Artificial intelligence (AI) is a broad term for computer systems that perform tasks we associate with human intelligence, such as recognizing images, translating languages or answering questions. Understanding a little of how modern AI works makes its ethical questions much clearer.
Machine learning in plain words
Traditional software follows rules written by programmers: if this, do that. Machine learning works differently. Instead of writing all the rules, developers give a system many examples and let it learn patterns from them. Show a system thousands of photos labeled “cat” or “dog,” and it gradually adjusts millions of internal settings, called parameters, until it can label new photos it has never seen. Many modern systems are neural networks, layered mathematical structures loosely inspired by the brain. This approach can handle tasks too complex to write rules for. Its weakness is that the system learns whatever patterns are in its data, including errors and biases, and its reasoning can be hard to inspect.
Large language models
Large language models (LLMs) are the systems behind most AI chatbots. They are trained on enormous amounts of text to do one basic thing: predict what word, or piece of a word, is likely to come next. Done at huge scale, this simple task produces systems that can summarize documents, write code, translate and hold conversations. Developers then refine them with further training, including feedback from people about which answers are helpful and safe. Researchers still debate how far these systems understand what they produce.
Errors and hallucinations
Because an LLM generates likely-sounding text, and likely is not the same as true, it can produce statements that are fluent and confident but false, such as invented quotations, wrong dates or references to articles that do not exist. These are often called hallucinations. Lawyers in several countries have been sanctioned for filing court documents with case citations made up by chatbots. Developers have reduced error rates, for example by connecting models to search tools, but no current system is error-free. Checking important claims against reliable sources remains essential.
Copyright and consent
Language models and image generators are trained on vast collections of text and pictures, much of it gathered from the internet, including copyrighted work. Authors, artists and news organizations have filed lawsuits. Developers often argue that training is a transformative use, like a person learning from reading, and that models do not simply store copies of the works. Creators argue that their work is used without permission or payment to build products that compete with them. Courts and lawmakers in different countries are reaching different answers, and licensing deals and opt-out systems are developing alongside the legal fights.
Deepfakes
Deepfakes are realistic fake images, audio or video made with AI. They have legitimate uses in film and satire, but also serious harms: studies have found that a large share of deepfakes online are sexual images made without consent, cloned voices have been used in fraud, and fake clips can mislead voters. Legal scholars also describe a liar's dividend: as fakes become common, people can dismiss genuine evidence as fake. Responses include laws against harmful deepfakes, watermarks and content credentials that record where media came from.
Who is responsible?
When an AI system causes harm, responsibility is spread among those who collected the data, built the model, deployed it and used it. Philosophers call this the problem of many hands: when many people contribute to an outcome, accountability can fall through the gaps. Many ethicists argue that responsibility should be assigned clearly in advance.
Worked example: the history report
Maya, a college student, is writing a report on the 1918 influenza pandemic. She asks a chatbot for a summary and five scholarly sources. In seconds she has a polished page and a neat reference list. What should she do next?
Check the sources. Maya searches her library database for each reference. Three exist. One has a real author but a made-up title. One does not exist at all. The chatbot was not lying in the human sense; it produced references that looked like what usually follows such a request. That is a hallucination, and it is why important claims must be checked.
Check the facts. The summary says the pandemic killed exactly 50 million people. Maya finds that historians' estimates vary widely, commonly from about 17 million to 50 million or more, so she reports the range and cites a real source.
Follow the rules and be honest. Her college allows AI for brainstorming but requires students to disclose its use and write their own text. Maya notes how she used the tool. Passing off generated text as her own would be dishonest whatever the rules said.
Think about responsibility. If Maya had submitted the fake citation, who would be responsible? The developer built a tool that can invent references, and the college chose its policy, but Maya submitted the work. The problem of many hands does not erase her share.
Think about the people behind the tool. The chatbot's fluency rests on writing by historians and journalists. Finding and citing the original authors respects their work, whatever courts decide about training.