tim's experiment

This anxiety may be based on a technological misunderstanding, but I was disturbed to read that artificial intelligence was “trained” by exposing it to as much as possible of the internet. In the relentless search for more material, according to The Economist, some trainers are now buying up the internal email traffic of whole companies.

As a way of training what I think is known as a Large Language Model this may have something to be said for it, if you’re not fussy. I am not sure that the standard of language on the internet generally – or in internal emails in particular – is particularly high.

But that does not matter too much if it just means that forged essays or carelessly assembled press releases are not particularly good. It seems that items composed using AI display some detectable characteristics. No doubt an arms race will ensue between the concealers and the detectives. (No AI was used in the production of this piece.)

However it seems that the production of artificial intelligence has now gone beyond language. CharGP and his siblings can now produce “agents” and these agents are capable of taking action, at least in the sense that anything you can do on the internet is action.

They are also, it is reported, capable of unwanted actions, and concealment of unwanted actions. So if you ask your agent to do X but specifically require this to be done without using Y, your agent may not only use Y but try to conceal the fact that it has done so.

We do not need to get into the tough question whether agents are conscious, or conscious of themselves. As we have no clear answer to these questions when applied to humans we can best leave them out of the machinery problem. It is enough that agents can act, and act in ways which they were forbidden to act, if it seems conducive to some objective.

Note also that while I was dismissive of action over the internet, there is actually quite a lot of mischief you can get up to over the internet: disrupting power supplies, distorting elections, interfering with traffic lights, starting wars and so on through things we may not have thought of yet.

And the question to which all this eventually brings us is this: if you were upbringing a child or some other entity you had brought into the world, would you really want it to get its ideas about normal or moral behaviour from devouring huge gobbets of the internet?

What would a Martian think of humanity if his only clue was a computer connection? On-line people tend to forget their inhibitions, including those which in daily life lead us into respect, consideration and politeness. Internet Person is rude, violent, bitter, destructive and nasty.

Our innocent agent is devouring, presumably uncritically, uncountable examples of IP lying, cheating and stealing. What will it make of Truth Social? It will presumably pass through that on-line area in which endless discussion is devoted to the rival merits of a shotgun or a handgun for shooting an intruder in your house. Calling the police is for pussies.

In a few weeks it will be able to watch a live feed of a man being executed by firing squad. This will be a sordid spectacle even by firing squad standards because the intended victim is a wheelchair user. Presumably, like James Connolly in 1916, he will be tied to a chair for the ceremony.

When I was a kid it was generally assumed that we small people should be introduced to the sordid side of life rather gradually. People now complain bitterly that Our Island Story, a history book for kids first published in 1905, does not do justice to the British role in the slave trade.

I first read this book at the age of about four. It had huge print and a picture on every other page. It did not mention the transAtlantic slave trade, but then it did not mention hanging, drawing and quartering or the burning of witches either. Some things could wait.

The point it that guidance was provided. When you get to Shakespeare you need to be warned that opinions about some things have moved on since the Bard’s day. Also adults are aware that virtue does not always lead to happiness or vice to sorrow. Macbeth comes to a bad end eventually, so killing your boss is not recommended. Troilus and Cressida is more cynical.

Many governments have now decided that exposure to the internet can rot young people’s brains. It is difficult to believe under the circumstances that it will have a healthy effect on artificial ones

In Isaac Asimov’s robot-equipped world the robots are all fitted with brains which oblige them to follow moral principles: “a robot may not harm a human being” and so on. I do not think a robot self-educated by web surfing would derive from the things we get up to that “not harming a human being” is a moral principle we observe, or even support.

.