19% of workers may see more than 50% of their work wholly stolen by Generative AI.

In some cases, that number increases to 100%.

Your job, dead.

These are among the conclusions from OpenAI’s latest study that’s making headlines around the world regarding the future of human work alongside Generative Pre-trained Transformers, or Large Language Models.

This study is especially relevant because it also includes something we’ve never seen before; we also asked GPT-4 its opinion on the matter.

With that, the results include some of the most controversial and unprecedented takes we’ve ever seen regarding our future job prospects with AI, some of which will leave you very worried… or happy.

But should we take these results seriously or is it just overhyped propaganda?

A Threatening Result

A team of researchers from OpenAI and the University of Pennsylvania has collaborated in a combined effort to measure how impacted are certain jobs or occupations to Large Language Models (LLMs) like GPT-4 or LaMDa.

And the results are genuinely unexpected.

But first, how have they measured such impact?

The exposure rubric

The researchers define the concept of exposure, the main scoring unit to measure the impact of AI in a job, as a measure of whether access to a GPT or GPT-powered system would reduce the time required for a human to perform a specific work activity or complete a task by at least 50 percent.

But how do we define what’s a job?

Using the O*NET US labor database, this database divides a job, or occupation, into tasks, which are then formed by different Detailed Work Activities, or DWAs.

For instance, taking a literal data point from the O*NET database, a Computer System Engineer could have ‘monitor computer system performance to ensure proper operation’ as a DWA, the task being ‘monitor system operation to detect potential problems.’

Simple, right?

But how do we measure exposure?

In short, a task was defined to have four different levels of exposure:

• E0: No exposure; LLM doesn’t reduce task time by at least half
• E1: Direct exposure; LLM reduces task time by at least half
• E2: Exposure by LLM-powered apps; additional software could reduce task time
• E3: Exposure with image capabilities; LLM + image system significantly reduces task time

Then, after having all these occupations fragmented into different tasks and detailed activities, as well as the common guideline to measure exposure, the researchers asked human experts and GPT-4 to define the exposure of certain DWAs and tasks to LLMs.

And with that, the results came, some of which are unprecedented.

Seeing this, did humans and AI agree?

A common agreement

The short answer is yes.

Both GPT-4 and human annotators shared similar opinions regarding how exposed a certain job is to LLMs, both from a standalone perspective (direct exposure, or E1) with +80% agreement, and when using LLM-powered software (E2 and E3) exposures, also sitting around 80% agreement.

This can be seen in the image below. As the majority of points follow a trend line, human and AI predictions are positively correlated (they share similar answers):

And with this, we enter the danger zone.

What were the most impressive results?

Whole lotta’ exposure

First and foremost, and probably the least surprising of all results, is that there are many jobs exposed to AI.

Specifically, 49% of workers could have half or more of their tasks exposed to LLMs in one way or another.

Naturally, you would be inclined to think that this will probably affect lower-wage jobs the most, right?

Well, if you have a high-paying job, brace yourself.

Your boss could be fired

You read that right, one of the most insane predictions from this study is that higher-paying jobs are more exposed to AI.

Unexpected, huh?

This can be seen in the image below, where both humans and GPT-4 have similar conclusions; the higher your wage, the higher your exposure (although it decreases at a certain point).

Middle-managers beware?

But what about specific occupations… should you be scared?

Usual suspects and crazy predictions

When I saw the results for some occupations, I really wasn’t surprised, as the usual suspects like programming and writing were looking hella’ worrying.

For instance, writers have more than 80% of their tasks exposed to a combination of direct and complementary exposure, which in a nutshell means that the great majority of writers are potentially exposed to LLMs.

And if we consider more high-end complementary technologies, or E2/E3 exposure, the number increases to 100%, according to human annotators.

Not looking good.

But other results truly surprised me.

Amazingly, mathematicians could have 100% of their tasks exposed to LLMs. Yes, mathematicians.

What are your thoughts on that?

Moving forward, among other very controversial claims, this research also “confirms” some other expected outcomes.

Demographic discrepancies

The research also concludes that Black and Hispanic groups have negative exposure to AI, even though demographic groups are evenly distributed across occupations.

In my opinion, this just confirms that these groups have access to more routine, physical and, thus, lower-paying jobs, which just signals how much work is there to do in the US regarding this matter.

And up to here is where most Twitter threads, LinkedIn posts, and Youtube Shorts will reach in search of clout.

But the real question is, should you really start looking for a new job?

Limitations, Hope, & a Prediction

I’ve seen a lot of people announcing the arrival of doom’s day cause of this research, but I bet the majority of them haven’t read the full paper.

Because some elements have to be taken into account before reaching any conclusions.

Potential bias

The researchers are very clear on one thing: the people that participated in the predictions aren’t experts in every occupation that they measured.

Hence, a significant drawback of the approach is the subjective aspect of the labeling process.

In the study, the human annotator group had limited occupational diversity, potentially leading to biased evaluations regarding the performance and efficiency of AI when executing tasks related to unfamiliar professions.

For example, can an engineer accurately describe all tasks a mathematician does in their job?

Of course not.

In fact, one of the skills most negatively correlated to AI — not exposed — was critical thinking. Therefore, if GPT-4 claims a 100% exposure to tasks performed by a mathematician, does this mean that mathematicians don’t use critical thinking as a part of their daily job routines?

Come on, let’s be real.

This simply clarifies that these results should be validated using diverse human annotators.

But besides the question about how well were tasks identified for each occupation, we also need to consider the actual framework.

Can all activities in a job be described in tasks?

Soft skills and physicality

Plenty of jobs include a decent part of their outcome in human relationships and physical activities, unnegotiable barriers to AI today.

Sure, multimodality with visual perception will allow machines to understand tone, expressions, and gestures as well as humans, but there’s a clear hurdle when it comes to relationships.

For instance, salespeople and other very emotionally-focused jobs include activities that aren’t going to be exposed to AI anytime soon.

But then, how do we define a job and the importance of hard skills and soft skills?

This task-based framework surely doesn’t seem like the ideal way, probably because humans haven’t really figured out still how to measure the importance of trust, relationships, and other soft skills in our daily work in a numerical and measurable manner.

Conclusively, what should we take from these results?

A biased but based conclusion

I feel this research brings light to a “problem” that workers are going to face over the following years.

The “new technology era” that Bill Gates has stated AI is bringing to our lives will surely have a huge impact on how we work.

However, the results are potentially very biased and inclined to increase exposure way above the threat that LLMs really represent today.

But that doesn’t mean we shouldn’t be afraid.

Technologies like Robotic Process Automation (RPA), OCRs, or rule-based Chatbots are already taking their fair share of work from humans, and LLMs are an extremely easy way to enhance the capabilities of these technologies to extend to newer processes and jobs, further declining the relevance of humans in routine-based or predictable jobs.

And we still need to account for the impact multimodal models like GPT-4 will have once they become the norm, models that not only have amazing linguistic capabilities but also have human-level perception skills.

And adding insult to injury, we can’t forget about emergent behaviors, new capabilities that LLMs develop as they increase in size, outcomes that not even the creators of these models, like OpenAI, can predict.

In a worst-case scenario, when these technologies become usual suspects in our lives, the threat not only won’t be smaller than what OpenAI predicts, but could potentially be much, much bigger.

A final word

If you’ve read this article, you’re now ahead of 95% of society when it comes to AI, and probably 95% more afraid of it also.

But with AI, being prepared can watertight your future and even allow you to benefit from it.

By subscribing to my weekly newsletter, you’ll deep dive into innovation topics in an easy-to-read, digested manner, providing you with the latest news and insightful innovations on Crypto and AI.

That way, you’re one step closer to benefiting from AI, not being affected by it.

Become aware of the technologies that are shaping our future. Be ready to profit from that knowledge.

You also have the option to become a Medium member, opening yourself to millions of articles related to any topic of your interest and, in the meantime, supporting this humble writer.

Join here! Don’t be afraid to level up your game.