You are currently in guest mode. Log in or register to fully use the HOOU platform.
Your advantages when registering:
- You can use chat and forums to exchange ideas.
- We will confirm your participation in some courses.
Ethics by Design
Topic outline
-
‘Ethics by Design’ is a pioneering learning programme that emphasises the central role of ethics in the development and application of technology. Instead of only looking at ethics retrospectively, it is integrated into every phase of development. This course shows you how to build technology that is not only innovative but also fair, just, and ethically sound.
In interactive modules, you will make decisions in everyday scenarios and deal intensively with topics such as fairness, justice and data protection. This hands-on content sensitises you to the relevance of ethical considerations and the positive effects of their implementation in the development processes of technology or the negative consequences if ethics do not play a role in the development process.
Basically, it's like with food: Our consumer behaviour changes when we know the origin of our food and the consequences of our consumption. The same applies to technology: the more we know about how it is developed and what the consequences are, the more you will view and develop technology differently. The course provides the necessary tools and knowledge to develop technology that is fair, transparent and in the best interests of all stakeholders.
-
Glossary and Community Forum
If you come across important terms in the course, you find short definitions in the glossary. In the forum, you can connect with fellow learners, ask questions, share insights, and engage in discussions.
You can access them at any time
-
-
Let's start our dive into Truth and Ethics by Design by getting to know Emma. In this episode, we will follow her on her journey.
Emma works in the marketing department of a company. Every day, she is burdened with numerous tedious and monotonous tasks, leaving her with barely any time to focus on campaigns – the very reason she entered this field. Recently, she has been hearing about a new technology in the news, something akin to a chat bot but purportedly much more useful. A few of her colleagues have tried it and are enthusiastic. She searches for it online and finds the one generating significant buzz: ChatGPT.
Emma decides to try it out, just for fun. After a few trials, she notices how well it can perform small tasks. Feeling confident, she decides to delegate one of her daily, repetitive tasks to it. Emma frequently receives emails from potential customers, which she needs to read through to understand their campaign requirements and deadlines. These emails often contain excessive information, making the process tiresome. She instructs ChatGPT to summarise the emails and identify the due dates.
-
Emma uses ChatGPT at Work.
Let's see how it goes.
Click on the two elements below to find out more.After Emma discovers that using ChatGPT
to summarise emails and finding due dates in them is effective, her confidence in the technology grows, and she continues to use ChatGPT for these tasks. However, one day, Emma notices that ChatGPT provided the wrong due date, even though the correct date was included in the email. Doubts crept in. "Is ChatGPT producing misinformation?", she asked herself frustrated about having to deal with now having to double check ChatGPT's work.
-
Large Language Models: Definition
Large language models (LLMs) like GPT or BERT are advanced computer programs designed to predict and generate text by spotting patterns in large amounts of language data.
They work by training on massive datasets of text (think millions of books, websites, articles, etc.), learning how words, phrases, and sentences connect. This training allows them to generate responses that sound human-like and relevant to the input text.Large Language Models: Transformers
The technology behind LLMs is called a transformer. A transformer is a type of neural network designed to process sequential data, like sentences, by focusing on relationships between words over long passages. It doesn’t just read text word-by-word; it has a way to “pay attention” to different parts of the text all at once. This approach, called self-attention, allows the transformer to weigh the importance of each word in a sentence and understand how words are connected across sentences.
Examples: Imagine a transformer reading the sentence, “The cat chased the mouse because it was hungry.” The model can understand that “it” likely refers to “the cat” rather than “the mouse.” This type of understanding, applied across massive amounts of text, allows transformers to build up complex language patterns.
Learn more about transformers in this short video from Google Cloud Tech:
Large Language Models: Stochastic Parrots
LLMs are sometimes called “Stochastic Parrot”. The term "stochastic parrot" was introduced in a 2021 paper by Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. It highlights the idea that, at the end of the day, large language models aren’t actually understanding the text they produce. They’re like very sophisticated parrots: they repeat patterns, sequences, and phrases that sound human because they’ve seen them before in their training data. "Stochastic" refers to randomness in these models—they don’t think about what’s true or accurate. They pick words based on probabilities calculated from patterns in the training data.
-
Let's quickly recap what we have learned about LLMs so far
-
From Misinformation to Hallucination
Emma is also concerned with misinformation and she comes across a term often mentioned in this context: hallucination.
In LLMs, a hallucinations refers to when the model generates text that sounds plausible and fluent but is factually incorrect, misleading, or simply made up. This happens because the model is designed to predict the next word in a sequence based on patterns in the data it has seen during training, rather than on a grounded understanding of factual accuracy.
Let's watch this short video from IBM Technology about LLM hallucinations to get a bit more context:
-
What is Truth?
Intrigued by the LLM hallucinations, Emma felt a deeper question pressing on her mind: What exactly is truth, that elusive concept these powerful models so often miss? She realized that if LLMs, built on massive datasets and cutting-edge algorithms, could still “hallucinate” and mislead, then perhaps truth itself was more complex and layered than she’d ever considered.
Determined to explore this idea, Emma dived into philosophical texts and found herself fascinated by different theories of truth—each offering a unique way of understanding how we define what’s real, accurate, and meaningful.Please explore what Emma uncovered:
-
Quick Recap
Please read four chapters of the following activity carefully by clicking here or on the title.
-
Truth Filters for LLMs?
Emma's curiosity deepened as she wondered how these philosophical perspectives on truth could be practically applied to help both users and developers interact more responsibly with LLMs. If truth could be seen through different lenses, then each lens might offer a way to improve the way people work with LLMs. She began to think of these theories as 'truth filters' - different ways of evaluating, interpreting and even designing model responses with specific goals in mind. These filters, she thought, could form the basis of an ethics by design framework, a system that could make interactions with LLMs more transparent, grounded, and ultimately safer for all.
She sketched out ideas for how users and developers could apply each truth filter to guide their approach:- Correspondence Filter
- Coherence Filter
- Pragmatic Filter
- Epistemic Filter
-
Excited by her insights, Emma decided to share her ideas with Alex, an AI developer. She hoped she could help her understand how practical it would be to integrate these truth filters and find ways to control hallucinations in LLMs. As the conversation deepened, Alex, brought up a technical concept that piqued her curiosity.Let's follow their conversation to learn more by clicking here or on the title of this activity.
-
Practical Application: Emma returns to Work
Feeling better informed after her deep dive with the developers, Emma returned to her workspace, ready to make a pivotal decision about how her team would implement their AI assistant. She wanted the tool to be both useful and responsible, but she was faced with two distinct options for how the assistant should handle information and respond to users.Please explore the two options below.
-
She paused, considering the pros and cons of each option. Emma looked at her notes, wondering which path would best serve her users and meet her own ethical standards.Emma's Choice
Emma paused, considering the pros and cons of each option. She understood the stakes - how each choice could shape the users’ experience with the AI and, potentially, how responsible her team would be seen in deploying this technology.Emma has two options:
- Option A: Allow unverified, creative responses, enabling a flexible and imaginative AI that inspires and brainstorms but may generate inaccurate or fictional information.
- Option B: Restrict AI outputs to fact-checked content, creating a reliable and grounded AI focused on accuracy, with the trade-off of a more reserved, factual tone.
What should Emma choose? Please feel free to explore both options.
-
-
-
From Truth to Knowledge
If you have already worked through the episode on truth, you are already familiar with Emma. However, having finished the episode on truth is not necessary to tackle this episode.
After pondering over the topic of truth for a long time, Emma leaned back in her chair, feeling the weight of the challenges. Her journey had begun with questions of truth - how to make her AI accurate and reliable in its responses.But as she delved deeper, these questions naturally evolved into a more profound challenge: what does it mean to ‘know’?
If truth itself was elusive in the realm of AI, could an AI ever truly know anything, let alone understand it?
She considered how casually we speak of what an AI "knows," "understands," or even "thinks," as if it were capable of human-like comprehension. We might even regard it as a "source of knowledge." But this led her to question: does the ability to generate plausible responses equate to knowledge or understanding? Or is the AI simply simulating the appearance of knowledge without genuinely grasping the meaning behind its words?
In this episode, we explore whether LLMs can 'know' anything at all and what is required for humans to have knowledge.
Click on the two options below to find out whether LLMs can 'know' anything at all, and what criteria need to be met for humans to know. These are central questions of this episode. You will also be introduced to some interesting philosophical thought experiments related to these questions.
-
AI as a Tool for Knowledge Acquisition
Emma considered the implications of the transfer of Searle's thought experiment and Gettier's criticism of a definition of knowledge centred on justified beliefs and truth: If users accepted AI responses at face value, mistaking fluent language for true knowledge, they might assume they were “learning” something meaningful when, in reality, they were interacting with text devoid of understanding. Worse, the confidence of an AI’s responses could obscure the fact that it might be “right” only by coincidence or luck.If the AI functioned as a bullshitter, uninterested in truth, then any knowledge gained from it would be tenuous, relying on users’ critical thinking rather than on the AI’s reliability.
In the episode on truth, this is discussed at length.
So could humans gain knowledge from an indifferent entity? Emma realised the answer was complex. Users might extract valuable information by critically engaging with the AI’s output, fact-checking, and applying their judgment. In other words, humans could create knowledge from AI responses by scrutinising them carefully, but the AI itself was no reliable source of knowledge. It couldn’t impart understanding directly; it could only offer language patterns that users would have to dissect and validate for truth.
The AI, then, could be a tool, a prompt, or a starting point for knowledge, but not a true knower or teacher.
And this, Emma decided, was a crucial insight. If she proceeded with the design, she’d need to make it clear to users that the AI was not an authority but an assistant—an entity that could aid in exploration but should never be mistaken for a source of true understanding.
Let's see how it goes for Emma.
-
Emma's Conversation: Designing for Knowledge
Emma gathered her team to discuss the implications of Searle’s Chinese Room and Gettier cases on their AI’s design. She wanted to explore practical solutions to make the AI responsible and transparent, and she was excited to see her team’s reactions to these ideas.Read through Emma's conversation to find out what practical solutions she and her team have come up with. Luckily she took screenshots of the conversation. Use the navigational arrows to view all screenshots.
-
-
-
-
<
Welcome to our episode on Justice. Using COMPAS, you will learn about the non-neutrality of technology and think about questions like: data should be included to achieve fair outcomes? Should we use all data or exclude sensitive information to avoid bias?
Through the case study on COMPAS, you will key concepts associated with justice such as bias, discrimination, and equality, how these principles apply to technology. Finally, you will get to see how philosophical theories of justice can guide us when technical are unavoidable. This episode provides a solid foundation thinking critically about justice and fairness in design and practical ways to approach it. -
Let's dive straight into a real-world scenario!
Imagine a courtroom where a judge has to decide whether to release a defendant on bail. The judge considers not only the facts of the case but also a risk score generated by an AI system called COMPAS. This score, produced by analysing various data points, claims to predict the likelihood that a defendant will reoffend if released. The higher the score, the more “risky” the system deems that person to be.
When COMPAS was actually used, it provided risk scores that looked something like this:
Foto credit: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
This score raised concerns among some, as it assigned a low-risk rating to a white man who turned out to be quite dangerous, while giving a high-risk rating to a woman of colour whose criminal record was far less severe and who did not reoffend. It is, of course, unreasonable to judge an algorithm based on a single instance. Nonetheless, questions emerged: could it harbour bias against individuals of certain ethnic backgrounds or genders?
-
COMPAS is Real
COMPAS is not a fictional example. It was and is actually in use. ProPublica, a non-profit investigative journalism organisation, published an article about COMPAS in 2016 that went viral and claimed that the algorithm was biased against black people. If you want to learn more about this real-world case and ProPublica's stance, please read their original article, 'Machine Bias'. They also explain their methodology in a separate article that is also worth reading called 'How We Analyzed the COMPAS Recidivism Algorithm'.
Hany Farid discussed the COMPAS case in detail in his TEDx talk. Please go watch it:
The COMPAS case can be seen as a reminder that algorithms, far from being purely objective, can replicate and even magnify human biases. But let's take a step back and break down some of the aspects from the video and our previous claims, starting with the following question: AI systems are basically computer systems working with maths and numbers. So, why aren't they morally neutral anyway?
-
The Non-Neutrality of Technology
Many people assume that technology, especially algorithms, is inherently objective and free from human bias - after all, isn’t it just math?
However, technology is rarely as neutral as it appears.
Every AI system, including COMPAS, reflects a series of human decisions:
- what data to include,
- how to weigh different factors,
- and what outcomes to prioritise
In the case of COMPAS, these decisions have serious consequences, as the algorithm’s risk scores can determine whether someone goes to jail, receives bail, or is released.
Yet the data driving these predictions - such as past arrest records or demographics - often contains existing social biases, which the algorithm can then reproduce or even amplify. This is where ethics by design becomes essential.
Example:
Social media platforms appear to be neutral since they appear to simply provide a way for users to create, share, and react to content. But the algorithms that prioritise and recommend content based on engagement metrics such as clicks, shares and comments are not neutral - they reflect the platform's goal of maximising user attention and advertising revenue. As a result, they often amplify sensational or polarising content because it generates more engagement, even if it spreads misinformation or fuels division.
This demonstrates how technology embeds specific values and agendas, influencing social behaviour and shaping public discourse in ways that are far from neutral.Algorithms won’t automatically be fair or unbiased just because they’re created by computers.
Instead, developers must consciously address and mitigate potential biases during the design process. The stakes are high, and so is the responsibility: in a world where algorithms influence real lives and liberties, we must approach their development with ethics and fairness at the forefront.
Having read that: Let's again look at COMPAS. How does it challenge the belief that technology is neutral? Please share your thoughts below.
-
Now that you know that technology is not neutral, how would you train your software?
The Scope of Training Data
Imagine you’re tasked with designing an AI system like COMPAS, assuming that there is at least the prospect that it could outperform humans on a specific prediction task. But in addition to high performance, you also want the system to be fair, and you’re thinking about the data it will use.
Would you include all data available, including factors like sex and ethnicity (Option A)?
Or would you exclude certain factors to guard against potential bias (Option B)?
Please explore both options.
-
-
Selecting the Right Data is Not Enough
As shown with these two approaches of informational maximisation and Fairness Through Unawareness, simply including or excluding certain types of data won’t ensure justice or fairness. We must further investigate fairness metrics to understand how to design systems that genuinely avoid bias.
Before we dive deeper, however, we need to clarify some essential terms.
Understanding these concepts will equip us to evaluate and design AI systems with greater ethical awareness and precision.
-
-
How to ensure Fairness?
In the section on the scope of training data, we explored two attempts to ensure fairness, but saw that this was not enough. Now that we've clarified some important concepts, let's go back to looking for a way to ensure fairness in the development process.
Fairness Metrics
In the COMPAS system, Brisha Borden and Vernon Prater were two individuals who received very different risk scores. Borden, a young woman of African-American descent, was rated high risk, while Prater, a middle-aged white man with a longer criminal record, was rated lower risk. But we can’t really say much on the basis of one prediction. So, how can we actually test for bias in an algorithm?
Let's dive deeper into the tools and concepts we can use to assess whether an AI system, like COMPAS, is actually fair.
These tools are known as fairness metrics, and they help us examine how the algorithm performs across different groups.
Fairness metrics are essential in identifying potential discrimination and helping us understand where an algorithm might be biased. Here are the two most common fairness metrics we use in AI ethics:
The first metric is accuracy, or how often the algorithm makes correct predictions.
For COMPAS, this would mean the percentage of times it correctly predicts whether or not someone will reoffend.
Error Rate Balance is about comparing the types of mistakes the algorithm makes across different groups. Specifically, we look at false positives and false negatives for each group.
False Positive: When COMPAS predicts someone will reoffend, but they don’t.
False Negative: When COMPAS predicts someone won’t reoffend, but they do.This analysis shows that even if the algorithm is very accurate for all groups, it could still make different types of mistakes for different people. For one group, it might mostly label people as high risk when they aren’t dangerous (false positives), while for another, it might label people as low risk when they are dangerous (false negatives).
This difference really matters on a personal level: the effect on someone’s life is very different if they are wrongly released or wrongly kept in detention.
-
Calculating Fairness
Let's look at a dataset from COMPAS to see how discrimination can creep into a technology unnoticed, and to see that discrimination, at least in this case, can be determined mathematically.
The following table shows in the first line ("reality") whether or not the persons assessed by COMPAS actually recidivised.
The second and third lines ("test core") show the COMPAS scores for blacks and whites (high or low risk of reoffending).
So, let's apply the two metrics: accuracy and error rate balance.
-
Calculating Fairness
1. Accuracy
Let’s begin with accuracy by asking: how often does the algorithm make correct predictions for each group? To measure this, we use conditional probabilities, starting with the positive predictive value (PPV). This refers to the probability that someone actually reoffends (recidivates) given that they have been assigned a high-risk score by the algorithm.
We can also examine this question from the opposite angle, using the negative predictive value (NPV). This measures the probability that someone does not reoffend, given that they have been assigned a low-risk score.
By applying these calculations to the data in the table above, we can determine how well the algorithm performs in these two specific areas. The results give us insight into whether the algorithm is equally accurate for different groups or if there are discrepancies that need to be addressed.
Here are the calculations:
Positive Predictive Value
Black: P (Recidivism | High Risk): 60/(60+20), 0.75
White: P (Recidivism | High Risk): 16/(16+5), 0.76
Negative Predictive Value
Black: P (No Recidivism | Low Risk): 6/(6+14), 0.70
White: P (No Recidivism | Low Risk): 1- (22/(22+57), 0.72
It seems the algorithm performs similarly for both groups, with no clear indication of bias. But is that truly the whole story?
2. Error Rate Balance
Now let’s explore a second important metric: error rate balance. This helps us understand the kinds of mistakes the algorithm makes when it gets things wrong. Are its errors more often false positives (labeling someone as high risk when they are not) or false negatives (labeling someone as low risk when they are actually dangerous)?
To calculate these numbers, we take a slightly different approach to the conditional probabilities we used earlier. Previously, we asked about the probability of recidivism (reoffending) given a high-risk score. Now, we reverse this perspective: we ask about the likelihood of a high-risk score, given that someone actually did reoffend.
Similarly, from the negative perspective, instead of asking about the likelihood of no recidivism based on a low-risk score, we now consider the likelihood of receiving a low-risk score, given that someone did not reoffend.
By looking at these reversed probabilities, we can better understand whether the algorithm treats mistakes equally across groups or if there are imbalances that might reflect underlying biases. When use our data from above, we get the following results:
Positive Error Balance
Black: P (High Risk | Recidivism): 60/(60+6), 0.91
White: P (High Risk | Recidivism): 16/(16+22), 0.42
Negative Error Balance
Black: P (Low Risk | No Recidivism): 14/(20+14), 0.41
White: P (Low Risk | No Recidivism): 57/(57+5), 0.91
Now we can see a significant difference between the groups. For the Black group, the most frequent errors are false positives—being labelled as high risk when they are not. In contrast, for the White group, the most common errors are false negatives—being labelled as low risk when they are actually dangerous.
As we discussed earlier, this difference has a profound impact on individuals. Imagine being detained without justification—falsely labelled as a danger to society. Now compare that to being wrongly released when you actually posed a risk. The consequences, both emotionally and practically, are vastly different depending on which mistake is made.
So, COMPAS had a significant bias in the end: even though the accuracy rate was equal across group, the error rate balance was not.
Unfortunately, we cannot maximise accuracy and equal error rates at the same time:
-
-
Impossibility Theorem
It is mathematically impossible for an algorithm to meet all fairness goals at the same time. In the case of COMPAS, this means we cannot achieve both high accuracy for all groups and an equal ratio of false positives (wrongly labelled as high risk) and false negatives (wrongly labelled as low risk) for everyone. We must decide what to prioritise: overall accuracy or fairness in error rates.
This challenge is called the Impossibility Theorem.
It happens because of differences in what are called base rates in the data used to train the algorithm. Base rates show how often something, like reoffending, actually happens in each group. However, these rates aren’t always a reflection of individual behaviour alone—they can be shaped by structural injustices.
For example, if one group has a higher reoffending rate, it may not be because of personal choices alone. It could be linked to systemic issues, like biased policing, unequal economic opportunities, or access to education, which unfairly impact certain communities. Since AI systems rely on historical data, these injustices can end up embedded in the algorithm, making it harder to ensure fairness across groups with different base rates.
As a result, trying to meet one fairness goal (like equal error rates) may come at the cost of another goal (like accuracy), because of these built-in inequalities.
This means that the people who design algorithms face difficult ethical decisions. Under the conditions of our morally imperfect world, they must choose which fairness goals to prioritise, knowing that it is impossible to achieve a perfect balance while the data itself is shaped by wider inequalities in society.
-
-
-
-
In this episode, we delve into the complex world of work in the age of AI. We'll start with an opening case, exploring a business challenge in algorithmic scheduling. From there, we'll consider questions like "What is work?", examine the boundaries of AI, and reflect on two visions of the future of work and purpose. -
Business Challenge in Algorithmic Scheduling
To investigate some of the aspects mentioned so far, let us explore a fictitious but still realistic case, in which a café, call it ShiftMaster Café, faces a recurring challenge: managing unpredictable customer demand during peak times. Mornings and late afternoons see customer surges that strain staff capacity, leading to long wait times and customer dissatisfaction. Conversely, off-peak hours often leave employees underutilised, resulting in inefficiencies and missed opportunities for optimisation. This inconsistency not only affects customer satisfaction but also impacts employee morale, with lower ratings, lost revenue, and potential risks to job security.
Management sees an opportunity to leverage AI-driven shift management to dynamically adjust staff schedules in real-time, ensuring that team members are available when demand is highest. But how can they effectively make this vision a reality?
There are two options for algorithmic scheduling and job structuring. Explore the two options. Please explore them both by clicking on them.
-
Ethics by Design
In an ethics-by-design approach, deciding between Option A (Demand-Driven Scheduling) and Option B (Employee-Centric Scheduling) requires determining what the AI should optimise for—either operational efficiency or employee well-being. This choice is especially relevant for a reinforcement learning (RL) algorithm, where the AI's reward function will drive its learning.
With Option A, the RL algorithm could be designed to maximise revenue and efficiency by aligning staffing with fluctuating demand, though it might disregard employee preferences, risking burnout or dissatisfaction. Option B would instead prioritise employee autonomy and well-being, rewarding the AI for balancing operational needs with worker satisfaction, fostering a positive work environment but potentially limiting flexibility in high-demand moments.
This ethical choice shapes the AI’s long-term behavior: should it prioritise short-term profits or invest in sustainable, people-centered practices? This distinction not only defines the AI’s functionality but also determines its broader impact on the workforce, illustrating a key ethical decision in reinforcement learning design.
What would you do if you were the manager?
-
-
What is Work?
Let's take a step back. Before we dive deeper, we need to clarify what work is.
Work could be understood as any purposeful activity that involves effort, skill and time to achieve specific goals or produce goods and services. It may encompass a wide range of activities in different sectors and can be characterised by different dimensions, including physical, intellectual and emotional contributions. For the purpose of our episode, however, we may want to concentrate on a more narrow understanding, proposed by John Danaher:
“Any activity (physical, cognitive, emotional etc.) performed in exchange for an economic reward, or in the ultimate hope of receiving an economic reward.”
This definition has limits: it excludes things like domestic work. But it is wide enough to include things like internships (activity inn the hope of receiving an economic reward later) and focuses our attention on those areas where AI's impact will be significant.
The Impact of AI on Work
Building on this definition of work, it becomes clear that the advent of AI is fundamentally changing the nature of work. This transformation unfolds along three key dimensions:- displacement
- monitoring
- assistance
-
Let's quickly recap
-
Let's jump back to our fictional example.
After the successful implementation of the AI-powered scheduling system, ShiftMaster Café experiences improved efficiency and employee satisfaction. The management and staff are excited about the possibilities of further delegating tasks to AI, leading to brainstorming sessions on how to effectively split responsibilities between humans and machines.
They discuss the importance of defining clear roles for AI and human employees, ensuring that both can complement each other's strengths while maintaining a positive interaction. In a meeting, they now consider two further options for AI.
Let's look at the two options and the team's ethical analysis of them by clicking on the buttons below.´
-
Interim Conclusion
By applying an Ethics by Design approach to the development of BrewGenius and Smart Inventory, ShiftMaster Café can proactively address the ethical challenges associated with automation. This involves asking critical questions that focus on creativity, safety, employment impacts, and social responsibility, while also integrating specific development considerations that promote a balanced and ethical integration of AI technologies in the workplace. -
Work and the Quest for Meaning: Designing AI with Purpose and Autonomy in Mind
Work has long been a cornerstone of human identity, shaping our sense of meaning, purpose, and connection to the world. Through work, individuals often find purpose, achieve goals, and contribute to their communities. It serves not only as a means of economic survival but also as a way to express creativity, develop mastery, and engage in something larger than oneself.
As AI continues to transform the nature of work, understanding how people derive meaning from work becomes essential for designing AI systems that support, rather than detract from, human fulfillment, autonomy, and dignity.
-
