Google's AI Doesn't Understand Restaurant Menus


In the glorious future, every website will be chock-full of semantic metadata. Restaurants won't have a 50MB PDF explaining the chef's vision for organic cuisine - instead, they'll have little scraps of data on the HTML page like:

"hasMenuItem":{
  "@type":"MenuItem",
   "name":"Dodo In A Bun",
   "description":"The legendary extinct bird cooked in tomato sauce, served in a gluten-free bun.",
     "offers":{
       "@type":"Offer",
       "price":"7.99",
       "priceCurrency":"GBP"
     },

But, for now, they don't. So Google uses MACHINE LEARNING and DEEP AI to scrape the menus out of PDFs and photos submitted by patrons. It works about as well as you'd expect.

Examples

Here's a bunch of screenshots from Google Maps. Would a human have made these mistakes? (No they'd make much more interesting mistakes!)

Chicken £0.86. Gourmet Burger £810.76. Veggie US$10.76.

Are those realistic prices for burgers? Would a place that sells a gold-covered burger also sell second-hand chicken? Is it usual that UK menus offer items in different currencies? What about this:

A side dish called "plain paper" for £45

Is that a likely name for a side dish? You wouldn't find that even in trendy Shoreditch. And, again with the weird pricing. Side dishes rarely cost that much. Still, it could be worse…

Chicken Tika Massala listed for £1,990.

I guarantee you the little curry shop I was looking at didn't have anything on the menu that pricey!

Other times, it seems to just pick random words in a vague attempt to be "useful".

List of salads including one called "the". And another called "middle".

Perhaps my favourite is this British restaurant which offers spam, spam, and more spam:
A menu with the word "egg" written three times.

There were dozens (dozens!) of examples I found within a few minutes of looking for somewhere nearby to grab a cheap lunch. To be fair, when the website has supplied data in an appropriate format, Google is good at displaying it. But when Google tries to be clever... ah.

There is, of course, no way for a user to contact Google and say "I dunno, man. This looks kinda off."

Whose Fault Is It Anyway?

Obviously, restaurateurs should provide machine-readable descriptions of their menus. That way search engines, smart speakers, and all sorts of funky computers will be able to parse the food available.

But they don't. Because they're busy cooking.

Google has discovered that it takes 90% of the effort to get 90% of the way there - but the last 10% takes the next 90% of the effort.

As ever, a feature has launched and it is of dubious value to business owners and customers - but I'm sure a product manager at Google got their wings.


13 thoughts on “Google's AI Doesn't Understand Restaurant Menus

  1. Google really should not be allowed to misrepresent businesses like this. It could be perceived that they are giving other businesses an advantage by putting potential customers off.


      1. Bud Ryerson says:

        It's 90% of the last 10% or, in other words, 9%. So it takes 9% of the effort to get the last 10% of "the way there." That means it only takes 99% of "the effort" to get 100% of "the way there."

        It's a world of wonder, innit?

  2. says:

    I have mixed feelings about this, I’d be inclined to think that it’s still a net positive to highlight business like this.

    I really doubt a potential customer will actually think that the curry costs 2 grand.



    1. says:

      But Google thinks it costs two grand. So when someone searches for 'cheap lunch' or 'good value curry' - there's not much hope of this curry house showing up, or it'll be shown on the map as 'expensive' for no good reason.

  3. DinoNerd says:

    I wonder how many people will continue to use the feature after its initial, relatively useless release, and notice if and when Google gets to the point where it's functional enough to be usable - let's say it's only wrong 1% of the time, so it gets some menus 100% correct.

    In my experience, vendors who release new features at this quality level essentially never announce "Heh folks, we've fixed it; you should come back and try it again". There are Google features that were at least as useless as this one at first release, which I never use, because I "know" they don't work. Every once in a while some youngster tells me that they use the feature, and find it usable and convenient, and I look back and realize it's been years since I tried it.

    One example is Google Translate. When I last used it, its many flaws included the practice of omitting words it didn't understand, without any ellipsis. It only took one example where the word omitted basically meant "not", reversing the meaning, to cause me to file Google Translate as "useful ONLY if you basically already understand the text". I relatively recently discovered it's now good enough that people with no language in common routinely use it to communicate on MMORPGs; another player wondered why I wasn't interacting with an in-game neighbour whose chats always used the Cyrillic alphabet, and didn't think "because none of the languages I know use that alphabet" was a reasonable response.

  4. It's funny because if you asked me who I thought the leader in AI was, I would probably think 'Google' without a second thought.

    I suppose the reality is that it's easier to train ML models when you have complete control, like how Google knows how an Android phone will see text, numbers or take photos.

Leave a Reply

Your email address will not be published.

%d bloggers like this: