There's a new paper arguing, contra Pinker, that the world is not getting more peaceful:

On the tail risk of violent conflict and its underestimation

Pasquale Cirillo and Nassim Nicholas Taleb

Abstract—We examine all possible statistical pictures of violent conflicts over common era history with a focus on dealing with incompleteness and unreliability of data. We apply methods from extreme value theory on log-transformed data to remove compact support, then, owing to the boundedness of maximum casualties, retransform the data and derive expected means. We find the estimated mean likely to be at least three times larger than the sample mean, meaning severe underestimation of the severity of conflicts from naive observation. We check for robustness by sampling between high and low estimates and jackknifing the data. We study inter-arrival times between tail events and find (first-order) memorylessless of events. The statistical pictures obtained are at variance with the claims about "long peace".

Every claim in the abstract is supported by the data - with the exception of the last claim. Which is the important one, as it's the only one really contradicting the "long peace" thesis.

Most of the paper is an analysis of trends in peace and war that establish that what we see throughout conflict history is consistent with a memoryless powerlaw process whose mean we underestimate from the sample. That is useful and interesting.

However, the paper does not compare the hypothesis that the world is getting peaceful with the alternative hypothesis that it's business as usual. Note that it's not cherry-picking to suggest that the world might be getting more peaceful since 1945 (or 1953). We've had the development of nuclear weapons, the creation of the UN, and the complete end of direct great power wars (a rather unprecedented development). It would be good to test this hypothesis; unfortunately this paper, while informative, does not do so.

The only part of the analysis that could be applied here is the claim that:

For an events with more than 10 million victims, if we refer to actual estimates, the average time delay is 101.58 years, with a mean absolute deviation of 144.47 years

This could mean that the peace since the second world war is not unusual, but could be quite typical. But this ignores the "per capita" aspect of violence: the more people, the more deadly events we expect at same per capita violence. Since the current population is so much larger than it's ever been, the average time delay is certainly lower that 101.58 years. They do have a per capita average time delay - table III. Though this seems to predict events with 10 million casualties (per 7.2 billion people) every 37 years or so. That's 3.3 million casualties just after WW2, rising to 10 million today. This has never happened so far (unless one accepts the highest death toll estimate of the Korean war; as usual, it is unclear whether 1945 or 1953 was the real transition).

This does not prove that the "long peace" is right, but at least shows the paper has failed to prove it wrong.