That statement refers to the fact that you loose (automatic) referential integrity when you start using Immutable data.
Take the following scenario: you have an app with Users and Tasks. Tasks can be assigned to users. When you express this using immutable data you have two problems. The first one is linking tasks to users. If you would link Task1 to Person1 and after that change the name of Person1, you would get a new person, Person1v2. However Task1 would still refer to the old, unmodified Person1, so then you have suddenly two versions of the same person in memory. A fresh and a stale one. In other words, you lost referential integrity.
Now the usual way to fix is, is to not use real references, but to normalize data and store only a key to the person. The effect of this is that you have to always lookup the person related to Task1 again from the state tree. If you have a reference (variable) to that person somewhere, make sure it is short-lived because you won't know if you are still referring to the correct one.
The second problem is that when you have correctly normalized your data and always do fresh lookups, you will still not know _when_ to make these lookups. If you have a TaskView that renders the name of the related Person, and the related person changes, the TaskView will not detect this automatically. The related person is not even a prop of the TaskView component so pushing a new state tree through your app won't repaint the TaskView if you are using PureRenderMixin (which is usually the purpose of using immutable state trees in the first place). For that issue there is a solution as well; you can introduce selectors (or lenses) that query the state tree to detect if the relevant person is changed. So now we already have three new concepts (no refs, normalization, lookups, selectors) to solve a problem that mutable data doesn't have in the first place.
By using mutable data and observing it using MobX these problems are solved (imho) more elegantly: references are never stale because if you would modify Person1, Task1 would still refer to the 'latest' person1, because it is still the same object. Secondly, because it is the same object, fine grained object observers can be established automatically for you, making sure the TaskView gets updated as soon as something relevant changes. This means that you have less concepts to learn and maintain (no copy-on-write, no need to assign a unique key to everything, no data normalization, no lookups, no selectors to make sure your views are always in sync with the state). That saves a significant amount of boring yet error prone work when your application's state model becomes more complex than the average Todo app.
So that is the long story behind that short comment. I hope it clarifies the statement!