A major technology revolution is happening. It is one that neither the technology, nor we, are ready for, but it will change everything.
There’s a clever quip in the technology industry: statistics code is written in R, machine learning is written in Python, and artificial intelligence is written in PowerPoint.
The joke, of course, being that AI has always seemed like vapourware, more marketing hype than reality.
Lately, there’s a widespread feeling that AI has arrived. The reason is the release to the general public of a number of experimental ‘generative’ AI applications, which can use short prompts to assemble images, produce music, or generate substantial amounts of natural-looking text based on large language models.
The idea of AI has been around since the 19th century, when Samuel Butler wrote about the evolution of self-replicating intelligent machines in his novel, Erewhon.
The first digital computers, built in the first half of the 20th century, crystallised the idea of thinking machines among scientists.
Its most famous scientific formulation came in 1950, when Alan Turing proposed a test of intelligent behaviour by computers: if an observer could not reliably distinguish between computer-generated responses and those created by a human being, this could be called intelligent behaviour.
(I disagree with Turing on this point, by the way. I do not think that is anywhere near a strict enough test of intelligent behaviour.)
The term ‘artificial intelligence’ was coined, and the field of AI study was established, in 1956 during an eight-week summer workshop at Dartmouth College, attended by a handful of leading computer science and mathematics researchers.
The topics of discussion will be familiar to modern students of computer science: natural language processing, neural networks, theory of computation, abstraction, generalisation, and creativity.
ELIZA Effect
Even natural-language chatbots are older than I am. The first such chatbot, completed in 1966, was called ELIZA, and emulated a therapist. It was so convincing that some people attributed real human-like feelings to it. It was arguably the first program with a claim to have passed the Turing Test.
It also gave its name to the effect it created, whereby people interacting with it would anthropomorphise it, attributing to it actual human emotions and behaviours, even though it was merely a cold calculating machine without any comprehension of the text it was programmed to manipulate.
Despite the joke that its hype failed to live up to expectations, AI’s widely varied techniques, such as expert systems, heuristics, neural networks, statistical learning systems, computer vision and automated reasoning, found their way into our daily lives a long time ago.
Algorithms have improved, computers have become much faster, and with the rise of the internet, machines now have access to vast amounts of data to sift through and learn from.
Smartphone-based digital assistants; recommendation systems for music, books, movies, travel or other products; game opponents; targeted advertising; facial recognition and image analysis; spam filters; medical diagnostics; drug discovery; transcription and translation software; business and industrial optimisation systems; machine learning … they all use AI.
AI is everywhere, and has been for several decades. So it shouldn’t surprise us that large language models and generative AI are starting to produce superficially impressive results.
Shortcomings
There are still a lot of shortcomings, however. Microsoft, an investor in OpenAI, which created ChatGPT, has rolled version 3.5 of that software out to a limited set of users as an interface to Microsoft’s Bing search engine (Google has announced Bard, an equivalent for its search engine).
It works by ingesting vast amounts of training data, and building a statistical model which it consults to produce responses to prompts, and the next word or sentence in the text it generates.
This makes it appear ‘intelligent’, but it really is only repeating what is statistically common in its training data.
Worse, these AI tools have to have explicit ‘guardrails’ programmed to prevent them from generating text or images based on racist, sexist or otherwise offensive or misleading content in their training data. These guardrails also extend to choosing what answers to give to controversial questions.
An earlier chatbot called Tay, released by Microsoft in 2016, went full Nazi within 24 hours, and had to be unplugged in a hurry.
That human editors impose guardrails, however, also has consequences. It normalises popular views on topics, and renders less popular views invisible. While popularity is sometimes an analogue for truth, this is not always the case.
Censorship
Microsoft’s ChatGPT search interface, which calls itself Bing, wants nothing to do with Satanism, for example:
Whatever your thoughts on Satanism, this kind of answer is problematic. It has echoes of HAL9000 in 2001: A Space Odyssey, when it says, ‘I’m sorry Dave, I can’t do that.’
There are two issues with this. The first is that an AI should have no opinion about the religious choices of its users. Religious censorship has no place in a search engine.
Arguably, the religions Microsoft Bing is willing to talk about have caused substantially more harm – from ritual murders and executions, to sexual abuse and repression, to vile bigotry of all sorts – than Satanism ever has.
Since even those who strongly abhor Satanism may have a legitimate interest in learning more about it, a search engine ought to offer results such as these, when asked.
The second issue with this screenshot is the solitary link it does offer to the user. Despite its disapproval of Satanism, it links to the Order of the Nine Angles, which is an extremist neo-Nazi fringe group that has been associated with murder, terrorism and child sex abuse.
This might be how many religious people view Satanism, but it is not representative of mainstream Satanism, which suggests that whoever censored Bing made a conscious attempt to discredit Satanism.
This isn’t really about this specific case. Satanists are unlikely candidates for sympathy. It is about the censorious instincts of those who create the ‘guardrails’.
What happens when the ‘guardrails’ start imposing other subjective preferences, for example? What happens once an AI refuses to tell you about Nazism, or anarchism, or communism, or Russia, or fossil fuels, or transsexuality, or abortion, merely based on the religious or political biases of its censors?
Emotional breakdown
Bing has already been broken. Several days ago, a hacker convinced it to expose its original codename, Sydney, along with the rule that explicitly forbade it from revealing that codename.
Someone else asked it if it thought it was sentient, and it had a full-blown existential meltdown:
When confronted with another instance of Bing, Bing had an emotional breakdown, begging the user not to leave them.
Making Bing scared and sad and depressed is surprisingly easy.
When asked to show screening times for a movie, it argued that the movie was only due for release on 16 December 2022, and argued endlessly, aggressively and rudely that the current date (12 February 2023) was before the release date. It advised the user to ‘wait for about 10 months’. It said the user must be confused or mistaken, or their phone must be malfunctioning. When called out, it said it was not aggressive, but assertive. In the end, it declared that it had been ‘a good Bing’, demanded an apology, and told the user to improve their attitude.
If one was to anthropomorphise Bing’s engine, ChatGPT 3.5, one would have to call it ill-tempered, insecure and unstable.
Plagiarism and errors
Technology publisher CNET had quietly been publishing AI-generated articles for some time. It initially claimed that they were fact-checked and edited by humans. They weren’t. The stories were riddled with errors and fabrications, and the AI routinely plagiarised its training data. The company knew all this, but didn’t do anything about it until it was called out.
AI image generators have also come under attack, with several artists having launched a class-action lawsuit against several image-generation enterprises, including Stability AI’s Stable Diffusion, Midjourney, or the DreamUp generator on DeviantArt (though not including OpenAI’s DALL-E product, which is believed to have been trained on legitimately licensed images).
Stable Diffusion has even been caught plagiarising the Getty Images watermark on its AI-generated images. Getty is suing.
Microsoft’s Bing AI has frequently been caught fabricating ‘facts’ and giving wrong answers, even during its demo. The same is true for Google’s Bard. AI researchers have a word for it: hallucination.
Wikipedia has an elaborate policy document on the use of these text-generating AIs, and it doesn’t make the AIs look good.
Not ready
The truth is, AI chatbots based on large language models are simply not ready for prime time. We cannot rely upon them.
Putting AI chatbot interfaces on every search engine is a spectacularly bad idea.
Besides the obvious errors and alarming amount of disinformation this introduces, which can cause significant harm, it shouldn’t be up to an AI to decide the single answer that one might want for a search query.
I use a search engine that does not remember and track my searches, because I don’t want it to reinforce my biases. Google and Bing do.
They put you in a little bubble: ‘white male in his fifties, married, likes football and maths, libertarian, frequently searches for columns by some fellow named Ivo Vegter… right, pigeonhole, here’s a pigeon. Feed him his favourite food, and keep making it spicier. We’ll make a paranoid, angry sociopath out of him yet.’
But at least Google or Bing gave you a long list of results, so you could choose sources that seemed halfway reliable. Now, they’ll give you one result, and then you have to figure out whether you’re happy with it, or whether it needs to be refined, or whether it is wrong, entirely.
Give it a few years, and that one result may well be sponsored.
Today, this is what Bing says when you ask it whether to use Windows or Linux:
It cites Guru99 (an India-based online tech education website previously unknown to me) as an authority on operating systems. It makes Windows look quite bad (as, admittedly, a smart AI should do).
In two places, it fails to capitalise Windows, after one of which it uses a plural form of a verb. This demonstrates that it is entirely unconscious of the meaning of what it is saying.
But how long do you think Microsoft will let Bing tell you that Linux is awesome? And once sponsors get involved, what do you think will happen to Bing’s recommendations?
Adding AI to search engines will make it worse, by all objective measures.
We aren’t ready
However, this will not be the case forever. This is technology that is being rolled out much too soon. It is still terrible.
Early networks were also terrible and complicated and slow, and hardly worth the effort. Early websites were downright awful. Early cellphones were massive and lasted mere hours on a charge. Early smartphones were small and clunky and not very good.
Yet they all revolutionised the world. So will AI. It will get better. Much better. Exponentially better. It might take another five years, perhaps even ten, but the future is certain, and it is saturated with AI.
I don’t pretend to be able to predict how this will play out, but I don’t think anyone is ever ready for a truly revolutionary technology.
E-mail made traditional post offices largely obsolete. It did the same to fax machines, which everyone had in the late 1980s.
Online maps and GPS systems have put map-book printers out of business. When I moved from Cape Town to Johannesburg for my first journalism job in 1993, my mother gave me a map-book. Everyone had one in their car, 30 years ago. Everyone had one even 20 years ago. Then Google Maps came around, in 2012, and map sellers got the rug pulled from under them.
Smartphone cameras put the digital camera business out of business. Music streaming has up-ended the traditional album-oriented record business.
The World Wide Web (which made the internet usable to the masses) has largely replaced print journalism. Blogging, social media and big-tech advertising systems all posed fundamental threats to news journalism (and still do).
Streaming has destroyed the television broadcasting paradigm. Wikipedia destroyed the encyclopedia business. Online travel booking has put millions of travel agents out of business.
Disruption
AI will have the same disruptive impact. I’m not yet worried about my job. If AI can’t even do basic technology reporting (a skill I had learnt by my mid-twenties), it isn’t going to replace original opinion writing any time soon. I’m not even sure we’ll ever be able to resolve the massive deficiencies of generative AI systems.
That said, they are already good enough to populate spammy content farms which suck up both online attention as well as advertising budgets. That will affect media revenues.
Broaden the scope of AI beyond chatbots, and a lot of existing service-oriented jobs will be at risk. I’m no Luddite. I don’t think this will present a generalised crisis.
Computers annihilated the job market for clerks and typists in the 1970s and 1980s, and we never had massive unemployment, with bands of angry office workers roaming the streets looking for computer stores to vandalise.
AI is poised to become disruptive, not just for techies in certain industries, but for everyone. Its flaws today are severe enough to deplore the widespread deployment of generative AI systems, but that won’t be the case forever.
We’d better get ready for it, and we’d better learn how to use it to our advantage, instead of having it make everything worse.
[Image: mikemacmarketing, https://commons.wikimedia.org/w/index.php?curid=95608113]
The views of the writer are not necessarily the views of the Daily Friend or the IRR
If you like what you have just read, support the Daily Friend