What is item discrimination?

Item discrimination, in simple terms, indicates how well a question is able to differentiate between test-takers who have a high(er) IQ and those who have a low(er) IQ. A highly discriminatory question is one that only those with a high IQ are able to answer correctly, while those with a lower IQ would be likely to get it wrong.
The item discrimination can be seen in the Item Characteristic Curve (ICC). The ICC shows the chance that someone with a certain IQ will give the correct answer to the question. If the chance that someone with a higher IQ gives the correct answer is a lot higher than for someone with a lower IQ, the curve is very steep. The item discrimination can be seen by looking at the slope of the item characteristic curve.
In the ICC above, it can be seen that someone with an IQ of 120 has a much higher chance of correctly answering the question than someone with an IQ of 110. This means that the question discriminates highly between people with an IQ of 120 and an IQ of 110. If someone gives the correct answer to the question, there is a high chance that their IQ is higher than 110. If someone gives the incorrect answer, there is a high chance that their IQ is below 120.
On the other hand, the question does not discriminate well between people with an IQ of, for example, 100 and 80, because both have a chance of around 20% to get the correct answer. This is because the question is too difficult for people with an IQ below approximately 110. There is no use in asking this question to someone with an IQ below 110, because it does not provide much information about the precise IQ of this person. It is better to ask an easier question that discriminates well between people with, for example, an IQ of 100 and 110. The IQ level at which discrimination is highest is referred to as the Point of Maximum Information.
It is possible that a question does not have a steep ICC. For example, if we hypothetically included a question such as "What is the most popular song of Mariah Carey" in an IQ test, there would be a percentage of people who get the correct answer ("All I Want For Christmas Is You"). However, the likelihood that someone with an IQ of 130 would give the correct answer is not much higher than that of someone with an IQ of 100.
The ICC of that hypothetical "IQ test question" would most likely look similar to the graph below, which shows that the question is not discriminating well between people with different IQ levels.
There is a percentage of people who do not know the answer, but whether or not one knows the answer does not say much about their IQ. Therefore, it would be ineffective to include this question in an IQ test. This is an extreme example, but at BRGHT, many questions are eliminated because they do not have enough discriminatory power to be useful in measuring IQ, or there are better options available that are more effective at measuring IQ.