HelloAI
L0 Chapter 3 🥚 🕒 14 min

What AI Can and Can't Do — Talking With Receipts

Stop listening to hype and FUD. Item by item: what AI already beats you at, and what it absolutely can't do yet.

H
HelloAI Editors
5/29/2026

On social media, two camps fight about AI.

One side: “AI replaced another profession! Designers/programmers/translators/teachers are about to be jobless!” — with a scary screenshot.

The other side: “AI is still useless—can’t even tell 9.11 from 9.9” — with another screenshot.

Neither camp is solving anything. Because AI’s capabilities are uneven: superhuman at some things, sub-elementary at others. You need a trustworthy checklist to know what to delegate and what to keep.

Let’s make that checklist.

✅ AI is already great at (better than you, even)

1. Generic writing

Emails, product copy, status reports, blog posts, resume edits, translation, polishing—AI reaches 70-90 points of a professional writer’s level on these.

Real impact: Most Zhihu/Xiaohongshu/Substack “content account” operators have been AI’d. 1 human + 1 AI now produces what 10 humans did.

But “written fluently” doesn’t equal “written correctly”—AI might misstate your product features. Always verify facts yourself.

2. Summarization and extraction

Compress a 5000-word paper to 300 words. Extract key clauses from a contract. Turn a 1-hour meeting recording into minutes—AI’s strong suit.

Real impact: Law firms now widely use AI for “contract first-pass review.” A 4-hour job becomes 20 minutes; the lawyer only checks the key points.

3. Programming assistance

Boilerplate code, debugging errors, explaining others’ code, suggesting refactors—AI here is like a tireless senior coworker.

Real impact: Cursor / GitHub Copilot boost experienced engineers’ productivity 30-100%. AI is now de facto standard for engineering.

Caveat: limited help if you can’t code at all—you don’t know if what it writes is correct.

4. Translation

Chinese↔English, English↔Japanese, especially business/technical text—AI translation surpasses 95% of human translators. Literary translation still doesn’t work.

5. Certain medical judgments

Radiology AI reading CT/X-ray now matches experienced radiologists for malignancy detection. Skin cancer image recognition matches dermatologists on some tasks.

Real impact: Some major Chinese hospitals already use AI as “second reader”—AI flags suspicious regions, human reviews flagged ones.

But: doctor still owns final decision—they take the malpractice risk.

6. Customer service front-line

Common questions, user guidance, initial issue triage—AI handles these fully.

Real impact: 80% of e-commerce and fintech app support is now AI-handled, only complex issues escalate.

7. Brainstorming

100 names, 100 headlines, 30 event ideas—AI’s breadth is superhuman. Its ideas might not be the best, but they break your mental ruts.

8. Visual creation

Midjourney / Stable Diffusion / Sora / DALL·E reach 70-80 points of a commercial illustrator’s level. Logos, social media images, concept art, marketing visuals—often “good enough.”

9. Games with explicit rules

Go, chess, esports—AlphaGo settled this. Crushed.

10. Some scientific discovery

AlphaFold predicted the 3D structure of every known protein (200M+). Took biologists 50 years to figure out 170K manually. Earned DeepMind the 2024 Nobel Prize in Chemistry.

💡 The pattern

Notice the common thread? All these domains have massive training data. Text, code, medical images, board games, protein structures—all are decades-old large-scale datasets.

⚠️ AI can do, but tread carefully

1. Anything involving “facts”

It hallucinates. That’s “hallucination”—sounds confident, content is invented.

Example: ask “famous calligraphers of China’s Southern and Northern Dynasties,” it gives 5 names that sound real—but 2-3 might be made up.

Mitigation: ① never let AI answer factual queries you can’t verify; ② give it source material so it answers “based on text” not “based on memory”—this is the RAG idea.

2. Math and arithmetic

GPT-4 often gets 245 × 367 wrong (not because it’s dumb—its mechanism is bad at exact arithmetic).

Mitigation: have it write code to calculate rather than calculate directly. "Use Python to compute 245 × 367" — it writes one line, result is correct.

3. Real-time information

It has a training cutoff date. “Did the US stock market open today?” “Who won last night’s NBA game?” It either doesn’t know or makes one up.

Mitigation: use web-enabled AI (Perplexity, web-search ChatGPT/Claude).

4. Ethical judgment

“Should I divorce?” “Should I call the police?” “Should I fire this employee?” AI’s answer might sound reasonable—but it’s a statistical average of what humans tend to say, not real ethical thinking.

Mitigation: treat as “an ordinary friend’s opinion,” never as final advice.

5. Your private information

Pasting company secrets, customer data, personal medical records into ChatGPT—this data may be logged, may be used for training.

Mitigation: paid tiers typically promise no-training; sensitive data → private deployment (build in-house); most sensitive things—just don’t.

❌ AI currently can’t do

1. True causal reasoning

AI is a “correlation master, causation blind man.” It knows “rooster crows then sun rises” but can’t distinguish whether crowing causes sunrise or sunrise causes crowing.

Real example: AI making medical diagnoses “why is this patient feverish”—lists possible causes but can’t tell cause from symptom.

2. Long-term planning and execution

“Make a 3-month diet plan”—AI can write one. But having it track 3 months on its own, adjust based on progress, remind you—nope (without engineering wrappers, at least).

Current AI Agents are advancing fast, but stable multi-step long-period autonomous execution is still hard.

3. Emotional resonance

You vent your heartbreak to AI; it can produce extremely appropriate comforting words—but it isn’t feeling. Its “understanding” is simulation.

The ethical implications are deeper than they seem.

4. Creating truly “new” things

AI can endlessly combine within existing styles (“paint a cat in Van Gogh’s style”), but it can’t become the next Van Gogh—all its “understanding” comes from existing human work.

Nobel-Physics-level breakthrough ideas: AI can’t yet.

5. Embodied tasks in the physical world

Let a robot get you a glass of water in an unfamiliar home? Still very hard. Vision + motor control + common-sense reasoning is far from mature.

Tesla’s Optimus, Boston Dynamics’ Atlas—still demo stage, far from practical.

6. Working in data-free domains

AI to diagnose a brand-new disease (not in training data)—it’ll give a confident wrong answer. Great in known patterns, terrible in genuinely unknown territory.

7. Actually understanding what it says

This is philosophical but has real impact. When AI says “I understand how you feel,” it’s understanding nothing—just outputting statistically likely characters.

This doesn’t make it useless, but use it with this awareness.

A cheat sheet

TaskRecommendation
Email/copy/translation✅ Use freely
Summarize/extract✅ Use freely
Programming (you know code)✅ Use freely
Brainstorming✅ Use freely
Title/name generation✅ Use freely
Learning/research⚠️ Use, but request sources
Math/computation⚠️ Have it write code, not compute directly
Programming (you don’t know code)⚠️ Dangerous
Medical advice⚠️ OK to ask, but must see doctor
Legal advice⚠️ OK to ask, but must find lawyer
Investment advice❌ Don’t
Major life decisions❌ Don’t
Confidential information❌ Don’t

One-line summary

AI is a broad-but-not-deep, fluent-but-not-rigorous, confident-but-not-reliable assistant.

Use its breadth and fluency to multiply your productivity, but everything requiring rigor and reliability is on you.

💡 A mantra

Always treat AI like a smart intern, not an authoritative expert. It can do many things for you, but you’re accountable for its output.

Next: “What Is ‘Hallucination’? Why It Happens, How to Mitigate”