Generative AI: Is this the End of Human Creativity?

Image credit: Alexa Steinbrück / Better Images of AI / Explainable AI / CC-BY 4.0

It looks like we’re at the start of a new cycle of AI hysteria; current Limited-Memory technology is not yet delivering anything like Artificial General Intelligence (AGI), but suddenly Generative AI is here to take all our jobs and make human thought redundant. I suspect the average person, not being an AI geek, very quickly loses interest after the latest ‘advance’, usually described as ‘ground-breaking’, in the field of artificial intelligence is announced in the non-academic press. Until ChatGPT, a form of Internet search tool that can provide lucid answers to complex questions arrived with huge fanfare, the public were made aware of new developments in AI with headlines like: ‘Computer beats chess grand master’, or more recently ‘computer beats world champion in the game of Go’.

Limited-Memory AI

In recent years, much research effort has gone into making AI systems ‘clever enough’ to enable full automation of transportation (‘driverless’ vehicles) and factories (Industry 4.0). Key to achieving a level of automation sufficient to justify the removal human operator or supervisor is the development of sensor systems, usually video, and a ‘brain’ that can characterise what it sees and react accordingly. A neural network is ‘trained’ to recognise a limited range of relevant objects and sends this information to perhaps a more conventional computer system for control action. This form of AI, characterised by phrases such as ‘Deep Learning’, is basically a glorified pattern recognition system. This type of AI is not creative; once trained it doesn’t learn anything new, so is very unlikely to become an existential threat to the human race. All it does is find patterns in data - signatures of the objects it’s been taught to recognise. A big problem is that when a recognition mistake is made at ’run-time’, it’s impossible to trace the weakness in the trained dataset that caused the failure. It can never be 100% reliable when working in an environment with an infinite number of possibilities, such as a vision system for an autonomous vehicle trying to identify everything it sees out on the road. In real-time. With its point of view and the objects to be identified in constant motion. It has no experience to guide its decision making, few ‘skills’, no ‘common-sense’, no ‘lateral thinking’, and no ability to handle unexpected situations. In fact, it’s pretty stupid. Wiping out humanity? Fat chance.

Generative AI

Work on developing Large-Language Models (LLM) has been quietly going on in the background for years. Their purpose is to mechanise the concept of understanding natural language, so that robots can be instructed to perform tasks verbally by untrained staff with poor diction or strong accents. In return, the robot should speak in a human way and not sound like er.., a robot. Taking this further, the machine may be able to identify emotional states and respond accordingly. Another use would be reliable instant translation of one language to another. They are similar to Limited-Memory AI in basic structure, consisting of huge neural-networks trained on equally huge quantities of text gathered from databases such as Wikipedia. Whole words, parts of words and even phrases are converted to numbers (tokenized) to speed up subsequent processing and to keep mass-storage requirements to a manageable level. What makes this type of AI so powerful is that the tokenized words have a vast collection of parametric data calculated at the training stage attached to each token. This data consists of statistics of how often a token appears in the dataset, how often it appears with another word as a pair, and so on. ChatGPT uses all this data to come up with a perfectly crafted answer to a question or ‘prompt’, even a complete essay. The basic algorithm is quite simple: it builds sentences token by token, asking the question each time, “What word should come next?” Of course, it’s not that simple in practice, but that’s the principle. Earlier versions of the software looked at the words of the input question and the sentence constructed so far, then scanned the vast dataset for words with the highest probability of coming next. The token with the highest probability was then selected and the cycle repeated. The result: accurate and ‘safe’ if dull answers, using rather flat, uninspiring language. Results would be identical if the same input was used again. Good for search-engine results, but not very ‘creative’. Not very ‘natural’. The solution they came up with was simple, and could be called a stroke of genius, but as it turns out it could be the equivalent of ‘Letting the Genie out of the Bottle’ or ‘Opening Pandora’s Box’. Pick your favourite literary reference. The dull, repetitive output has gone thanks to the software not always selecting the highest probability token on the list, instead one tagged with a lower probability. Careful tweaking of this feature can create pleasant, non-robotic text, and maybe take the reader down fresh avenues of thought. But remember, it doesn’t understand anything, it is not having ideas. It’s up to you, the human reader, to interpret what appears to be original thought.

ChatGPT Plus is based on Generative Pre-trained Transformer 4 (GPT-4), a multimodal (it can handle both images and text) Large Language Model. If you want to know more about how it works then Stephen Wolfram’s book [1] is a very good starting point. But you still need a basic knowledge of AI techniques and statistical mathematics to avoid crashing your natural brain. There lies the question, why has this particular development sent academics and politicians into a frenzy, demanding that AI applications be heavily regulated in law?

The Good, the Bad and the Ugly

Internet content can be broadly divided into these three categories: the Good, containing useful factual data which is at least harmlessly entertaining if not educational; the Bad, error-strewn or unintentionally misleading information (‘anti-knowledge’) that has a negative impact on understanding; and finally the Ugly, that designed to defraud the unwary, poison the minds of the gullible or satisfy criminal urges deemed unacceptable in a modern society. There is a big worry, in my opinion wholly justified, that unrestricted access to Generative AI applications will lead to a massive increase in the amount of damaging content on the Web. Not only that, but illegal use of copyrighted material and privacy (and security) violations will become commonplace. These concerns have led to ‘tech-bosses’ such as Elon Musk signing an open letter to world governments asking for a pause on further AI development. Even the Deep Learning pioneer Geoffrey Hinton has expressed his fears for the future.

We are all doomed (again)

The doom-laden outcry that dominates the news media whenever there is a new AI ‘breakthrough’ is always the same:

• Huge numbers of jobs will be lost and
• Artificial intelligence will soon exceed that of the human race, and ‘they’ will no longer need us…

This is not the first occasion when creators of a new technology have worried that their discoveries will be misused. Remember Alfred Nobel and Dynamite, or Robert Oppenheimer and the atomic bomb? But it must be a first for modern creatives like Sam Altman, the boss of OpenAI, the company that launched GPT-4, to demand government regulation of their own products.

A possibility is that the news media having ‘cried wolf’ so many times with each new development in computing/AI/robotics, suggesting imminent disaster that ultimately never happened has made AI developers and engineers complacent about what could happen one day.

The Human Factor

What makes Generative AI so exciting, or terrifying, depending on your attitude towards robots, is that it appears to exhibit original thought. That’s a problem: up until now we have generally assumed that only the human brain is capable of having novel ‘ideas’. GPT-4 has exhibited some emergent behaviours, some expected, many not. The most obvious one when attached to a search-engine is to provide answers to requests for information, not just lists of URLs where the information may be found. This is of course intended, and when it works properly and delivers a correct answer with clear explanation, it’s brilliant. It can also write computer code and construct essays on any subject providing it’s given the right prompts. With their lucid, natural and supremely confident style, ChatGPT responses are so convincing. Just like the best human con artists, it makes you suspend all scepticism – you want to believe it’s true. Especially when only a few keystrokes takes you to a ready-packaged answer apparently saving hours of normal search engine activity. Unfortunately, just like a con artist, it can’t respond with “I don’t know the answer”; instead, it seems to ‘make up’ a response. The output is still well-formed, but it’s factually rubbish. The developers call this effect ‘Hallucinating’, in an attempt to humanise a major software failure. Perhaps a better word should be ‘lying’. An excellent example of uncritical acceptance of ChatGPT results concerns the humiliation in court of lawyers attempting to justify a ridiculous claim for damages. Schadenfreude or what? The lesson here is that the user should know enough about the real answer to avoid accepting garbage from the computer. Ideally, you should just use the tool to confirm the facts already in your mind.

A big concern amongst teaching academics is that students will use these tools to do all their work for them, not just research the facts but create the whole essay or homework test ‘oven-ready’, so bypassing the need to understand the subject, and as a result avoid learning anything. Every college will deny this is a problem, but if tutors are not up to speed with the latest technology, how do they know? Recently, a college lecturer had a class of students use ChatGPT to prepare essays and then assess them for accuracy and relevance. The results were a real eye-opener, with all essays containing significant numbers of ‘hallucinations’ and factual inaccuracies. Fake text is bad enough, but this type of software can also generate fake images and even fake video. Deepfake technology has already fallen into the hands of blackmailers and even extortionists extracting money by posing as a friend or relative on a real-time video call. I can see a time when no-one will trust anything they see, hear or read on the Internet – and then where will we be?

The bottom line here is simple: generative AI is creative and can produce amazing artwork, poems and fiction. When dealing with facts, the human brain must have the last word. It’s rather like the current state of autonomous car technology. Switch it on, but keep your hands near the steering wheel and your eyes on the road ahead. The human factor still needs to be part of the equation.

Finally

Is the human society at a turning point, when the process of handing over control of our destiny to machines with vastly superior intelligence becomes unstoppable? Not until there’s a real breakthrough, not just an incremental improvement in existing technology, which is what we have here. Albert Einstein was responsible for the classic equation E=mc², a real breakthrough which led straight to the development of both the peaceful and destructive use of nuclear power. He spent the rest of his life tinkering with his General Relativity equations, almost but never quite succeeding in his quest for a ‘theory of everything’. That’s where we are with AI, trying to make infallible vision systems with limited-memory neural networks, and now an oracle, a fount of all knowledge which will remove the need for human creativity and thought. The idea might seem attractive assuming the machine is benevolent. But it won’t be; it will be based on a very imperfect database of past human ideas and actions, good, bad and positively evil. We all know how that’s going to turn out. In the meantime, companies continue their attempts to perfect generative AI, adding features to prevent offensive (racist, sexist, etc.) or downright dangerous outputs (how to build a bomb, make napalm, etc.) from chatbots. They won’t succeed. Just like Einstein and his successors, they are attempting to describe the infinite: in the case of generative AI, developers are up against the seemingly infinite ability of the human mind to twist a discovery or invention from beneficial application to one that’s ultimately destructive. The future is in our hands; we humans still have a choice. Casually releasing software with the capacity for emergent behaviour onto the Internet may be analogous to letting a virus with no known antidote escape from a laboratory into the wild. Let’s not find out the hard way.

Reference

[1] What is ChatGPT Doing…and Why Does It Work? Stephen Wolfram. Also available to read on-line as a blog post with code-links: https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/

If you're stuck for something to do, follow my posts on Twitter. I link to interesting articles on new electronics and related technologies, retweeting posts I spot about robots, space exploration and other issues. To see my back catalogue of recent DesignSpark blog posts type “billsblog” into the Search box above.