Where error bounds are stated, they are generally given for a 95% confidence interval, by convention.

Errorbars are usually 68% (1-sigma), unless stated otherwise. In physics, confidence intervals are usually 95% (2-sigma), though 90% is becoming increasingly common.

I was talking about error bounds expressed as a confidence interval, e.g., in political polling. Not error bars used in various plot types (e.g., box or whisker plots), which are based on standard deviation.

As I noted, it's a convention, and really, the CI should be explicitly stated.