The line on Justice Kennedy is that he’s tough to predict. Ever since Justice O’Connor’s retirement, he’s been the Supreme Court’s swing vote, and the swing of the swing vote introduces an element of suspense to many of the most contentious Supreme Court cases. His vote, to a greater extent than any of his colleagues, is thought to be up for grabs. Advocates write “Kennedy briefs” in an effort to curry his favor. At oral arguments all eyes are on Justice Kennedy, his often opaque remarks parsed in the post-argument recaps.

This unpredictability has always been somewhat exaggerated. Although Kennedy has surely been the swing vote on many of the Court’s recent major cases, he has a pretty well established pattern. On questions of federal authority, he swings conservative. On gay rights, liberal. On First and Second Amendment claims, conservative. On the death penalty, mostly liberal.

And until today, when it came to the constitutionality of racial preferences, he was a predictably conservative vote. He dissented in Grutter v. Bollinger, the 2003 decision in which Justice O’Connor, writing for the Court, upheld flexible, “holistic” racial preferences in university admissions. He voted with the majority in Parents Involved v. Seattle Schools, in which the Court struck down the use of racial preferences in public school assignments. He was never as strict an adherent to the “color-blind Constitution” ideal as his conservative colleagues. He did not see government classifications based on race as categorically prohibited. He saw some role for taking racial factors into account when deciding where to draw school district lines, for example. But he made clear that using race as a factor in individualized assessments of people was hard to square with his reading of the Fourteenth Amendment’s Equal Protection Clause. “Preferment by race, when resorted to by the State,” he wrote in his Grutter dissent, “can be the most divisive of all policies, containing within it the potential to destroy confidence in the Constitution and in the idea of equality.”

The Justice Kennedy who today wrote the majority opinion in Fisher v. University of Texas (since this was the second time this case came before the Court, this case was know as Fisher II) struck a different tone. He clearly retains his sense of discomfort with using race in admissions. He wants any school that uses race as a factor in its admission process to proceed with caution, to use race no more than necessary, and to constantly assess its program to make sure it’s actually serving a necessary role in increasing diversity. But he did not focus on what he had previously described as the costs of race conscious decision making on its beneficiaries as well as those who did not receive the preferences. This could make a significant turn in his approach to this volatile civil rights issue.

Kennedy’s jurisprudence has developed over time, but usually along predicable paths. Fisher II was less predictable. It was a shift of direction. Not an about face by any means, but it was a shift. (Yes, I called this one wrong.)

“Something strange has happened since our prior decision in this case,” wrote Justice Alito in the opening of his dissent. He argues that Kennedy, who wrote the first Fisher opinion, made a demand on the University of Texas, but then failed to follow through on that demand. So what strange thing happened? Maybe Kennedy’s triumphant performance in last term’s gay marriage decision pulled him a bit closer to his liberal colleagues. Maybe the slashing dissents in that case pushed him a bit further from his conservative colleagues. Maybe national events—the renewed focus on racial inequality in the wake of Ferguson and Black Lives Matter; the stark racial divides that the Trump campaign is exacerbating—played a role. We may never know the answer. (Kennedy himself may not have really known the answer.) The key question now is whether this decision marks the beginning of a new path for Justice Kennedy’s affirmative action jurisprudence, or whether he has just become less predictable.