What AI still gets wrong about flood risk

AI-driven risk scores show up next to the address now — a number from one to ten sitting beside the beds, baths and square footage on Zillow, Redfin and realtor.com, quietly implying it knows something about this house the listing itself doesn't say. Insurers are increasingly pricing flood coverage with similar machine-learning models. None of that is a flood zone determination, and the gap between the two matters most exactly where the stakes are highest: waterfront.

The confusion is understandable. Both a FEMA flood map and an AI risk score claim to answer whether a property will flood, and both hand you something that looks like a definitive number. They're built for different purposes, from different data, on different update schedules — and on a Great Lakes shoreline, where the water itself moves through decade-long cycles measured in feet, the gap between them can be the difference between an accurate picture and a misleading one.

A score is not a flood zone

FEMA's Flood Insurance Rate Maps, or FIRMs, are the regulatory instrument. They define Special Flood Hazard Areas, they're what a mortgage lender checks to decide whether flood insurance is federally required, and they're revised through a formal process that can take years to catch up with a changing shoreline. First Street Foundation's Flood Factor — the model behind the scores now embedded on the major listing portals — is something else: a privately built, probabilistic model that scores cumulative flood risk over a 30-year period using rainfall, streamflow, storm surge and tidal data, refreshed more often than FEMA's maps and covering ground FEMA hasn't mapped at all. It's a genuinely useful data point. It is not a legal determination, and it doesn't decide whether a lender requires a policy.

• FEMA's FIRM: whether the parcel sits in a mapped Special Flood Hazard Area, and whether a federally backed mortgage requires flood insurance
• A portal risk score: a private model's estimate of cumulative flood probability over 30 years, built for general awareness, not underwriting or permitting
• An insurer's own AI pricing model: proprietary, unpublished, and under no obligation to agree with either of the above
• None of the three has to match the others, and none replaces a site-specific look at a waterfront parcel

Where a national model runs into the Great Lakes

Most flood-risk models are built and validated against river gauges, coastal storm surge and precipitation records — the inputs that drive flooding across most of the country. Lake Michigan doesn't work quite like that. The lake moves through multi-year water-level cycles stacked on top of its normal seasonal swing, and the gap between its record low and record high runs more than six feet. Shoreline that sits comfortably dry in a low-water year was underwater, or actively eroding, the last time the lake ran high. A model tuned to national flood patterns can easily miss that a Great Lakes lot's real exposure has less to do with any single storm than with where the lake happens to sit in its own multi-decade cycle.

Erosion compounds the problem, because it isn't a flood risk in FEMA's sense at all — Michigan tracks it separately. EGLE designates roughly 250 miles of Lakes Michigan, Huron and Superior shoreline as High Risk Erosion Areas, based on shorelines receding a foot or more a year over at least 15 years, and calculates construction setbacks from the erosion projected 30 and 60 years out. Lake levels across the Great Lakes have settled back to roughly their long-term average this year after the extreme highs of 2020, and shoreline researchers at Michigan State describe the moment as a calm window rather than a resolved risk — the lake is expected to cycle back up eventually, on a timeline nobody can pin down. None of that shows up in a score built around a ten-point scale.

What to actually pull before you write the offer

The generic advice is to check the flood zone. On a Great Lakes lot, that's one document out of five.

• The FEMA FIRM for the parcel, pulled directly from the Flood Map Service Center — not summarized secondhand
• EGLE's High Risk Erosion Area and Critical Dune Area maps, to see whether state permitting applies to construction or earth-moving on the site
• An elevation certificate, if one exists, showing the lowest floor's elevation relative to the base flood elevation
• A CLUE report on the property's insurance claims history, which can surface prior water damage a listing never mentions
• Current lake-level data from NOAA or the Army Corps of Engineers, read against the parcel's own historical aerial photography

None of this is a reason to distrust the tools — a risk score is a fast, useful first signal, and most buyers never used to see anything at all before the offer stage. It's a reason to treat the score the way you'd treat a home inspector's first walk-through: a prompt to look closer, not the inspection itself. We keep the FEMA map, the state's erosion data and the claims history in the Property Research Center for exactly that reason — a number on a listing page was never built to carry that much weight on its own.