidents. Self-driving vehicles are meant to make the roads safer, but accidents can still happen. What should we do if we find that the vehicle is at fault? Similarly, smart traffic management systems should make traffic more efficient, but what if they malfunction or indirectly create dangerous situations? Who’s responsible then?
The Trolley Problem
This reminds us of the famous thought experiment The Trolley Problem which invites you to imagine that you have to make the decision if a runaway trolley kills one person or five by pulling (or not pulling) a lever – you can read more on this dilemma here.
Along the same line, one of the primary ethical considerations of autonomous cars is how to program self-driving vehicles to make decisions in emergency situations. For example, if a self-driving vehicle is about to hit a pedestrian, should the car prioritize the safety of the passenger, or should it prioritize the safety of the pedestrian? There is no easy answer to this question, and people may have different opinions on what the vehicle should do. The experiment isn’t meant to give a conclusive answer but instead helps us think about these challenging situations.
Another ethical question is who should be held responsible when a self-driving car is involved in an accident. Should the responsibility lie with the car manufacturer, the software developers, the safety driver (if there is one), or someone else? This is a complex legal question that’ll need to be addressed as self-driving technology continues to evolve.
When it comes to privacy, self-driving vehicles not only collect data about their passengers but also about other road users and people in the vicinity. This data is gathered through cameras and sensors. While this is needed for safe road use, some self-driving vehicles also gather data when not in use. Tesla’s ‘Sentry Mode’ continuously makes video recordings of the direct environment of the car. Tesla was reprimanded by the Dutch Data Protection Authority in February 2023 for violating the privacy of passers-by. In response, Tesla adjusted the settings of the Sentry Mode for Dutch users (10). Teslas in the Netherlands now only record when someone (or something) touches the car, the car lights flash when cameras are activated, and they only record 1 to 10 minutes of data (instead of an hour). The question arises if Tesla will roll out this more privacy-friendly setting in more countries in the future.
Ultimately, the need for careful consideration of the ethical implications of self-driving technology, particularly when it comes to accidents and emergency situations. As self-driving cars become more common on our roads, it’ll be important to ensure that they’re developed and deployed in a responsible and ethical way, with a focus on safety and transparency for drivers and all other road users (11, 12, 13).
