B5 · Generative AI and Conversational Assistants

Benefits and Limitations of Generative AI

Generative AI offers real benefits: fast drafts, patient explanations, help across languages and wider access to skills. It also has real limitations: invented facts, bias, sameness, out-of-date knowledge, privacy risks, misuse and a heavy use of energy. A sensible user holds both in mind.

6 min Beginner Jump to the activity

What you need: Nothing. This page starts from the beginning, with no programming and no formulas.

Definition

Is generative AI the most useful tool of our time, or the most over-praised? Could both be true?

Generative AI offers real benefits: fast drafts, patient explanations, help across languages and wider access to skills. It also has real limitations: invented facts, bias, sameness, out-of-date knowledge, privacy risks, misuse and a heavy use of energy. A sensible user holds both in mind.

In simple words

A pressure cooker cooks dal in a fraction of the time, and every kitchen is glad to have one. It can also scald you if you open it carelessly. Nobody concludes that pressure cookers are wonderful or that they are terrible. They learn what it is good for, and where the danger lies. Generative AI deserves the same level-headed treatment.

Why it matters

Public talk about generative AI swings between excitement and alarm. Neither helps you decide whether to use it for a particular task. A clear list of what it does well and badly does.

Key terms

Benefit
A helpful result of using a tool.
Limitation
Something a tool cannot do or does poorly.
Bias
A consistent slant in outputs that treats some groups or views unfairly.
Stereotype
A fixed, oversimplified idea about a group of people.
Misuse
Using a tool to cause harm, such as cheating or deceiving.
Data privacy
Control over who can see and use information about you.

Explained in detail

Benefits

Speed is the most obvious. A first draft of a letter, a lesson plan or a summary appears in seconds, and it is far easier to improve a draft than to face an empty page. Patience is another. A chat assistant will explain the same idea ten times in ten different ways without sighing, at midnight, to a learner who is too shy to ask a teacher again. Reach across languages matters greatly in India: a person can ask in Hinglish and receive an explanation of an English document. And the tools lower barriers to skills. Someone who has never studied design can produce a decent poster. Someone who cannot yet program can build a simple working tool.

For people with disabilities, the benefits can be larger still: descriptions of images for those who cannot see, clear rewording for those who find dense text difficult, and a voice for those who cannot speak.

Limitations

LimitationWhat happensWhat to do
Invented factsFalse statements and sources, stated confidentlyVerify anything that matters
Out-of-date knowledgeKnows nothing after its training date unless it can searchCheck recent matters elsewhere
Bias and stereotypesRepeats the slants of its training dataNotice, question and correct them
SamenessProduces typical, average workAdd your own ideas and voice
Weak at exactnessSlips in arithmetic, counting and precise logicCheck calculations yourself
PrivacyWhat you type may be stored and usedNever enter private details

Bias you can see

Image generators make bias visible. Researchers who asked popular tools for pictures of people in different jobs found that requests for a "doctor" or a "chief executive" mostly produced men, while a "nurse" or a "housekeeper" mostly produced women, and that people in highly paid jobs were shown with lighter skin than people in poorly paid ones. The tools were not instructed to do this. They reproduced the patterns in the pictures they learned from, and in some cases exaggerated them. Text tools carry similar slants, less visibly: in the examples they choose, the names they give characters and the culture they treat as normal.

Costs to society

  • Misuse: the same tools that draft a letter can draft a thousand scam messages or produce a fake video of a real person.
  • A flood of low-quality content: the internet is filling with generated articles and pictures made only to attract clicks.
  • Effects on creative workers: writers, illustrators, translators and voice artists find some of their work replaced, often by models trained on their own output.
  • Energy and water: generating text and especially images and video uses much more computing than a simple search.
  • Dependence: skills fade when a tool always does the task.

Deciding when to use it

Generative AI is a good choice when a draft is useful, when you can judge the result yourself, and when a mistake would be minor and easy to fix: brainstorming, rewording, summarising something you will also read, practising a language. It is a poor choice when exact truth matters and you cannot check it, when the task is the learning itself, when private information is involved, or when a person's feelings, rights or safety depend on the outcome. Most real tasks fall in between, and there the right approach is to use the tool for a first version and bring your own knowledge and checking to the final one.

How it works

Good usesPoor uses
Getting ideas when you are stuckGetting facts you cannot check
A first draft you will rewriteFinished work handed in as your own
A simpler explanation of something you are studyingMedical, legal or financial decisions
Practising a conversation in another languageAnything involving private information
Feedback on your own writingMaking content that deceives or hurts someone

Try it yourself

The same tool on a good day and a bad day

A small shop owner uses a chat assistant twice in one week.

Benefit or limitation?

What does each situation show about generative AI?

  • Asked to draw "an engineer", it draws only men

  • It turns rough notes into a tidy first draft in seconds

  • It describes a photo aloud for a blind user

  • It explains the same idea five different ways without tiring

  • It does not know about a law passed last month

  • It gives a book title that does not exist

Common mistakes

  • Treating generative AI as either brilliant or worthless. It is useful for some tasks and unsuitable for others.
  • Using it where you cannot check the result and a mistake would matter.
  • Ignoring bias because the output looks neutral and polished.

Check yourself

Question 1

1. Which is a genuine benefit of generative AI?

Question 2

2. An image tool shows mostly men when asked for "a scientist". What is this?

Question 3

3. For which task is generative AI the poorest choice?

Summary

  • Benefits: speed, patience, help across languages and wider access to skills.
  • Limitations: invented facts, bias, sameness, out-of-date knowledge, privacy risks and misuse.
  • Use it where you can judge the result and a mistake would be small.

Finished reading?

Sources: UNESCO: Guidance for generative AI in education and research · AI4K12: Five Big Ideas in AI