What happens when What3Words gets lost in translation?
UK emergency services organizations urged to consider alternatives
What3Words, the website and app that translates physical coordinates into short memorable combinations of words, has been praised and criticized over the years.
Now a computer scientist at the University of Exeter in the UK has formally described in a paper how confusion can arise from the geocoding algorithm used by What3Words, and questioned its fitness for directing emergency services to incident locations.
What3Words (W3W) dates back to 2013. It's an app that uses a proprietary geocoding system to map geographic coordinates to a set of three words, because long numbers tend to be more difficult for people to remember and communicate than words.
The system works by dividing the world into a grid of three meter squares and assigning three words from a data set of approximately 40,000 words to each square. Thus the latitude and longitude of New York City Hall (40.712772, -74.006058) maps to clip.apples.leap, which typically would be easier to recall, say, or type into a mobile device.
The app, according to the company, is used in 170 countries by thousands of businesses as well as governments and NGOs. It has received endorsements in the UK from the Automobile Association (AA) and some emergency service groups, which have advised travelers to use the W3W app in case they need to communicate their whereabouts. The makers of the app itself have described W3W as an "additional tool" that emergency services around the world can use to locate callers.
But the app is also the subject of ongoing criticism, documented on the website What 3 Words is a Mess. The site lists various news and social media reports that cite problems arising from the use of W3W addresses.
Pointing to these concerns, Rudy Arthur, a senior lecturer in data science at the University of Exeter, has analyzed the W3W geocoding algorithm, as described in the company's patent filing because the company's code is closed to scrutiny.
"If W3W is widely adopted by emergency services it must be subject to rigorous evaluation," he explained in a preprint paper titled: "A Critical Analysis of the What3Words Geocoding Algorithm."
The Register contacted Arthur to talk about his findings but he declined, stating that he hopes the paper will be published in an academic journal and that he would rather wait until the peer review process is complete before discussing the work.
Arthur in his paper says that while he's not aware of prior formal academic analysis of W3W, there have been public posts from technical experts who have raised concerns about the geocoding scheme. For example, two years ago, penetration tester Andrew "Cybergibbons" Tierney argued the algorithm can create confusion if people incorrectly pluralize/depluralize words or misinterpret spoken words for homophone spellings.
Tierney noted that W3W has claimed "people confuse plurals only about 5 percent of the time when hearing them read out loud," which he said means that the chance of a W3W address being confused just on the basis of plurals is one in 27.
W3W, on the other hand, says the possibility of confusion is much more remote.
"Let's say there are 10,000 potentially confusable combinations ambiguously close to each other in the UK," the company said when addressing these concerns in 2021. "That is 1 in 2.5 million (0.00004 percent) overall chance of hitting a square that could be considered to have a nearby square with a confusably similar address."
- Tired: What3Words. Wired: A clone location-tracking service based on FOUR words – and they are all extremely rude
- Say what you see: Four-letter fun on a late-night support call
- UK's National Cyber Security Centre recommends password generation idea suggested by El Reg commenter
W3W also asserted: "The overwhelming proportion of similar-sounding 3 word combinations will be so far apart that an error is obvious."
The consequence of such confusion in the context of emergency services includes scenarios like ambulances directed to the wrong location, to the potential detriment of those in need of assistance.
Arthur's paper identifies 14 distinct modes in which errors can be introduced in the oral or written transmission of word triplets. These include typos, homophones, autocorrect substitutions, regional spelling variations, uncertainty about word boundaries ("dogs.tart" or "dog.start"), and so on.
As an example taken from Arthur's paper of just homophone error possibilities, the location "arose.recede.home" has a confusion set that consists of "arrows.recede.home," "arose.reseed.home," and "arrows.reseed.home." That is to say, these are potential soundalike errors that could be introduced when attempting to convey a W3W address to another person.
I estimate 20-25 percent of addresses have more than three other addresses with which they could be confused
It's difficult to evaluate the frequency of these error modes, the data scientist says, due to the restrictive terms of the W3W license. He has, however, made the code for his data analysis public.
"I estimate 20-25 percent of addresses have more than three other addresses with which they could be confused, considering only homophones and typing errors," Arthur says in his paper. "Thus it may be very difficult, if not impossible, to decipher a miscommunicated address with the current system."
"Contrary to what W3W claim, this work has shown that many confusable W3W addresses are likely to exist and has identified some serious problems with the W3W algorithm. However, while showing that many confusable pairs exist, this work does not show that they will be confused."
Pointing to the various other geocoding systems that are being used – GeoHash, Natural Area Codes, Plus Codes (a.k.a. Open Location Code), the Military Grid Reference System, and the Maidenhead Locator System – Arthur suggests the Ordnance Survey National Grid reference system would be more suitable for UK emergency services applications.
What3Words did not respond to requests for comment but has previously told the BBC that human error is "a possibility with any type of tool." ®