Bits are correlated: we care about the order of the bits, not just their population count.

Furthermore, <50% is only an antipredictor for a single coin flip. If you have two bits - two coin flips where you care about the order - then there's four values, and you have to be <25% to be an antipredictor. Even then, an antipredictor only removes one possibility[0], so flipping the bits will only slightly improve your terrible chances of recovering the data.

[0] https://en.wikipedia.org/wiki/Monty_Hall_problem

Unbelievably, the paper actually makes that claim:

> That is, if you toss a coin, you have a 50% chance of correctly choosing the value. In many instances, using a MFM to determine the prior value written to the hard drive was less successful than a simple coin toss.

More believably, the paper was written by Dr Craig Wright.

The article isn't super clear, but I think what it's saying is that the data recovery method they were testing (when looking at recovering a single bit) was as good as 92%, and as bad as 32%, depending on how the data was overwritten and other factors like the specific drive. Thus you likely can't just flip the result to get better odds because in a real scenario you wouldn't know which category you fall into (whether it's really good, or really bad), and overall if you don't know that information then the odds are close to 50%.

Edit: And of course even a high of 92% is abysmal for a single bit, since the combined odds of multiple bits quickly goes to zero for any sizable amount of data.

I don't think that follows? Just because a better method exists (the inverse) doesn't make the original perform better.

Imagine the following method: take whatever the bit was, and set it to the opposite of that. 0% < 50%. Of course you can invert and get 100%, but that's a different method. It doesn't make the original any better.