How Does AI Work? · 3 min read
AI in Practice – Strengths and Limitations
Where AI shines and where it struggles – a realistic view to help you make good decisions about when and how to use AI.
AI in Practice – Strengths and Limitations
Section titled “AI in Practice – Strengths and Limitations”Learning objectives
- Know which tasks AI typically handles well
- Identify risks and limitations
- Build a mental model for when to use AI and when to be careful
You now have a good grasp of how AI works technically. But the most practical question is: What can AI actually do well – and where does it stumble? Here we take a realistic look at strengths and limitations.
Where AI is strongest
Text-based tasks
Section titled “Text-based tasks”- Summarization – Extract the essentials from a long document, meeting notes, or report.
- Translation – High quality for most major languages; nuances may need review.
- First drafts – Emails, reports, descriptions, and job postings in minutes.
- Rewriting and improving – Adjust tone, shorten, restructure, or clarify.
Analysis and pattern recognition
Section titled “Analysis and pattern recognition”- Data analysis – Identify trends, categorize feedback, detect anomalies.
- Comparison – Go through multiple alternatives and summarize pros and cons.
- Classification – Sort texts into categories (support cases, sentiment, topic).
Availability and scale
Section titled “Availability and scale”- Speed – Processes thousands of documents in the time a human needs for one.
- Availability – Available 24/7, responds immediately.
- Consistency – Doesn’t lose focus, doesn’t have bad days (though results do vary).
Where AI struggles
Hallucinations and factual accuracy
Section titled “Hallucinations and factual accuracy”As we covered in Section 2: the model generates plausible text, not necessarily true text. Without good sources in the context, it can fill gaps with confident-sounding guesses.
Practical implication: Always verify facts that will be used in decisions or public communications.
Long causal and logical chains
Section titled “Long causal and logical chains”AI can solve isolated logical problems, but complex reasoning with many interdependencies can trip it up. This is an active research area – models are getting better, but are not reliable for intricate calculations.
Cultural nuance and social context
Section titled “Cultural nuance and social context”Irony, humor, subtle power dynamics, regional variations, and implicit meaning in language are areas where humans have a natural advantage. AI often misses what’s between the lines.
Sensitive and high-stakes decisions
Section titled “Sensitive and high-stakes decisions”- Legal decisions about individuals
- Medical diagnoses
- Decisions affecting rights, employment, finances
These areas require human oversight – AI can be a support, but not a replacement for decision-makers.
Real-world knowledge
Section titled “Real-world knowledge”The model has a knowledge cutoff and no real-time access to the world (without external tools). It doesn’t know what happened yesterday unless you tell it.
Risk: Bias and unfair outcomes
AI learns from historical data – and history contains inequalities.
If a model is trained on hiring decisions that historically favored a certain group, it can amplify those patterns. This can result in systematic discrimination that is harder to spot than a human making the same mistake, since the AI’s decisions can feel “objective.”
This is called bias and is one of AI’s most serious social problems.
What you should know:
- AI is not neutral just because it’s a machine.
- Check whether the model has been evaluated for bias in your use case.
- Be especially vigilant when AI is used for decisions about people.
The human-AI framework
Think of it as a team where both human and AI play to their respective strengths:
| Task | AI | Human |
|---|---|---|
| Process large text volumes | ✅ Fast, consistent | ⬜ Time-consuming |
| Fact accuracy | ⚠️ Verify always | ✅ Critical review |
| Creative first draft | ✅ Quick starting point | ✅ Edit and refine |
| Ethical judgments | ⬜ Not reliable | ✅ Essential |
| Social context | ⬜ Often misses it | ✅ Strong |
| 24/7 availability | ✅ Always | ⬜ Limited |
The most effective model: AI generates, human evaluates and decides.
Summary
Section titled “Summary”- AI is strong at: text-based tasks, analysis, speed, and scale.
- AI struggles with: hallucinations, complex logic, social nuance, and high-stakes decisions.
- Bias is a real risk when AI is used for decisions about people.
- Best practice: AI + human in the loop – AI for first drafts and heavy lifting, humans for quality control and final decisions.
Test your knowledge
Question 1 of 3
In which situation is AI LEAST suitable to use without strong human oversight?