Machine learning is a subfield of artificial intelligence, but really what that means is a very sophisticated method of pattern analysis. It is a process where computers use algorithms and models to make data-driven decisions and inferences, without explicit instructions. Instead of being specifically programmed, the system 'learns' from patterns in the training data to predict outcomes, like which donors will respond to a direct mail appeal.
So how is this different to current approaches? The short answer is that machine learning is much better at predicting donor behaviour.
Most not-for-profits use outdated rule-base methods or 'segmentation' techniques like RFM (recency, frequency, monetary value) analysis to generate campaign lists for their appeals. These processes are useful for clustering donors together based on common traits, such as recent giving or the time elapsed since acquisition, but have many limitations. For instance, they only consider the recency, frequency and value of gifts, but ignore other useful data such as communications histories, demographics, activities, etc. They also are not predictive, are difficult to test, and don't allow you to treat donors as individuals. If you want to include one donor from a segment, you have to include the whole segment! This is expensive and inefficient.
With machine learning, we can identify much more nuanced patterns in very complex data sets and to generate projections about how individual donors are likely to behave. In essence, rather than simply using a limited amount of information, machine learning takes into account everything you know about each donor and identifies patterns in that data to predict future behaviours with much greater accuracy. This allows you to run more targeted campaigns with a better chance of maximising returns.