Amara's Law: Too Much in Two Years, Too Little in Twenty

Roy Amara, a researcher who spent his career on forecasting, left behind one observation that has held up better than most forecasts: we tend to overestimate the effect of a technology in the short run and underestimate it in the long run.

The useful part is that this is a claim about being wrong in two opposite directions at different times, which is more interesting than simply being wrong.

Someone who declares a technology overhyped after two years of little visible change is frequently right about the two years and wrong about the technology. Someone who extrapolates early excitement into imminent transformation is right about the direction and badly wrong about the timing.

Both errors come from the same source: drawing a straight line through a process that is not straight.

Why the early years disappoint

A new technology arriving in usable form is only the first requirement. Several other things have to change before it matters, and they move much more slowly.

Supporting infrastructure. Electric motors existed decades before factories were reorganised around them — and the productivity gains only arrived after that reorganisation, because factories had been designed around a single central power source and simply swapping the source changed little. The technology was ready long before the buildings were.

Cost has to fall far enough. Most technologies work first as expensive curiosities. The transition to ordinary use waits on cost curves, which take years.

Complementary technologies. Many things only become useful in combination. A capability may be genuinely available and useless until something else exists to pair with it.

Skills and institutions. People have to learn to use it, organisations have to reorganise around it, and rules have to be written. These move on the timescale of careers and legislatures rather than product cycles.

Habits. The slowest of all. People continue doing things the way they know, and the transition often waits on generational replacement rather than persuasion.

So the early period looks like failure. Demonstrations work; adoption does not follow. The reasonable conclusion from that evidence is that it was overhyped — and the evidence is genuinely consistent with that, which is what makes the error so easy to make.

Why the later years surprise

Then the supporting pieces arrive, costs fall past a threshold, and adoption accelerates. The visible change over a decade can exceed everything predicted during the disappointing years.

The underestimate has a specific cause: people project the current use rather than the uses that become possible.

Early forecasts describe the new thing doing the old thing slightly better — faster horses, in the familiar phrase. What actually happens is that the technology enables activities that were previously impossible, and those are much harder to imagine because there is nothing to extrapolate from.

This is the same point made in the AI essay about cost curves: a tenfold fall in the cost of something does not produce a slightly cheaper version of existing products. It produces categories that could not previously exist. And nobody forecasts a category that has no precedent.

There is also a compounding effect. Each technology that becomes cheap and reliable becomes a building block for the next, which is why the pace can appear to accelerate even when each individual step is ordinary.

Using it without deploying it as an excuse

The obvious misuse is treating the law as a defence of any struggling technology: it is not failing, it is in the underestimated phase. Plenty of technologies genuinely do not work out, and the law offers no way to tell which is which by itself.

Two questions do reasonable work at separating them.

Is anyone getting genuine value from it today, however narrow? A technology in the slow phase usually has a small number of users for whom it already works properly — not pilots and demonstrations, but real dependence. A technology with no such users after several years is a different situation.

Is the cost falling on a consistent trend? A capability that is improving in price and performance year after year will eventually cross thresholds that make new uses viable. One that has plateaued may simply have found its level.

Beyond that, the honest position is that timing is genuinely hard and confident dates are worth little. The useful stance is directional: assume that a technology which works for someone and is getting cheaper will matter more than currently expected, and that it will take longer than currently expected — while accepting you cannot say how much longer.

That is a less satisfying conclusion than a forecast. It is also considerably more likely to be right, which is the trade Amara's observation is really pointing at.