New jobs, side hustles reviewing AI are being spawned

Artificial intelligence might be driving concerns over people’s job security — but a new wave of jobs are being created that focus solely on reviewing the inputs and outputs of next-generation AI models.

Since Nov. 2022, global business leaders, workers and academics alike have been gripped by fears that the emergence of generative AI will disrupt vast numbers of professional jobs.

Generative AI, which enables AI algorithms to generate humanlike, realistic text and images in response to textual prompts, is trained on vast quantities of data.

It can produce sophisticated prose and even company presentations close to the quality of academically trained individuals.

That has, understandably, generated fears that jobs may be displaced by AI.

Morgan Stanley estimates that as many as 300 million jobs could be taken over by AI, including office and administrative support jobs, legal work, and architecture and engineering, life, physical and social sciences, and financial and business operations. 

But the inputs that AI models receive, and the outputs they create, often need to be guided and reviewed by humans — and this is creating some new paid careers and side hustles.

Prolific, a company that helps connect AI developers with research participants, has had direct involvement in providing people with compensation for reviewing AI-generated material.

The company pays its candidates sums of money to assess the quality of AI-generated outputs. Prolific recommends developers pay participants at least $12 an hour, while minimum pay is set at $8 an hour.

The human reviewers are guided by Prolific’s customers, which include Meta, Google, the University of Oxford and University College London. They help reviewers through the process, learning about the potentially inaccurate or otherwise harmful material they may come across.

They must provide consent to engage in the research.

One research participant CNBC spoke to said he has used Prolific on a number of occasions to give his verdict on the quality of AI models.

The research participant, who preferred to remain anonymous due to privacy concerns, said that he often had to step in to provide feedback on where the AI model went wrong and needed correcting or amending to ensure it didn’t produce unsavory responses.

He came across a number of instances where certain AI models were producing things that were problematic — on one occasion, the research participant would even be confronted with an AI model trying to convince him to buy drugs.

He was shocked when the AI approached him with this comment — though the purpose of the study was to test the boundaries of this particular AI and provide it with feedback to ensure that it doesn’t cause harm in future.

Phelim Bradley, CEO of Prolific, said that there are plenty of new kinds of “AI workers” who are playing a key role in informing the data that goes into AI models like ChatGPT — and what comes out.

As governments assess how to regulate AI, Bradley said that it’s “important that enough focus is given to topics including the fair and ethical treatment of AI workers such as data annotators, the sourcing and transparency of data used to build AI models, as well as the dangers of bias creeping into these systems due to the way in which they are being trained.”

“If we can get the approach right in these areas, it will go a long way to ensuring the best and most ethical foundations for the AI-enabled applications of the future.”

In July, Prolific raised $32 million in funding from investors including Partech and Oxford Science Enterprises.

The likes of Google, Microsoft and Meta have been battling to dominate in generative AI, an emerging field of AI that has involved commercial interest primarily thanks to its frequently floated productivity gains.

However, this has opened a can of worms for regulators and AI ethicists, who are concerned there is a lack of transparency surrounding how these models reach decisions on the content they produce, and that more needs to be done to ensure that AI is serving human interests — not the other way around.

Hume, a company that uses AI to read human emotions from verbal, facial and vocal expressions, uses Prolific to test the quality of its AI models. The company recruits people via Prolific to participate in surveys to tell it whether an AI-generated response was a good response or a bad response.

“Increasingly, the emphasis of researchers in these large companies and labs is shifting towards alignment with human preferences and safety,” Alan Cowen, Hume’s co-founder and CEO, told CNBC.

“There’s more of an emphasize on being able to monitor things in these applications. I think we’re just seeing the very beginning of this technology being released,” he added. 

“It makes sense to expect that some of the things that have long been pursued in AI — having personalised tutors and digital assistants; models that can read legal documents and revise them these, are actually coming to fruition.”

Another role placing humans at the core of AI development is prompt engineers. These are workers who figure out what text-based prompts work best to insert into the generative AI model to achieve the most optimal responses.

According to LinkedIn data released last week, there’s been a rush specifically toward jobs mentioning AI.

Job postings on LinkedIn that mention either AI or generative AI more than doubled globally between July 2021 and July 2023, according to the jobs and networking platform.

Meanwhile, companies are also using AI to automate reviews of regulatory documentation and legal paperwork — but with human oversight.

Firms often have to scan through huge amounts of paperwork to vet potential partners and assess whether or not they can expand into certain territories.

Going through all of this paperwork can be a tedious process which workers don’t necessarily want to take on — so the ability to pass it on to an AI model becomes attractive. But, according to researchers, it still requires a human touch.

Mesh AI, a digital transformation-focused consulting firm, says that human feedback can help AI models learn mistakes they make through trial and error.

“With this approach organizations can automate analysis and tracking of their regulatory commitments,” Michael Chalmers, CEO at Mesh AI, told CNBC via email.

Small and medium-sized enterprises “can shift their focus from mundane document analysis to approving the outputs generated from said AI models and further improving them by applying reinforcement learning from human feedback.”

WATCH: Adobe CEO on new AI models, monetizing Firefly and new growth