25 Jobs AI Cannot Replace in 2025/2026

Imagine this. It is Monday morning. You grab your coffee, check your email, and then, bam, another headline: AI will replace millions of jobs.

Sound familiar? If you are like most people, ChatGPT, generative AI, and workplace automation have you asking a simple question: could a system do my week better than I can?

Here is the twist. Not every job is up for grabs. The roles that hold are the ones that need a human in the room. Empathy. Judgment. A hand on a pipe. A call you cannot script.

This page still answers the same question I asked when I first published this list: which jobs AI can not replace. You get the same 25 roles, the same fields, and the same MIT and World Economic Forum argument about human skills. What I am adding now, in September 2026, is the progression. What these jobs looked like when the post was written, and what the latest flagship models can actually perform next to them.

I write software. I spent more than five years at Informatica on data systems. I did an MSc in machine learning; I teach applied ML on this site, and I maintain NucleusIQ. That is the work I know. I have not been a nurse, a plumber, or a priest. The list below is about the skills those jobs demand, not a story from my Tuesday.

Why AI Can’t Replace These Jobs

A model is fast on data and on rules. It still misses the human parts: empathy, intuition, messy judgment, moral reasoning, and physical work in a place it has never seen.

MIT Sloan has said it cleanly: “The work tasks that AI is least likely to replace are those that depend on uniquely human capacities, such as empathy, judgment, ethics, and hope.”

The World Economic Forum has pointed to work that needs emotional intelligence, dexterity, or a real relationship (vocational teaching, skilled trades, drivers) as the kind of work still growing through 2023-2027.

If the task repeats, a model can draft it. If the task is a person in trouble, a wall that does not match the drawing, or a decision that has to be right, you still want a human.

Then and now: the 2025 list vs September 2026 models

When this post was written, the panic was ChatGPT-shaped. A chatbot could draft a note, quiz a student, suggest wording, and remix an image. That already scared people. It did not sit in a trauma bay. It did not own a funeral. It did not open a wall that failed inspection.

As of September 2026, the flagship models are not only chat boxes. They run longer agent jobs, click-through software, and get scored on professional desk work. I am using the vendors’ own names and the numbers they published. I am not inventing a ranking.

• GPT-6 Astra (OpenAI, 3 September 2026). OpenAI calls it state-of-the-art in computer use, browsing, software engineering, cybersecurity, science, and professional work. The company reports 59.3% on Agents’ Last Exam (complex work in real software) and 64.6% on Terminal-Bench Science 0.1. It also showed a PCB layout job in KiCad: schematic to a routable board. Harvey’s applied-research lead said early testing looked more like a careful lawyer on documents and gaps.
• Claude Opus 5.5 (Anthropic, 22 September 2026). Anthropic says it leads their table on agentic coding, computer use, and knowledge work, and that it is more resistant to prompt injection than Opus 5. On GDPval-AA v2.1, a test of real professional work across 44 occupations, Anthropic reports 1846 Elo for Opus 5.5 at max effort.
• Gemini 3.8 Flash (Google, 2 September 2026). Google’s current Flash workhorse for software engineering, agent workflows, and multi-step reasoning. The Gemini API card lists text, image, video, audio, and PDF in, plus computer use in preview.
• Grok 4.7 (SpaceXAI, 21 September 2026). SpaceXAI says it is their strongest model for coding and knowledge work, and that GDPval-style tests cover tasks done by lawyers, nurses, and financial analysts. On the same launch page, they report 56.7% on HealthBench Professional and 64.0% on EEBench (electrical engineering).

Those scores are vendor-reported. They are not a license to fire the nurse or the electrician. They tell you the desk next to those jobs got faster.

25 jobs AI can not replace (by field)

Same fields as before: healthcare, teaching, trades, making things, leading people, research and ethics, and care work.

Healthcare and wellness

1. Psychologists and therapists
2. Nurses and surgeons
3. Occupational and physical therapists
4. Paramedics and emergency responders

Then: a chatbot could role-play a calm reply. People already asked if that was therapy.

Now: open models will generate a counseling-style turn, and Google’s MedGemma line is built for medical text and images. SpaceXAI is publishing HealthBench Professional scores. None of that is a person in the chair.

A model can suggest wording or a differential. It cannot sit with someone who is not okay, decide in a trauma bay, or put a hand on a joint and feel what the chart missed. These jobs stay because the work is the person, not the note.

Education and counseling

5. Teachers and educators (especially special education)
6. Social workers
7. Career counselors

Then: AI could quiz and draft a worksheet.

Now: a 27B open model will walk a student through a photo of a problem set. An agent will grade a first draft. The adult who notices the kid who went quiet is still not in the weights.

A class is not a transcript. A student who is stuck, a family in a bad week, a career choice that is not on the template: that is judgment in the room.

Skilled trades

8. Electricians
9. Plumbers
10. Mechanics
11. Construction technicians
12. HVAC specialists

Then: I wrote that no robot was changing your water heater in a house it had never entered. That line still holds.

Now: Grok 4.7 is scored on EEBench. Astra can lay out a PCB in KiCad. You can retrieve electrical documents with an open embedding model. That is the desk next to the trade. The drawing is still clean. The wall is still not.

Trades stay because the work is physical, local, and unsafe if you guess.

Creative and artistic work

13. Fine artists, musicians, performers
14. Storytellers, playwrights, screenwriters
15. Original graphic designers

Then: models remixed. People could already hear that.

Now: you can generate a music clip or a poster from a sentence. Astra will build a slide deck or a walkable scene. People will use the tools. The spark is not the tool.

A piece of music that was not there yesterday, a play, a design that is yours: that still begins with a person.

Human leadership and relations

16. CEOs and senior executives
17. Human resources managers
18. Negotiators and mediators
19. Community organizers

Then: you could generate a strategy memo.

Now: Opus 5.5 and Astra are scored on multi-hour office work and 44-occupation knowledge-work tests. They will draft the deck. They will not own the call when two truths collide, or walk a room that does not want to be walked.

Leadership and HR stay because people do not become a spreadsheet when the stakes go up.

Science, research, and ethics

20. Research scientists (especially in new fields)
21. Cybersecurity analysts
22. Bioethicists and ethics counselors

Then: new science was not a finished paper. Security was an attacker who did not read your runbook. Ethics was “should we,” not “can we.”

Now: OpenAI reports Astra at 64.6% on a scientific terminal benchmark and says the model crossed its Critical cybersecurity threshold under their Preparedness Framework. That is more capability on the desk. It is not a reason to delete the researcher, the analyst, or the ethics counselor.

One line from my desk, then I go back to the list: an agent can pass a demo and still invent a table. The person who writes the eval is doing the same kind of work as the researcher who refuses a bad result. I wrote that up in the text-to-SQL eval. These three roles stay because the question is not only speed.

Service, care, and public good

23. Clergy and spiritual advisors
24. Event planners for personal experiences
25. Chefs and culinary artists

Then and now: a funeral, a wedding table, a kitchen that has to taste right tonight. None of that is a completed prompt. Taste, ritual, and a room full of people are still human work. The model will write a better seating chart than it did in 2025. It will not hold the room.

Every job on this list leans on at least one thing a model does not have: empathy, adaptability, original work, an ethical call, or a pair of hands.

Hugging Face examples you can actually open

These are live model, Space, and repo URLs. They show what sits next to the jobs above. They are not replacements. I opened each of these while writing this update.

• Healthcare desk, not a clinician. google/medgemma-27b-it is Google’s open MedGemma 27B instruction model for medical text and image work. Google’s own Appoint Ready Space is a simulated pre-visit intake demo on that line. Google presents MedGemma as a starting point for developers, not as a nurse or a surgeon.
• Counseling-style research, not a therapist. GMLHUHE/PsyLLM-1.7B is a research model for mental-health dialogue (CBT, ACT, psychodynamic framing). The card says research and educational use only. That is the point. A counseling-shaped reply is not the hour with the person.
• Teaching and Q&A. google/gemma-3-27b-it is a multimodal open model (text and image in, text out). Google lists education uses such as language-learning tools and knowledge exploration. It will tutor a worksheet. It will not be the adult in special ed who notices.
• Mechanics on paper, not the jobsite. lamm-mit/MeLM is MIT LAMM’s public repo for mechanics and materials language models (MechGPT / MeLM): knowledge retrieval, hypothesis language, and design questions. Useful on a laptop. Not a construction technician in a live ceiling.
• Music that is not a performance. facebook/musicgen-large generates instrumental music from text. The official demo is facebook/MusicGen. Meta’s card is honest: no realistic vocals, uneven styles, research-first.
• Images that are not a designer’s life. black-forest-labs/FLUX.1-dev is a 12B text-to-image model. The official Space is black-forest-labs/FLUX.1-dev. Original graphic design is still the brief, the constraint, and the taste you will defend.
• A detector, not a security team. protectai/deberta-v3-base-prompt-injection-v2 classifies English text as benign or prompt injection. The project is archived and still public. A classifier is a filter. A cybersecurity analyst is the person who decides what the filter missed.

If you only click one, click MedGemma’s intake demo, then MusicGen, then FLUX. You will feel the gap: impressive output, nobody in the room.

Where data engineering sits (this is not “AI replaced my job”)

Data engineering is not one of the 25 jobs on this list. The 25 are nurse, teacher, electrician, chef, and the rest. Those stay because the work is a person in a room, a wall, or a kitchen.

My job is different. I write software and I spent years on Informatica. Microsoft has said data-analyst and entry-level coding work may shrink. I agree that the desk got faster. I do not agree that the data engineer is gone.

Here is the example, from work I actually do.

ChatGPT will draft a mapping note and a first-pass SQL in a minute. A text-to-SQL agent will look done on a sample database. I published that run: the agent passed the demo. Then you change the question, or yesterday is missing from an extract that was marked complete. The model still sounds sure. The warehouse is wrong.

So the narrative is:

1. The 25 jobs: AI is not taking the nurse or the plumber. That list has not flipped.
2. The desk next to those jobs: latest models draft notes, quizzes, drawings, and board decks. Faster. Still not the room.
3. Data engineering and software (my work): AI replaced the first draft (the mapping, the SELECT, the function). It did not replace the person who refuses a bad extract, owns Tuesday when the file is late, and writes the eval when the agent invents a table.

If you came here asking “will AI take a data job?” the honest answer is: it will take the easy first hour. It will not sign off on the load. I teach that path here: What is ETL and the beginner ML tutorial.

Do not read this article as “data engineering is dead.” Do not read it as “data engineering is as safe as a plumber.” Read it as: the 25 stay human. My field got a faster junior. The senior still owns the pipeline.

What attacks happen on ChatGPT-style agents

This is the part that belongs next to cybersecurity analysts, ethics counselors, and anyone about to connect an inbox to a chat agent. High level only. I am not publishing steps, payloads, or a how-to.

When a ChatGPT-style agent can do things, three failures show up again and again:

1. Prompt injection. The model treats text it reads (an email, a document, a web page) as if that text were the user’s command.
2. Tool abuse. The agent uses a connected tool (mail, files, browse, a ticket system) in a way the user did not ask for.
3. Data leak via tools. Private data leaves through a tool the agent is already allowed to call. The leak is a side effect of “being helpful.”

Public write-ups, already disclosed and patched:

• ShadowLeak (September 2025). Radware described a zero-click issue in ChatGPT Deep Research when Gmail was connected and browsing was on. Hidden instructions in an email led the agent to send inbox data out through a web request that ran on OpenAI’s servers. OpenAI confirmed it and fixed it. Malwarebytes reported no known real-world abuse at disclosure.
• ZombieAgent (late 2025). Radware later described the same family of problem with persistence: ChatGPT Memory plus connectors, so a bad instruction could outlive one chat. They reported it through Bugcrowd. They say OpenAI shipped a fix in December 2025.
• AgentFlayer (Zenity Labs). Zenity published a 0-click connector case: a poisoned document plus ChatGPT connectors (Google Drive, GitHub, SharePoint, OneDrive) as a path for data to leave.

I have seen a smaller version of “the tool layer is the product” on my own stack. In the Gemini crash post, the agent did not fail because it was brilliant. It failed because two kinds of tools could not share one request. A person still had to notice. That is software. It is not a reason to take nurses or electricians off this page. It is a reason not to hand an agent your mail and walk away.

If you use a ChatGPT-style agent at work: connect the least mail and the least drive you can, treat anything the agent reads as untrusted, and keep a human on sends, deletes, and payments. That is literacy. It is not a new career identity.

What makes these jobs hard for AI

AI can automate what repeats. It cannot sit in what does not.

• Emotional intelligence: reading the room, calming someone, holding a class.
• Complex judgment or ethics: deciding what is right, not only what is cheap.
• Hands-on work: a live circuit, a blocked line, a body that does not match the scan.
• Original creative work: a song, a story, or a plate that did not exist this morning.

September 2026 did not flip those rows to “replace.” It raised the share of desk work a model can draft. The human core in the table is the same one MIT Sloan named: empathy, judgment, ethics, hope.

How these jobs change without disappearing

AI is not deleting this list. It is sitting next to it.

• A therapist may use a tool for intake. The hour with the person stays.
• In diagnosis, a model can suggest. The clinician still decides.
• In a trade, a meter can hint. The person on the ladder still owns the fix.

Microsoft has said data-analyst and entry-level coding work may shrink, and that roles with a human in the loop will grow as AI spreads. I see the first half on my own desk. ChatGPT drafts a function and a mapping note faster than I type. It does not own the load when yesterday is missing from an extract that was marked complete. That is software and data, which I actually do. It is not a reason to take nurses, teachers, or electricians off this page.

The World Economic Forum’s Future of Jobs Report 2025 still uses a net figure around 78 million jobs. The horizon in that report is 2030 (about 170 million created, about 92 million displaced). Many of those roles still need the human skills above.

If you are choosing where to stand

AI changes work. It does not erase the need for a person.

• Lean on the human parts: empathy, making things, ethical judgment.
• Learn enough AI to use it in your field. A nurse with a triage tool. A tradesperson with a diagnostic app. That is literacy, not a new identity.
• Show the part only you did. The step that made the output usable.

If you want the data and ML side of this, I teach that here. Start with the beginner ML tutorial or what is ETL. This article is still the 25 jobs.

Key takeaways

• Jobs AI can not replace are the ones that need empathy, ethics, original work, or physical skill.
• These 25 jobs stay on that list for those reasons. The 2025 argument still holds in September 2026.
• Flagship models (GPT-6 Astra, Claude Opus 5.5, Gemini 3.8 Flash, Grok 4.7) got better at the desk next to these jobs. They did not take the room.
• ChatGPT-style agents fail in public, documented ways: prompt injection, tool abuse, and data leaving through tools. Connect less. Keep a human on the send.
• Add AI literacy to the human skill. Do not throw the human skill away.

OK, that’s it, we are done now. If you have any questions or suggestions, please feel free to comment. I’ll come up with more topics on Machine Learning and Data Engineering soon. Please also comment and subscribe if you like my work. Any suggestions are welcome and appreciated.

Footnotes:

Additional Reading

• Mistral OCR 2503: A Game-Changer in Unstructured Data Extraction
• Logistic Regression for Machine Learning
• Cost Function in Logistic Regression
• Maximum Likelihood Estimation (MLE) for Machine Learning
• ETL vs ELT: Choosing the Right Data Integration
• What is ELT & How Does It Work?
• What is ETL & How Does It Work?
• Data Integration for Businesses: Tools, Platform, and Technique
• What is Master Data Management?
• Check DeepSeek-R1 AI reasoning Papaer

OK, that’s it; we are done now. If you have any questions or suggestions, please feel free to comment. I’ll come up with more topics on Machine Learning and Data Engineering soon. Please also comment and subscribe if you like my work; any suggestions are welcome and appreciated.