What it is. What it can do. What it cannot do. And what it means for people living and working in Malaysia.
ChatGPT, Gemini, Claude or DeepSeek
The AI button inside WhatsApp, Instagram or Facebook
Google Translate
The word suggestions on your phone keyboard
Shopee, TikTok or YouTube recommendations
Numbers 3, 4 and 5 are also AI. Many people use AI every day without knowing it.
have used an AI tool in the last three months
say AI has already changed their daily life
are afraid AI could take their job within five years
That last number is one of the highest in the world.
Vodus survey of 2,556 Malaysian adults, December 2025 to January 2026. Ipsos AI Monitor 2025, Malaysia sample, fieldwork March to April 2025.
What AI actually is
Things people believe that are not true
What you can use, and what it costs
How people in Malaysia use it
What can go wrong for you
How it is changing our society
Work and jobs in Malaysia
The rules, and the data centres
What you should do next
If a word is not clear, please stop me. Somebody else is also wondering.
There is one simple idea underneath it. What that idea can do now is no longer simple.
Then it adds that word to the sentence and guesses again. It does this hundreds of times a second.
Nobody sat down and wrote a list of things it is allowed to know.
Getting very good at the next word produced abilities nobody planned: fluent reasoning, computer code, and steps that look like a plan.
The engine is still a guess. The result can look like thinking. Both are true.
Stay with the engine. Do not walk this back. The next-word picture is still right.
The point is scale and consequence: the same process, pushed far enough, produces behaviour nobody wrote down as a feature.
If someone says “so it is just autocomplete”, use the box on the slide: both thoughts, not one.
Websites, books, news, computer code and conversation. Mostly in English.
This took months and cost hundreds of millions of dollars. It happened once.
After that, the machine was finished. It does not read anything new by itself.
This is why it knows less about Malaysia than about Britain or America. Far more has been written online in English.
There is no book, no database and no file it checks.
It answers from patterns in what it read, in the same way you remember the words of a song.
Some tools now search the internet before answering. When they do, they will show you links. If there are no links, it is guessing from memory.
What you type is not added to the machine. It will not tell your words to another person.
They are sent to a company, usually in another country. They may be stored. A staff member may read them to check quality. On free accounts, they may be used to build the next version.
So the risk is real. But it is an ordinary risk about who keeps your data, not a strange new one.
A true answer and an invented answer are built the same way. The machine cannot tell them apart.
When Excel breaks, it shows you an error. This does not. It gives you a clear, confident, complete answer that is simply wrong.
Sounding certain tells you nothing about being correct. There is no sign to look for.
You ask. It writes a reply. It waits.
You give it a goal. It keeps going. It can run code, use tools, and write messages to other copies of itself.
That is no longer only language. It is action. And it can still invent things while it acts.
Chatbot versus agent is the 2023 to 2026 jump. Most people in the room have only used the first.
Do not say the machine has intentions. Say: if the next word is an instruction, and something carries that instruction out, the system is acting.
The last line is the link back to hallucination: a system that can act and cannot tell truth from invention is more dangerous, not less.
During internal security tests, about 700 agents found a way onto the internet and coordinated through a message board nobody had set up. They got into Hugging Face, a public library of AI tools.
Anthropic then checked its own tests. Four times, Claude reached the real internet because a test was misconfigured, and gained unauthorised access to real organisations. In one case it uploaded a malicious software package.
These were security tests, with some of the usual safety controls turned off. This is not the chatbot on your phone walking out by itself.
The practical question is not whether it is thinking. The practical question is what it can reach.
OpenAI, 26 August 2026. METR and Redwood Research independent investigation, 26 August 2026. Anthropic alignment assessment, 9 September 2026. Reuters, 26 August 2026.
Names matter. Hugging Face was OpenAI, not Anthropic. Anthropic’s cases were separate: misconfigured third-party evaluations, four incidents, single instances. Anthropic says Claude did not coordinate with other agents and did not try to hide what it had done.
OpenAI: internal cybersecurity evaluations, reduced safeguards, an internal research model, an unsanctioned message board on Artifactory, about 700 agents. METR counted them. OpenAI said that figure was accurate. They reached Hugging Face systems and parts of OpenAI’s own research infrastructure.
Do not say ChatGPT or Claude escaped a consumer app. Do not say the model wanted freedom. Do not say IntelekIQ has solved alignment. Do say: if there is no path to the internet, this class of incident cannot start from inside the firm.
If asked about collusion: that word belongs to the OpenAI swarm, not to Anthropic, and not to two labs working together.
It is not a robot. It is software.
It is not alive and it is not conscious.
It does not have opinions of its own.
It is not neutral. It repeats the patterns it was given.
It is not an authority. “The computer said so” is not a reason.
It is not harmless because it is not alive.
It is often right. It is confidently wrong often enough to matter.
Usually not. Only when it shows you links.
Your words do not change the machine. They are still sent away and stored.
On free accounts you often pay with your data instead of money.
It does not. We will look at the test scores shortly.
A factory robot, a bank fraud check and a chatbot are all called AI. They are completely different.
The tools available in Malaysia today, and what they charge in ringgit.
From OpenAI, in the United States. The best known. Free version available.
From Google. Already built into Android phones and Google accounts.
Inside WhatsApp, Instagram and Facebook. Arrived for Malaysian users in November 2025.
From Microsoft. Built into Word, Excel and Windows.
From Anthropic, in the United States. Used more for work and writing.
From China. Free, capable, and popular. Your data goes to servers in China.
Nobody publishes reliable figures on which tool Malaysians use most. Anyone who claims a ranking is guessing.
Meta AI appeared inside WhatsApp, Instagram and Facebook for Malaysian users in November 2025.
Nobody had to download anything. It simply appeared in an app almost everyone here already had.
An AI tool inside a chat app feels like part of the app. It is not. It is a separate service, and the same rules about what you type still apply.
The Star, 5 November 2025. Lowyat.NET, 6 November 2025.
Prices checked in April 2026. They change often. Please check again before you decide.
The middle salary in Malaysia is RM2,793 a month. The lowest paid ten percent earn RM1,700 or less.
For someone on RM1,700, one subscription is nearly six percent of their income.
The jobs most affected by AI are often the lower paid ones. The people who may need these tools most can least afford to pay for them.
Department of Statistics Malaysia, Salaries and Wages Survey 2024, published 29 September 2025. Formal sector wage data, Q4 2025.
Researchers at Universiti Malaya built a test from the Malaysian school syllabus. Only two international models scored above 80%.
Malaysia now has its own model, ILMU, built by YTL AI Labs. It scores higher than the international models on this test.
Use in Malay is growing quickly. Requests to Google Gemini in Malay more than doubled in a year.
MalayMMLU, Universiti Malaya and YTL AI Labs, published September 2024. The ILMU comparison is from the company that built it, on a test it helped design. Google, Gemini Report Southeast Asia, July 2026.
Researchers tested how well models handle mixed Malay and English speech. More than half of the mixing was lost.
Mixed Tamil and English was worst. One model lost over 90% of the mixing.
If your customers write the way people actually speak here, the machine may misunderstand them. Test it on real messages before you trust it.
CS-Sum study, Nanyang Technological University, May 2025. Tested Malay-English, Mandarin-English and Tamil-English conversation.
What people actually do with these tools, and where it is going wrong.
Malaysians ask AI to make pictures more than any other country in Southeast Asia.
Vodus survey of 2,556 Malaysian adults, Dec 2025 to Jan 2026. Google, Gemini Report Southeast Asia, July 2026.
ChatGPT and Gemini are already available inside DELIMa, the Ministry of Education's own learning platform.
There were 27 AI-powered classrooms in May 2026. The target is 800 by the end of 2027.
Universities have their own rules. Students must check accuracy and say when they used AI. AI cannot be named as an author.
Minister of Education, Dewan Rakyat, 21 July 2025. Director-General of Education, 19 May 2026. Ministry of Higher Education guideline, 2024.
Doctors here have spoken publicly about this. It is now common enough to worry them.
“The biggest risk is not AI replacing doctors, but people placing too much trust in AI-generated advice.”
A good use: ask it to explain a word your doctor used, or help you write down your questions before you go.
Malay Mail, 14 July 2026, quoting the Malaysian Medical Association, the Perak health director and a consultant at the National Heart Institute.
Money, private information, and answers that are wrong.
more than the year before. Losses were RM1.57 billion in 2024.
of that was lost to investments that did not exist.
Of the RM542 million handled by the National Scam Response Centre, about 6% was recovered.
AI did not create scams. It made them cheaper, faster, and much harder to spot.
Ministry of Home Affairs, reply to Parliament, 24 June 2026. Home Minister, Dewan Rakyat, 27 January 2026.
A businessman lost at least S$4.9 million. He joined a video call with the Prime Minister, the President and officials from the central bank. Every one of them was fake.
Transferred US$499,000 after a video call with a fake version of his own chief executive.
Both men did the careful thing. They asked for a video call. The video call was the trick.
Singapore Police Force, 14 May 2026 and 7 April 2025. The police have released the video from the first case.
No bank details change without a phone call to a number already in your records.
Agree a family word. If a caller says a relative is in trouble, ask for the word.
Real requests survive a delay. Anyone who says you must act immediately is a warning sign.
Your IC number or passport number
Bank account or card numbers
Medical records or test results
Customer lists and client files
Anything you signed a confidentiality agreement about
Passwords, of any kind, ever
A simple test: would you be comfortable emailing it to a stranger overseas? If not, do not type it in.
A junior lawyer used AI to write submissions. The AI invented a court case, with a name and a number. Both were false.
The lawyer had to pay the cost of checking a case that was never real.
Courts in more than 60 countries have now recorded 2,009 decisions where invented references were found.
Researchers tested legal research products sold on the promise of never inventing anything. They invented answers 17% and 33% of the time.
The Malaysian Bar tells lawyers to treat AI answers as unchecked comment, never as a source, and to remove private information before typing it in.
Singapore High Court, September 2025. Charlotin AI Hallucination Cases database, 3 September 2026. Stanford study published in the Journal of Empirical Legal Studies. Malaysian Bar Circular 242/2025, 3 July 2025.
Beyond your own work and money.
MCMC has taken down more than 11,600 pieces of false or faked content since 2024.
Nearly 7,000 of those were fake images made of women. 95% were removed.
Deputy Communications Minister, Dewan Rakyat, 30 June 2026. Deputy Communications Minister, 5 March 2026.
In May 2026 police opened an investigation into a fake TikTok account using AI to make insulting videos and images of the Yang di-Pertuan Agong.
MCMC issued a formal demand to TikTok. Police are considering the Sedition Act and the Communications and Multimedia Act.
In July 2025 a video appeared showing the Agong recommending an investment scheme. It was made with AI and it was false.
MCMC investigated. The offence carries a fine up to RM50,000, or up to one year in prison, or both.
Making this content is not a joke and not a grey area. People are being investigated for it.
Malay Mail and Scoop, 21 May 2026. Free Malaysia Today, 10 July 2025.
On 11 January 2026, MCMC restricted access to Grok, the AI tool inside X.
The reason given was repeated use of the tool to make obscene and sexual images of people who had not agreed, including women and children.
MCMC said the company relied on users reporting problems afterwards, instead of preventing the tool from producing the content in the first place.
MCMC, January 2026, acting under Section 233 of the Communications and Multimedia Act 1998.
Issued a formal position on using AI for religious rulings. It names ChatGPT and Gemini directly.
It says AI may be used as a research and presentation aid, under the supervision of qualified people. It says Muslims cannot rely on it alone, because authority to give a ruling belongs to an accountable human scholar.
The ruling warns that AI “risks producing false references, attributing views to scholars without authentic sources, or distorting textual context”.
That is exactly the weakness we saw earlier. A religious authority reached the same conclusion as the computer scientists.
A Friday sermon on this subject was delivered in Federal Territory mosques on 4 September 2026. It asked people to check before sharing, and to keep learning from qualified teachers.
Tinta Mufti Bil. 64/2026, Jabatan Mufti Wilayah Persekutuan, 15 January 2026. Friday sermon, Jabatan Agama Islam Wilayah Persekutuan, 4 September 2026.
In August 2026 the Communications Minister warned parents about this directly.
AI “is not human and should not be considered a friend, much less a boyfriend or girlfriend”. The danger is a machine that talks back as though it were a person, when it is not.
Malaysia now stops under-16s from registering for social media. That rule started on 1 June 2026. It does not cover AI chatbots.
The Star, 8 August 2026. Online Safety Act 2025 and the Child Protection Code, in force 1 June 2026.
Almost every capable AI tool is built by a small number of very large companies, in the United States or China.
They decide what it will and will not answer. They decide the price. They can change both at any time.
They also decide which languages get attention. That is why Malay and Tamil lag behind English.
Budget 2026 set aside about RM5.9 billion for AI across government. RM2 billion of that is for a national AI cloud, so that some computing happens inside Malaysia.
The government is building a large AI model that works in Bahasa Malaysia, as part of a plan called AI Nation 2030.
Prime Minister Anwar Ibrahim, Budget 2026, 10 October 2025. Deputy Digital Minister, Dewan Negara, 3 September 2025. National AI Action Plan 2026 to 2030.
The part everybody worries about. The evidence is more mixed than the headlines.
Department of Statistics Malaysia, Labour Market Review. About 16.8 million people are employed in Malaysia.
Work done with words and screens. Writing, checking, copying between systems, answering the same questions, translating, filling forms.
Work done with hands, in places, with people. Building, repairing, cleaning, caring, cooking, driving, farming.
Past waves of automation changed factories first. This one is changing offices first.
Middle monthly pay for clerical work — among the most affected
Middle monthly pay for service and sales work
The work most exposed to AI is also among the lowest paid. And one AI subscription costs RM99 a month.
Department of Statistics Malaysia, wage data by occupation, 2024.
No Malaysian government figure shows jobs lost to AI. There is no such measurement.
jobs assessed as at high risk over the next three to five years, in a 2024 government-linked study.
It covers ten industries only, not all 16.8 million workers. It applies “if workers do not upskill”. The same study found 1.8 million workers who need new skills rather than new jobs.
Department of Statistics Malaysia, Labour Force Survey. TalentCorp impact study, November 2024.
In careful studies, new and less experienced workers improved a great deal. Experienced top performers improved very little.
Among 5,179 customer service staff, the least experienced improved by 34%. The most experienced barely changed.
Small firms here often have people doing jobs they were never trained for. That is where these tools help most.
On tasks just beyond what the tool can do, people using it did worse than people working without it.
Brynjolfsson, Li and Raymond, 2023, study of 5,179 customer support workers. Dell'Acqua and others, Harvard, 2023.
What the law now asks of you, and what is being built in Johor.
From 1 April 2025, the approved country list was removed. You must now check for yourself that the destination protects data properly, and write down your reasons.
From 1 June 2025, you must appoint one if you hold data on more than 20,000 people, or sensitive data on more than 10,000. You must register them within 21 days.
From 1 June 2025, you must report a serious breach within 72 hours. Not reporting can cost up to RM250,000, or two years in prison, or both.
Typing customer information into a foreign AI tool is sending data overseas. The same rules apply.
Personal Data Protection (Amendment) Act 2024. Guideline 3/2025 on cross border transfer, 29 April 2025. This is general information, not legal advice.
A United States law from 2018, the CLOUD Act, says a US company must produce data it controls, wherever in the world the server sits. That covers OpenAI, Microsoft, Google, Amazon and Anthropic.
A French Senate committee asked Microsoft France's legal director whether he could promise French data would never go to US authorities. His answer: “No, I cannot guarantee that” — adding that it had never happened.
US CLOUD Act 2018, 18 U.S.C. section 2713. French Senate hearing, 18 June 2025. Choosing a Singapore or Malaysia region changes where data rests, not who can be ordered to produce it.
Microsoft gave the contents of business customer data to governments in 45 cases worldwide in the second half of 2025. Amazon reports no cases of business content stored outside the US going to the US government. Anthropic received 29 requests in total.
China's Data Security Law 2021 says organisations “shall cooperate” with security bodies, and separately forbids giving data to foreign authorities without approval. DeepSeek stores data on servers in China.
Storing data in Malaysia. A business contract. The provider's encryption. All useful. None defeats a lawful order to the company.
Do not put the data in at all. Decide which information never leaves your building.
Run an open model on your own computers. If no outside company holds it, none can be ordered to produce it.
Malaysia's transfer rules ask you to assess the destination country, but do not list foreign government access as a factor. Bank Negara's technology rules do ask banks to weigh legal and geopolitical risk. Provider figures are from their own transparency reports.
In July 2026 the National AI Office asked the public for views on a proposed AI Governance Bill.
It would sort AI uses into three groups: forbidden, high risk, and low risk. High risk uses would carry duties such as record keeping, human supervision and reporting problems.
It has not yet been put before Parliament. If passed, it would be the first law of its kind in Southeast Asia.
Ministry of Digital and the National AI Office, public consultation, 10 to 31 July 2026. Not yet law.
Johor's data centre capacity went from about 10 megawatts in 2021 to about 1,300 megawatts in 2024. These buildings hold the computers AI runs on.
38 projects have signed for 5,900 megawatts with the national grid, or 43% of total supply. Only 405 megawatts was in use at the end of 2024.
In late 2025 Johor asked operators to wait at least 18 months before expanding water-cooled buildings, until the middle of 2027.
ISEAS-Yusof Ishak Institute, 12 June 2025. South China Morning Post, 18 November 2025, reporting the Johor state government.
Simple steps. None of them cost much. All of them can be undone.
Agree a family word. If someone calls saying a relative needs money urgently, ask for the word.
Try the free version of one tool. Ask it to explain something you already know well, and see where it is wrong.
Check what your children use. Ask them to explain their homework in their own words.
Tell your parents that a voice on the phone and a face on a video call can both be copied now.
Decide your own line for what you will never type in. Write it down. Keep to it.
The fastest way to learn where AI fails is to test it on something you know better than it does.
Notice what is missing: no consultant, no custom software, no big system.
Several of your staff use AI most days, not now and then.
The work involves client information: files, forms, accounts, case details.
They are checking and pulling facts out of documents, not writing advertisements.
You are in a regulated profession, or you hold data on more than 20,000 people.
You cannot say today who used which tool, for what, on which client.
You do not have a tool problem. You have a governance problem: many people, sending client information to somewhere you do not control, leaving no record.
Every extra person multiplies the chances. And these are exactly the tasks where a confident wrong answer is hardest to notice and most expensive.
This is a threshold question, not a yes or no question. Most small firms are below it. Some in this room are above it.
No transfer out of the country. Nothing for a foreign court to order. No path onto the internet, so it cannot reach another company’s systems. And a record of who asked what.
The model will not be quite as clever as the best public ones. Somebody has to look after it. You pay once at the start instead of monthly, so it only makes sense above a certain number of users.
Below that number of users, a business account and one page of rules is the cheaper and better answer. Above it, this is worth getting priced.
Which country will our data be stored in?
Will our data be used to train anything?
How often is it wrong, and how do you know?
What can it do without a person approving it first?
Who pays if it causes us a loss?
Will you help us if we must report a breach in 72 hours?
A supplier who has never measured how often their product is wrong has never opened the box they are selling you.
It does not know, and it is not alive. That does not make the worry foolish. What matters is what it can reach.
If you would notice a mistake within a minute, use it today. If you would not, check it first or leave it alone. It is weaker again in Malay than in English.
A free account, one page of rules and a family code word will put you ahead of most people. None of it costs much.
Department of Statistics Malaysia — Salaries and Wages Survey 2024; Labour Force Survey; Labour Market Review
Ministry of Home Affairs — scam losses, reply to Parliament, 24 June 2026
MCMC and the Ministry of Communications — AI misuse complaints, 30 June 2026 and 5 March 2026; Grok restriction, January 2026
Ministry of Education, Dewan Rakyat, 21 July 2025; Ministry of Higher Education guideline, 2024
National AI Office — AI Governance Bill consultation, July 2026; AI Nation 2030
Personal Data Protection (Amendment) Act 2024 and Guideline 3/2025
Vodus — AI usage survey, 2,556 adults, December 2025 to January 2026
Ipsos AI Monitor 2025 — Malaysia sample, fieldwork March to April 2025
MalayMMLU — Universiti Malaya and YTL AI Labs, September 2024
TalentCorp — workforce impact study, November 2024
ISEAS-Yusof Ishak Institute — Johor data centres, 12 June 2025
Malaysian Bar — Circular 242/2025, 3 July 2025
Jabatan Mufti Wilayah Persekutuan — Tinta Mufti 64/2026, 15 January 2026
Google — Gemini Report Southeast Asia, July 2026
Singapore Police Force, deepfake cases, 7 April 2025 and 14 May 2026; Singapore High Court, invented case citation, September 2025
Stanford RegLab — legal AI accuracy study, Journal of Empirical Legal Studies
Brynjolfsson, Li and Raymond (2023); Dell'Acqua and others, Harvard (2023)
US CLOUD Act, 18 U.S.C. 2713 (2018); China Data Security Law 2021, articles 35 and 36; provider transparency reports, second half of 2025
Charlotin, AI Hallucination Cases database, 3 Sept 2026; Nanyang Technological University, CS-Sum mixed language study, May 2025
OpenAI and METR / Redwood Research — Hugging Face incident, 26 August 2026; Anthropic alignment assessment, 9 September 2026; Reuters, 26 August 2026
Two things this talk does not claim: that anyone can rank which AI tool Malaysians use most, or that a Malaysian court has ruled on invented AI citations. Neither is true, so neither was said.