The growing applications of Generative AI Large Language Models (LLM) has kindled uncertainty and anxiety about jobs and occupations for the future. Students and Professionals in all fields of study and occupations are in a dilemma if their education, experience, and expertise will be replaced by AI. Here in this blog, I put forth my points and discussion on why Artificial Intelligence (AI) will not replace jobs. On the contrary, I justify that the need and expertise of people from all occupations will be required even more than ever with the growth of AI and LLMs.
The Anthropic team on March 5, 2026, published a paper from their study of Labor market impacts of AI. The paper is accessible through the URL Labor market impacts of AI: A new measure and early evidence \ Anthropic1. This paper states 8 occupations – Computer & Match, Business and Finance, Office & Admin, Management, Legal, Arts & Media, Architecture & Engineering, Life & Social Services to have high theoretical susceptibility (greater than 80%) of being replaced by LLMs. The below radar graph taken from the paper shows the theoretical AI coverage (in blue) and Observed AI coverage by Anthropic (in red). The theoretical AI coverage is taken from another study by Eloundou, et.al2
Below is the same information presented as a bar chart.
As per this chart,
- Computer & Math has the highest theoretical susceptibility percentage at 96%. Out of this 96%, Anthropic measure of job coverage is 32% which says roughly one-third of the occupation can be covered with Anthropic LLMs.
- Business & Finance has the second highest theoretical susceptibility percentage at 94%. Out of this 94%, Anthropic measure of job coverage is 28% which says roughly little less than one-third of the occupation can be covered with Anthropic LLMs.
- Office & Admin has the third highest theoretical susceptibility percentage at 94%. Out of this 94%, Anthropic measure of job coverage is 42% which says roughly half of this occupation can be covered with Anthropic LLMs. That’s a high percentage!!!
This list is followed by Management, Legal, Arts & Media with high theoretical susceptibility.
A quick overview of the chart shows those occupations which have a high mix of digital work are marked with high theoretical susceptibility of LLM replacement. Those occupations which require both physical actions and digital outputs are marked with medium susceptibility such as Healthcare, Social Services, Healthcare Practitioners. Occupations such as Installation & Repair, Construction, Agriculture, Food & Serving, Grounds Maintenance require physical presence and are least susceptible to LLM replacement. This is intuitive since LLMs generate only digital outputs and cannot perform physical actions.
This study gives an interesting perspective for analysing the employment prospects in various occupations. Before getting into this analysis, it is worthwhile to briefly mention the evolution of such AI technologies that has brought us to this current situation. Further, providing a big picture of the model building process in AI technologies will help to discuss and analyse the participation of human workforce in their respective occupations.
Evolution of Digital and AI technologies
Digital transformation
As data was organized and stored in structured format such as relational databases, spreadsheets, etc.., it became easier to come up with rules to categorize data and attach labels. To minimize manual work, computer programs and applications were developed to execute well-understood rules paving the way to Automation of work. Businesses went through digital transformations. With clearly defined rules, it was possible to classify if a record is a valid or invalid record, and so on. It was possible to calculate the prices based on the data available in the record.
AI/ML transformation
While rules were well-defined and definitions were concrete for digital transformation, there were many scenarios which had gray areas. It was not possible to determine exactly the outcomes such as classifications or prices. This brought up the opportunity to apply statistical models in business applications. These were probabilistic models. The outcomes from these models were probabilities for each category or range – e.g., the probability that this transaction is a fraudulent transaction is 80%, not a fraudulent transaction is 20%, the probability the price will be in the range $50-$60 is 72%, the price will be in the range $80-$9 is 18%, the probability that this image is a box is 90%, and so on.
This transformation was made possible by the rapid deployment of Machine Learning packages and development of Python, Spark, Cloud and other technologies. Several Natural Language Processing (NLP) packages were enriched with advancements which allowed for efficient parsing, contextualization, and annotation of text data. Similarly Image, audio, video augmentation packages and models enabled robust handling of data in these unstructured formats. Advanced NLP technologies such as BERT and others automated summarizations of huge texts.
LLM transformation
OpenAI released the first version of ChatGPT in late 2022 and this has been an amazing disruptor for the workforce. This was based on LLM which stands for Large Language Models. LLMs go further by mimicking the human way of working. It does not just summarize text or data but does reasoning, decision-making, multiple-step executions which are characteristics of the human way of approaching work. LLMs are becoming robust and intelligent with each release from major technology vendors. The LLM models currently available include OpenAI’s ChatGPT, Google’s Gemini, Anthropic’s Claude, Meta’s Llama, Amazon’s Nova and a few more. When these models can mimic human way of reasoning and working, they obviously pose a threat to the workforce as many tasks in various occupations can be done by the LLMs.
Building AI technologies
The core ingredient with the evolution of each technology is “data”. At every stage, data has been the ingredient that was used to come up with business applications and to evolve such technologies. Thus availability of data, quality of data, processing of data, and communicating data are crucial for successful model-building. The goal in building these technologies is to make the machine learn the data, process the data and provide decisions/answers with accuracy. reliability and reproducibility.
If the machine has to learn the data, then it has to be fed with the data. The machine then builds a model using the data and stores the metadata ((data about the data) and various parameters and at this state the machine is said to have acquired “Intelligence”. Later on, when some information or answer is sought from the machine, it refers to its model to provide the outputs. This process of machine gaining intelligence begins from the humans by feeding curated, high-quality data. Only if high-quality, curated data is fed by professionals (from now I will refer to humans as professionals), the machine will gain the correct intelligence. So what is this high quality, curated data that is fed into the machines as the first step in the model-building process – “training dataset”.
What is a “training dataset”? It is a dataset which has both questions and answers. By providing both the questions and answers, the machine is made to learn about the data and to also relate the answers with the data so that when a completely new data without answers is fed, the machine can use the intelligence it acquired, to provide the correct answer just as professionals would do.
In this diagram, I have given a high-level schematic overview of the “Intelligence” acquiring process. It starts with an input dataset.
As the diagram shows, model-building is an iterative process. All those action items highlighted in orange (yes, as correctly noticed they are at the beginning and at the end) requires active participation of professionals. It is just not possible to build such models without involvement of professionals. As it is an iterative process, professionals have to make the decision if more rounds of iterations are needed.
As businesses evolve or new data arrives, professionals have to be involved in building the models and assessing if it is successful or needs more work. They also have to figure out how to improve the model and how to rectify any incorrect answers or insufficient responses that are received as outputs from the models. Thus comes the challenge, opportunity and evergreen need of professionals with domain knowledge/expertise.
Importance of high-quality input training datasets
In AI/ML model building, the training dataset in most situations is a labeled dataset, i.e. each input record will have a manually assigned label (answer or output). The machine while building the model finds the patterns within the dataset and tries to relate it to the assigned label so that in the future if it sees a similar pattern, it will predict an appropriate label to it. The phrase “Garbage-In, Garbage-Out” is an apt mention here. If the given data is not correct or of poor quality, so will be the output too.
In Generative AI LLMs, the training dataset consists of prompts, input materials and outputs. Unlike in non-generative AI/ML model building, LLM prompts are free-flowing natural conversational style of inputs. The prompt is given in such a way that it is like assigning the task to the LLM. It should contain all the background information, context, goal, directions to execute the task and all input materials that the LLM will require to generate the output. The input material could be files in any format or free text and the output too can be a free text or an image(s) or video(es) or audio(s) or file(s) or a combination of all output types. It all depends on what is instructed to the LLM. As said earlier, the ultimate goal of the LLM is to behave as a professional by being able to do reasoning, decision-making, judging, and/or executing multiple steps.
Job outlook for professionals across occupations
It appears AI, ML, NLP, Image processing technologies have been for decades and are stable and matured. But then why is there anxiety or uncertainty now? It is because of Generative AI where LLMs are trained to behave just like professionals. Hence if a task in an occupation can be done by a LLM, then there is a high risk of that task becoming extinct for professionals. But here is the reasoning that such occupations may not become extinct:
- As described in the “Intelligence” acquiring process, the crucial steps such as annotating input training datasets, assessing model outputs, improving correctness and quality of training datasets will always require the domain knowledge and expertise of professionals.
- E.g., if the LLM outputs a software code, a professional has to check the code, debug the code if necessary. Imagine debugging someone else’s code (here it is LLM) when it is challenging as such to debug one’s own code!
- LLMs can only create digital outputs. Not all tasks in an occupation will need only digital outputs. It will also require physical actions such as conversation, experimentation, discussion, explanation, and so on. Even if a task is only about generating digital output, a professional is needed to maintain the quality of the model and to upgrade the model to newer versions with evolving/changing business needs.
- E.g., if a new version of Java or Python or some web-programming language is released, then the LLM will not be aware of the newer version. A professional will be needed to impart the intelligence to the LLM.
- Some occupations will need both digital and physical work. For such occupations, e.g. teaching, content or lesson plan can be created digitally using LLMs. But again, the correctness has to be verified, and the teacher has to be present physically to deliver the lessons to the students.
These reasons drive me to justify that Generative AI and professionals will be working alongside and will be complementing each other in the future. Generative AI is to be seen as an opportunity to refine/develop skills to derive maximum benefit out of work. It is to be capitalized as a medium for professional growth and productivity enhancer. It will redefine the way the work is done but will not essentially replace an occupation. It may replace certain tasks done within that occupation but will not entirely replace the occupation. The future workforce will have LLM-powered occupations just as we have AI-powered digital tools/software.
Conclusion
In this blog, I have put forth my points and reasoning to justify that the future workforce will require professionals from all occupations irrespective of their characteristic risk of AI susceptibilities. Professionals will continue to contribute to their roles by executing traditional responsibilities as well as capitalizing on LLMs to increase quality and productivity in their work. Hence it is important that one chooses their career path and occupation based on their interest and aspirations rather than by evaluating based on AI susceptibilities and to be creative and innovative to work alongside LLMs.
Richard Scarry’s Best Busy People –
“We can be anything we want to be
Just look around and you will see
It’s a busy world and there’s lots to do
And this busy world needs me and you”
Bibliography
(1)Maxim Massenkoff and Peter McCrory, “Labor market impacts of AI: A new measure and early evidence”, 2026-03-05, URL: https://www.anthropic.com/research/labor-market-impacts
(2)Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock, “Gpts are gpts: An early look at the labor market impact potential of large language models,” arXiv preprint arXiv:2303.10130, 2023, 10.
(3) Ramaa Sivashankar, “Automation in Manufacturing leads to high-wage job creation at all skill levels”, 2025-07-22, URL: https://www.foxtail-research.org/automation-in-manufacturing-leads-to-high-wage-job-creation-at-all-skill-levels/
(I am citing this in bibliography as I took some points for this blog from my own blog article published a few months ago)
Image courtesy: Courtesy: Image from freepik
Leave a Reply