About twelve years ago, my team and I were working on how autonomous vehicles might matter for regional planning.

This was around 2014, when AVs were neither science fiction nor a near-term certainty. Our question was not “will autonomous vehicles happen,” but how planners should reason when a technology might alter some of the core assumptions embedded in travel models. What happens if driving becomes easier or more pleasant? What happens if pricing shifts from ownership to usage?

We wrote it up — the paper is here, and the regional model we were using at the time was SoundCast, the Puget Sound activity-based travel model.

Looking back now, what I find most interesting is not whether the scenarios were correct, but how differently various assumptions aged.

What aged well

Travel models did need to evolve, and many regions invested in doing exactly that. Shared-ride and taxi-like modes did become central to how AV impacts were discussed. Pricing — particularly per-mile charges — consistently showed up as one of the few mechanisms capable of offsetting increased vehicle travel when driving became easier. That finding has held up.

What we underweighted

Other aspects were less fully anticipated. Safety became a much larger part of the AV conversation than we emphasized at the time. AVs also emerged in reality as a premium, highly private mode, attractive in part because they reduce social interaction — something we did not model well. We did not fully appreciate how fragile automated systems could be in complex, human edge cases: the recent snafu when traffic lights went out, how they would handle emergency zones, pedestrian ambiguity. We also did not spend enough time on the potential value of this technology for aging populations and people with mobility limitations.

What mostly didn’t happen

Most of the outcomes we worried about most have not really materialized. Large-scale private AV ownership radically reshaping land use and dramatically increasing VMT has not happened — at least not yet. Those futures remain plausible, but they have proven more contingent and slower to unfold than many early narratives suggested.

The value of written-down assumptions

What I value most about this old work is that the assumptions were written down clearly enough to be revisited. Institutions are often very good at producing forecasts and much less good at returning to them, asking what aged well, what did not, and why.

I feel lucky to have lived long enough to see my own assumptions play out in real time. I can see where the mental models were useful, where they were incomplete, and where entire dimensions of the problem were missing.

A single forecast never tells the whole story. But revisiting them — together and across disciplines — is one of the ways judgment actually gets built.