An option's price depends on the strike, the time to expiry, the underlying price, rates, and how much the underlying is expected to move. All of those except the last are observable. So the convention is to invert the relationship: take the traded option price, assume a pricing model, and solve for the volatility that makes the model reproduce that price. That number is the implied volatility.
This makes implied volatility a quoting convention as much as a quantity. Traders quote options in volatility terms rather than in dollars because volatility is comparable across strikes and expiries in a way that premium is not: a fifty-cent premium means nothing without knowing the strike and the tenor, whereas 30 percent vol is immediately interpretable.
The critical consequence is that implied volatility is model-dependent by construction. Change the model and the same option price implies a different volatility. The market has settled on Black-Scholes-style conventions for quoting precisely because everyone agreeing on a convention matters more than the convention being correct, and the model's known deficiencies then show up as structure in the surface rather than as pricing errors.
Implied volatility is also a price, which means it responds to supply and demand for options, not only to expectations about movement. A large hedging programme that needs to buy downside protection will lift the implied volatility of those strikes regardless of anyone's forecast. Reading a rise in implied volatility as new information about future movement, when it was a flow, is a routine error.
The systematic gap between implied and subsequently realised volatility is the variance risk premium. Selling volatility has historically earned it, and the reason is structural: the seller takes an asymmetric loss profile and requires compensation. That it has historically paid is not the same as it being a free return, and the losses arrive concentrated in exactly the episodes when everything else is also losing.