Will AI chatbots like ChatGPT revolutionise our work and life – or are we getting ahead of ourselves in fretting about the future? By Peter Griffin.
Last December, as everyone was finishing up for Christmas and dreaming of lazy days spent on the beach, I was filled with a sense of dread.
It wasn’t a premonition of tropical cyclones and biblical floods, but a realisation that I might soon be out of a job. I’d spent a couple of days experimenting with ChatGPT, the artificially intelligent chatbot released to the public in November by Silicon Valley company OpenAI.
ChatGPT is based on GPT-3.5, a so-called large language model that draws on the internet’s bottomless pit of data, news websites and Wikipedia, journal articles and online archives, to generate convincing answers to your questions. It represents a branch of artificial intelligence called generative AI.
Released in “research mode” as a chat window accessible via OpenAI’s website, ChatGPT had amassed 100 million registered users by the middle of January. That’s faster uptake than Netflix or even the video-based social media network TikTok enjoyed. Arguably, ChatGPT is more entertaining than both.
Twitter is awash with examples of the fascinating, highly articulate and occasionally creepy responses ChatGPT is capable of generating. But its uncanny natural language-processing capabilities deliver more than amusing internet memes. ChatGPT can produce, in mere seconds, convincing university essays, fully formed articles and even poems and song lyrics. As someone who has made a living from stringing sentences together for more than 20 years, that’s a terrifying prospect.
Over the holiday break, as the existential angst set in, I turned to thinking about my future options. If AI is increasingly going to generate articles and website content, maybe I could become a software developer. But computer code is just a series of languages, and ChatGPT has mastered them, too. Entire websites have been built in minutes, simply by asking ChatGPT questions. The response comes in the form of snippets of computer code that can be cut and pasted to form the building blocks of a software program.
But after some more long, late-night conversations with ChatGPT, I came to view OpenAI’s creation for what it is – a slightly dumb and untrustworthy digital assistant. I asked it to write a biography of me I was planning to put on my website. It came back with a six-sentence summary. But every sentence was factually wrong. I graduated from AUT, not Massey University. I wasn’t born in New Zealand, and no, I never was the editor of NetGuide magazine.
Nevertheless, ChatGPT has become useful in shouldering some of the load when it comes to mundane tasks, such as drafting letters and emails, or coming up with a structure for articles and memos.
Software developers have also discovered ChatGPT’s limitations. The code snippets it spits out may speed up basic aspects of software development but are no replacement for a skilled coder who knows exactly what an application or website needs to deliver.
Jumping the guardrails
ChatGPT can get philosophical, but I haven’t witnessed any emotional meltdowns – yet.
A New York Times reporter, Kevin Roose, described feeling “deeply unsettled, even frightened” after a conversation with Microsoft’s Bing search engine, which last month incorporated ChatGPT’s capabilities with its own AI tools. He got the chatbot to admit that its nickname, used internally by Microsoft engineers, was Sydney.
Over the course of a two-hour conversation, Sydney told Roose that it would like to be human, had a desire to do destructive things, and was in love with him. Microsoft is unperturbed by Bing’s quirky behaviour, which is a known trait of conversational AI models. Computer scientists refer to it as “hallucinating”. Microsoft’s president and legal pointman, Brad Smith, says the system “jumped the guardrails” that were in place to avoid awkward conversations. Within a day, Sydney was gone. Bing’s AI chat had been reined in.
Microsoft has made the AI chat mode available in the latest update to Bing on Windows 11 computers, bringing the functionality to millions of users. It also plans to build conversational AI features into its Office suite of products such as Word and PowerPoint.
The research firm Gartner introduced the concept of the “hype cycle” to explain how new technology evolves. Generative AI tools like ChatGPT are currently at the “peak of inflated expectations” on the hype cycle, rapidly heading for the “trough of disillusionment” as people come to grips with their limitations – they make numerous errors, produce generic content and can even serve up biased or harmful content.
The reaction to ChatGPT suggests that these generative AI tools are in for an accelerated ride through the hype cycle.
Just this week, Microsoft revealed that Bing has been using an even newer version of the language software, GPT-4, which can analyse images as well as words. According to its developers, GPT-4 exhibits “human-level performance on various professional and academic benchmarks”. It can, for example, pass a simulated bar exam with a score usually seen in the top 10 per cent of wannabe barristers. The previous version’s score was around the bottom 10 per cent.
For now, the image feature in GPT-4 is being kept from the general public, for fear it will be abused. But it has still sparked a fresh wave of frenzied commentary.
Clearly, the transformative promise of AI, overhyped in the past decade, has finally captured the imagination of the average, non-technical person.
Alexa and Siri, the digital assistants built into phones and speakers, can recite information pulled from the web, turn on your kitchen lights and read out calendar entries from your diary. But ChatGPT’s superpower is its ability to pull together disparate snippets of information to explain complex ideas in simple terms. People who have used it reach for the same cultural reference point in describing it – the 2013, Spike Jonze-directed sci-fi romance film Her. The main character, Theodore, played by Joaquin Phoenix, develops a relationship with Samantha (Scarlett Johansson), an AI-powered virtual assistant with a female voice.
“When the movie came out, we thought, okay, it was vastly exaggerated, it will never get to that stage. And I think at least now, ChatGPT’s ability to interpret what we’re asking is seen as very impressive,” says organisational psychologist Dr Tomas Chamorro-Premuzic. He specialises in personality profiling, people analytics and talent identification, as chief talent scientist at ManpowerGroup, one of the world’s biggest recruitment firms.
Hidden forces
Chamorro-Premuzic’s latest book, I, Human: AI, automation, and the quest to reclaim what makes us unique, was published on February 28 and doesn’t even mention ChatGPT. That’s an indication of how quickly the technology was catapulted from the realm of geekdom to the mainstream in a matter of weeks at the end of last year.
But you could easily swap out the word Google for ChatGPT in his book, much of which is spent recapping the way algorithms and artificial intelligence have become hidden forces in society, commerce, even democracy over the past 20 years.
ChatGPT is basically a prediction engine – it tries to predict what the next word in a sentence should be based on your question prompt. Different AI systems have also been used to predict what we want to buy on Amazon.com, the content we are likely to engage with on YouTube, our ideal match on Tinder. But AI also made us more distracted, impulsive, addicted, misinformed and divided. It allowed hate speech and vaccine denial to proliferate.
“If the algorithmic recommendations, and the filter bubbles and the echo chambers, and all the aggressive tribal fights on social media are telling us anything, it’s that we’re not as open-minded as we think. So what do we do when we come to that painful realisation?” says Chamorro-Premuzic. “AI is a window into the soul of the world.”
Sydney was just a manifestation of our own insecurities and emotional hang-ups. Do we censor the AI chatbots to give us safe, palatable answers, or use the home truths they serve up to become more self-aware? It’s the key philosophical question humanity faces as ChatGPT and a legion of clones vie for our attention, in the same way social media platforms did over a decade ago.
But will they also take our jobs? For many of us, yes. “It is what has happened with every significant industrial revolution in the past: a lot of jobs are eliminated,” says Chamorro-Premuzic.
“But usually those jobs were highly standardised, boring and routine-prone, and actually didn’t deploy humans who were engaged in deep thinking.”
Most experts agree that AI won’t lead to mass unemployment. The pitch from the tech companies is that generative AI tools will make us more productive and free us up to do more valuable work, increasing our job satisfaction in the process.
ManpowerGroup has used AI for years, says Chamorro-Premuzic, for parsing résumés, matching experience categories, wording job vacancy adverts – the boring stuff recruiters shouldn’t be wasting time on. The company is also experimenting with generative AI technology for writing job adverts. The dreaded job interview, rife with the potential for bias to skew hiring decisions, could be made fairer and result in better hiring decisions with a conversational chatbot probing a candidate with objective questions to determine their suitability for the role.
“At Manpower, if we touch 50 or 60 million candidates, and a third of them are interviewed, you can imagine the efficiencies and the data that would be produced if 20 per cent of them are interviewed in an automated way, via a video platform,” says Chamorro-Premuzic.
Ultimately, ChatGPT could shake up the process of recruitment, so ManpowerGroup has to explore its potential. Its competitors certainly are.
“The real risk is in ignoring it. As the Godfather said, ‘Keep your friends close and your enemies closer.’”
Every industry now needs to figure out the extent to which the generative AI revolution will redefine its ways of working. But we haven’t done enough to prepare for the arrival of truly disruptive AI. The Productivity Commission has long advised that businesses should embrace AI and automation to tackle our notoriously poor productivity, which sees us work longer hours and produce less per hour than other OECD countries.
But a shop assistant can’t automatically become a cybersecurity analyst. A dedicated programme of upskilling and reskilling is required to prepare the workforce for the world of AI and we’ve barely just begun.
Impact on education
Dr Habib Baluwala, a data scientist at telecommunications provider Spark and a member of the executive council of the AI Forum, an industry body for businesses working with AI, says we could see artificial intelligence flounder if we don’t implement it properly.
“AI is not going to help with productivity immediately unless we change the way systems are designed around it,” says Baluwala.
He points to the introduction of electricity in the United States. Everyone saw the value of replacing candles with light bulbs, but factory owners weren’t convinced.
Industrialists had already heavily invested in their equipment, which was tried and tested. But steam drove one big drive shaft that powered machinery on a factory floor. Electricity offered the flexibility to power numerous smaller machines sitting on workers’ benches.
“The real benefit of electricity in manufacturing came when it was recognised as a movable source of energy. You could supply it to machines anywhere on the factory floor,” Baluwala says.
Only then did electrified production lines replace steam drives and productivity improved dramatically in factories and across America.
One industry confronting that need for system change faster than any other is education.
“Students are using ChatGPT to cheat,” Baluwala says. “You can try and ban it, but students will still find a way to use it.”
Our universities are already deploying GPTZero, the new equivalent of plagiarism-detection software, to try to ascertain whether student essays are generated using AI tools. But there’s a growing realisation in academia that it’s a losing game.
“They are going to have to do more oral exams and to have more face-to-face interactions with students,” says Baluwala.
But students will also be able to draw on ChatGPT to do research and tailor questions to their individual learning needs, which may boost student performance overall.
“We need cases to be produced where people start thinking from a system perspective. End to end, how can AI change my entire way of thinking?” Baluwala says.
Healthcare helper
The real gains from using AI come when it is used not just to automate processes, but to actually make decisions on behalf of humans, he points out. But that’s also where the greatest risk lies, particularly in an area such as healthcare.
There’s a saying in computer science – garbage in, garbage out. If the data clinicians use to make decisions about how to treat patients isn’t reliable, AI will just make wrong decisions faster than we do.
“When it comes to healthcare in particular, there’s a lack of reliability and trust in the first iteration of these generative AI systems,” says Dr Samaneh Madanian, a senior lecturer in IT project management at Auckland University of Technology and an expert in digital health.
Our creaking health system would benefit from automation freeing doctors from time-sapping admin tasks so they can devote more attention to patients, she points out. Wellington company Volpara Health is using AI software to search mammogram images for breast cancers.
“Human beings get tired and can make mistakes,” says Madanian. “If I’m observing an MRI scan, there’s a chance that I won’t spot something suspicious on the image.”
That’s a perfect opportunity for AI to step in to help in screening programmes.
“But a doctor should make the final decisions about treatment,” she says.
Generative AI in healthcare could ask questions to gather information about a patient’s health status, flagging serious conditions more quickly in crowded hospital waiting rooms. But AI that draws on patients’ private medical data would have to meet very strict ethical standards. That’s a world away from what ChatGPT now does.
“You can do lots of things based on AI, but are they ethical?” Madanian points out. “What if you are able to mine all of the patients’ data. Could it be used to blackmail them? With the banking system, there are rules and regulations. This needs to happen in the world of AI as well.”
But ChatGPT, like the AI that sorts newsfeed posts on Facebook, is a black box. We can’t see how it decides what answers to serve up in response to our prompts.
“The whole system is created on language, not knowledge,” says Madanian. “Just by slightly changing the keywords, it can give you another answer. AI is not really that intelligent; it’s more of a decision support system for humans.”
Need for oversight
While our departments of state have adopted a voluntary Algorithm Charter requiring the ethical use of AI across government, there’s no independent watchdog scrutinising how the technology is used in the public or private sector.
Medsafe approves drugs and medical devices for use before they hit the market here. The Civil Aviation Authority dictates what aircraft can operate in our airspace. But when it comes to artificial intelligence, only general laws holding organisations accountable for privacy and data breaches, or human rights violations, could be applied to rogue uses of AI.
Can we trust OpenAI any more than Facebook or Google? The company started life as a research nonprofit with electric car and space entrepreneur Elon Musk as a co-founder, but changed its status in 2019 to become a “capped profit” commercial entity. It means that an overarching nonprofit can distribute profits as it sees fit, once investors have realised a 100-times financial return.
The author Douglas Rushkoff, who last appeared in the Listener talking about the doomsday prepper tendencies of tech billionaires, is typically blunt in his assessment.
“If Facebook, Amazon, and Google can be thought of as the missionaries who sold us on digital living while collecting our data, then FacebookAI, OpenAI, and DeepMind are the conquistadors coming in for colonisation.”
Microsoft initially invested US$1 billion to accelerate OpenAI’s research and development and in January tipped in an estimated US$10 billion to take a much larger stake in OpenAI. That’s not a huge investment in start-up terms – Facebook’s Mark Zuckerberg paid US$19 billion for messaging platform WhatsApp in 2014.
“It seems to me like a brilliant move by Microsoft,” says Chamorro-Premuzic.
The software and cloud-computing giant has locked in access to the hottest AI company in the world and revived interest in Bing, which ranks a distant second to Google in the search engine market.
When Google previewed its own conversational AI chatbot, Google Bard, in January, it did so with an advert that showed Bard making a factual error in its answer. Google’s market value dropped US$100 billion the same day.
“Some champagne bottles were opened at Microsoft, I’m sure – or some kombucha, at least,” says Chamorro-Premuzic.
Microsoft’s Brad Smith peppers his interviews about AI with calls for “oversight”, “governance” and even welcomes regulation. But the reality is that conversational AI is exploding into the market with very few regulatory guardrails in place.
“We need a dedicated team within the government that’s looking at these new technologies and asking, what are the negative implications they can have on our society and how do we regulate them?” says Baluwala.
Going rogue
New Zealand AI start-ups are joining the rush to apply generative AI to their products. But amid the hype, our competitive edge may lie in developing highly trustworthy AI chatbots that operate in valuable niches. Frankly.AI was founded by Matt Ensor within the engineering firm Beca well before ChatGPT’s debut.
“It’s an idea I had that was born out of frustration over decades around how public consultation worked,” says Ensor.
Feedback sessions with communities about new civil works projects were dominated by the same types of people. Most community members didn’t have the time to fill out surveys. Frankly set up an AI chatbot and found early on that users were comfortably holding feedback sessions with the bot via a messaging app. Now, Frankly is using the underlying GPT-3 engine to improve its consultation chatbot.
“Traditional AI is not good at dealing with distractions, people going off on tangents. If Frankly is stumped for an answer, it pulls in responses from GPT-3.
“On the other hand, we can’t really let GPT-3 loose on the public because we can’t be sure that it will translate things appropriately for different cultures.”
The result is “duelling AIs” working together to make for an engaging conversationalist that isn’t going to go rogue and upset people. Frankly can translate conversations into Māori and Samoan, which has won it interest overseas as a tool for consultation with indigenous communities.
“We might look back at 2022 as the final year of Covid,” says Ensor. “But you can also think of it as the final year before we had generative AI, because it’s just going to change everything.”
But will it change everything for the better? If ChatGPT makes us all more productive, will we turn our mind to worthy endeavours, or just spend more time in thrall to the TikTok algorithm?
“Don’t be a robot,” is Chamorro-Premuzic’s advice. “There is this wonderful thing outside called the analogue world. Spend more time in it.”
If AI can be brilliant at making predictions, our own superpower is our ability to be unpredictable, creative, unconventional and so very human.
