In news headlines, artificial intelligence is sold as miraculous: somehow better at diagnosing than doctors, more effective at design than architects.
Key points:
- AI tech company Appen boasts of having 1 million "flexible contributors" in more than 170 countries
- The term "ghost workers" was coined to describe these contributors, some of whom are complaining about being underpaid
- The company says it uses location-based minimum wage to ensure they're "paying above"
In the narratives spun by start-ups and industry behemoths, AI is the mysterious creation of those few with six- to seven-figure salaries and degrees from ivy-covered universities.
But behind our increasingly smooth experiences online are a vast number of people that label the data that trains these systems, often precariously employed as contractors without sick leave or bargaining power.
Sometimes called "ghost workers", they may be students in America's Midwest, stay-at-home workers in Canada or, as MIT Technology Review reported, Venezuelan immigrants in Colombia doing digital "piece work".
This includes manually tagging photos and video, transcribing audio and categorising text, so when we use a search engine or speak to a voice assistant, the process feels effortless.
This is the "sleight of hand" behind terms like AI or machine learning, according to Sarah T. Roberts, an associate professor of information studies at UCLA, who has spent time with workers in the Philippines who do content moderation for companies like Facebook.
"Nothing in any of those terms suggest that there are legions of humans that are needed," she said.
Some of these legions work for Appen.
A globally significant and long-term player in the AI supply chain, the Sydney-headquartered company boasts of having 1 million "flexible contributors" in more than 170 countries who do these tasks for clients including Amazon, Google and Boeing.
Missed invoices and dead ends
Dispersed across the world, these workers don't typically meet in person.
But in online forums, Appen's "crowd" chats about promising projects, the benefits of flexibility and which data-labelling companies pay the most. They also detail frustrations: missed invoices, slow communication from the company and projects that disappear or don't make sense.
On a Reddit forum dedicated to working online, posts advise "Beware of Appen!!" and "Appen, save yourself time and frustration".
Several Appen contractors who spoke with the ABC under condition of anonymity, due to fears of repercussions from the company, claimed invoices were sometimes missed for a whole month or delayed.
One American worker showed the ABC an invoice tab on Appen's dashboard that showed no invoice generated for the month of June, which they said amounted to hundreds of hours of missing pay.
The company did have an issue with invoicing for one project this year that has now been fixed, according to Appen's spokesperson Christina Golden.
"We have always kept open communication with our contributors.
"If they have a need to share feedback anonymously, we have a whistleblower hotline that is given to all contributors."
But getting in touch with someone who can actually help at Appen is increasingly difficult, according to some of its workers, with slow and seemingly boilerplate responses to problems with jobs or pay.
"I used to get generic 'please allow three business days for tickets to be settled' emails," one said.
"The last three support tickets I sent were automatically closed as soon as I sent them with no reply or action taken on them."
Another said he was invited to participate in projects that simply never materialised.
Appen has turned up frequently in the business press after a recent downturn in financial performance.
In its first half of 2022 results call, chief executive Mark Brayan said revenue had fallen in part due to a slowdown in spending from some of its large customers.
Mr Brayan ascribed this to weaker digital advertising revenue, affecting its ad-related programs.
Some have speculated this will impact how the company treats its contractor workforce — especially when it's competing with competitors that chase workers in locations where they can pay the least.
As Appen's former boss Lisa Braden-Harder recently told the Sydney Morning Herald, "you can definitely do a race to the bottom".
The ABC spoke with workers who said their pay was around or just above minimum wage in their location.
Appen spokesperson Ms Golden said the company used location-based minimum wage to ensure they were "paying above".
"Rates vary by the project due to the different levels of difficulties or expertise required to complete the task."
In fact, "crowd conditions" are named as a potential risk in the company's 2021 annual report, with customers requesting information on worker conditions — particularly around content moderation work, which can expose people to extreme and disturbing content.
'One of the maids sweeping their nursery room'
Technology companies rarely acknowledge the work they contract through companies like Appen, much less the army of people that train their algorithms.
Monash University's Jathan Sadowski, who studies the social impact of digital technology, says this kind of data labour is barely discussed in public. And when it is, it is not granted anywhere near the same level of respect or remuneration as other kinds of high-tech work.
"You don't want to look at the wizard behind the curtain."
Appen's workers described tasks like making voice recordings or taking short videos of certain objects to help develop what they thought was voice command technology. Others seemed to be focused on improving mapping technology or search engine results.
Arden*, an Appen contractor from the US Midwest, spoke about one project in which they would be shown a piece of content — an image, video or string of text.
They would then have to summarise the purpose of the post in one sentence, and find a link to a trustworthy source that touched on the same subject. They would be given a minute or less to review text, but more time for a long video or streamed content.
"[There were] many health-related scams, election fraud, COVID misinformation, fake quotes, faked UFO photos," they said.
"These were our specialty, but I'd also sometimes see calls to violence, self-harm-related content, violence, gore, abuse material and more."
Arden said it was "both fascinating and unnerving to literally train a computer" to do their own job.
"It's very engaging work, but once these computers learn from us, that job is gone. Possibly forever."
Dr Sadowski suggested data annotation work will only become more specialised as companies attempt to build AI systems in industries like medicine or insurance.
But will workers that develop expertise — such as the knowledge of how to spot a melanoma or assess a car damage claim and correctly label it — be recognised as essential to the industries for which they help build these systems, and be rewarded for it?
"I tell my friends and family that while the engineers at Google, Amazon, Apple, etc are like birth parents and private tutors of AI babies, I'm one of the maids sweeping their nursery room and cleaning their clothes," one worker based in East Asia told the ABC.
"I know that I am a replaceable workforce in this business, but Appen had shown such little respect to its contributors that I was constantly reminded of that fact."
*Not their real name.