- Explore why AI ethics are essential for responsible innovation.
- Understand common ethical challenges such as bias, privacy, transparency, etc.
- Understand Responsible Research and Innovation (RRI) framework and how it can be integrated your project’s ethical plan
- Learn ways to ensure AI systems remain fair, accurate, and accountable.
- Reflect on how ethical thinking can be built into your own AI project from the start and consider it an ongoing process.
Ethics in AI
As AI becomes more powerful, we must understand the scope of the tools we have and think about the wider impact of what we create. AI can be unfair, make mistakes, and, if misused, cause harm. This is why we, as developers and entrepreneurs, need to approach our work with thoughtfulness.
When we hear the word “ethics,” it might sound a bit formal or complicated, but it’s really about asking questions.
“Is this fair? Who might this help, and who might it hurt?”
The bottom line: AI is a technology created by humans, along with our biases on what actions should be prioritized, what data should it learn from, and what is defined as success. Even with the best of intentions, AI is inherently imperfect. The key is to recognize where it can go wrong and use it responsibly.
AI Ethics in Your Project
While you are in the early phases of your project, it is not too early to begin thinking about how ethics will play into your final solution. Here are some principles you can use as a guide to think deeply and specifically about the ethics of AI projects, and your project in particular. Read through and reflect on the questions as they relate to your project.
Do No Harm
Simple but powerful. Even though most creators never intend to cause damage, the impact of AI often extends beyond what’s expected. A system might make decisions or recommendations that seem helpful on the surface but, when scaled, could unintentionally disadvantage or mislead people.
“Do no harm” means continuously asking how your technology might go wrong, who might be affected, and what systems are in place to minimize risk.
Example: Community Safety App Gone Wrong
❌ The Problem
A team develops an AI-powered community safety app that alerts users to “potentially unsafe situations” nearby, using data from social media and local reports. The goal is to make neighborhoods safer by helping people stay aware and avoid risk.
But users and community groups notice a troubling trend: the app sends more alerts for neighborhoods with larger marginalized populations, even when no incidents occur. Trained on biased historical data and language patterns, the AI learned to associate certain areas and words with danger.
Though the developers didn’t intend to reinforce stereotypes, the harm was done—the app stigmatized communities and fueled social division and mistrust.
✅ Better Approach
This team involved community members from diverse neighborhoods in the design process—not just data scientists. They tested their model across demographic areas and measured for bias.
Instead of relying on social media sentiment, they used transparent data sources like official police reports. Each alert included an explanation (e.g., “Alert based on: 3 reported break-ins this week”) and a feature for users to flag false or biased alerts.
As a result, the app improved community awareness WITHOUT stigmatizing neighborhoods. When bias was detected (alerts concentrated in certain areas), they investigated and corrected the data and updated the model.
- What specific, built-in systems are in place to detect these failures quickly and minimize the resulting risk?
Bias and Fairness
AI is ultimately a human creation and can reflect the assumptions and priorities of its creators. If the data used to train AI is collected through a biased lens, it can reinforce incorrect assumptions, and results can be skewed in ways that are difficult to anticipate.
Example: Bias in Job Recruitment
In this study, it was discovered that AI tools meant to ease the processing of job applications had biases relating to perceived race and gender. The AI tools here heavily favored Caucasian male-associated names. With the immense adoption of AI across the hiring process, this discovery throws into question the fairness of these tools.
Addressing bias requires awareness, reflection, and intentional action. Users need to be critical and skeptical of what AI produces. Developers must examine the data they use, question the assumptions behind their models, and test outputs for fairness across different groups. Sometimes this means bringing in diverse perspectives during design, auditing models for disparate outcomes, or adjusting algorithms to correct for imbalance.
- What specific steps will be taken during the data collection and labeling phase?
- How will you ensure that diverse perspectives are meaningfully integrated into the design, testing, and auditing phases of the AI model?
Privacy and Data
AI often relies on large amounts of personal data to function effectively. While this data helps AI provide useful recommendations or services, it also raises important ethical questions about privacy.
For instance, an app that tracks user activity to suggest local services might inadvertently expose sensitive details about a person’s routines, preferences, or beliefs. These details could be misused if they fall into the wrong hands, or even subtly influence how the user is treated by algorithms elsewhere.
-
Collect only what you need
Limit data collection to the information necessary for your app to function.
-
Obtain clear consent
Ensure users understand what data is collected and why, and offer the option to opt in or out.
-
Anonymize data when possible
Remove personally identifiable information to reduce risk if the data is exposed.
-
Store data securely
Use encryption, secure servers, and strong access controls to protect user information.
-
Be transparent
Clearly communicate what data is collected, how it’s used, and who can access it.
-
Give users control
Allow users to review, edit, or delete their data.
-
Consider downstream consequences
Think about how collected data could be used beyond your app and any potential impacts.
- How will you design the user interface and consent process to ensure users give truly informed consent?
- What layered security measures will you implement to protect the data?
Accuracy and Human Oversight
AI can be incredibly powerful, but even the most sophisticated models can make mistakes. One example is “AI hallucinations”, where the system confidently produces information that is entirely false or fabricated.
These errors can occur because the AI is trained on incomplete data, built on flawed assumptions, or biased toward generating outputs that satisfy the user, even if they’re incorrect. This is why human oversight is essential. AI outputs should be approached with a mindset of critical engagement: evaluating, questioning, and contextualizing what the system produces.
Instead of treating the AI as an infallible authority, it should be treated as a tool that augments judgment, providing insights while leaving final responsibility in human hands. For example, imagine an AI system that generates summaries of medical articles for a healthcare app. If the AI hallucinates details not found in the source material, users could make decisions based on false information. A human reviewer who catches and corrects these errors can prevent serious harm before it affects real people.
- Who will check the outputs?
- How will mistakes be flagged and addressed?
Transparency and Explainability
Both the people building it and the people using it should be able to understand how the AI made a particular recommendation or decision — the logic of the algorithms, how the AI was trained, how data was processed and labeled, and how the model’s performance is evaluated.
As AI is increasingly being used for high-stakes decisions (such as hiring, medical diagnoses, loan approvals, or investment recommendations), inaccurate or biased outputs can have serious consequences, affecting careers, finances, or even health and safety.
You could use simple visualizations, short text explanations, or interactive demos to show how your system works. The key is to give users enough insight so they can make informed decisions, ask questions, and understand the reasoning behind your AI’s outputs.
- How can you make your AI’s decisions clear and understandable to your users?
Accessibility and Inclusion
AI should work for everyone, including people with disabilities, people who don’t speak English, people without high-speed internet, and people unfamiliar with technology.
- Will there be visual interfaces that might exclude blind users?
- What about audio input that might exclude deaf users?
- Do your users have the technical knowledge to easily navigate and use your product efficiently?
- Will you take into account non-English speakers who use your app?
Environmental Impact
While AI is amazing and can transform people’s lives, it also comes at a cost. Training AI models requires significant computational power, which affects the environment. Even though your initial project may not significantly affect the environment, you should still consider this issue could come to the forefront as your project scales up.
945 TWh by 2030
By 2030, data centers powering AI are projected to use around 945 terawatt-hours of electricity — more than Japan’s entire annual consumption. According to the International Energy Agency, 60% of that energy will still rely on fossil fuels.
- How might you use existing machine learning models instead of training your own?
- How might you optimize your model and use less data?
Labor Impact
AI automates many tasks previously performed by humans. Your project could potentially cause unintended harm by causing people to lose their jobs and their income.
- Does it replace human jobs, and if so, what will happen to those workers?
- Who will benefit economically, and who might become disadvantaged by it?
Regulations You Should Know
Ethical design goes beyond what is technically legal to ensure your product is fair and trustworthy. However, there are regulations and laws you must follow when launching new technologies.
If your AI handles personal data:
- GDPR (Europe): strict privacy and data protection rules
- CCPA (California): consumer privacy rights
- COPPA (US): children’s data protections
- HIPAA (US): healthcare data protection
If your project uses AI:
- EU AI Act (2024): Categorizes AI systems by risk (low, high, unacceptable). High-risk systems—used in hiring, education, healthcare—must meet strict transparency, testing, and documentation rules.y
- U.S. AI Bill of Rights: Not law yet, but outlines principles like protection from algorithmic bias, data privacy, and human alternatives for critical decisions.
- Other frameworks: Canada’s AI and Data Act, Singapore’s Model AI Governance Framework, and OECD AI Principles all stress accountability, fairness, and explainability.
- Intellectual Property (IP): Use caution with generative AI that may produce copyrighted content. Avoid using copyrighted materials in training datasets for your own models.
Being an ethical creator is a journey
AI is powerful, but it is never perfect. Models evolve, learn from new data, and interact with users in ways that can produce unexpected outcomes. This means that ethical considerations are not a one-time checklist but instead require continuous attention.
Remember, the goal isn’t just to build an AI solution that works today — it’s to build one that continues to work fairly, safely, and beneficially tomorrow and beyond.
You and your team should always be asking: “What can we do now to make our product even more responsible?” Your values or the context might change over time, so you need to do this reflection continuously.
Keep these tips in mind:
- Keep learning: Improve your skills, ask questions, and think about how to make your product more responsible.
- Always aim to do good: It’s not just about avoiding bad things — it’s about having an intentional mission to help your community in a kind and fair way.
- Make it part of the plan: Set aside time and energy for ethics, just like you would for designing your product’s look or writing the code.
- Have a Team RRI Leader: Someone to remind everyone to think about being responsible, especially when you’re all busy and excited about new ideas.
- Seek support: Don’t be afraid to ask for help from peers, mentors, users, or professionals if you question whether something is ethical about your business.
ACTIVITY 1
Your Continuous AI Ethics Action Plan
Estimated Time: 30 minutes
Building ethically is an ongoing journey. Use the questions in the worksheet to think about how you’ll manage ethical concerns in your project.
Remember to stay curious! The AI world is always changing, and being a great innovator means always being a learner. You’re now equipped to not only build something amazing but also to guide it ethically into the future.
Additional Resources
In this talk by Professor Stuart Russel, he dives deeper into the ethics of AI and further highlights the importance of responsible design, oversight, and consideration of societal impacts.
