{
  "id": 3902040,
  "title": "Why AI Struggles to Guess Your Age — and Why That's a Bias Problem, Not Just an Accuracy One",
  "url": "https://urgent.news/2026/08/28/why-ai-struggles-to-guess-your-age-and-why-thats-a-bias-problem-not",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-28T05:28:41.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/theaidownside/why-ai-struggles-to-guess-your-age-and-why-thats-a-bias-problem-not-just-an-accuracy-one-742"
  },
  "original_language": "en",
  "account": "Artificial intelligence systems are increasingly being asked to guess a person's age without explicit information, such as a birth certificate. This process, known as age estimation, relies on machine learning models trained on large datasets of faces paired with corresponding ages. While these models can often make decent age predictions in the middle of the age range, they struggle at the edges that matter most, particularly for determining if someone is 18 or older. The quality of these guesses varies widely among different demographic groups, leading to significant errors—particularly near the ages where laws impose restrictions.\n\nThe two main approaches to age estimation are facial estimation, which attempts to predict age from someone's appearance, and behavioral inference, which infers age from how a user interacts with a platform. Both methods rely on statistical models that produce probabilities rather than definitive answers. Mean absolute error—a standard metric for assessing prediction accuracy—is often used to evaluate the performance of these models. However, even the most advanced facial systems, as demonstrated by Australia's 2025 Age Assurance Technology Trial, can have mean absolute errors of around one year in ideal conditions. Yet, the law is rarely concerned about the obvious cases; it focuses on the boundary ages, such as 13, 16, and 18, where estimation systems tend to be least reliable.\n\nThese errors do not distribute evenly across different demographic groups. Independent testing has shown that facial age estimation systems can have error \"buffer zones\" of roughly two to three years around each age threshold, particularly for non-Caucasian users, female-presenting individuals, and older adults. This demographic disparity in error rates is a clear example of algorithmic bias. The underlying data used to train these models may be biased or lacking in representation for certain groups, and even the way age is labeled and perceived can introduce bias into the models. This combination of poorly represented data and uneven evaluation contributes to the problem of bias in age estimation.\n\nThe consequences of biased age estimation can be significant. If a system systematically misjudges the age of certain demographic groups, it may disproportionately affect those groups, leading to unjust restrictions or additional scrutiny. For instance, at the ages where age verification is critical—such as accessing adult content or services—these errors can have real-world impacts, including denying access to services or triggering unnecessary identity checks. Moreover, when age estimation is applied to vulnerable populations like asylum seekers, the stakes are even higher. Misjudging someone's age in such scenarios can have life-altering consequences, exacerbating existing unfairness rather than mitigating it. Thus, while age estimation may not be an accurate replacement for age verification, its implications extend far beyond mere accuracy, touching on issues of fairness and equity.",
  "summary": "You are being asked, more and more, to prove your age to a machine. A porn site wants to know you’re 18. A social platform wants to know you’re 16. Now even a general-purpose chatbot wants to form a view about whether you’re a minor . In almost none of these cases does the machine actually know how old you are. It guesses — and the quality of that guess, and the way its errors are shared out, is…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}