Fill the new edges with a blur of your own photo

Nothing, which is the point. The blur happens to a second copy, drawn underneath, and your photograph is put on top of it afterwards at one pixel to one pixel. Nobody looking at the result is looking at a blurred version of your picture — they are looking at your picture, in front of something soft made out of the same colours.

It works because it answers the question the eye asks about a frame that is the wrong shape. Black bars say "this picture did not fit". A blurred continuation of the same scene says "this is the frame, and the picture is in it". Every social platform has spent a decade teaching people to read it that way, which is the only reason it feels like a native part of a photograph rather than an effect.

A blur is the most forgiving of the five fills and it is still not universal. A photograph with one strong subject against a plain background gives a backdrop that is essentially one colour, and at that point a flat fill would have been simpler and sharper. A photograph that is mostly white gives a backdrop that is mostly white, and your picture floats on it with no edge at all unless you use the tone control.

The other thing worth knowing is what it costs in file size. A soft backdrop compresses very well, so the extra pixels are cheaper than they look — but there are three times as many of them on a 9:16 to 16:9 conversion, and the estimate under the Output panel will tell you what that comes to before you commit.

Where does the blurred part actually come from?

From your photograph and nothing else. A copy of it is enlarged until it covers the whole target canvas — cropping whatever hangs over the sides — then softened until no detail survives, then optionally taken up or down in tone. Your original is laid over the middle of that at its own size and full sharpness. So the colours behind the picture are the colours in the picture, which is exactly why the effect sits so comfortably.

What does the radius slider really change?

How far each pixel's colour is smeared into its neighbours before the backdrop is drawn. Low values leave recognisable shapes behind the photograph, which usually looks like a mistake rather than a choice. High values reduce the backdrop to soft fields of colour, which is what you want. The tool picks a starting value from the size of the canvas — a fortieth of the short edge — because a radius that suits a small picture is invisible on a large one.

Why does the tone slider exist?

Because a backdrop made from a bright photograph can be almost as bright as the photograph, and then the edge between them vanishes and the whole thing reads as a smudge. Taking the backdrop down ten or fifteen percent puts a little air around your picture without adding a border. Going the other way, lighter, does the same job for a dark photograph heading onto a white page.

Is this blur the same one my browser does?

No, and deliberately so. Canvas has a built-in blur and it is faster, but the three major browser engines disagree about what it does at the edges of a surface, which shows up most on exactly the wide soft backdrop this page spends its time drawing. So the blur is computed here instead, as three box passes on a small proxy. That is close enough to a Gaussian for a backdrop nobody is meant to study, and being plain arithmetic rather than a browser feature, it comes out the same wherever it runs.