Is AI Job-Thief or Superpower? An Insight into the Ground-Level Effect of Generative AI on Work

The rumors became roar: Generative AI is here and revolutionizing the fabric of our work life. But beyond the headlines and armchair speculation, what is happening in the trenches of the everyday work? Are we on the brink of large-scale automation or an unprecedented extension of human capabilities? An unprecedented study from massive corpus of real-world interactions with the Microsoft Bing Copilot - Working with AI: Measuring the Occupational Implications of Generative AI by Kiran Tomlinson, Sonia Jaffe, Will Wang, Scott Counts, and Siddharth Suri (arXiv:2507.07935) offers some of the most unequivocal answers yet, revealing the veil on the work activity people actually do with the AI and the ways in which the AI itself works. This is not theory but close look at hundreds of thousands of anonymized conversations that give us glimpse into the nascent dance between humans and their co-pilot AIs. What we end up with is nuanced profile, one that defies popular narratives and needs nuanced treatment of the work of the future.

From Hype to Hard Data: Unpacking AI' Role

To have insight into the impacts of AI, the researchers looked into corpus of 200,000 anonymized and privacy-cleared dialogues between Microsoft Bing Copilot and its users. Their state-of-the-art approach deviates from two crucial components of every interaction between humans and AIs: the user goal and the action of the AI. Just think about it: the user goal is what the human seeks to accomplish courtesy of the AI. So, for instance, if you're experiencing technical issue, then your user goal would be something like "resolve computer problems". The AI action, conversely, is the activity the AI actually performs in response – in the same case above, the AI’s action would be to "provide technical support".

The reason we draw this line between user goal and AI action is significant is that we can better recognize just how much the former is complementing human activity and yet in some cases actually doing the human’s work. So what do people do most with generative AI? The research reveals the most common user aims in Copilot dialogs are information gathering, writing, and communicating with other people. These are the things people are most turning to the aid of the AI for. And what about the AI itself? The most common behaviors of the AI most often place it in the service role. It' most often providing information and assisting, writing, teaching, and advising. Look at the following illustration: while the user is perhaps trying to pull out information for report (user goal), the AI is perhaps providing the information (AI action). Notice this lack of correlation: in 40% of the dialogs the user goal and the AI action applied disjointed sets of work activity. The AI is more adept at advising and teaching and the user is using it for information gathering, reading, and research.

Work activities with the most extreme ratios between user goal and AI action activity share

Source 1: Working with AI:Measuring the Occupational Implications of Generative AI

To quantify this effect on spectrum of occupations, the researcher developed an AI applicability score per job. This is not just raw usage in vacuum; the score includes all the following:

• If the work of an occupation is performed using AI in "non-trivial" proportion.
• How well the activities are completed aided by the use of AI (according to user input and classifier of task accomplishment).
• AI’s range of influence – the distance from negligible to significant contributions to an activity of work.

These highest-scoring jobs on this applicability of AI scale are mostly knowledge work and communication-intensive jobs. Those would be jobs like:

• Interpreters and Translators,
• Authors and Writers,
• Sales Professionals of Services,
• Customer Service Reps,
• Computer and Mathematical jobs,
• Occupations in Office and Administrative Support.

Top 40 occupations with highest AI applicability score

Source 2 Working with AI:Measuring the Occupational Implications of Generative AI

Conversely, least applicability on the basis of AI is the case of jobs requiring manual work, operation of machines, or body contact with humans. Think about jobs like Nursing Assistants, Massage Therapists, Dishwashers, or Pile Driver Operators. This is not surprise – an LLM bot is essentially not suited for manual tasks or direct control over machinery.

Bottom 40 occupations with lowest AI applicability score

Source 3 Working with AI:Measuring the Occupational Implications of Generative AI

Interestingly, the paper' results follow previous estimates of the effects of work from AI fairly closely, .g., from Eloundou et al. (Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock. GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702):1306–1308, 2024.) and show very high correlation (0.73 on the occupation level and remarkable 0.91 on the most aggregated level of the broadest occupation group). This suggests that our intuition about the potential scope of the coverage of AI is about right on the average.

Past the Productivity Paradox: Augmentation Rather Than Automation

Common debate often frames the impact of AI in zero-sum modes: augmentation (AI works with humans to do things better) or automation (AI takes over tasks entirely). But the sources note that the capability of new tech has no direct impact on downstream business decisions. The venerable analogy of the ATM reveals the point just so: even as the ATM automated the very core function of the bank teller, jobs for bank tellers actually increased as branches proliferated and the tellers became relationship-builders.

This work substantiates the nuance. Even though action is taken by the AI, its core value today lies in amplification. Research and writing work, for instance, not only represent the most common user goals but also have the most positive user satisfaction and the most work completed. That means Copilot is actually proving useful in the above work, indicating that it is being used as an incredibly efficient productivity amplifier. It' more capable on the writing and research sides of knowledge work than on advanced data analysis or visualization design. "The activities that AI performs are very different from the user goals the AI assists: in 40% of conversations, these sets are disjoint." This alignment of user objectives and the activity of the AI works to further highlight this complement. In support of the user, the AI assists the user on more work than it accomplishes directly. The AI then moves into the mode of coaching, training, and advising and helps the user accomplish his or her knowledge work activity. For example, Copilot might help the person prepare meal by recommending recipes but actually doing the work of preparing the meal. This suggests an interactive rather than purely substitutable relation.

The Shifting Sands of Skill and Income: Socioeconomic Correlates

One of the most popular questions about the effect of AI on the labour market is about wages and schooling. Will the effect be greater on low-wage labour or on high-skilled jobs? Analysis found "weak and inconsistent relationship" between applicability score and average occupation wage for AI. Whereas prior work forecasted substantial impact on well-paying jobs, this usage data provides the inverse scenario and suggests the relationship is less direct than previously thought. Sales and office support jobs with high employment rates have relatively low wages but possess high applicability for AI. If we look at the schooling requirements, there' slightly more evident pattern: Bachelor' jobs have higher applicability score for AI than do jobs requiring less schooling. But the researchers note that there still exists massive overlap and extensive distribution of potential impact even in every wage and every education bucket. This suggests the narrative about the impact of AI only being driven in some parts of the workforce is incomplete.

What this suggests has important implications for adaptation and reskilling policies. If the impact of AI cannot be perfectly divided according to wage or schooling tiers, then the adaptation issue becomes more pervasive. This makes the argument for cross-cutting policies over group-oriented policies even more important.

Navigating the Uncharted Waters: What We Still Don' Know

Even as this work provides unprecedented insights, it also signals the vastness of the unknowns that lie ahead. The researchers candidly note the following limitations:

• Single AI Platform: The data is only from Microsoft Bing Copilot. The other AI tools may possess different usage behavior and user base.
• U.S.-Centric Focus: The research relies on the O*NET database and thus has U.S.-centric focus. The results may not generalize globally.
• Snapshots in Time: The potential of AI is constantly changing. This research is snapshot and the "frontier of the applicability of AI to work" is constantly in motion.
• Work and Leisure: It is fundamentally challenging to be entirely sure whether discussion pertains to work or leisure.
• More Than Tasks: Work divided into tasks, while popular, does not capture the "glue of connections" – the fuzzy work of value added.

These limitations point out key questions still unanswered and signal key future research directions:

• Work of Refactoring Duties: How will work tasks overall be transformed once organizations redefine what people and AIs do?
• New Professions Emergence: In the past, new technologies have given rise to completely new professions. What professions will be created by AI?
• Shifting Abilities of AI: How will the dynamic boundary of the abilities of AI continue to alter its work applicability over time?

A Call to Reflective Action

This is not only tech phenomenon; it' societal one. This work provides convincing data-driven proof that today' AI is embedded in much of today' knowledge work and communication. It' even helping us to gather information, write, and communicate better, almost entirely in the role of helper. The "human-in-the-loop" is far from disappearing, but the nature of the loop is essentially changing. Organisations must be tactically considerate in the ways they integrate AI so as to maximize augmentation and unlock the potential of the workforce rather than target and try to replace them. Policymakers face the near-term obligation of building lively reskilling programmes and safety nets that account for the widespread, not narrowly designed, impacts of AI.

For individuals, the take-away is straightforward:

• lifetime learning,
• flexibility and
• predisposition to take on new models of human-AI co-work will be essential.

Work' future is not predetermined point but landscape we ourselves are forging. If we understand the practical ground impacts rather than react to exaggerated reports, we collectively can steer this massive technology towards future that elevates human abilities and supports inclusive economic growth. The debate has begun; the difficult work of thoughtful adaptation remains.