Let me start with what a large language model actually is, because most of the noise around it quietly disappears once you understand the simple thing underneath. An LLM is basically a very sophisticated graph. It looks a bit like a tree, and its one job is to predict the next letter it will show you. That is the whole trick.
So an LLM only pretends to have a memory. It only pretends to reason. What it is genuinely, almost startlingly good at is showing you the result of a very binary analysis. The next letter is either a good fit or it is not, and everything else branches out from that single decision, made over and over.
Once that clicks, you can predict which jobs get handed to these models first. The easiest ones to automate, and the soonest, are the jobs that produce fairly binary answers.
Watch on Youtube
The Jobs With a Clear Right Answer
Take junior coders. A line of code either runs or it does not. You can check it in a second. That clarity is exactly why those roles were among the first to be handed over.
The same logic reaches further than most people expect. Think about diagnosticians, the doctors whose work is reaching a diagnosis. They sit in a similar spot, and I would say sooner rather than later, because the outcome of what they do is also binary. The diagnosis is either right or it is wrong.
Financial analysts are in the same boat. The analysis is either correct, mathematically, or it is not. And lawyers belong there too. There is a code of law, and almost all legal advice is built on top of it, so you can check whether the reasoning holds and whether what the lawyer proposes is legally sound.
Try It With a Contract
If you think this will not actually happen, picture writing a contract. You really have two options. You can hire a lawyer who will probably spend ten days drafting it, going back and forth with you, making the small decisions along the way. Or you can start right now with a good LLM and get something of very similar quality almost instantly.
There is one real difference, and it is worth sitting with. You cannot sue the model. If it hands you bad advice, there is no one to hold responsible. You cannot go back and audit how it reasoned, because there is no real trail to follow.
That missing audit trail is exactly the kind of thing that should make a company pause before it leans on these tools, and right now almost nobody is talking about it. The encouraging part is that this is a known and solvable problem, and a whole field of researchers is busy prying the black box open, so that one day we will be able to trace what a model did and why.
The Economics Are Hard to Argue With
These jobs get automated first not because someone made a careless or unkind decision, but because it simply adds up. From a pure budget point of view, it is very hard to defend paying for a process that is ten times slower, that needs to rest, and that can fall sick, when you can get comparable quality in seconds.
I want to be clear that this is not the end of these professions. Lawyers will not disappear, in the same way translators did not disappear once DeepL started to perform adequately. What changes is the shape of the work.
Lawyers will be pushed toward the parts a machine cannot touch: visibility, judgment, and real relationships with clients. It is the same move translators made when they became proofreaders and editors, polishing and checking the output rather than starting every time from a blank page. Some, of course, still translate literature, and that is genuinely valuable, because a few of them carry rare and beautiful skills.
I Was a Translator Once
I can say all of this plainly because I was a translator myself when I was younger. So I am not describing this from the outside.
And honestly, the bulk of that work was not inspiring at all. It was technical documentation and other dry material, page after page of it. It paid my way through my studies, and I am grateful for that. But would I have wanted to do it for the rest of my life? No.
So I am glad a tool like DeepL or Claude can carry that load now, rather than me. These jobs are being handed over to AI as I write this sentence, and that is not something to fear so much as something to read clearly. Learn what an LLM really is. Get comfortable using one. Then put your energy into the things it cannot fake: taste, judgment, trust, and the human relationships that sit on top of all the right answers. That is where the interesting work is moving, and it is wide open.
Do Your Own Research
If you want to check any of this for yourself, here are solid, peer-reviewed places to start, grouped by topic.
What a language model actually is
- Vaswani et al., "Attention Is All You Need" (2017), the paper that introduced the transformer, the architecture behind today's models: https://arxiv.org/abs/1706.03762
- Li et al., "Mechanics of Next Token Prediction with Self-Attention" (2024), a close look at what a model is really doing when it guesses the next token: https://arxiv.org/abs/2403.08081
Which jobs are most exposed
- Eloundou, Manning, Mishkin and Rock, "GPTs are GPTs," published in Science (2024). It finds that higher-paid jobs may be more exposed to these models, not less: https://www.science.org/doi/10.1126/science.adj0998 (open version: https://arxiv.org/abs/2303.10130)
Coding
- Peng et al., "The Impact of AI on Developer Productivity: Evidence from GitHub Copilot" (2023), a controlled experiment measuring how much faster developers work with an AI assistant: https://arxiv.org/abs/2302.06590
Medical diagnosis
- Goh et al., "Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial," JAMA Network Open (2024): https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11519755/
Translation
- Zouhar et al., "Neural Machine Translation Quality and Post-Editing Performance" (2021), on how the job has shifted toward editing machine output rather than translating from scratch: https://aclanthology.org/2021.emnlp-main.801/
The black box and the missing audit trail
- "Bridging the Black Box: A Survey on Mechanistic Interpretability in AI," ACM Computing Surveys (2025): https://dl.acm.org/doi/10.1145/3787104
- "On Reading and Interpreting Black Box Deep Neural Networks," International Journal of Digital Humanities (2023): https://link.springer.com/article/10.1007/s42803-023-00075-w
