Here, I'm referring purely to the mechanics of how a user can believe what is on the screen, have faith that passwords can't be cracked, and so on. They say that "justice must not only be done, but must be seen to be done"; by the same token, security must be seen to done, or we are asking users to place blind faith in the good will and competence of those who the user is obliged to trust.
The core problem is that humans are weaker than computers when it comes to the amount of pure and arbitrary data they can perceive and remember.
No matter how tightly-coded the security validation logic, and how strong the key strength, what the user will eventually see (and typically, pay cursory attention to) will be a bunch of pixels on the screen. Anything that can fake those pixels, will get trusted.
We know that to prevent an attacker brute-forcing or forging something, there has to be a minimum amount of information present. We like passwords to be randomized across a minimum number of bits, and we provide hard-to-copy cues in forgery-resistent material. So we have foil strips and watermarks in bank notes, hard-to-manufacture copper-on-aluminium software installation CD-ROMs, and so on.
When it comes to displaying something in a forgery-resistent manner, we are restricted to pixels that have 16 million possible color values, of which users may distinguish 100 or so at best. The entire screen area may be as low as 640 x 480 pixels. Anything can set any pixel to any color, so there's no way to prevent forgery.
Even if they were, humans cannot perceive and appreciate arbitrary pixel patterns. The brain will derive patterns from the raw data and the mind will evaluate these patterns. The raw data itself will not be fully "seen"; only a limited number of derived patterns.
A large number of bits is the best-case strength for a key, applicable only if possible values are randomized over the key space, and if "cribs" (encrypted information for which the plain-text can be guessd) are not available. WEP failed both of these criteria; key strength was devalued by OEMs who left several bits in the key to known default values, and WEP traffic included a lot of stereotypical packets that provide "cribs".
Humans usually don't remember raw data; instead, they remember algorithms that can create this data. This skews values within the key space from a truly random spread, to preferred values that match the way humans think, and thus weakens the key strength.
So if it's easy to remember, it's easy to guess. If it's not easy to remember, then the user will write it down (or worse, enter it into a file on the system) and your strong password system becomes a weak and unmanaged token system. If you're going to use a token system anyway, then it's better to do this properly (e.g. biometrics, USB fobs, etc.).
User-managed passwords may be acceptable if you just want the semblance of due dilligance. You can point to your password policy, shrug about bad workers who break the policy, and seek a scapegoat whenever things go wrong. But once something that is essential to make things work is also disallowed, you lose management control, and examples of that abound.
Most assessments of key strength against brute-force or weighted-guess attacks assume that only one particular system is being targeted. The odds change considerably if you don't care what target you penetrate, and have millions of targets to choose from.
Instead of having to back off after 10 attempts due to some sort of password failure lock-out, you can simply make 9 attempts on a few thousand systems every hour or so. Eventually, you'll break into something, somewhere, and all stolen money is equally good.
Obviously, consumer ecommerce on the Internet presents this opportunity for a one-to-many relationship between attacker and victim. Slightly less obviously, it also facilitates a many-to-many relationship (the bane of database design) when the attacker can use multiple arbitrary malware-infected PCs as zombies from which to launch the attacks.