Learn AI / Compare /Bias vs variance (and neither is discrimination)

Bias vs variance (and neither is discrimination)

Why does "bias" mean three different things?

The short answer

Statistical bias means a model is systematically off in one direction; variance means it swings wildly with small changes in the data. Neither is the same as algorithmic bias, which is about unfair outcomes for people. The word is overloaded, so always ask which one is meant.

Bias (statistical)

Model Quality and Evaluation · Deeper

In one sentence: Statistical bias is a systematic gap between a model’s expected prediction across possible training samples and the true target, often caused by restrictive assumptions.
Why it matters: A model can produce tidy scores and still fail in a predictable way because it is too simple for the real problem. If you do not recognise statistical bias, you may choose a system that looks stable but keeps missing important patterns. That can lead to poor decisions about procurement, staffing, risk, or service quality. For non-specialists, the key question is not only ‘How accurate is it?’ but also ‘What kinds of mistakes does it make again and again, and why?’ Understanding bias helps you ask whether the model is underfitting, whether a benchmark is too easy, and whether a more suitable design or richer data is needed.
The everyday version: A high-bias model is too simplistic, so it keeps making the same kind of mistake.
The one thing to remember: Statistical bias is systematic error in the expected prediction across possible training samples. It often appears when restrictive assumptions keep the model from representing the true relationship.
The commonest misconception: Statistical bias means unfair discrimination against a social group. That is a different concept, often called algorithmic bias or fairness bias. Statistical bias is about systematic prediction error caused by model assumptions or estimation procedures. The two can interact in practice, but they are not the same thing.

Read the full page

Variance

Model Quality and Evaluation · Deeper

In one sentence: Variance describes how much a fitted model’s predictions change when the same learning procedure is trained on slightly different samples.
Why it matters: A model can look strong because it scored well on one test, yet still be unreliable in practice if that score changes a lot with small data changes. Variance helps you ask a better question: not only “How high is the score?” but also “How stable is it?” This matters when choosing whether to deploy a system, compare vendors, set monitoring thresholds or trust a benchmark result. For non-specialists, variance is a warning sign against over-reading a single number. Stable performance is usually easier to defend, monitor and improve than performance that depends on a lucky split.
The everyday version: Variance means the model’s results do not stay steady when you test it on slightly different data.
The one thing to remember: A strong average score is not enough. If refitting the same procedure on slightly different samples changes predictions substantially, the model has high variance and may not generalise reliably.
The commonest misconception: Variance just means the model makes many mistakes. Not exactly. A model can have a reasonable average error rate and still have high variance if that error rate changes a lot across different samples or runs. Variance is about instability, not simply the number of errors.

Read the full page

Algorithmic bias

Responsible AI, Human Rights and Safety · Starter

In one sentence: Systematic unfair patterns in data, rules or model outputs can disadvantage some people or groups, especially when tools are used in important decisions.
Why it matters: Bias matters because AI tools can shape access, attention and opportunity without making the risk obvious. A system may look efficient overall while still failing certain groups more often. For a non-specialist, the key question is not only 'Does it work?' but also 'For whom does it work, who is missed, and who checks the result?' In public services, humanitarian work, hiring, education, health, finance or online platforms, unequal errors can create real harm, weaken trust, and expose an organisation to legal, ethical and reputational risk. Good practice means testing for uneven impacts, documenting limits, keeping meaningful human oversight, and being ready not to use the tool when the risk is too high.
The everyday version: An AI system can treat people unfairly because of the way it was designed, trained, or used, even if no one intended that result.
The one thing to remember: Good average performance can conceal systematic harm. Examine the data, labels, proxy variables and thresholds, compare errors across relevant groups, and keep people accountable for consequential decisions.
The commonest misconception: If the system does not use a protected characteristic directly, it cannot be biased. A system can still produce unfair outcomes through proxy variables, such as postcode, school, language pattern or browsing behaviour, that correlate with protected characteristics. Bias can also come from missing groups, poor labels, or unequal error rates even when the same formal rule is applied to everyone.

Read the full page

Bias (in a neural network)

Neural Networks · Deeper

In one sentence: A bias is an extra adjustable number added to a neuron's weighted input so the neuron's output can shift even when all inputs are zero.
Why it matters: If you understand bias as an offset parameter, you can better interpret why a neural network responds too often, too rarely, or at the wrong threshold. That matters when choosing models, checking errors, or discussing calibration and behaviour with technical teams.
The everyday version: A bias is the neuron's learned starting value. It shifts the result up or down before the neuron produces its output.
The one thing to remember: Weights determine how each input changes a neuron's score. The bias is a learned offset that shifts where the neuron activates, including when every input is zero.
The commonest misconception: Bias means the system is unfair. In this context, bias is a mathematical parameter inside a neuron. Unfairness is a different issue, often discussed as algorithmic bias, though model parameters can still contribute to harmful outcomes.

Read the full page