As it can be easily pointed out that true random numbers cannot be generated fully by programming and some random seed is required.
This is true. In fact, it is impossible to solve using software. No software-only technique can generate randomness without an initial random seed or support from hardware.
This is also true for AI software. No AI design that uses deterministic software can do this - e.g. any Turing machine without a magic unexplained "random" function can be shown to remain deterministic, no matter how complex. That's because any combination of deterministic functions is deterministic. It may not be predictable without following the process exactly, it may be "chaotic" and depend critically on initial conditions, but it is 100% deterministic and repeatable.
On the other hand, humans can easily generate any random number independently of other factors.
Typically not high quality randomness. It is not clear how we make random decisions, but entirely possible that internally we rely effectively on noisiness from the environment or our own internal "hardware" doing something simple such as racing decisions between neurons (timing will vary due to speed of electrical impulses, diffusion time of neuro-transmitters across synapses etc)
Does this suggest that absolute random number generation an AI cocept?
I think it is orthogonal issue. We can already produce very high quality artificial randomness - better than human quality used by humans for conscious decisions (such as "choose a random number between 1 and 10"). These artificial random number generating systems are part of modern cryptography and are tested thoroughly.
Essentially "true" artificial randomness is a solved problem using hardware, and does not involve anything that has traditionally been called AI.
In reverse, AI systems often rely on stochastic functions in order to break symmetry, break ties, regularise models etc. So it does look like some kind of RNG is necessary within an artificial agent. However, even pseudo-random number generators (PRNGs) seem to be fine for this purpose. Mersenne Twister is a very common choice for generating random numbers inside neural networks for weight initialisation, dataset shuffling, dropout regularisation, or when simulating environments for RL, or taking exploratory actions. Despite the fact that it is not "true" random, a PRNG will work just fine for these purposes.
Working definition of "true" randomness that I use: Even if you know the state of a system as accurately as modern physics allows, the output cannot be predicted with better accuracy than a fixed guess.
Human randomness already fails this test. If you ask someone to choose a number randomly between 1 and 9, you will generally have a better than 1 in 9 chance of guessing the correct value, based on statistical analysis. If we were able to take good state measurements of brains, it might be possible to predict with high accuracy - although this is unknown and not possible with current technology.