To quickly be able to assess what consumers might do in a host of future scenarios, we need to simplify their behavior down to a mathematical model. The multinomial logit (MNL) model was chosen because it captures decisions patterns very well.
However it carries with it a simplifying assumption called IIA.
Here is the brief idea. You have 2 options to get to work. Car which you value at a utility of 5, and bus which you value with a utility of 5. In this scenario, the odds of driving your care are (1/2,1/2).
But one day your town launches a new line of blue buses to complement the red ones you are used to. You don't care about the color so your utility for the 3 options are all 5.
The dumb thing is that according to the MNL with 3 options your odds of each are now (1/3,1/3,1/3). Even though you would would treat the red and blue buses as a single options with a joint utility. The model doesn't know this.
This was a big deal in the early days of choice modeling, but not so much any more.