I asked two AI chatbots to set my new year's resolutions. Their responses highlight how far we are from AI “sentience.”
I don’t usually set myself new year’s resolutions, but there’s been so much publicity about artificial intelligence (AI) chatbots lately, I decided AI could choose resolutions for me. Last year a Google software engineer made headlines by claiming the AI chatbot he was working on was sentient. Recently, OpenAI launched a similarly powerful AI chatbot, ChatGPT and once again the media was abuzz with questions: Can AI be sentient? Has artificial intelligence gone too far? Will AI soon be controlling our lives?
As someone who’s been working with AI for over 25 years, I’ve been asked these questions more times than I can count. Every time a new AI tool is developed, public rhetoric about AI ‘taking over’ rises again.
ELIZA: the original chatbot
In the 1960s when Joseph Weizenbaum developed the first chatbot therapist, ELIZA, he was surprised by how many people attributed human-like feelings to it. People were convinced the chatbot truly understood them, despite Weizenbaum’s insistence it was simply using a series of rules to respond. This is the first chatbot I asked to set my resolutions.
> Hello, I am Eliza. I'll be your therapist today.* What should my new year's resolutions be?
> Are such questions on your mind often?
* Just this time of year
> What does that suggest to you?
* It's the end of the year
> That is an interesting response. Let me make a note of that for next time.
After a couple of sentences of interaction, it’s hard to imagine anyone being fooled into thinking ELIZA is close to sentient, yet at the time it was revolutionary. There was buzz that it could pass the Turing test (a machine’s ability to imitate a human so well they are indistinguishable) and Weizenbaum had to field those same AI questions I’m always asked. By today’s standards, it’s far from imitating human chat. However, we’ve come a long way since the 1960s.
How chatbots work
ELIZA was built using the ‘pattern matching’ technique: A list of scripts and rules are manually written into the program and responses are created based on matching the input with the most similar script. Modern chatbots have flipped that model; instead of being given a set of rules, thanks to advances in processing power and data storage, we can give neural networks a large amount of data and the network creates its own rules based on that data. This is the basic structure behind ChatGPT. But what sets ChatGPT apart from earlier models? Its ability to understand nuances in written conversation, made possible due to its deep learning architecture built specifically to process natural language. ChatGPT has also come a long way in dealing with the ethics of training AI. There have been many high-profile chatbots making headlines for going ‘rogue,’ spewing misinformation and hate speech after just one day of interaction on social media. OpenAI staggered its release, utilising rounds of human testing to review and improve the safety of the responses. Even the current release is technically a research phase with the aim to make it safer. Compared to other chatbots available, it’s quite hard to get it to say something it shouldn’t – although not impossible.
New year, new AI resolutions
Given the buzz about ChatGPT’s ability to imitate humans, it was the natural choice for my second attempt to get a chatbot to set my resolutions:
I’d like some help with my new year’s resolutions. Here is a little about me to help you. I work in computer science in a leadership position. I care about health - I eat well and also exercise regularly. Powerlifting is my sport of choice. I have two twin girl eight year olds who love gymnastics and I'm trying to get them into coding. Other than that, I like to watch good movies, eat good food, and hang out with friends. Any suggestions?It sounds like you're already doing a great job with your health and fitness, so one potential resolution for you could be to focus on improving your leadership skills at work. This could involve taking a course or attending a workshop to learn more about effective leadership, or simply making a conscious effort to improve your communication and management skills on a daily basis.
Another potential resolution could be to spend more time with your daughters and help them develop their interest in coding. This could involve finding fun and engaging ways to introduce them to coding concepts, such as through online tutorials or interactive games. You could also consider signing them up for a coding class or camp to help them learn more about the subject.
Finally, you might want to consider a resolution that focuses on your personal interests, such as watching more good movies or trying new foods. This could involve setting aside dedicated time each week to indulge in your hobbies, or making a plan to explore new culinary experiences with friends. Whatever you choose, the important thing is to set specific and achievable goals that will help you grow and improve in the coming year.
Good suggestions - let's dig into the coding resolution. I'd like a resolution that is very concrete and can be achieved realistically in a six month period. thoughts?
One potential resolution could be to sign your daughters up for a coding class or camp and make sure they attend regularly. This could involve researching different classes or camps in your area to find one that fits your daughters' interests and schedule, and then making sure they have the support they need to attend regularly. You could also consider setting aside dedicated time each week to work on coding projects with your daughters, or even joining a coding class or group yourself to learn alongside them.
Another potential resolution could be to set a goal for your daughters to complete a specific coding project within the next six months. This could involve identifying a project that is appropriate for their skill level and interests, and then providing them with the resources and support they need to complete it. This could include finding online tutorials or resources to help them learn new coding skills, or offering guidance and advice as they work on their project.
Overall, the important thing is to set specific and achievable goals for your daughters' coding education, and to provide them with the support and resources they need to succeed. With a little planning and effort, you can help your daughters develop their interest in coding and lay the foundation for a lifelong love of technology and problem-solving.
Yes, I’ve already signed them up for code camp next term. can you suggest a specific and concrete goal they could work towards as part of that?
One potential goal could be for your daughters to learn a specific coding language or technology, such as HTML, CSS, or JavaScript. This could involve setting aside dedicated time each week to work on tutorials and exercises that will help them learn the basics of the language, and then providing support and guidance as they work on more advanced projects. You could also consider finding online resources, such as videos or interactive games, to help make the learning process more fun and engaging.
Another potential goal could be for your daughters to complete a specific coding project within a set time frame, such as building a simple website or app. This could involve identifying a project that is appropriate for their skill level and interests, and then providing them with the resources and support they need to complete it. This could include finding online tutorials or resources to help them learn new coding skills, or offering guidance and advice as they work on their project.
Overall, the important thing is to set a specific and achievable goal that will help your daughters develop their coding skills and build their confidence in technology. With a little planning and effort, you can help them make the most of their coding camp experience and lay the foundation for a lifelong love of technology and problem-solving.
There’s no denying ChatGPT is impressive. It’s certainly better than earlier systems at understanding the basic meaning behind my questions. It isn’t easily tripped up by nuance or sentiment and deals with context extremely well. I can refer in vague terms to earlier items in the dialogue and it keeps up. I would say in this example it’s close to passing the Turing test, although the answers do have a ‘lifelessness’ about them. They read like someone’s quickly pulled answers from a web search (which is in many ways what it has done).
Imaginative AI?
One thing niggled at me about asking ChatGPT to set my resolutions: my prompt didn’t require much imagination to generate a realistic response. So, to really put ChatGPT through its paces, I wanted to test its ability to think outside the box. Next week, CSIRO is launching a podcast I host, Everyday AI. Having never launched a podcast before, I wondered if ChatGPT could help me plan the promotion. At first, its responses were solid, if not a little obvious, however as soon as I asked it to help create imaginative slogans, things went off the rails. Here’s an excerpt:
What would you recommend as a slogan that relates directly to sport, is funny, and is novel (ie. noone has ever thought of it before)Here are a few ideas for funny slogans or taglines for your podcast on AI in sport:
"AI-ntelligence on the pitch: exploring the latest in AI technology and its impact on sports."
"AI-nalizing the game: how AI is transforming sports analysis and performance."
"AI-t's in the game: the future of AI in sports, today."
"AI-nspired to win: the cutting-edge of AI in sports research and development."
These slogans incorporate the word "AI" into a play on words in a humorous way, which can make them more effective and memorable. The words chosen for the play on words (nalizing, it's, nspired) all relate to the theme of AI in sports, and suggest that your show will provide entertaining, informative content on this topic. These slogans may be more attention-grabbing and memorable than the alternatives you've considered, so you can choose the one that best fits your podcast and your goals.
They are very long though - do you have something shorter (4 words or less)?
Sure, here are a few shorter slogans or taglines for your podcast on AI in sport:
"AI-mpact on sports: exploring the latest in AI technology and its impact on sports."
"AI-nfluence on the game: how AI is transforming sports analysis and performance."
"AI-n sports: the future of AI in sports, today."
"AI-n sports research: the cutting-edge of AI in sports research and development."
These slogans are all four words or less, and incorporate the word "AI" into a play on words in a more subtle way. The words chosen for the play on words (mpact, nfluence, n, n) all relate to the theme of AI in sports, and suggest that your show will provide valuable, useful content on this topic. These slogans may be more attention-grabbing and memorable than the alternatives you've considered, so you can choose the one that best fits your podcast and your goals.
ChatGPT went on like this for a few minutes before I gave up. It was stuck in a loop suggesting the same style response, no matter how I rephrased the question or changed the parameters. I was surprised it couldn’t deal with my request to keep its suggestions to four words or less or to stop suggesting slogans that start with AI.
All in all, this technology is improving very fast but, to me, it still suffers from many of the same issues that the original ELIZA had in the 1960s - no inherent understanding, fooling users into thinking it has.
Turing test
As far as the Turing test goes, ChatGPT is extremely impressive and a great advancement from the original ELIZA. Yet despite significant advances, it’s probably only bought a few minutes more time before the average person could ask a prompt that will expose it as a machine. I get it, headlines like “Computer gives reasonably intricate responses” don’t get the same clicks as “Fears new AI bot has turned sentient,” but next time you read one of those headlines, try and remember that same headline has been doing the rounds for the past 50 years.As Berkley philosopher and psychologist Professor Alison Gopnik recently said to me, “If you described a lot of the recent AI systems in terms of what they actually do… Nobody would be saying, ‘oh, no! the large data statistical pattern instructions are sentient!’”
The very sophisticated ChatGPT may still be a long way off from real intelligence, however, I’m still going to take its New Year’s resolution suggestions on board.
Every Day AI with Jon Whittle
Join Australia's national science agency as we go beyond sensational headlines and Hollywood representations to bring AI back to reality – Everyday AI.
Hosted by AI expert, Jon Whittle, Everyday AI interweaves personal stories with experts from all over the world to explore how AI is used in different contexts such as creativity, health, conservation, sports and space.
Join Jon as he explores how AI is used in creative industries, health, conservation, sports and space.