Sir, – If I don’t put the handbrake on a car and it rolls away, and I then talk about it going rogue and rolling away without my permission, people would laugh at me, and so they should. The car isn’t being disobedient, it’s just a machine doing what it was built to do.
Large language models (LLMs) are the same. They are machines, not some sort of disobedient animal. They do not have desires. They do not go rogue. They are chatbots. They repeat versions of the text they were trained on.
If an LLM was trained on a human hacker’s interaction with a computer, it can replicate a version of that interaction. As with any machine, if the behaviour surprises you, that’s your mistake either for not using the machine properly, or for choosing the wrong machine for the job. – Yours, etc,
BEN AVELING,
READ MORE
Ranelagh,
Dublin.
Sir, – Kevin T Ryan is right to warn against deploying artificial intelligence systems in life-critical areas when their correctness cannot be guaranteed (Letters, September 17th).
The EU’s revised Product Liability Directive is welcome, but liability operates after something has gone wrong. That may be sufficient where damage can be compensated. It is of limited comfort where the possible consequence is irreversible catastrophe.
If a pharmaceutical company sought to release a medicine while potentially disastrous side effects remained inadequately investigated, unrestricted approval would not ordinarily be granted.
Why should AI be treated differently? Its developers should have to demonstrate that each big advance is acceptably safe before, not after, releasing it.
Ireland, meanwhile, is being encouraged to accommodate enormous new data centres required to develop ever more powerful AI. Should further planning permissions be granted, or construction permitted to continue, while leading figures within the industry acknowledge that adequate safeguards do not yet exist?
Applications for large AI-related data centres should at least be held in abeyance until binding regulation requires independent safety testing, disclosure of serious incidents, legal liability for harm, and internationally agreed safeguards ensuring that any decision to use nuclear weapons remains under meaningful human control.
Ryan is right that we should hope the first big failure is not catastrophic. Public policy must offer something stronger than hope. – Yours, etc,
PAUL O’SHEA,
Shankill,
Dublin 18.
