Quick answer: The biggest ethical issues in AI are algorithmic bias, gender stereotyping in chatbots and digital assistants, misuse of AI for profiling and surveillance, and autonomous weapons. Governments are now acting: the EU AI Act is phasing in, and UN member states are debating rules for lethal autonomous weapons.
Key takeaways
- AI learns from data, so biased data produces biased decisions, from hiring and lending to healthcare.
- Chatbots and voice assistants can reinforce gender stereotypes through the personas they are given.
- Capable AI can be misused for profiling, and the same technology powers both self-driving cars and weapons.
- Regulation is arriving, but unevenly. Businesses building AI today should plan for it now.
As ventures for profit, it is hard to guess how companies can be made to hold ethical standards. Technology moves faster than the law, and companies with deep resources can often find loopholes. Products designed with good intentions can still have unintended consequences. In the rush to build tech that is more appealing, and therefore more profitable, ethics is often the last thing the creator or the consumer thinks about.
This is a four-part series on the ethical grey areas the tech industry has faced, is facing and will face. We begin with artificial intelligence. When you finish, continue with Part 2, which looks at tech and authoritarian regimes: the Great Firewall of China, internet shutdowns and facial recognition.
What are the ethical issues of AI?
Debates about AI range from the technical (is strong AI possible, as in the Chinese room problem?), to the philosophical (the trolley problem), to the political (will machines take our jobs?). Advances in AI keep adding new ethical questions to the list. The most pressing today fall into four groups: bias, gender stereotyping, misuse and warfare.
How does AI bias happen?
AI systems are trained on data and then apply what they learned to similar data. If the training data is flawed, the system is flawed too. This is the “garbage in, garbage out” principle. A well-known early example: researchers fed IBM Watson the entire Urban Dictionary, and had to delete what it learned because the system started swearing. That one was funny. The consequences are not funny when AI is used to make decisions about people.
A widely cited 2019 study in Science found that a healthcare algorithm used across the US underestimated how sick Black patients were compared with white patients at the same risk score. Hiring tools and credit-scoring models face the same risk, because the historical data they learn from reflects historical discrimination. Systems meant to remove human bias can end up amplifying it.
You can see this yourself. Try the word2vec demo with different professions and watch it follow traditional gender roles. Representation gaps matter too. Early commercial cameras with face detection sometimes read Asian faces as blinking, and the 2018 Gender Shades study found commercial facial-analysis systems made far more errors on darker-skinned women than on lighter-skinned men. Improving diversity in data and teams is not only a moral issue; it is how you build AI that works.
What changed since 2020: generative AI. Large language models and image generators are trained on web-scale data, so they can reproduce the same stereotypes at much greater volume. Regulators have noticed. Under the EU AI Act, AI used in areas such as hiring, education and credit scoring is classed as high-risk. The Digital Omnibus, which entered into force in July 2026, pushed those obligations back to 2 December 2027 for standalone systems (and 2 August 2028 for AI built into regulated products). It delayed the deadline, not the requirement.
Do chatbots and voice assistants perpetuate gender stereotypes?
Often, yes. Most digital assistants have historically been given female voices and personas. Because their job is to serve, this can reinforce stereotypes about women in service roles. A 2019 UNESCO report raised exactly this concern. The choice is usually based on market research, which means it may simply mirror society, but as assistants become more human-like, the effect compounds. Steve Worswick, creator of the Loebner Prize-winning chatbot Mitsuku, discussed the ethics of giving his bot a female persona and noted that many users treated it as a living person.
Chatbots also reflect the behaviour of the people who talk to them. Microsoft’s Twitter bot Tay learned from trolls and began posting racist and sexist tweets within a day of launch. Users regularly harass chatbots too, which tells us something about human nature that designers must plan for.
What changed since 2020: many assistants now offer a choice of voices, and AI companions and agents are far more widespread. A related rule is now arriving in law. Most of the EU AI Act’s transparency obligations (including telling people when they are interacting with an AI system) apply from August 2026.
When capability outruns ethics: self-driving cars and profiling
AI can spot patterns humans miss, never tires, and learns from more data than a person could review in a lifetime. Self-driving cars show the scale. In 2020 we cited Waymo’s billions of simulated miles. By March 2026, Waymo reported about 220 million fully driverless “rider-only” miles on public roads and roughly 500,000 paid rides per week, with company data claiming far fewer serious-injury crashes than human drivers over the same distance. These are Waymo’s own figures, but they show why many believe self-driving cars will eventually be safer than human drivers. The ethical questions remain: who is liable when one fails, and who decides how it should behave in a dilemma?
Capability also creates room for misuse. In 2017, researchers published a study claiming AI could infer sexual orientation from photos. It was presented as a warning, but it was widely criticised for its methods and showed how such tools could be used to profile and expose people. Facial recognition has since become a flashpoint worldwide, which we explore in Part 2. The lesson is simple: when a system can classify people, someone will eventually use it against them, and there is not much to stop them.
Is AI in warfare an ethical problem?
Many think so. Systems that identify targets and guide munitions already exist, and the US Navy’s autonomous vessel Sea Hunter and Israel’s Harpy loitering munition are well-known examples. Today most systems keep a human operator in the loop before force is used, but there is no guarantee future ones will. Project Maven, the Pentagon’s AI drone-vision programme involving major tech firms, triggered an employee backlash at Google in 2018 and became a defining case of workers challenging military AI work.
What changed since 2020: the debate has moved from the margins to the UN. In December 2024 the UN General Assembly backed a resolution on lethal autonomous weapons by 166 votes to 3. In early September 2026, nearly 130 countries agreed on a text defining lethal autonomous weapons systems, a first step toward a possible treaty, and member states are due to decide at the Convention on Certain Conventional Weapons meeting in November whether to begin negotiations. Reporting indicates that the US and Russia pushed to remove language on human review of AI-selected targets from the draft, so the outcome is far from settled. Whether humans must stay in the loop is now the central question.
What does this mean for businesses building AI today?
You do not need to build weapons to face these issues. Any company deploying AI in hiring, lending, healthcare, customer service or operations should:
- Audit training data for gaps and skew before training, and test outputs across groups.
- Keep a human in the loop for decisions that affect people’s rights, money or safety.
- Be transparent when users are interacting with AI.
- Protect data with strong security and clear governance.
- Track regulation such as the EU AI Act, whose timelines have just shifted but whose obligations remain.
Frequently asked questions
What are the main ethical issues in AI? The main issues are algorithmic bias, lack of transparency, privacy and surveillance, gender and other stereotypes in chatbots, misuse of AI for profiling, and autonomous weapons.
How does AI bias happen? AI learns patterns from training data. If that data underrepresents some groups or reflects past discrimination, the model reproduces and can amplify those patterns in its decisions.
Is the EU AI Act in force? Yes. It entered into force on 1 August 2024 and applies in phases. Rules on prohibited practices and general-purpose AI already apply, and the high-risk rules were delayed by the Digital Omnibus to 2 December 2027 (standalone systems) and 2 August 2028 (AI embedded in regulated products).
Are autonomous weapons banned? Not yet. UN member states are negotiating how to define and regulate lethal autonomous weapons, but there is no binding global treaty as of November 2026.
How can companies reduce AI bias? Use representative data, test models across user groups, keep human oversight for high-impact decisions, and monitor models after launch.
Build AI you can trust
Ethical AI is built, not declared: it starts with the right data, sound engineering and security from day one. If you are planning an AI product, Nuventure is an AI development company that helps businesses design and build intelligent solutions, with ISO 27001 and SOC 2 certified processes. If your use case depends on training models on your own data, explore our machine learning solutions. Talk to us about AI development services for your next project.
This is the first part of a four-part series. Continue to Part 2: tech and authoritarian regimes.
