An AI apocalypse? Hong Kong’s real challenge is much less dramatic

For many of us, artificial intelligence (AI) started out as something to play with. We asked it to write poems and arrange our calendars; we shared its more entertaining mistakes. Then it became something to believe in: an escape from drudgery, an assistant to every profession, even perhaps the start of a new industrial revolution. Now it is something to fear.

The progression has been astonishing. Governments are scrambling to draw up AI strategies, companies are reorganising operations and investors are financing the infrastructure. The world seems to be mobilising at warp speed to join the AI rush, although there is rather less agreement about where we are going.

Hong Kong’s sensitivity to this race was illuminated by a recent report that the Monetary Authority had questioned a major bank about its decision to establish its AI centre in Singapore and hire over 100 AI specialists there. The competitive anxiety is understandable. Attracting the investment, however, is not the same as preparing for the risks that accompany it.

This gap should concern all of us. Investors may tolerate uncertainty, provided their AI investments deliver financial returns. A regulator cannot. Bringing AI into a business is one decision, but understanding what it may affect is another. The second decision is the one most easily skipped in a rush.

Alongside this rush came the warnings from leading AI developers about the dangers of the technology they are advancing.

There is an obvious irony. Some of those most concerned about the technology’s advance are also doing the most to advance it. Their businesses benefit from making their products more capable. While the conflict of interest is worth noting, it does not dispose of the argument. A warning can be self-serving without being wrong.

03:02

Chinese researchers question calls to slow down AI development as US tech leaders warn of catastrophe

But hearing a warning is not enough. For Hong Kong, the useful question is what companies should do. If companies adopt overseas tools carelessly, they risk importing someone else’s account of the problem along with the technology. This is where the government can step in to provide guidance.

It is important to note that a model that helps an attacker has not, by that fact alone, defeated any organisational defence. The attacker still needs to identify the vulnerabilities and the opportunity to act.

But some threats are not so forgiving. Familiar safeguards cannot be assumed to remain adequate simply because the threat sounds familiar. Fraud is old; the means of making it convincing are not. Besides, AI tools are improving by leaps and bounds, and while AI did not invent identity fraud, it can make age-old deception far more scalable.

A case reported by the Hong Kong police last month illustrates the point. Investigators alleged that a syndicate used identity-card images, photo-editing software and AI tools to create false documents for online account applications. Most were automatically rejected, but some slipped through.

The problem is not confined to bad actors. Last month, it was reported that unrepresented litigants were making error-strewn legal references, including citing non-existent ordinances – mistakes thought to be due to the use of generative AI. There was no malice but the risk to trust is real. The court was not fooled but those without legal training might have been easily persuaded.

01:33

Hongkongers lose HK$200 million to scams in a week, AI voice-cloning used

The danger is not merely that the machine gets something wrong. It is that users wield the tool irresponsibly. A large-language-model assistant on a phone can make expertise look deceptively easy, but true professional qualifications often require years of training. The time experts have to spend separating the wheat from the chaff is only part of the cost; others may act on the errors before anyone catches them.

Hong Kong’s government is examining the legal implications. The Department of Justice is leading an interdepartmental review of legislation to support wider AI use, while government policy also addresses AI-enabled crime, liability and the management of AI agents.

That is welcome. But the law is one thing; equipping people to understand it is another. Hong Kong’s judicial guidance on generative AI, for example, stresses the importance of understanding the technology’s limitations and verifying its output. If even the legal profession requires such supervision, what hope has the layman?

“Human oversight” is a reassuring phrase, but it means little if the person supposedly in charge is kept in the dark about how the system reached its conclusion. Oversight is meaningful only where the person can actually interrogate the result, and that capacity matters most in high-stakes domains.

03:33

Will AI replace conventional methods in forecasting typhoons in Hong Kong?

None of this requires us to treat every application as dangerous. Translating a routine document is not equivalent to approving a loan. Where AI accelerates an existing threat, augmented defences may suffice. However, where AI changes the chain of authority, something new is needed.

For example, AI agents may be handy, but what happens when an agent acts on faulty instructions? Suppose an AI agent wrongfully signs a contract on a principal’s behalf. Would that bind the parties? The question is no longer hypothetical.

Talk of an AI apocalypse attracts attention because it is dramatic. The challenge is less theatrical: preventing the gradual transfer of judgment to systems whose answers are easier to accept than to challenge.