Vitalii - stock.adobe.com

AI interview: Krystal Kauffman, lead organiser, Turkopticon

Remote Mechanical Turk workers are responsible for training artificial intelligence algorithms and completing other data-related business processes - we hear about the workplace issues they face

Globally dispersed digital gig workers tasked with training artificial intelligence (AI) algorithms are in the process of coordinating collective responses to common challenges they face in the workplace, says Turkopticon lead organiser Krystal Kauffman, who has been involved in the transnational organising process with other Mechanical Turk workers since 2020.

While the popular perception of AI revolves around the idea of an autodidactic machine that can act and learn with complete autonomy, the reality is that the technology requires a significant amount of human labour to complete even the most basic functions.

Otherwise known as ghost, micro or click work, this labour is used to train and assure AI algorithms by disseminating the discrete tasks that make up the AI development pipeline to a globally distributed pool of workers.

While many companies use the unpaid activity of their customers to complete such work – you do this whenever you solve a ReCaptcha online, for example – many others outsource the work to online platforms such as Mechanical Turk, a crowdsourcing marketplace owned and operated by Amazon that allows businesses, or “requesters”, to outsource various processes to a “distributed workforce”, who then complete tasks virtually from wherever they are based in the world.

Although a large part of Mechanical Turk’s work revolves around accelerating the development of AI and machine learning algorithms, it also extends to completing surveys, validating or deduplicating data, moderating content and conducting research.

However, despite bringing some benefits to workers, particularly around flexibility, the work on such platforms is frequently defined by low wages, long hours and poor conditions.

One reason why conditions remain poor is the complete geographical separation of workers from one another, which makes it difficult to organise collectively. Still, efforts are underway to do exactly that.

An example of this is Turkopticon, a Mechanical Turk “review site” that originally started as a way for workers to swap notes on different tasks or requesters, but has since morphed into an “advocacy organisation” for their collective interests.

Speaking with Computer Weekly, Krystal Kauffman, a lead organiser at Turkopticon, outlines the workplace issues facing “Mechanical Turkers” and other digital gig workers, and their attempts to create collective solutions.

A typical working day

Having worked for Mechanical Turk since 2015, Kauffman says while there is a degree of predictability in the job, especially when a “large batch of work” comes through from a requester that she knows will take a few days to complete, it’s usually quite varied, with the pace of the work changing from day to day.

“One day can be very, very slow. The next day can be very, very fast paced. You’ve always got something in your queue to work on, so it varies a lot,” she says. “You just don’t know what you’ll be working on, which enhances the flexibility of the job, but it also ties you to your computer watching for better paying work to come through.

“If you’re itching to get out of the house, then you don’t want to wait for work to come through. There’s also always a feeling when you’re outside of the house [that you might be] missing something right now.”

Kauffman says one way Mechanical Turk workers, or Turkers, manage the varied pace and volume of work – as well as to avoid waiting around and screen watching – is to self-generate daily, weekly and monthly goals.

“I might be working one day and I’ve passed my daily goal, but I know that this work is there today and it might not be there tomorrow, so I might go above my goal today to kind of anticipate that next slow day,” she says.

It’s not a perfect approach, but many workers find it helpful in managing the daily fluctuations of their workload. “There’s always a good side and a bad side to these things. A lot of hours in front of the computer is not great for everybody, but having that flexibility is super helpful for a lot of different reasons, for a lot of people,” says Kauffman.

She says that, from her experience, this includes people raising children or caring for loved ones, those suffering from different forms of anxiety, and people who find it too expensive to constantly travel and work outside the home.

In Kauffman’s own case, the flexibility gave her a way to earn money after she became too ill to leave the house while studying for a geology degree. “I was nearing the end of my degree and I got sick, and nobody knew what was going on with me, and before I knew it I couldn’t work or go to school or do anything outside of the home, but I still needed to pay my bills,” she says.

She found Amazon Mechanical Turk when Googling work-from-home opportunities. “I signed up, they approved me within a day or two… and I was able to support myself and pay my rent and bills for the next two years while my health was being sorted out. It was easier for me to work from home and set my own schedule because I didn’t know what was going to be a good health day or a bad health day,” she says.

Workplace issues with requesters

Despite the benefits, however, Kauffman points out there are also a number of workplace issues that Turkers are attempting to deal with collectively.

Giving the example of when she had to label aerial photos of border crossings, Kauffman says one of the major issues they face is the lack of transparency about what they’re working on, and who for.

“These weren’t traditional border crossings, but footpaths or tyre tracks. I didn’t know where the area was. I didn’t know why I was marking them. But it really didn’t feel right because I didn’t know what was happening to those people,” she says.

“There are requesters that change their name after a few bad reviews, and then they look like a brand new requester. Right now, Amazon allows that, which is not a good business practice”
Krystal Kauffman, Turkopticon

“Was I contributing to them being picked up and detained? Was I contributing to them getting help? I didn’t know. I eventually stopped doing them and I did this [other] company’s lighter tasks, like marking nesting cranes.”

In another instance, Kauffman says she is aware of a requester who “repeatedly” asks for pictures of people’s feet. “You want to hope this person is a podiatrist and they have a really great reason for this, but they don’t let you know,” she says. “I don’t know a single person who has sent in a picture of their foot, but it really makes you wonder what you’re doing and why you’re doing it.”

She adds that while workers will sometimes find out what they’re working on either upfront or afterwards, other times they don’t find out at all, and “those are the ones that kind of make you think twice”.

Part of the issue here is that the requesters, which could be anyone from private businesses to individual academics conducting research, are not obliged by Amazon to provide their real identities, and can change their identification at any given time.

While Turkopticon provides a forum for workers to discuss incoming requests, Kauffman says this means workers are largely left to their own devices in figuring out what they’re working on and evaluating whether it’s something they’re comfortable with.

“There are requesters that change their name after a few bad reviews, and then they look like a brand new requester. Right now, Amazon allows that, which is not a good business practice,” she says, adding it could easily step in to require the requesters to use their real names or limit the number of name changes allowed.

“They could put into their terms of service that some of these things needs to be better explained, or at the end of a project [make sure] we’re told how our data was used and then have the option to remove our data if we don’t want it used in that way,” says Kauffman.

“But ultimately, once we put it out there, we don’t have any control over what is done with it. And that’s kind of a helpless feeling, because you’re thinking, ‘I’m a good person. I wouldn’t ever do anything to hurt anybody else’, and you really hope the work you’re doing isn’t.”

Responding to questions from Computer Weekly, Amazon says it is not the employer of those who complete tasks, and that the reward amounts for tasks are set entirely at the discretion of the requester while workers are free to accept any tasks that fit their needs. It adds the platform also does not play a role in price setting, and that all work and price agreements are taken at the sole discretion of the worker.

Power imbalances

Another issue is the power imbalance between Turkers and requesters, which Kauffman says has led to situations where people aren’t being paid for work they’ve completed because of the latter’s ability to mass reject work.

“You [as a requester] could reject all of my work even though it was done properly. You could change your name and disappear. And now you’ve got free data,” she says. “There are workers who [have lost] time, money, but worse yet, every time a worker has a rejected task it lowers their approval rating.

“If you have 100 or 200 tasks that are mass rejected, that can really damage an approval rating, especially for a newer worker. That can take a worker out of the workforce entirely, as most requesters require a 99% approval rating, some of them 95%. But once you drop below that, you’re not going to be able to find much work, and it’s not going to be good quality work.”

Kauffman adds dropping below the 95% approval rating is essentially a “death sentence” when working for Mechanical Turk, unless you have the time or patience to turn it around.

Giving the example of a newcomer who received a mass rejection of around 200 tasks after having completed a few thousand others without issue, Kauffman says that worker is now stuck doing the least desirable, lowest-paying tasks due to something that was completely out of her control.

In Kauffman’s estimation, the worker will now have to complete over 200,000 tasks to repair her approval rating and get it back up to the 99% mark. “I can tell you, as someone who did this day in and day out for so long, that is a feat that takes more than a year to do, if not several years,” she says.

“She is just plugging away hoping that she’ll be able to hit it, and I admire her so much for sticking to it because most people would throw up their hands and walk away – she’s the exception.”

Not all workplace issues affect all workers equally, with Kauffman adding that the geographic distribution of work and rewards creates a sense of unfairness. Some requesters, for example, may choose for the work to be exclusively completed in the US. Sometimes this can be for valid reasons, like the funding coming from a US government grant, but other times Kauffman says many think they’ll get better quality work by selecting US workers, “and that’s not the case”.

While some workers from global south countries may only have four or five tasks available to them at any given time, Kauffman herself, living in the US, has several hundred available to her.

“There definitely is a difference with that. Also, there are still countries in which Amazon pays earnings to people in Amazon gift cards,” she says. “Amazon will say they’ve been trying to get everybody paid into their bank accounts, and we’ve seen some improvement over the years, but there are still some countries [where it happens]. Somebody I know very, very well, a friend of mine located in the global south, still gets the Amazon gift cards. When you’re trying to make a living, that doesn’t really pay the bills.”

Kauffman adds that Amazon also restricts the use of gift cards, so those handed out to workers cannot be sold on or used to purchase other gift cards, such as those offered by financial services firms like Visa that could be used to pay bills.

Responding to Computer Weekly, Amazon says there is a Participation Agreement and an Acceptable Use Policy to ensure there is no abuse in the marketplace by either those requesting work or those agreeing to do tasks. It adds that the platform also does not play a role in price setting, and that all work and price agreements are taken at the sole discretion of the worker.

‘The union treatment’

For the Turker dealing with the mass rejection and others facing similar issues, Kauffman says “there is no recourse whatsoever” available to them from Amazon. While the company has previously told Turkopticon that it is implementing changes to limit requester identity issues, it was unwilling to share details of its actions. 

“Whatever Amazon is doing to solve the problem, it isn’t working,” she says, adding it’s not about preventing requesters from rejecting legitimately poor work, but reaching a power balance between the parties.

“What Amazon does not seem to understand is that when you have happy requesters and happy workers working together, the data quality is better, the work gets done faster, and there are fewer times that the requester may have to relist their tasks”
Krystal Kauffman, Turkopticon

“What Amazon does not seem to understand is that when you have happy requesters and happy workers working together, the data quality is better, the work gets done faster, and there are fewer times that the requester may have to relist their tasks.”

Kauffman adds while Turkopticon does have meetings with Amazon from time to time about fixing various workplace issues, the company tends to use the same tactics it uses against its unionised warehouse workers.

“That is to drag their feet very, very slowly in hopes that we’ll give up,” she says, adding that while Amazon does offer very small concessions, they have nothing to do with the demands of Turkers. “It’s the same way they treat other organisations that are unionising within Amazon, but we’re getting the unionisation treatment without unionising, which is really quite frustrating.”

Unlike other workers who share specific workplaces or are in relatively close physical proximity, such as those working for ride-hailing apps in the same city or region, click workers are much more geographically dispersed, making unionisation a much more difficult task.

While part of the issue is their employment classification – in the US, you have to be classified as a full-time employee to unionise – Kauffman says the fact that Mechanical Turk operates in roughly 190 countries, each with their own employment laws and protections, makes forming a union for workers incredibly difficult.

“Where do you even start? The other threat is if one country says, ‘We’re going to demand better conditions, better wages’, it’s really easy for Amazon to say, ‘Well guess what, we’re not going to work in this country any more’, as they have 189 others to pull data from.”

While many workers can gain leverage over their employer by threatening to withhold their labour and ultimately striking, putting pressure on Mechanical Turk in this way is difficult due to the algorithm-based nature of the work, whereby the system can simply reroute the task to someone potentially halfway across the world who is either willing to do it, completely unaware a strike is even happening, or unable to engage in the action due to their circumstances.

“It isn’t that the people who won’t strike are bad people if, especially in other countries, this is the only way of making a living, the only way to put food on the table for them,” she says.

Responding to Computer Weekly, Amazon said because the work available on the platform is not physical and does not come with the commitment or requirements of a traditional job, it is therefore inaccurate to compare Mechanical Turk work to a traditional job, especially given its task-based nature and since people who complete tasks on the platform are not Amazon employees.

Global virtual organising: ‘It takes time’

In lieu of being able to form recognised unions with formal bargaining power and the ability to coordinate effective strikes, Turkopticon is instead engaged in what Kauffman describes as “global virtual organising”, which entails a slow process of building connections transnationally through “every corner of the internet”.

This means being active on almost every conceivable social media platform, given that not every platform is available in every country, creating blogs and newsletters to spread awareness of problems, and opening worker-only community forums to discuss anything from specific work tasks to workplace issues.

As a result of being able to communicate with more and more workers, Turkopticon identified mass rejection as a top issue, and has been able to organise a campaign against the practice.

Kauffman says that initially, a big challenge was building trust between people who had never met each other in person, especially given fears about Amazon’s surveillance of their activities.

“We had a really hard time reaching international workers to begin with. A lot of people we approached were very nervous to even speak because they were afraid their account would get suspended if it was found out by Amazon they were speaking to us,” she says, adding that to protect people’s privacy from potential infiltrators, Turkopticon doesn’t use people’s names or any other specific identifiers.

“Without a doubt, I would say that in 98% of our forums, we have someone from Amazon trying to sneak in. Luckily, we know the signs, catch it, and we boot them within a couple of minutes if they do make it in.”

As a result of this four-year-long connection-building process, Kauffman says more and more workers from outside the US are getting involved.

“A lot of people are sitting at home thinking they’re being mistreated. They don’t have any recourse. They don’t have a voice. They don’t have any power. When in reality that couldn’t be further from the case,” says Kauffman, adding she attends monthly meetings with other worker-focused Amazon groups in which the struggles faced by different kinds of workers, and how they can work together to overcome them, are discussed.

“We just need enough people to come together from wherever they are, and join us in fighting for some of these things. I truly believe that these past few years, we’ve been building an organisation, and now we feel like we have the strength to really go forward with this and make a difference for workers around the world.”

Computer Weekly contacted Amazon regarding every claim made by Kauffman about working on the Mechanical Turk platform.

Amazon spokesperson Brad Glasser said: “Mechanical Turk is an online crowdsourcing marketplace that connects those seeking help on virtual tasks with individuals who might be interested in completing those tasks. While many of these claims lack context and important details, we look forward to continuing our dialogue with MTurk users as we look for opportunities to further improve the marketplace.”

Computer Weekly further asked Amazon for clarification on which claims lacked the necessary detail, but received no response on this point.

Read more about artificial intelligence

Read more on Business intelligence and analytics