Data reliability is highly dependent on the type of data you're working with, and the procedures, processes, and checks on that.
I've worked with scientific, engineering, survey, business, medical, financial, government, internet ("web traffic" and equivalents), and behavioural data (e.g., measured experiences / behavour, not self-reported). Each has ... its interesting quirks.
Self-reported survey data is notoriously bad, and there's a huge set of tricks and assumptions that are used to scrub that. Those insisting on "uncensored" data would likely scream.
(TL;DR: multiple views on the same underlying phenomenon help a lot --- not necessarily from the same source. Some will lie, but they'll tend to lie differently and in somewhat predictable ways.)
Engineering and science data tend to suffer from pre-measurement assumptions (e.g., what you instrumented for vs. what you got. "Not great. Not terrible" from the series Chernobyl is a brilliant example of this (the instruments simply couldn't read the actual amount of radiation).
In online data, distinguishing "authentic" from all other traffic (users vs. bots) is the challenge. And that involves numerous dark arts.
Financial data tends to have strong incentives to provide something, but also a strong incentive to game the system.
I've seen field data where the interests of the field reporters outweighed the subsequent interest of analysts, resulting in wonderfully-specified databases with very little useful data.
Experiential data are great, but you're limited, again, to what you can quantify and measure (as well has having major privacy and surveillance concerns, often other ethical considerations).
Government data are often quite excellent, at least within competent organisations. For some flavour of just how widely standards can vary, though, look at reports of Covid cases, hospitalisations, recoveries, and deaths from different jurisdictions. Some measures (especially excess deaths) are far more robust, though they also lag considerably from direct experience. (Cost, lag, number of datapoints, sampling concerns, etc., all become considerations.)
It's complicated.