Right, time to step in here.
First of all, there exists a number of known physical and chemical pathways leading to the generation of mutations. But it's
precisely because such a large number of possible pathways exist, that it's extremely difficult to know
which of the possibilities is going to be active in any one instance. This is partly due to the sheer volume of data to wade through, before we start thinking about the difficulties of collecting any of that data. For example, there are approximately 7 billion human beings on the planet. Each of these human beings is comprised of a large number of cells, it's safe to take 10
10 cells as a ballpark figure. Let's concentrate for a moment on germ cells, namely, the cells that are responsible for producing gametes, i.e., sperm and egg cells. In the case of sperm, the average human male produces 100 million sperm prior to each instance of sexual intercourse, so that's 100 million cell replications
per male human being per instance of sex. Multiply this by 3.5 billion male human beings, and factor in here that the typical male experiences something like 3,000 ejaculations in a lifetime, and we have, just for our species alone, the following number of cell replications
just to produce sperm:
(3.5 × 10
9) × 10
8 × 3,000 = 1,050,000,000,000,000,000,000 cell replications
This is around 10
21 cell replications,
just to produce human sperm in one generation. So the magnitude of the data to be gathered, if one is going to cover the entire biosphere, very quickly becomes apparent. Every second on this planet, something like the same number of cell replications is taking place throughout the biosphere, and whilst performing the detailed breakdown into such processes as reproductive mitosis for single celled organisms, versus reproductive meiosis for multicellular eukaryotes, would itself be a pretty tedious exercise for the dedicated masochist, we can safely regard something like 0.1% of those cell divisions as being directly reproductive. Which still leaves us with 10
18 reproductive cell divisions
per second taking place on the planet.
Now, because the numbers involved are of this order of magnitude, there's no way that scientists are going to gather detailed molecular biological data on
every one of those cell divisions. They are generally restricted to watching what happens on a very tiny subset of those cells in the laboratory. But, those observed events are pretty typical of what's happening within cells outside the laboratory too.
Now comes the fun part. The chemistry of DNA replication that is taking place within those cells, is
not taking place in isolation from the activity of other chemicals in the call. Although in the case of eukaryotic cells, the cell nucleus acts as a reasonably effective container isolating that DNA replication from the most dangerous of the other chemicals in the cell, it isn't a perfect barrier by any means, and some of the chemical compounds required for the life processes of the cell can themselves interfere with replication if they turn up alongside a DNA strand being replicated at the wrong moment. We know that this can happen, because this has been observed in the laboratory, but what we don't know, is which of the possible interfering reactions is taking place in any given instance outside the laboratory.
In short, that's what "random" means in the context of evolutionary biology. Numerous possibilities potentially affecting a vast number of replications, and no way of accessing the data directly as it happens. But because we can observe what happens in small but representative samples in the laboratory, what we
can do, is assign probability values to different classes of mutation, and construct probability distributions which can then be applied to a statistical random variable, used to model the behaviour of the wider system. When we do this, it works. it reproduces data sets that match the data sets arrived at
a posteriori via observations of wild populations.
Needless to say, a certain amount of fidelity is required for successful DNA replication. A replication process too prone to errors, will pretty quickly grind to a halt. Typically, one error is observed in DNA replication per 10
9 nucleotide copying actions, though in some systems, the mutation rate is considerably lower (germ cells tend to have additional mechanisms to correct such errors), and some systems (the HIV virus being a classic example) have much higher mutation rates. Feed this into the above statistics for cell replications, and you eventually wind up alighting upon the fact that, in humans, for example, each new human being that is born, acquires something like 100 mutations from the germ cells of its parents. This isn't a problem, because the human genome has wide expanses of non-coding DNA, and usually, mutations affecting these stretches don't have major knock on effects. Additionally, even if a mutation
does hit a coding region, it's only going to be a problem if it hits an exon - if it hits an intron, which is spliced out post-translation, then again, it's not going to exert much influence.
Now we turn our attention to what those mutations actually
do. The majority of them aren't actually particularly significant in the short term. A good many of them don't actually do anything of note - examples being so-called "synonymous" mutations, that have the effect of producing a new codon in the stretch of DNA, that codes for the same amino acid as the codon it replaced. Even a non-synonymous mutation, which results in a different amino acid being coded for, may not have a significant effect, and as evidence for this, I cite the existence of
many hundreds, if not thousands, of different variations on the insulin molecule across the vertebrates. All of these variations on the insulin theme work, and indeed, it's the reason humans were able to use bovine insulin from cows to keep diabetics alive, until the advent of a much more elegant biotehnology solution, involving inserting human insulin genes into bacteria, and then culturing the bacteria
en masse to produce human insulin in quantity.
I chose insulin as an example deliberately, because, as you're probably already aware, insulin is a pretty critical molecule for us human beings, and for that matter, is critical for the rest of the vertebrate clades as well. If an organism doesn't produce a working insulin molecule, it's headed for an early death. Consequently, any mutation that destroys the functionality of the insulin protein, is going to be an evolutionary dead end, quite literally, in a short space of time. But, the fact that
many different variations of insulin exist, all of which work, teaches us an important lesson, namely,
exact codon sequence is not of primary importance. Some
parts of that sequence may be critical, but the entire sequence as a whole isn't. All that matters, ultimately, is whether or not any variation that arises still works. If that variation does work, organisms inheriting it will continue to operate as normal, live and reproduce, and pass on that variation. Variations that don't work will come crashing to a halt when the organisms inheriting them die.
Which brings us to another important lesson. Some mutations will be instantly lethal to those organisms inheriting them, and won't go any further. After all, dead organisms can't reproduce, though if you listen to some of the more fatuous creationist assertions on the subject, you end up concluding that they think dead organisms
can pass on lethal mutations. But I digress. On the other hand, some mutations aren't lethal, but act as a handicap to reproductive fitness. Some of these may be passed on to a future generation, if the handicap doesn't stop the inheritors thereof from mating and reproducing altogether, but the bigger the handicap, the more likely those mutations are to disappear from the gene pool after a few generations. However, provided that handicap doesn't take the organism below the minimum level of fitness required to be reproductively competent, that mutation stands a chance of being passed on.
Of course, any mutation emerging that
increases reproductive competence, is instantly going to be favoured. If you end up inheriting a trait that makes you either more fecund, or more of a shag magnet to others of your species, then you're going to be the one passing genes on to the next generation on a large scale. Likewise, if you acquire the ability to make use of a new food source, one that for the time being remains exclusively yours, then you're going to boost the fitness of your population by passing that on to future generations, and because you're the first to acquire that mutation, you get to be the one reaping the benefits in terms of being well fed and healthy, and therefore making yourself more of a shag magnet once again.
In short, what decides if a mutation is going to enjoy some persistence in your species, is [1] whether or not it works, and [2] whether or not it works
better than before. The moment you obtain even a
slight advantage, that advantage is there to be built upon. Which basically, is how evolution works:
[1] Generate lots of variations;
[2] Discard the failures;
[3] Build upon the successes.
There doesn't need to be
any "intelligence" directing all of this, let alone any magic entities. Indeed, human beings have pressed this basic algorithm into service to "design" via an evolutionary process, entities
outside the biosphere. I have a nice paper on the application of this algorithm to the "design" of a spacecraft communications antenna, which was eventually flown on a real spacecraft, tested, and found to work.
Have fun reading the above.
by 
-----> I Need that glass of wine for digestion of information
