We don’t believe the productivity numbers about AI

Hal Varian pointed out that GDP doesn’t deal well with free:

So much for that AI-driven productivity boom, says Apollo’s Torsten Slok.

Or as he puts it “the AI boom is clearly visible in investment data and in equity valuations, but it is not yet visible in the productivity statistics, which means the productivity payoff from AI remains a forecast rather than an observation.”

This somewhat echoes Robert Solow’s famous quip “you can see the computer age everywhere but in the productivity statistics “.

A certain amount of computing output is indeed free. Free email, free search, free telecoms with VoIP and so on. This all turns up in productivity as a decline in productivity. For we still have the hours of labour to produce it all but no measured output on the other side. As productivity is value of output at market prices divided by labour hours in its production this is indeed a reduction. One of us once badgered Facebook into telling us how many people they had working on VoIP for Messenger and WhatsApp. “A couple of hundred engineers” was the level of accuracy they’d give us. That gives free telecoms to users - free, as in buckshee. That now applies to some 4 billion people, half the species. In our productivity statistics we have the couple of hundred engineers - which may well be rather more than that now - but on the output side we’ve just nothing, zip. Which is odd for a couple of hundred people providing some or all of the telecoms to half the species but is true, we record that as a fall in labour productivity.

We do not find it hard to believe that some to much of whatever AI is doing is hiding in the same place.

If we assume that this is true - and we would absolutely insist that it is true in part - then this has knock-on effects in many of our other economic numbers. A tripartite discussion between Marc Andreessen, Brad Delong and one of us ended up with the economist (obviously, Delong) telling the other two that if it is true then this value is ending up in the consumer surplus. The value consumers gain but do not have to pay for at market prices. This is usually estimated at some 100% of GDP - consumers gain, in what they’ve not got to pay for, about the same value as is recorded in GDP - but when we come to these free services that might climb rapidly to 10x, hey why not 25x, the recorded value in GDP. This ended up being named “Mokyrian value” - all of this took place before that well deserved Nobel was awarded.

Further, if that’s true then all estimations of inequality go out the window too. For example, email, search and telecoms inequality has certainly declined markedly these past few decades and yet that’s absolutely nowhere in anyone’s estimates of income or wealth inequality.

We, generally, insist that understanding an economic measurement requires grasping what is actually being measured. This, all too often, doesn’t happen. Sure, sure, productivity isn’t everything even if in the long run it’s pretty much everything. But the way it’s measured means that if we get free stuff then that - in the first instance at least - gets recorded as a fall in productivity, not the rise it so obviously is. Cleaving to the one single number as the measure of success might not be quite the right plan therefore.

Tim Worstall