There's actually no need to speculate on the answer to this question since scientists have published their estimates and methodology, as is their way. The following paper is a good review:
Fields C, Adams MD, White O, Venter JC. 1994. How many genes in the human genome? Nature Genetics 7:345-346.
Below are some truncated excerpts from the paper but, if possible, I recommend reading the whole thing and the references therein.
In genomic sequencing pilot projects... we found... an average of about one gene in 23.4 kb... Extrapolated for the whole genome, we would predict about 129,000 genes; however, the regions that we sequenced were chosen for high-GC content and hence gene richness. At most, half of the genome, in the GC rich... bands, is likely to have high gene density; if the rest has half of the density we observed, the human genome might contain 97,000 genes. But the gene-poor fraction of the genome probably has much less than half the density of the gene-rich fraction... If we assume that the genome comprises a gene-rich half with [23.4 kb per gene] and a gene-poor half with a tenth that density, we obtain an estimate of about 71,000 genes...
Making estimations based on average gene size has been discussed in another answer. The varying estimates result from the different assumptions made: there were a lot of unknowns at this time. You can read about the correlation between GC-content and gene density in this answer.
Estimates giving much lower gene content are, however, easy to come by... Wagner and colleagues note that only about 12% of a typical mammalian genome... is transcribed. Using an average gene size of 18 kb obtained from a list of characterized genes... they estimate a total of 20,000 genes. By assuming that 2,500 housekeeping genes (from estimates of Escherichia coli) constitute 18% of the total number of genes the same authors obtained an even smaller figure of some 14,000 human genes.
I don't have access to this reference so it's hard too delve to deeply into their methodology, but the number of genes expressed really depends on the cell-type. Thymic medullary cells, for example, express 85% of the coding genome. RNAseq has also suggested that over 90% of the genome is transcribed, though this controversial. All that said, their first estimate was rather spot-on.
Measurement of RNA reassociation kinetics suggest that approximately 10,000 distinct genes are expressed in a typical mammalian cell, from which Lewin estimates a total gene number of 20,000 to 40,000.
This is C0t analysis with RNA instead of DNA (called R0t). You can read more about this here. This estimate, too, has proved reasonably accurate.
Using restriction analysis with the methylation-sensitive enzyme HpaII, Antequerra and Bird estimated that the human genome contains 45,000 CpG islands. They also report that about 56% of sequenced genes contain CpG islands, and hence estimate a total gene number of about 80,000... This number, however, may be an overestimate, as even "complete" gene sequences rarely include extensive 5' or 3' flanking sequence and hence may muss associated CpG islands.
We now know that there about 30,000 CpG islands in the genome, of which about 9,000 are intragenic and that 72% of genes have CpG islands. This would revise their estimate to 30,000 genes.
We have used a collection of 3,483 nonredundant coding sequences as an effective genome against which to compare a collection of human ESTs... If [this] set of complete cDNA... is representative of human genes in general, the fraction of known cDNAs matched by randomly-selected ESTs should equal the fraction of novel sequences matched by randomly-selected ESTs. Our human EST sequencing project has so far identified ESTs matching 1,877 of the 3,483 unique coding regions (54%). We can, therefore, estimate that the novel ESTs that we have sequenced represent about 54% of previously-unknown human genes... To estimate how many genes these novel ESTs identify, we... [clustered] the ESTs. This step reduced 65,297 ESTs to 40,077 clusters... indicating that the novel EST set was 40% redundant. We can then calculate an expected number of human genes as: 40,077 / 0.54 + 3,483 = 77,700 genes. This calculation is an over-estimate, since the clustering procedure cannot identify ESTs from the same transcript unless they overlap. If the true average redundancy is 50%, we predict about 64,000 genes; if the true average redundancy is 60%, we predict 52,000 genes.
Although they accounted for alternative splicing, assumptions were made about how representative their collection of coding sequences was. It seems that one of the problems at this time was that many of the overestimates made, using different methods, more or less agree with each other. Unfortunately, many of the assumptions did not stand up.