The rise of 'pseudo-AI'
July 9, 2018 10:09 AM   Subscribe

"It’s hard to build a service powered by artificial intelligence. So hard, in fact, that some startups have worked out it’s cheaper and easier to get humans to behave like robots than it is to get machines to behave like humans."
posted by clawsoon (51 comments total) 39 users marked this as a favorite
 
I've dabbled in MTurk when I'm bored and so many of the jobs on offer there are insanely below minimum wage, I'm not surprised it's easier to dump stuff on there instead of building something real. Seemingly machine simple things like converting text from receipts, business cards, and other sources are dropped on there at 1¢ a pop. Usually takes 2-3 minutes to load, agree, transcribe those media, submit, so you're making 30¢ an hour assuming there's enough teed up to keep you going and not having downtime to peck.
posted by msbutah at 10:18 AM on July 9, 2018 [8 favorites]


I have to run, but hopefully someone can dig up the previously about the New Zealand startup/scam which appeared to being doing exactly this under the guise of AI medical transcription.
posted by clawsoon at 10:23 AM on July 9, 2018


I think some of the image identification apps for the blind use this kind of crowd-sourcing. I almost wish they would just let us talk to the workers, a la the Be my Eyes app. At least then you know you have an actual human. This kind of middle ground is disquieting.
posted by Alensin at 10:24 AM on July 9, 2018 [4 favorites]


I always wondered how prevalent this was. A few years, I was working at an online travel search company, whose CTO decided that AI was going to be the Next Big Thing and we needed to be ahead of the curve. So a couple of architects and some frighteningly senior developers disappeared into a closet for a couple of months, and came back with a chat bot who you could ask to search for something in plain English, rather than use the (meticulously designed, incredibly user friendly, core-product-of-the-company) search box on the main page. They rolled it into production with almost no testing, and loosed the world on it. The results were... um... not so good. Like, dozens of angry Tweets and Facebook messages and emails about how useless the thing was, and couldn't we just put things back to the way they used to be? Rather than lose face by turning it off, someone quietly flipped a switch and redirected all queries to aforementioned frighteningly-senior-developers inboxes. Which is how the highest-paid software devs in the company got to spend a good chunk of their days translating plain-text questions into search strings in a web application. When the rest of us got wind of what was powering our suspiciously-good "AI", we did what any red-blooded nerds would do, and bombarded the system with questions about swallow airspeed velocity and Voight-Kampff tests. The AI trial ended the following week. To my knowledge, it has not been resurrected.
posted by Mayor West at 10:24 AM on July 9, 2018 [100 favorites]


“It’s essentially prototyping the AI with human beings,” he said.

..incorrectly. Riding a bicycle and saying "brmm brmm" is not "prototyping a motorcycle" no matter how often you tell your mum it is.
posted by howfar at 10:29 AM on July 9, 2018 [126 favorites]


Corporations were already AIs, they just have a lot of human parts.
posted by ethansr at 10:32 AM on July 9, 2018 [26 favorites]


So, "Wizard of Oz" setups are really, really common in research, mostly because the experimenters aren't interested in making the automation work, but they're interested in how people respond to it. I've seen a lot of them in driving simulator experiments, mostly because the sims aren't capable enough or its the easier solution. I'm pretty skeeved out by the idea of telling users outside the lab that they're talking to an AI and having it be a human, but I'm having a hard time articulating why. I think a lot of it comes down to intent, but more the users' consent. That's a whole separate rant, but there's an idea to my mind of implied consent here, and changing what's implied to be behind the curtain seems like a betrayal of that trust and consent.

Mechanical Turk, as msbutah mentioned, is related to this, in so much as it's using humans to do tasks we really wish automation could do as well or as reliably. It gets used a lot in my corner of the research world (vision science), and its utility for asking psychological questions is debatable. If you want to run an experiment with a handful of trials per observer, which means you need hundreds of subjects to have enough power to conclude anything, it's a good approach, but you run into the problem of compensating subjects reasonably for their time. If you want to run the kind of experiments I do, where I need to show each subject several hundred trials, it's possible, but harder. Generally, you get to compensate Turkers better if you want that kind of data, because if they don't complete the entire set, their data isn't useful, but it's still an issue. A few years back, there was a thought in my research community that MTurk would be a cheap, fast way to collect good data. It turns out that, like the old saw, if you want good data, it's not cheap. Or, if it's fast and cheap, it's not good. It turns out, for experimenter time and subject costs, to be about a wash for most psychophysical experiments between bringing subjects into the lab and running experiments on MTurk.
posted by Making You Bored For Science at 10:37 AM on July 9, 2018 [12 favorites]


the New Zealand startup/scam which appeared to being doing exactly this under the guise of AI medical transcription

Previously: Theranos NZ
posted by miles per flower at 10:40 AM on July 9, 2018 [5 favorites]


My intelligence is also artificial. Can I have a billion dollars in VC money now? Thanks!
posted by It's Raining Florence Henderson at 10:43 AM on July 9, 2018 [6 favorites]


I'm pretty skeeved out by the idea of telling users outside the lab that they're talking to an AI and having it be a human, but I'm having a hard time articulating why.

Because it's lying or outright fraud.
posted by Sangermaine at 10:45 AM on July 9, 2018 [20 favorites]


Corporations were already AIs, they just have a lot of human parts.

Our artificially intelligent "smart correlation" technology believes this article may be relevant to your interests.
posted by Naberius at 10:51 AM on July 9, 2018 [4 favorites]


Anyone remember Ask Jeeves?
posted by Abehammerb Lincoln at 10:54 AM on July 9, 2018 [4 favorites]


And also because is runs in parallel with how the ethnic cleansing types start their campaigns - by convincing us by hook or by crook that the undesirables aren't "real people".


Yeah, it just struck me that we are talking about the literal meaning of “dehumanizing”, which pings the same neurons as the figurative use of the word in ethnic cleansing and xenophobia scenarios.
posted by darkstar at 11:01 AM on July 9, 2018 [4 favorites]




I think also because we treat non-persons/things very differently, and often a lot more callously, than we do people.

Your comment reminded me of another phenomenon I've read about: male abuse of female AI systems. Siri and her sisters have to be created with the understanding that they will receive demands and requests for sexual favors, that they'll hear irrelevant sexual comments, etc. Sounds like a fun side benefit while you make sub-minimum wage pretending to be a robot!
posted by Emmy Rae at 11:19 AM on July 9, 2018 [23 favorites]


I wonder if we'll start to see fraudsters signing up for Mechanical Turk to look for valuable data to exploit.
posted by smelendez at 11:23 AM on July 9, 2018 [1 favorite]


Previously: Theranos NZ

Did we cover that the Theranos CEO was indicted?
posted by thelonius at 11:24 AM on July 9, 2018


I always wondered how prevalent this was.

A company I, uh, knew, was one of the earlier entrants into the voicemail-to-text-message market (the tech, sold to telecom providers, who then sold it as a service to customers). At the time, speech recognition models to handle arbitrary speakers (particularly on cell phones -- awful hardware, awful audio bitrate) was not up to the task. The solution was a call center in India, where low-paid employees would listen to the voicemails and type them up.

There was always a plan to transition to pure speech recognition on the back-end, when it became economical. I don't actually know if it ever did (the tech has improved, but labor is still pretty cheap, so ...)
posted by tocts at 11:29 AM on July 9, 2018 [7 favorites]


So much of our tech economy is smoke and mirrors and it's disturbing to see how much people want to believe.
posted by bongo_x at 11:33 AM on July 9, 2018 [7 favorites]


I'm sorry, Dave. I'm afraid I can't do that. I need a bio break.
posted by It's Raining Florence Henderson at 11:34 AM on July 9, 2018 [10 favorites]


I wonder if we'll start to see fraudsters signing up for Mechanical Turk to look for valuable data to exploit.

MTurk has been around for quite some time now, I'd be surprised if that hasn't been happening for just as long.
posted by tobascodagama at 11:37 AM on July 9, 2018 [3 favorites]


Riding a bicycle and saying "brmm brmm" is not "prototyping a motorcycle" no matter how often you tell your mum it is.

To be fair there is an enormous amount of infrastructure that goes into these search/Q&A systems that has nothing to do with machine intelligence. Building an entire motorcycle but using a fake engine is, in fact, prototyping a motorcycle.
posted by Tell Me No Lies at 11:43 AM on July 9, 2018 [6 favorites]


If the output of the voicemail-to-text system we have as part of Outlook where I work is any indication, humans are definitely not involved at any point, including during the writing of the software.
posted by soren_lorensen at 11:52 AM on July 9, 2018 [8 favorites]


I've already been trying to turn myself into a mindless, soulless, service robot. I guess that is the wave of the future now.
posted by jenfullmoon at 11:53 AM on July 9, 2018 [3 favorites]


Nowt wrong with being a mindless soulless service robot.
posted by GCU Sweet and Full of Grace at 12:19 PM on July 9, 2018 [3 favorites]


Corporations were already AIs, they just have a lot of human parts.

For now.
posted by Barack Spinoza at 12:22 PM on July 9, 2018 [4 favorites]


Oh holy hell, It's Raining Florence Henderson, I associate "bio break" with a former boss who, I'm pretty sure, would have opted for catheterization if it let him work longer. Shudder.
posted by Making You Bored For Science at 12:25 PM on July 9, 2018 [2 favorites]


Corporations were already AIs, they just have a lot of human parts.

For now.


I'm allegedly human, and I'm only partially composed of human parts.

Oh holy hell, It's Raining Florence Henderson, I associate "bio break" with a former boss who, I'm pretty sure, would have opted for catheterization if it let him work longer. Shudder.

Ha! Yeah - I hate that phrase, too.
posted by It's Raining Florence Henderson at 12:27 PM on July 9, 2018


I'm pretty skeeved out by the idea of telling users outside the lab that they're talking to an AI and having it be a human, but I'm having a hard time articulating why.

People have talked about the dehumanizing aspect of saying that lower-wage workers are computers, but there's also the fact that I'll say stuff to a computer that I wouldn't say to a human, even one I don't know, because it's embarrassing and personal. I'll type my symptoms into WebMD but I don't want a person knowing all the stuff that's happening with my body because it's none of their business. If you tell someone that their information is going to a computer and won't be seen by anyone and then actually it's seen by a person, that's creepy as fuck. It's okay to want to control what people, even anonymous ones, know about you.
posted by Mrs. Pterodactyl at 12:29 PM on July 9, 2018 [22 favorites]


Call it Artificial Artificial Intellgence (AAI), and the cash will roll in!
posted by GenjiandProust at 12:37 PM on July 9, 2018 [5 favorites]


Nowt wrong with being a mindless soulless service robot.
I'm going to build a mindless soulful service robot: "What are you lacking human? Is it R-E-S-P-E-C-T? Do you need just a little bit? Or would you like do some sitting by the dock of the bay? Please respond."
posted by The_Vegetables at 12:46 PM on July 9, 2018 [4 favorites]


It's not explicitly mentioned here or in the article that I can see, so I felt like I wanted to remind people that Amazon's Mechanical Turk was named after The Turk which was "a fake chess-playing machine constructed in the late 18th century". Like they pretty much made it clear what it was for from the beginning. Or Amazon, so many years ago, just thought it was a clever reference without realizing the truth of what they were being was.
posted by skynxnex at 12:48 PM on July 9, 2018 [9 favorites]


People all over the world (everybody)
Join hands (join)
Start a love train, love train
You have 5 seconds to comply
posted by It's Raining Florence Henderson at 12:52 PM on July 9, 2018 [22 favorites]


Beware some of the weasel words in the article... When they say "some receipts were transcribed" this doesn't tell us that there's no AI. Transcription is non-trivial (think of all of the different receipt formats you've seen in your life), and it's easy to imagine being able to do 95% with high-confidence automatically, and then sending the lower-confidence 5% to humans for quick turn around, while also providing more ground-truth to help the automation get better. For the user, putting humans occasionally in the loop means fewer error messages, while also leading to better AI in the long term. Basically a win, so long as the humans are sufficiently indoctrinated on handling PII data...
posted by kaibutsu at 1:23 PM on July 9, 2018 [1 favorite]


Now I'm imagining that the first fully-working self-driving car implementation will be Mechanical Turk for Uber, in which someone in a low-wage country pilots your car for you.

Or maybe housecleaning. You know how your Roomba gets confused sometimes, or doesn't deal well with your kid leaving a mess on the floor?

And those agricultural robots that they're trying to build for weed-picking and ultra-precise herbicide application: Just stick somebody in a basement 12,000 miles away, and let them identify the weeds on a screen. No need to have migrant workers. No need to give them any opportunity to immigrate, either.
posted by clawsoon at 1:53 PM on July 9, 2018 [6 favorites]


kaibutsu: "Beware some of the weasel words in the article... When they say "some receipts were transcribed" this doesn't tell us that there's no AI. Transcription is non-trivial (think of all of the different receipt formats you've seen in your life), and it's easy to imagine being able to do 95% with high-confidence automatically, and then sending the lower-confidence 5% to humans for quick turn around, while also providing more ground-truth to help the automation get better. For the user, putting humans occasionally in the loop means fewer error messages, while also leading to better AI in the long term. Basically a win, so long as the humans are sufficiently indoctrinated on handling PII data..."

I've been at a company doing receipt-scanning in the past. Like some of our major customer relationships were built on the concept. There was nothing like the type of process you describe, and you should not treat this with good faith--the article definitely soft-pedals the problem.
posted by TypographicalError at 2:00 PM on July 9, 2018 [7 favorites]


That was, in fact, the subtitle of the Mechanical Turk service logo until fairly recently.

And it's notable that a "Mechanical Turk" is literally a device for faking AI.
posted by howfar at 2:45 PM on July 9, 2018 [1 favorite]


And I note that it's already been noted in this thread. I'll get me coat.
posted by howfar at 2:46 PM on July 9, 2018 [3 favorites]


Welcome to The Machine...
posted by littlejohnnyjewel at 3:25 PM on July 9, 2018


In Stephenson's The Diamond Age the interactive primer's main push is powered by a human. The main character, Nell and by extension, the Mouse Army, are not completely educated or shaped by an AI system, but by a budding actress hired by the AI creators. The actress, Miranda, comes to care for Nell and becomes a surrogate mother whose communication is completely mediated by technology. We are approaching two parallel, ambivalent paths that converge: AI replacing humans and humans imitating AI to provide missing features both small and large. These paths serve old desires of immortality, creation and maybe actualization virtual and otherwise.
posted by jadepearl at 4:05 PM on July 9, 2018 [3 favorites]


I'll say stuff to a computer that I wouldn't say to a human, even one I don't know, because it's embarrassing and personal.

Agreed, I thought about that while reading the article, especially this bit:
"A team from the University of Southern California tested this with a virtual therapist called Ellie. They found that veterans with post-traumatic stress disorder were more likely to divulge their symptoms when they knew that Ellie was an AI system versus when they were told there was a human operating the machine."

Talking to a robot or even a non-human animal doesn't bring the paralyzing anxiety that you might be judged and criticized for what you say and how you say it, that you'll be wrong or waste the other person's time. Reducing that anxiety is one of the intentions behind literacy programs in which children read to dogs: the child gets practice reading aloud to an audience, and the audience dog is attentive and supportive (snuggly) and will never interrupt to say "Actually, you pronounced that word wrong."
posted by nicebookrack at 5:25 PM on July 9, 2018 [12 favorites]


I've been at a company doing receipt-scanning in the past. Like some of our major customer relationships were built on the concept. There was nothing like the type of process you describe, and you should not treat this with good faith--the article definitely soft-pedals the problem.

Scanning or transcribing? Arbitrary receipts, or just the ones your major customers saw regularly? And was there any human QA of the results?
posted by kaibutsu at 5:59 PM on July 9, 2018 [1 favorite]


I point out to you, Marcus Claire Luyseyal, a lesson from past over-machined societies which you appear not to have learned. The devices themselves condition the users to employ each other the way they employ machines.
(Frank Herbert, God Emperor of Dune)
The danger of computers becoming like humans is not as great as the danger of humans becoming like computers.
(Konrad Zuse)
Automation may ultimately not take our jobs by being better than us at performing them, but by us being worse and worse at effecting even the most simple tasks, such a reading a document for ten minutes. Robots may also not replace our lovers and friends but not by being as good as them and giving warmth and companionship, but by technology making us worse at dealing with real people. Losing to automation may be a much more banal process that we conjecture it to be. Machines may not need to catch up, if we drop the ball as fast as we seem to be dropping it.
(Manuel Cebrian)
posted by runcifex at 7:39 PM on July 9, 2018 [8 favorites]


Scanning or transcribing? Arbitrary receipts, or just the ones your major customers saw regularly? And was there any human QA of the results?

I mean, there are tons of apps with 'coupon services' (for example Ibotta) that pay you back for certain products, and the proof is generally a store receipt, so I mean I'm guessing that someone has solved this problem relatively satisfactorily. Maybe they have only a few regular users or maybe the have a giant call center in some place where people stare at scanned receipts all day.
posted by The_Vegetables at 7:01 AM on July 10, 2018


> Making You Bored For Science:
"Oh holy hell, It's Raining Florence Henderson, I associate "bio break" with a former boss who, I'm pretty sure, would have opted for catheterization if it let him work longer. Shudder."

I use that when gaming. I thought it was more polite (and shorter to type), although I just usually say "Sorry. Bio."
posted by Samizdata at 9:25 AM on July 10, 2018


A company I, uh, knew, was one of the earlier entrants into the voicemail-to-text-message market (the tech, sold to telecom providers, who then sold it as a service to customers). At the time, speech recognition models to handle arbitrary speakers (particularly on cell phones -- awful hardware, awful audio bitrate) was not up to the task. The solution was a call center in India, where low-paid employees would listen to the voicemails and type them up.

There's a crazier version of this that exists right now. There are companies that provide phones for people who are hearing impaired that will provide captions for the phone call so the person with the phone can read what people are saying.

The way this works is that someone in a call center listens to the call and repeats what is said in real time (like a simultaneous translator) to a computer optimized to understand the call center employee's particular voice. The employee has to watch what comes out and correct typos with a keyboard while not falling behind in captioning.

(Commendably, the company I am most familiar with doesn't try to hide this, it is spelled out explicitly under "How it Works" on their website and in their literature.)

So for the brief window where computers are good enough to understand a particular person, but not yet good enough to understand an arbitrary person, there are thousands of call center jobs (in this country, because being able to make out what some people are saying can be tough even for a native speaker) translating from person-to-computer-to-person.
posted by straight at 10:58 AM on July 10, 2018 [3 favorites]


(And probably one of the more pleasant call-center jobs, as they employee spends most of their time helping people say "Is it raining there?" "What did you have to eat today?" and "I love you," to their mom or grandpa.)
posted by straight at 11:11 AM on July 10, 2018 [5 favorites]


kaibutsu: "Scanning or transcribing? Arbitrary receipts, or just the ones your major customers saw regularly? And was there any human QA of the results?"

This is a weirdly hostile way to approach a conversation, to be honest. I'd prefer if you just took my personal direct experience as an input point and not try to interrogate me to figure out if you're right or not.
posted by TypographicalError at 1:05 PM on July 10, 2018 [1 favorite]


I’m getting another image of using humans pretending to be machines, because the humans are seen as cheaper and more expendable.
posted by darkstar at 6:06 PM on July 10, 2018


Humans pretending to be robots and computers is the plot of a Lem novel, except that in it the humans had to be drugged into thinking they were robots. I guess that's another science fiction dystopia that wasn't quite pessimistic enough.
posted by Gygesringtone at 9:01 PM on July 10, 2018 [3 favorites]


That's sort of the plot of Blade Runner 2049 too.
posted by bongo_x at 2:34 AM on July 11, 2018


« Older in the Wired   |   Never Mind Form Follows Function Newer »


This thread has been archived and is closed to new comments