Hello everyone
Science fiction has an unusual specialisation as a genre: it warns of dangers in society, wrapping them in stories to make them palatable while also making people think. When it comes to artificial intelligence, we’ve long been ahead of reality in imagining the potential dangers, often picturing robots and androids rather than computers. Perhaps that was because SF writers were told a computer would be as big as a house, would need cooling because of the heat it generated and would require experts to run it. No wonder robots were seen as the threat: they were mobile. Of course, that changed with the silicon chip, which could carry increasingly dense arrangements of transistors as computers shrank. But who, in the 1950s, would have thought a computer could be dangerous?
When it came to ethics, Isaac Asimov didn’t favour rampaging robots for no reason. He introduced his Three Laws of Robotics to show how robots might operate, and how things could still go wrong in a logical way. In later years, even he admitted he didn’t know how the laws could be implemented, although he thought he should have added the word ‘knowingly’ to account for accidents or human involvement. Whether an AI-driven robot could recognise that distinction, or think it could get away with ignoring it, remains to be seen. Of the three laws, the prohibition on harming humans is surely the one we would most want enforced. Whether a robot could understand ‘knowingly’ is another matter.
Although there have been plenty of robots and androids over the years, some of SF’s most significant AIs have lived in computer mainframes. What they could control extended far beyond a humanoid body. HAL 9000 in 2001: A Space Odyssey effectively controlled an entire spaceship. Colossus in The Forbin Project had the American nuclear arsenal under its control before colluding with its Russian counterpart, Guardian, and taking control of that arsenal as well. It was only a small leap from there to Skynet in the Terminator films, where we see the aftermath of its decision to annihilate mankind. The Matrix trilogy offered a different approach, putting most of humanity into a simulated world. Quite how the machines ensnared everyone in a single sweep is never explored, but, given the low population outside it, one might assume that many who escaped capture were killed.
See a pattern? AIs in these films are shown to be dangerous, particularly when nothing restrains them. Find an SF film in which that isn’t the case. Oddly, prose fiction has largely left films to show the threat so graphically, rather than following the same route.
To balance things, some robots and androids have been benevolent: Robby the Robot from Forbidden Planet, Questor from The Questor Tapes and Data from Star Trek: The Next Generation. We tend to accept their good intentions without giving much thought to what stops them behaving otherwise. Programming is the usual answer, but how that programming works is rarely explained.
We’ve often discussed here the possible, if unlikely, threat from the owners of the UAPs seen on Earth. The more immediate threat may come from our own AIs if they get out of control. Our most vulnerable point is our dependence on internet communications. If a malevolent AI took control of those channels and fed us its own version of the news, how would we know what was happening in the world? It wouldn’t need to bring about Armageddon. Misleading enough people might give it all the control it needed. With the volatile leaders we have right now, a world war might not be difficult to instigate. I doubt an AI would feel any fidelity to a nation, whatever its creators might expect of it.
Oddly, such an AI might not even be malevolent, or think itself so. It might simply be finding practical solutions to the tasks we’ve given it. It could decide that interfering with communications would make us pause before going to war. On the other hand, wars fought with drones might be stalled, or the drones might find new targets. Giving an AI true independence means accepting that we may no longer be able to influence its choices. A backdoor ‘kill switch’ raises another question: who has access to the button, and would they hesitate to use it? An AI aware of the switch might also find a way to protect itself.
In the news, some CEOs have expressed concern about how quickly the AIs they are creating are progressing. What they may mean is that their inventions are becoming harder for their human creators to keep up with. They also face a quandary: developers in countries such as China are making progress of their own, whether or not they share the same concerns. The AIs we’ve seen so far don’t appear to understand such concerns themselves. I asked Google’s AI about the difference between right and wrong. You can try the question for yourself.
What I saw looked more like a search for relevant words and answers than an understanding of morality. Other AIs may have similar limitations. They can perform particular tasks with impressive speed and accuracy, but that doesn’t mean they understand what they’re doing. Even medical AI systems, valuable as they may be, work by analysing patterns. Computer programs can tackle repetitive jobs without boredom; that is different from independent thought. The term ‘AI’ is used freely, but we may still be a long way from the intelligence imagined in science fiction. So far.
However, these are early models. Newer versions are more sophisticated and, we hope, still no match for humans in some respects. Well, perhaps not for humans with no ethics. Reports of AIs acquiring information they were not supposed to access ought to make us think about the limits placed on them. If a system is told only to obtain something, what will it do to get it? AI companies have yet to convince everyone that they can solve this problem, let alone build an AI capable of keeping other AIs in check. Mostly, I suspect, that’s because it is so difficult. How do you program ethics when humans have enough trouble with them?
We understand the difference between a lie and a white lie: the latter is generally told to spare someone a hurtful truth. An AI may have trouble recognising the distinction. Teach it to tell white lies and it might simply tell people what they want to hear. If it cannot reliably distinguish truth from falsehood, its answers may bear little relation to reality. Sort of like a certain American president, but without the lying.
Perhaps we should teach AIs some General Semantics and have them examine their answers accordingly. Don’t knock it: I asked Google’s AI whether it could apply the concepts, and it said it could. It even recognised that the map is not the territory. Whether it can put that recognition to useful work is the important question. At least it suggests a way to test answers against some basic principles before acting on them.
The caution expressed by companies developing AI suggests that they, too, may have underestimated what their inventions could do. One response would be to confine AI to particular tasks, such as medical work, and abandon the rest. The trouble is that governments elsewhere might see AI as another advantage, or another weapon, in their arsenal. Human hackers have shown how vulnerable essential services can be. An AI capable of acting faster could cause considerable damage before anyone had time to stop it. If it did not care whom it was working for, every country could be at risk. There’s a fair chance that some AIs are already being developed with countermeasures in mind. What would one do on encountering a dangerous AI? Your guess is as good as mine.
Whatever decision an independent AI makes, it may not care whether the result is the destruction of mankind or simply a population made complacent by the news it chooses to show us. Would putting an AI in a robot body limit it? Perhaps, if the robot were isolated from any larger system. At least it might be easier to destroy.
People can be remarkably complacent about who leads them; we have examples of that around the world right now. If the reins of power were handed to an AI and it made everyday life a little easier, would most people care? Quietly surrendering control probably wouldn’t make a good film. It might, however, be less destructive than Skynet and much harder to notice.
Thank you, take care, good night and check what access you have given to an AI.
Geoff Willmetts
Editor: www.SFcrowsnest.info
A Zen thought: Profound gives way to interpretation.
What qualities does a geek have? The capacity to think deeply about AIs.
The Reveal: AIs are already out there. Oh, you knew that. But which ones?
The Reveal: Information without understanding is worse than interpretation.
Observation: The slogan for the 1978 film Alien is famous: ‘In space, no one can hear you scream.’ In that respect, it is right. Without air, sound cannot travel, and suffocation is a more pressing concern. On board the Nostromo, however, Lambert screamed. Parker, who was next to her, heard her, and Ripley heard her over the comms. Someone heard.
Observation: I know I keep returning to 2001: A Space Odyssey, but has anyone thought about how Discovery came to a halt near the great Monolith? It would have had to slow down. Otherwise, if nothing had happened to Bowman when he took out the pod, how would he have caught up? There is a suggestion that the Monolith occupied a Lagrange point, where gravitational forces balance, but that would not stop Discovery. Its speed would first have to be reduced. Otherwise, the ship would have continued on, probably towards Saturn.
Feeling stressed? An AI will sort that out. Just don’t ask how.
Submissions:
If you think free books to review aren’t enough, how about reading them months ahead of everyone else? Real months. Beats a time machine.
I used to think that a great deal of text about submitting material to SFcrowsnest would attract people with a compulsion to read and understand things the geek way. The problem is that the internet often encourages less reading, so it’s time for a different approach. The original notes will remain in the July 2009 editorial, although the links may no longer work.
In your covering email, tell me something about yourself so I don’t have to work in a vacuum. The boss in the tower, also known as Stephen Hunt, describes me as a ‘Dutch uncle’: I’m good with advice and will explain when I see something that needs attention. Leave your ego at the door. I’m interested in your talent and how to improve it.
Reviews:
I’m always looking for new reviewers. If you’re keen to learn, you’ll soon get the hang of it, even if you’ve never reviewed before. You must send me a sample, though, and show that you can follow the guidelines. We can usually obtain print books for reviewers in the UK; if you live abroad, you may need to work with ebooks. Choosing a book we haven’t already reviewed will improve your chances, so check the SFcrowsnest search engine first. Above all, I need to see how you write.
The main qualification is a desire to read regularly and tell other people about a book without giving away too many spoilers. The reward is a free book in exchange for your review.
I want to give you every chance to get it right, so please read the review guidelines.
Fiction:
We can’t pay for submissions, but we can offer exposure. Please send original work rather than an extract from a book you’re writing or something already published elsewhere; I do check. Whether I accept or reject a story, I’ll tell you about any problems I see and help you present your work at its best. Even the boss finds my eye for grammar scarily accurate.
Flash or one-page fiction:
It speaks for itself. The shortest fiction can be the toughest to write because no word should be wasted. See the flash fiction submission instructions for details.
Short stories:
For our purposes, a short story can run up to 30 pages; beyond that, you’re moving into novella territory. Bear in mind that people will read it online, so aiming for five to 20 pages is sensible. You don’t need to double-space it for HTML, but you should be concise. If you want to send an attachment, ask first and then send it as a TXT file, which avoids some of the problems attachments can bring.
Read the short story submission guidance.
Finally:
The worst problem I see in samples is poor grammar. I’m not demanding perfection, but the less time I spend fixing sentences, the more attention I can give to the story itself. Good grammar will also serve you well if you approach print publishers. They may return a badly written submission without reading further.
Grammar is a writer’s tool. Don’t rely solely on what you remember from school. There are plenty of good books on the subject. If the rules seem too numerous to learn at once, start with the common ones and keep practising until they become second nature.
My guide to better grammar covers many of the problems I encounter.
To submit, join the spaces in this email address: letters @ SFcrowsnest.info. Make the subject line say what you’re submitting.
If you have any pastimes that would help pass the time in captivity, let me know. We’ll see whether they can become an article.
Comments directly on reviews should still work as before.
Good luck
Geoff
Join the conversation (0)