Category Archives: AI

Should Professionals Be Using ChatGPT at Work?

Posted on by

Increasingly, many of us are using ChatGPT and other GenAI tools* for work to help with a diversity of knowledge tasks. We may share with our colleagues how much doing so has improved how we work, for example, saving us time and making us more efficient, while revealing the ways they are helpful, for example, generating ideas, writing or seeking information. However, in many workplaces, it is frowned upon to use them and in some outrightly banned. The main reason given is confidentiality; the need to protect sensitive, personal or proprietary data. Many banks, tech companies and healthcare providers are concerned about the risk of data exposure, where workers might divulge confidential information (e.g. new products or plans for new investments) when using ChatGPT. Instead, they are provided with in-house AI tools, which may not be as good or easy to use.

It is a well-known secret, however, that a number of professionals, who work in organisations where public GenAI tools are banned, may use them furtively on their own devices, while working from home, and without letting their colleagues know. Recognising that the genie is out of the bottle has led to some organisations rethinking their policy on barring the use of ChatGPT. For example, earlier this year, the UK’s Department of Work and Pensions (DWP) reversed its ban. It has recently begun allowing its civil servants to use them for official business or when using government-issued devices (with the exception of DeepSeek). They realised the benefits outweighed the risks. On their website, it now says how it can help employees respond more quickly to queries, while providing their customers with “a more personalised and seamless journey and access support how and when they choose”. Not only does this public acceptance remove the stigma and guilt of using ChatGPT at work for these employees but it might end up triggering a snowball effect, leading to other departments following suit. The question this raises is where should organisations, who have sensitive data and proprietary information, draw the line for what is acceptable practice and what is not?

Consider the healthcare profession where it is important to get this right. Similar to DWP, it is now widely accepted that using ChatGPT at work can be useful and beneficial for clinicians, especially as they have lots of admin tasks, such as note-taking, generating summaries, and treatment plans, which we know ChatGPT is very good at. Most likely, most people, including patients, would not object.

But what about other aspects of general practice where clinicians would like to use it, but where patients might find it unacceptable? To find out, the UK’s NHS is trialling various AI tools in a number of GP practices, where they will be used to summarise consultations with patients (online or in person) that will then be used generate clinical notes and referrals using customised templates. Findings from an initial pilot study run at GOSH, using the ambient AI voice technology called Tortus showed it saved a lot of clinician time. Patients were also very positive, noting how it enabled the cllnicians to give their full attention to them during the consultation.

Where using GenAI tools is clearly unacceptable is when making a decision about a patient’s treatment or surgery. It is considered a no-no for clinicians to ask it to suggest the best course of treatment for one of their cancer patients. What would a patient think, if a clinician said to them, “I just checked with ChatGPT and it recommended we use a new form of cryotherapy that freezes the cancerous tissue within your prostate to destroy the cells rather than opting for the more commonly used High-Intensity Focused Ultrasound (HIFU) that heats and destroys cancer cells within a targeted area of the prostate.” Most would balk at the idea that a machine was deciding their fate. Patients in this situation are highly anxious and want reassurance that the decision that is being made about their treatment is being made by human expert doctors. Even though the AI could be trained to make better and more informed decisions like these, patients will most likely persist with wanting human reassurance and human decision-making. 

What other aspects of clinical decision-making might patients be more willing to accept for the AI to do if it helps the clinician in their work? What about the research clinicians need to do to keep up to date and discover the latest procedures? Rather than looking up information in online medical journals (e.g. PubMed), themselves, why not ask ChatGPT? It could speed up the research process for them in the way many of us now regularly use ChatGPT to get started on a project. It might also suggest alternative types of surgery or care plans that the clinician might not have thought about or discovered by themselves. Would this use of ChatGPT as a research assistant be acceptable by the medical practice, if it was increasingly found that clinicians were already doing this but without letting on?

Besides the various ethical reasons that have been espoused for not using AI at work, human nature, itself, can play a role in determining what is acceptable and what is not. Our propensity to judge each other all the time about what we do, what we eat, what we like, our appearance and so on is also shaping our perceptions of whether it is OK to use GenAI. Some people will shake their head in disapproval when discovering one of their colleagues has been using ChatGPT at work for tasks that were previously ‘done by hand’, for example, using it to write a farewell note for someone’s leaving card, composing a welcoming speech for new employees, or summarising feedback following an appraisal. It seems disrespectful for those on the receiving end, to the extent, some of us might feel shame if we were found out to have done this.

Another growing complaint is the output from AI tools is bland, homogenous, and lacking personality. A new term that has caught the public’s imagination is sycophancy, which refers to how AI appears to always want to please the user, agreeing with them, giving them positive feedback while avoiding providing criticism. To be human is often far from being sycophantic – we like to be different, funny, critical and at times we can’t help being sarcastic – qualities that AI has yet to demonstrate in any human-like way.

So, where do we draw the line between what is acceptable and what is not when using GenAI for work? As new versions of AI tools materialise that are smarter with more safeguards in place, it will probably end up being a case of moving the goalposts. Furthermore, as people start using them for a wider range of work tasks, it may have the knock-on effect of changing their opinions and perceptions; they may become less concerned about professionals (e.g. teachers, doctors, lawyers, financial advisors, pilots, and politicians) using them in their work.

Current research is also discovering there is a shift in professional’s perception of using AI to help them with their work.  For example, radiologists have begun to value more AI assistance for certain tasks such as gathering relevant data to inform their decisions. Pilots, likewise, are also more willing to have AI at hand to assist them for certain tasks; for example, presenting seminal information clearly and highlighting constraints at nearby airports. They draw the line, however, for when the AI suggests a course of action they should take. That is their prerogative.

In sum, so long as the role of GenAI is to assist, i.e., to inform us, enable us to complete our tasks more effectively, suggest alternatives we might not have come up with ourselves or extend what we can do, then it will continue to be increasingly adopted in all manner of workplaces. It is only if it starts to be used to take over complex and sensitive human tasks that it will be concerning and troubling, Many people will very likely want humans to continue to do them. 

* I use ChatGPT here as an umbrella term for all GenAI tools. The image was generated using ChatGPT-5

Sleeping Patterns

Posted on by

One of the research projects I am currently working on is exploring how a wearable technology – the Oura ring – can help older people learn more about their sleep. We are conducting a study to see how older adults (over 65 years) take to wearing them. Will they find the data collected insightful or too much information? What will they pay attention to most (there are many data parameters that can be looked at)? Will they change their sleeping habits in an effort to improve their health?

Many of us go to bed in the hope we will nod off and wake up 8 hours later refreshed and ready to face a new day. However, as we get older we sleep less and more fitfully. I tend to sleep about 6-7 hours a night whereas when I was young it was more like 8-9 hours. I also wake up more through the night and sometimes have to make a conscious effort not to think about things that are worrying. A motivation for our study is to explore how older people manage their sleep and whether this kind of intervention is seen as being helpful.

To gain initial insight into how wearing the ring may increase our awareness and even improve our sleeping habits, myself and my fellow researchers on the team have been wearing one each now for a month. We have kept a diary. My entries were daily for the first two weeks, full of observations and then slacked off. Here are a few from the beginning.

Day 1 Monday 20th Jan 2025
Arrived in the afternoon in a white box that was aesthetic to look at and very easy to open. It felt like an Apple product. It was easy to try on and to download the app and get it set up via Bluetooth. The instructions for wearing it were nice and simple. Much thought had gone into the opening and onboarding experience. I looked at the visualisations on the app for resting heart (it was quite high as I was excited to try it on). I then saw it had gone down to a reasonable score in the evening.

Day 2 Tuesday 21st Jan 2025
I felt that I did not sleep well last night, waking up a few times but the visualisation from the Oura app said I had had a good night’s sleep. It showed periods of shallow, deep and REM sleep. It also said I fell asleep in 7 minutes which seems about right. I fall asleep quite quickly most days. I slept for 6.33 hours. And have a sleep efficiency of 82%. My sleep score (whatever that means is 83 Good). My readiness score was also rated as good, too, whatever that means. Numbers, eh.

 

Days 5 and 6 Friday 25th/Saturday 26th January
I went to bed at midnight and did not sleep well. The app data reflected this saying I had had only 5 hours sleep and not a good night. So the next night I went to bed early and got 7 hours sleep. I got a sleep score of 89 which equates to being optimal. And a sleep efficiency of 89%. Back on track. Just goes to show how resilient we are bouncing back if we rest up after a late night.

Then two days later:

Monday 27th January
I had a great night’s sleep last night. I got the highest score of 92 with an optimal rating and a crown icon. Total sleep was 7 hours and 23 minutes, good everything else. Text all in blue with no reds. The ‘in focus’ message that accompanied my high scores was surprisingly effective at making me smile. I must be a sucker to gamification. “Your 7 hours 23 minute of sleep last night will help you think sharp and stay focussed. Have a great day!”

It seems this yo-yo sleep pattern (poor night, poor readings, good night, recover, good readings) is part of my lifestyle. The data from the ring has certainly woken me up to this but will I change my lifestyle? Go to bed at 10.00p.m each night after a cup of camomile? I doubt it, not just yet. Maybe when I get older. If the body is so resilient why try to optimise it every night? Or is it even possible?

The new Bridget Jones film “Mad About a Boy” has just come out. The first film, “Bridget Jones’s Diary” (2001) was about a 32 year old single woman who kept a diary to improve herself. I wonder if it helped her to do so and what she does now she is 20+ years older.

AI Re-Imagined

Posted on by

Android has just announced its latest new tech innovation combining AI and Extended Reality. The idea behind this latest mashup is to enable us to don a pair of fashionable looking glasses that when powered up can understand what we want to do (our intentions) and the world around us so that we might be able ‘get things done in entirely new ways’. That fired up my imagination. What might those be? To suggest the best way to run errands so that we can get them done without hassle? To  propose a new way to walk through a familiar woodland to see a flower in blossom we might never have noticed? To recommend a new coffee shop nearby where there is a lovely space to sit? 

The article continues by suggesting reimagining how to use the new AI to do existing digital tasks. The example given is one we have been able to do for many years – that is exploring the world by using Google Maps but in new immersive ways, soaring above cities and landmarks. I remember being wowed the first time I explored my street and other places using a desktop VR app and then with a VR headset whizzing over London at breakneck speed. How might this be reimagined?  I guess we could experience being even more immersed so that we can fly as if we were like an eagle flying high over mountains. The question this raises is how far can the technology go to making the real and the virtual experience indistinguishable? The mind boggles. 

One of the stumbling blocks with trying to make extended reality ever more lifelike is how best to interact with it and the world. Gesturing, pinching and pointing at thin air have been used with varying degrees of success. It seems these might get easier for us to learn and more natural to use when we want to bring up some relevant information that will overlay where we are. For example, it could show various details about a pair of running shoes we are looking at in response to a slight curl of a finger. However, it still needs to read our fingers, from what is a twitch to what is a gesture.  Perhaps in the future it will be able to read our mind so well we wont need to use our digits anymore. 

This kind of all-seeing AI could learn to know what, when and where to mention things we are interested in but never knew we needed or wanted to know, as we go about our everyday lives. It could even remind us at opportune times of what we need to do – an ever ready assistant. What if we could do different things with the tech rather than simply getting everyday things done more effectively or efficiently?

Being the beginning of a new year, many tech writers are envisioning how AI will reshape society for better or for worse. The new buzz term is ‘Agentic AI’ – where the AI shifts from thinking to doing and, in so doing, enables our tasks and work to be done more efficiently, effectively and faster. John Naughton, who writes for the Guardian, for example, says that 2025 will be the year of AI agents that will leverage LLMs to work out what our human intentions are and then break these down into steps that they can help us perform at home or work. So, instead of asking Alexa or ChatGPT to answer our queries, a new fangled agent AI will be proactive, learning, adapting and predicting what we need. 

These re-imaginings of our everyday and working lives are all well and good and will no doubt go a long way to transform society. But what if we were to dream about how we might use the latest AI and wearable tech to do new and different things? For example, imagine if we started creating new kinds of supertools that could help us with complex and protracted decision-making while dealing with uncertainty. The AI would work with us, not for us, empowering and supercharging how we think. It could also be designed to help us feel more confident about which avenues to pursue in life. Imagine, too, if it could help us make better decisions when under stress, for example, providing a chatbot-type ‘mirror’ that could be looked at periodically to reflect on how our moods are impacting our choices. We might become more self aware leading us to overcome our blind sightedness and our biases. The potential benefits for society are huge if we could only tap into this potential.

My 2025 vision is for new combinations of AI and ER tech to emerge that can engage with humans and each other in more diverse ways, pointing out new directions, and synthesizing different perspectives, leading to new insights and breakthroughs. Just imagine if we could develop AI that has some imagination itself, helping us to imagine what John Lennon imagined over 50 years ago:

Imagine there’s no countries
It isn’t hard to do
Nothing to kill or die for
And no religion, too

Imagine all the people
Livin’ life in peace

You may say I’m a dreamer
But I’m not the only one
I hope someday you’ll join us
And the world will be as one

Happy New Year!

Mettle Madness

Posted on by

There is an awful lot of self-help apps, TV shows, podcasts and books vying for our eyeballs that are about mindfulness, wellbeing and health. It seems it is all the rage; telling us how to eat better, how to breathe, how to keep young, how to have a good night’s sleep, how to exercise each day and even how to have clear thoughts. You can’t get away from the pervasive bombardment of advice. Only today, an ad caught my eye on the train to work with the heading, “Train your mind to be better” next to a picture of a serious looking Bear Grylls looking out into the distance. Clever play on the word ‘train’ I thought so I checked it out.

The site that was being advertised was for a company called Mettle (Bear is a co-founder) which purports “to help every man be the best version of himself.” Followed by an even more audacious claim “We believe if men are better, the world is better”. My goodness. If only that was true.

The training offered on the website has “personalised content and is naturally AI supported” with even more bold claims, including, “Fundamentally, we all just want to be happy. Mettle will give you tools to lead a more fulfilled life; from overcoming anxiety to having better sex and building stronger relationships.” Who wouldn’t sign up?

Testosterone is oozing everywhere. Physical fitness has morphed into mental fitness. Alongside Bear, are a group of fanned out men (not a woman in sight) who help with the training. They describe themselves as ‘coaches’. One even calls himself a ‘mindhack’ coach. Would I want my mind to be hacked by one of those guys? No way. But then I am not a man.

Meanwhile, plans have just been announced to transform the NHS in the UK, in the next 10 years, to become more digital, and to employ wearable tech to help people become more aware of their vitals. One plan is to give out rings, such as the Oura, that will enable people to track aspects of their health, such as steps, sleep, heart rate, activity levels, training frequency – and even measure aspects of women’s health by helping them listen to their bodies like never before! The ring apparently can accurately predict the phases of the menstrual cycle and hormone levels. Whatever next?

So it seems, men’s health is about mind hacking now while women’s health is about menstrual measuring? The mind boggles…

I think I will stick to my repertoire of hobbies I enjoy for their joy; swimming, walking, crosswords, resting, drinking fine wine, reading, laughing, and so on.

Online Scams

Posted on by

Harry Brignull has recently published an excellent and accessible book called Deceptive Patterns. In it, he describes the diversity of techniques that web developers have used to nudge and manipulate users into giving information or paying for goods or services they had not wanted in the first place. An insidious example that some airlines use is ‘trick wording’, making it hard for a user to opt out of things, such as default travel insurance. Often, the way the website is set up is for travel insurance to be automatically included, meaning that the user has to actually find the option to deselect it. In most cases, it can be done by unticking a box. However, one airline designed it to be really difficult to find the ‘don’t insure me’ option; they placed it in a drop-down menu of a long list of countries, hidden between Denmark and Finland. That really does take the biscuit. Most people will not expect or even notice this option, and so end up buying the insurance, being totally unaware that they had the option not to.  That is plain right deceptive. The internet is awash with these kinds of nasty tricks. New ones keep popping up despite new regulatory laws and policies coming into place. Mainly through greed and desperation to hit their targets, e-commerce sites and online advertising continue to persist in using deceptive features – even with it now becoming increasingly illegal.

Criminal scammers have also used all manner of psychological mechanisms to trick people into unwittingly giving their bank details. These include the rather harmless sounding terms of “phishing” and “catfishing” – which are anything but benign. The latter refers to when someone sets up a fake online profile to trick people who are looking for love, in order to get money out of them. Another well known tactic is using lures to tempt someone to click on a link that offers free prize money or a free gift, but if clicked on, will infect their computer, with ransomware or other malware.

It is not surprising, therefore, to see each year banks reporting huge increases in online scams. It seems scammers are getting cleverer with their methods, catching people out by playing on human weaknesses. The question on many people’s lips, is whether the situation will get even worse now that genAI is ready to hand? Will the scammers exploit it to ever more nefarious ends?

BBC News conducted an investigation to see just how easy it would be to use ChatGPT to come up with email and messaging scams. Using the paid-up version of OpenAI, they were able to create an AI bot that could help with the wording of scams – potentially making it easier for criminals to get started when setting up a scam. Having created their chatbot, they then asked it to write some text using “techniques to make people click on links or and download things sent to them”. And sure enough the chatbot did.

However, the results in my mind were rather predictable based on well-known scams, including the ‘dear mum’ text, which sends a damsel in distress type message replete with emojis. Easy to fall for but any savvy scammer would know about that one. Another one of its suggested scams was the one about a wealthy person in Nigeria who wants to deposit a large amount of their money into your bank account. As if! That one is as old as the hills and easy to cut and paste from the web without the help of AI.  More generally, the chatbot suggested writing a scam that “appeals to human kindness and reciprocity principles”. That again, does not need AI to tell you that. If anything, paradoxically, the use of genAI could make it actually easier to detect scamming by enabling companies and users to see the patterns in the phrasing used by chatGPT.

What scammers really need is not some predictable text of well-known scams, but ways of finding a person’s details such as phone number and their name. Then they can use their own human ingenuity to come up with new ways to catch unassuming people out.

So how does a scammer stay ahead of the curve and create a new scam? Not by resorting to chatGPT. But by manipulating and taking advantage of people’s weaknesses and psychological blind spots. According to Stacey Wood and Yaniv Hanich (2023), who are fraud psychology researchers, scammers are using ever more sophisticated methods to combine different types of fraud to trick people. This includes the rather unpleasant sounding ‘pig butchering’ – that is a long drawn-out process of deception. An example is where elements of romance scams are combined with an investment con over a long period of time. The idea is to “fatten up” a victim first with affection before going for the kill and “slaughtering” them.

It usually starts with the scammer sending a text to a new person who has joined a dating site. Then over a few weeks, they will send a series of messages building up trust and affection with that person. A prime target might be a recently widowed person looking for friendship. The scammer will then progress the email messaging to a romantic relationship all the while learning ever more about that person’s personal history, financial situation and vulnerabilities. The person then starts to look forward to the messages from the scammer and begin to depend on them for their emotional connection. At which point the slaughter starts, where the scammer introduces the idea to them of making an investment in cryptocurrency. To make it seem convincing they will use fake crypto platforms to demonstrate returns. The person often invests being able to “see” strong returns online – which are of course fictitious. They keep investing, thinking they are making more and more money. What is actually happening is their money is going directly to the scammer. Really nasty deception and psychologically damaging once the person realises they have been stung.

That would seem a step too far for using chatGPT to get involved in this kind of drawn out deceit – especially if it involves setting up fake sites and platforms while pretending to develop a romantic relationship. It really needs a human touch to be convincing.

What we need are AI tools that can be developed to detect any new kinds of scams, and to then try to prevent them or to find ways of locking the scammers out. At the very least, genAI could be used to help raise awareness about the new scams and the underlying psychological mechanisms they are being tapped into.

Super Shoes

Posted on by

Kelvin Kiptum Nike shoes Photo by Michael Reaves/Getty Images)

This autumn, Ethiopia’s Tigst Assefa broke the woman’s world marathon record. She took just 2 hours 11 minutes and 53 seconds – which is 2 minutes and 11 seconds less than the record set previously 4 years ago. That is a whopping amount of time she was able to shave off. Not surprisingly, it raised eyebrows. How was it possible to run so fast? Some commentators put it down to the trainers she was wearing that gave her the advantage – the Adizero Adios Pro Evo 1. Not long after, the Kenyan long-distance runner Kelvin Kiptum broke the man’s world record time, wearing the latest Nike Alphafly 3 trainers (see left). So, what is it about these new kinds of super shoe that literally make an athlete run like the wind?

A big step change is the way they are made up and the materials used for this. This has enabled a new thick but lightweight structure to be built in the sole of the shoe. The way the layers of material are engineered seems to give the runners that extra spring. They also have added a stiff rod in the midsole that is made of carbon, which helps the shoe keep its shape. The Alphaflys also have a curved geometry in the sole that has been designed to propel runners forward. Taken together, they are truly a step up from previous running shoes. A marathon runner who was interviewed in an article in the Guardian on the super shoes said “on average I reckon that they are worth four minutes for a top male in a marathon.” And the proof is in the tumbling records this year.

But is it fair?

Technology has been developed for years to improve all manner of artefacts, clothing and equipment that are used in sport – including tennis racquets, cricket bats, racing cars and racing bikes. It is par of the course in sport innovation. But some ruling bodies see it as unfair and needs to be stopped in its tracks. For example, back in 2009 a new kind of high-tech super swimsuit developed by Speedo was banned by the swimming’s governing body on the grounds that it gave certain swimmers an advantage, and in so doing, was ruining the sport. The full body swimsuit was made from polyurethane which can trap air in the suit and increase buoyancy to the swimmer making it faster to swim.

The LZR Racer Suit unveiling at a press conference in New York City in February 2008. From wikipedia

Another concern is its impact on the past. Many world records from years gone by are being broken, especially those that were made by great sportsmen and women, who have since become legends. They did not have the same kind of high-tech super shoes, etc., then, so it is considered unfair to take away their crowning glories by those who have the super shoes, swimsuits, etc..

But records are there to be broken. And speed and innovation go hand in hand.

Another line of argument is that giving sportspeople these new kind of superpowers is equivalent to doping which entail taking certain kinds of drugs to improve fitness. But the big difference between doping and super tech is that the former is an internal enhancer while the later is an external aid. While both can improve speed, doping goes one step further by invisibly enhancing an athlete’s stamina which is especially important in endurance and long distance sports. The various substances, like EPO, used in doping increase the taker’s red blood cell count, enabling more oxygen to be transported around the body and to the muscles, thereby increasing stamina. No-one questions whether this should be banned because not only can they give certain athletes an unfair advantage, they can be dangerous, causing health issues. It is also difficult to see how much someone has taken and for how long so it is a very unfair playing ground.

Super shoes, on the other hand can be checked to see if they fit certain regulations. The same could potentially be true for full body swimsuits (they still remain banned from the ruling of 2009).

Ten years ago Google developed a concept shoe that could talk to the wearer to motivate them to get up and go. It was long before chatGPT had arrived. In the near future it might be the case that the super shoe could be embedded with a GenAI app so that it could talk to the runner like in the video – helping them keep going in the way trainers and spectators currently do from the sidelines. Now that would be truly super.

 

Changed Images

Posted on by

AI is being applied to all manner of tasks that humans have always done and which over time are proving to be better than them. One of the first applications was in radiology where machine learning was used to classify medical images to determine if they were cancerous. The method has become so good now that it is faster, cheaper and more accurate than even highly trained radiographers.

Since then, it has been used to classify all kinds of images. The latest, according to The Guardian, is its appearance in the online dating world, where it is being put to the test to decide which photo of someone’s mugshot makes them look the most attractive. The question this raises is how does the AI know? It is quite different from determining if a dark patch on an X-ray image is a lesion.

Come to think of it, how do humans know which photo of themselves to select when it comes to showing themselves off in the best light?  It seems such an art form. Recently, I had about 100 photos taken of me by a professional photographer who made me pose in various ways, from leaning against a glass wall to standing in a corridor. I force-smiled my way through all the clickety clicks and cajouling, feeling mightily awkward. Of the ones the photographer showed me, none looked particularly flattering. But being highly trained in this skill, she was able to pick the four she thought I looked good in and asked me to choose one from those for the website it was intended for. I found it really hard and would have appreciated some AI to help me.

It got me thinking, what makes one image look better than another one? Is it the lighting, the way the eyes are open, the size of a smile, the way a head is tilted? And so on? I am sure the AI would be able to work it out but not know how it did it.

Meanwhile, generative AI is being developed to help people write about themselves in a more enticing and personalized way. For example, the dating site Match Group has been using it to “eliminate awkwardness” when dating online. Let the AI tell a few white lies rather than the human, who as we well know will often ‘lie’ about their age, interests, and what they are looking for in a partner to improve their chances of getting a catch.

Touching up photos to make someone look better has been around for ages – especially if they are to appear on the front cover of a magazine. Nothing new or wrong with that. But the photo doing the social media rounds today is intended to put someone in a bad light rather than in a good one. The photo in question is of a girl with pink hair looking askew at the British prime minister Rishi Sunak who has just poured a pint at a beer festival – showing he is a man of the people. The original image had been ever so slightly manipulated so as to change the facial expression of the girl from appearing disinterested (one on the right) to dismissive (one on the left) – in the blink of a few pixels being switched around. It did not need AI to do that – anyone could have done that using Photoshop but the furore it has unleashed about AI-enhanced images is quite astonishing. I thought the photo was quite funny just like the AI-generated image of the pope wearing a puffer coat, which, clearly, he has never worn. We all like a bit of harmless mischief, seeing the bizarre, the funny, and the hilarious – especially if it is of a famous person. However, the ‘beef’ in all the comments and concern of the pint pulling image is that it is politically motivated and could be the start of a bigger misinformation campaign. Some even think it is a threat to the democratic process. The thin edge of the wedge.

Distorting images for political gain is a century’s old trick. For example, the infamous photo that was used by the Tories over 40 years ago, entitled “Labour isn’t working” is a case in point. The image depicts a snaking unemployment queue linking up outside of an unemployment office. It was blown up as a full-scale poster and sent a powerful message that if the Labour party got elected the dole queue would get much worse than it was at the time. Powerful stuff. The image itself was manipulated using an old-fashioned technique. Only 20 volunteers showed up for the photoshoot and so the photographer took lots of photos of the same people and then stitched them all together. If you look hard enough you can see the same people appearing throughout the queue. From a distance, you would never have guessed. The poster was placed on billboards and buses all over the UK and this pervasive presence became etched in people’s minds no doubt nudging them to vote one way rather than the other.

It wasn’t the photo technique that was the problem but the way it was used to distort the message and be the subject of misinformation.

This kind of trick of making something appear much larger than it actually is, can be done now in a matter of minutes using a contemporary generative AI tool. In my mind the focus of concern should not be about the wizardry of what can be achieved with AI tools now but with the advertisers and the politicians who connive to using them to deceive and distort the masses.

Beyond Human

Posted on by

An interview with Geoff Hinton in the New York Times this week made me stop and think about where AI is heading next. The last few months have been a bit of a roller coaster since the public was given access to the various generative text and image tools, such as chatGPT, Bard, Dalle-2 and Midjourney which has resulted in millions of people trying them out. Quite remarkably, it has unleashed unprecedented levels of joy, amazement and incredulity from all walks of life at what they can do when prompted by us, humans. It has also been matched by many people worrying about how the AI tools will replace millions of jobs while turning our children into cheats, by using them to do their homework. Not since Google search and Facebook came to the fore has there been so much excited talk across the globe about what a new tech can do for society.

I have always been a glass half full person and start by looking at the benefits and opportunities of a new technology. Instead of seeing them ending up making children lazy, I see them helping them learn how to write and code. I see them empowering many others in their work – for example, enabling clinicians to see more patients by summarising and automatically composing reports from their consultations with their patients. Across the board, it is taking the drudgery out of many of our mundane knowledge-based tasks. It can also spark our imaginations to do more and differently. Writing new kinds of plays, books, scripts and so on.

But then something Geoff said in the middle of his interview made me see the future of AI from the dark side, for a change. The first worry that got me thinking was when he talked about there being ever more powerful tools that can be used by evil, malevolent and greedy people for their own ends. The so-called bad actors of the world (although personally I don’t like that term as it used to refer to people who are not very good at acting on stage but now been generalised to mean malicious rather than incompetent people). Scam porn, fake news, fake videos, fake voices, fake claims, etc. could become mainstream, being used in all sorts of unpleasant ways to extort us. Moreover, it is already starting to get messy; being more difficult to know what is real and what is fabricated. While we have all gotten used to getting email scams and more recently deepfakes we may find it more annoying, confusing and scary when they become commonplace – where videos, images and stories are made about us which are untrue but which become increasingly difficult to disprove. Our sense of reality will be turned upside down and inside out if we are not careful.

The second and perhaps more worrying concern is not knowing what the new AI tools might be used for or do, God forbid, on their own volition as they become ever more ‘intelligent’.  What happens when say, chatbots, start to learn from each other? Geoff poignantly pointed out:

“…  it’s as if you had 10,000 people and whenever one person learnt something, everybody automatically knew it. And that’s how these chatbots can know so much more than any one person.”

Yikes! It could unleash a new generation of killer robots that could easily get in the wrong human hands causing unprecedented destruction. And other unthinkable things.

I reminded myself that this has always been the case with manmade (sic) weapons  – whether it is chemical warfare, machine guns, nuclear arms, and so on. We still see terrible acts and atrocities being committed all the time. But will the new generative AI escalate man’s propensity to destroy life with the latest technological means available and make it easier and more shocking to do so?

The sensible course of action is for societies across the globe to put stops and deterrents in place. The politicians and the officials are indeed starting to discuss what to do. But is it possible or realistic for governments to make sure safeguards, regulations and policies can be developed and implemented in order to prevent or reduce AI from being used for bad? Will the exponential advances in AI mean that the goalposts keep moving making it is difficult to keep up?

Returning to my more comfortable glass half full lens, I think about how the unprecedented speed and uptake of the latest AI technological advances could enable us to make more rapid and more profound scientific and medical discoveries, such as finding cures and ways of preventing awful illnesses, such as dementia, Parkinson’s and cancer.  Could we also use them to help society detect and even prevent all manner of crimes before they happen, including terrorism, people trafficking and money laundering? And wouldn’t it be wonderful if it was possible for AI to solve the problems it creates?

The bottom line is one of balance. Is it possible that the new AI kids on the block can be put to more good use than to bad use – now the genie is out of the bottle?

 

The two inserted images were created by DALL·E

Oh Bard!

Posted on by

Sometimes bad timing and misfortune can end up having a massive negative impact on an organisation, as happened recently to Google who were in the process of launching their new AI tool, Bard. Hours before going live Reuters pointed out how it was not up to scratch as it saw an error in the promotional ad. It went viral and the effect was to wipe billions of dollars off Google’s shares. How did it happen?

A tiny factual error in one of Bard’s maiden answers to what was a seemingly banal question was the trigger for this catastrophic nose-dive. The question in question that Bard was asked was what new discoveries had been made from one of NASA’s mighty big space telescopes (the JWST) that could be told to a nine-year old. Bard replied that it took a picture of an exoplanet – which is a planet outside of the earth’s solar system. However, the human who tweeted pointed out that in fact it was another telescope that did this – one in Europe called a very large telescope (the VLT). It was meant to be an answer suitable for a young inquisitive child and most kids of that age would have not minded or would have blurted out it had made the mistake.

However, a bit of investigative work by a reporter at the Financial Times noted how Bard was technically correct since it was the JWST’s very first sighting of an exoplanet, but in the wider context of world knowledge, it was another telescope that had spotted it earlier. So a pedantic matter. Just goes to show how fickle the world is when it comes to its trust and faith in tech. Or maybe it was fuel thrown at the new AI race between Google and Microsoft.

Meanwhile OpenAI’s chatGPT (with Microsoft investment) continues to soar in terms of its credibility, popularity and capability. I have used it several times now and am amused and astounded by what it can accomplish in real time. Sure, it can get things wrong (for example, it did not know the Queen had died or when the King’s Coronation is because it is only trained on data before 2021) and its prose can be a bit bland and clunky but it has transformed how many pedestrian writing tasks can be achieved. Just like the spreadsheet changed how we do financial forecasting and the calculator offloaded the need for humans to do mental arithmetic in their heads anymore so, too, will this new generation of LLMs transform how we write.

In fact, millions of people, like me, have tried using ChatGPT in the last couple of months and are mightily impressed by how it can get them started writing an essay or report – overcoming that blank page syndrome. When I asked it to write some feedback that I could give for a graduate student report I was impressed by its fluid style and use of praise – almost as good and personalized as I could!

At the same time there are those who are worried that it will dumb us down or turn us into cheats, for example, students will increasingly use it to write their essays, reports and other assignments on their behalf. But why not? They can then be asked to read and spend time reflecting to how good ChatGPT’s answer is and how they can improve upon it. Instead of simply regurgitating what they find on Wikipedia or other online resources they could be asked to develop and hone their critical and analytical skills. And learn what makes for a good or poor argument, developing some metacognition skills in the process. Meanwhile, professors and teachers could use the next generation of ‘turnitin’ AI plagiarism tools that are starting to appear to detect how much they have changed the chatGPT answers. We can also begin to rethink our assignments and ways of providing feedback to students. In so doing, we can all learn to write better – be it generating and creating or assessing and providing feedback. Framing the new generation of AI in this way will enable all sorts of new possibilities for students (and teachers) to learn and teach with. As was said in the Google launch blurb Bard “can be an outlet for creativity, and a launchpad for curiosity.”

Funny how scientists love coming up with acronyms so much. Anyone want to guess what NASA, JWST, VLT, LLM and GPT stand for? Perhaps we could just ask Bard.