Zillow, the online real estate portal and darling of the property-tech boom, shocked the markets last week by closing its online home buying business unit, taking a $500 million write-off, and trying to offload 7,000 unsold homes from its inventory. The CEO said he made the decision due to the unpredictability of forecasting home prices and the adverse effect that can have on Zillow’s balance sheet. Translated into plain language that will resonate with many homeowners, this means Zillow’s houses are now underwater: worth less than it paid for them.
Tech pundits (admittedly like us) jumped all over the CEO’s further explanation that Zillow’s Zestimate home pricing algorithm, used by Zillow to price its offers to buy homes, could not accurately predict future fluctuations in home prices. This seems to us simply a proof of something we instinctively knew about AI: the machine can be every bit as dumb as a human being.
American homeowners are very familiar with the Zestimate score, a brilliant marketing tool that estimates a home’s market value within a 5-10% range by comparing several home data points gathered by Zillow (such as comparable sales in the area, number of bedrooms, good schools etc.). We remember well our first Zestimates and the excited feeling that we made a bundle of money (on paper at least). We also remember the real estate agent’s eye roll when we said list it at the top end of the Zestimate!
A surprisingly high Zestimate score could encourage an owner to put her home on the market, which of course is what Zillow wants as that drives advertising revenue through its real estate listing portal. Zillow has also been accused by real estate agents of manipulating local property prices by over-pricing a neighborhood and driving irrational expectations from sellers. The company has defended itself in the past by saying it’s not us, we only report the market data.
So now in a moment of pure irony, it appears that Zillow has been hoisted on its own petard. We’ve always wanted to use that favorite old Shakespearean phrase in a post, which the Free Dictionary describes as meaning one has “fallen victim to one’s own trap or schemes.”
Does the Zillow algorithm fail have larger meaning for AI? Some tech pundits have used this story to rail against the blind adoption of AI by businesses or to preach against the sins of replacing human wisdom with AI bots. After reading a few posts, you’d be forgiven for thinking this could be the beginning of the end for evil AI.
At Deep Analysis, we thoroughly disagree with the fearmongers and naysayers. This is no reason to stop or even delay AI projects. Here are some lessons we take from the Zillow story for any business owner using or considering AI:
- Selling and buying homes is not a purely rational, data-driven process. When AI is used to automate such a complex market space where purchases are heavily influenced by human emotions, the machine must be supervised carefully by emotional humans lest it miss the tell. Redfin, Zillow’s competitor, was quick to say it has always included humans in the loop when making an offer to buy.
- AI is not infallible; it can mimic any human error in judgement. The problem is that, left unattended, AI multiplies the error at unprecedented speed and at scale. Zillow was bitten hard by this AI weakness. In a way this reminds us of when the Internet first came out, and we observed how it multiplied human ignorance at unbelievable speed and scale.
- Zillow is not abandoning its algorithm; far from it. The company says it will continue to tweak the model and work out the bugs.
This is, after all, a machine LEARNING.