I think its the score from 5,000 people though how they choose them I am not sure but its a fairly good barometer normally on how well an episode was liked. I have to admit I expected a low 80's score for episode 3.
Considering most surveys use a sample of closer to 1000 people, having a sample 5000 is actually statistically speaking very substantial, just ask anyone who studies statistics. I was quite surprised to see how large the sample is. By comparison, Nielsen Ratings used a 5000-strong sample prior to about 2007 (It's since substantially increased this number) - for a nation of about 300 million people. The UK has 60 million, so a 5000-person panel for the AI in the UK, percentage wise, dwarfs what Nielsen and most survey organizations in the US do.
I happened to like the pirates episode, so I'm glad to see it score in the high 80s. Incidentally, 85 is considered "excellent" and people tend not to get their knickers in a twist unless an episode scores below 70. Which I think has only happened once or twice since DW came back on the air.
Don't Scare the Hare scored a 46 on its first episode, which is a disaster.
Alex