The defining feature is a schedule that declines to zero. A back-end load is usually structured to fall by roughly one percentage point for each year the shares are held, reaching zero after a set number of years. A typical schedule might charge 5% for a sale in the first year, then 4%, 3%, 2%, 1%, and nothing after the fifth or sixth year. The effect is a penalty for leaving early that fades to nothing for an investor who stays, which is the behavior the charge is designed to encourage. Because the schedule and its length are specific to the share class, both need checking in the prospectus before relying on either.
It is charged on the lesser of two amounts, which usually spares your gains. The SEC states that a back-end load is typically calculated on the lesser of the value of the initial investment or the value of the investment at redemption. So a position that has grown is generally charged against the smaller original figure rather than against the appreciated value, and only a position that has fallen would be charged on the lower current value. This is a meaningful detail, because it means the charge is normally computed on what you put in, not on what the investment became.
B shares typically convert, which caps how long the higher fee lasts. Back-end loads are the signature of the traditional B share class, which also carries a higher annual 12b-1 distribution fee. Because a B share's back-end load falls to zero after the stated period and the class then usually converts automatically to the lower-cost A share, the higher ongoing fee is not permanent. That conversion is the mechanism meant to make a B share defensible for a patient holder, and it is also why a B share should never be evaluated on its current year's fee alone.
The cost you cannot know in advance is the point of difference. A front-end load is fixed and immediate; a back-end load depends on a decision you have not made yet, namely when you will sell. An investor who is confident of holding past the schedule may pay nothing, while one forced to sell in a bad year pays the most exactly when it stings. That contingency is a difference in kind from the front-end version, not merely in timing, and it is the reason the effective cost of a back-end load is genuinely unknown at purchase.