Common words about how AI can be unfair in unexpected ways

Smart computer programs that write human-like text are now everywhere. These programs learn from billions of sentences found online, so they should reflect how people really speak. But when researchers tested four advanced versions, they found the programs kept making the same mistakes over and over. The first mistake is called stereotype bias – basically, the programs decided certain jobs, hobbies, or beliefs belong to specific groups of people. When asked to describe a doctor, the program often used words like “he” or “she, ” while a nurse was usually called “she. ” When asked about religion, the program linked some faiths to specific countries more often than it should. These automatic links aren’t always true, yet the program kept repeating them.The second mistake is called deviation bias – this means the picture the program paints of society doesn’t match reality. If you ask the program to generate 100 random profiles, you might expect to see job titles, ages, and backgrounds that look like the real population. Instead, the programs created profiles that were too young, too male, or too concentrated in certain parts of the country. This mismatch matters because these programs are now used to help decide loan approvals, job candidates, and even prison sentences. Even the most advanced programs couldn’t shake these habits. Each one kept favoring the same groups and showing the same gaps. That shows the problem isn’t just in the training data – it’s also in how the programs are built to guess missing information.