A clear-eyed, evidence-based look at which tasks AI genuinely outperforms humans on today — and which ones still need a person. No hype, no fear-mongering.
The AI jobs conversation is broken. One side says AI will take every job by 2030. The other side says AI is a toy that hallucinates too much to matter. Both are wrong.
Here is a grounded, specific look at what AI genuinely does better than humans today — and where it still falls short — based on published benchmarks, research studies, and real deployments.
10 things AI does better than most humans, right now
### 1. Summarising long documents AI can read a 50-page research paper or a 200-page contract and produce an accurate, structured summary in seconds. A human would take hours — and miss things. Claude with its 200K context window is purpose-built for this.
Evidence: In a 2024 study published in NEJM AI, GPT-4 summarised medical research papers with accuracy comparable to human clinicians, but in under 30 seconds versus 45 minutes on average.
### 2. Translating between languages AI translation quality (DeepL, GPT-4, Google Translate) now matches professional human translation for common language pairs (English-Spanish, English-French, etc.) in blind evaluations.
Who should worry: Generalist translators doing business, legal, or technical documents. Who is safe: Literary translators, poetry, marketing copy that requires cultural nuance.
### 3. Data entry and extraction Give AI a scanned invoice, a handwritten form, or a screenshot of a spreadsheet, and it extracts structured data with error rates below 2% — better than the average human data entry worker.
### 4. Writing routine business communications Emails, meeting summaries, status reports, job descriptions, and policy drafts can be generated by AI in seconds. Human review is still needed for tone and accuracy, but the first draft cost drops to nearly zero.
### 5. Basic coding and debugging GitHub Copilot and Cursor can write boilerplate, fix straightforward bugs, explain unfamiliar code, and generate unit tests faster than most developers writing from scratch. On competitive programming tasks (Codeforces), GPT-4 scores in the 85th percentile.
Who should worry: Junior developers doing routine CRUD work. Who is safe: Senior engineers doing system architecture, debugging novel issues, or working on large, messy codebases.
### 6. Image and video generation Midjourney and DALL-E 3 produce images that in blind tests are rated as more aesthetically pleasing than human-created stock photography. Runway and Sora generate short video clips from text prompts.
Who should worry: Stock photographers, generic illustration work. Who is safe: Conceptual art directors, photographers with a distinct style, and anyone requiring precise control over output.
### 7. Customer service triage AI handles the first layer of customer support — identifying the issue, looking up account details, suggesting solutions, and escalating only when truly needed. Companies like Klarna reported AI handling two-thirds of customer service chats with higher satisfaction scores than human agents.
### 8. Language tutoring AI tutors provide unlimited, patient conversation practice, explain grammar rules with examples, and correct pronunciation — all for free or minimal cost. Duolingo Max uses GPT-4 for this.
### 9. Medical image analysis AI models (specifically deep learning models trained on radiology data) now match or exceed radiologists at detecting certain conditions — lung nodules, breast cancer on mammograms, diabetic retinopathy — in controlled studies.
Important caveat: AI assists radiologists rather than replacing them. It flags potential issues; a human makes the final call.
### 10. Generating marketing copy variants Need 20 versions of a Facebook ad headline? 50 variations of a product description for A/B testing? AI does this in seconds with quality comparable to junior copywriters.
5 things AI still cannot do well
### 1. Original scientific research AI can summarise existing research. It cannot design a novel experiment, form a genuinely new hypothesis, or interpret surprising results in a creative way. Science requires understanding causality — something LLMs do not have.
### 2. High-stakes negotiation Negotiating a salary, closing a complex B2B deal, or mediating a conflict requires reading the room, adapting tone in real time, and understanding unspoken power dynamics. AI has no theory of mind and no emotional intelligence.
### 3. Physical skilled trades Plumbing, electrical work, carpentry, surgery, cooking, gardening — anything requiring physical dexterity, spatial reasoning in unstructured environments, and adapting to novel physical situations. Robots exist but are decades from general-purpose capability.
### 4. Leadership and people management Motivating a team, resolving an interpersonal conflict, making a layoff decision with compassion, inspiring a group toward a vision — these require empathy, accountability, and moral reasoning that AI does not have and may never have.
### 5. Handling genuine novelty AI works by pattern matching against its training data. When a situation has no close precedent — a once-in-a-generation crisis, a novel legal question, a creative breakthrough — AI has nothing to match against. Humans can reason from first principles; current AI cannot.
The real answer: augmentation, not replacement
The most useful framing is not "AI vs humans" but "humans with AI vs humans without AI." A developer with Copilot writes code faster than one without. A marketer with ChatGPT drafts copy faster. A doctor with an AI second-opinion catches things they might miss alone.
The jobs most at risk are not the hardest or the simplest — they are the ones where the output is a document, a response, or a routine analysis, and where the quality bar is "good enough."
The jobs safest are ones requiring physical presence, genuine creativity, high-stakes judgement, or deep human connection.
If you work with information for a living, the question is not whether AI will affect your job. It is whether you will learn to use it before your competitors do.