Last week marked the start of landmark court action that a consortium of US states has brought against Meta over the effects of its social media platforms on young people. Much of the opening arguments focussed on the addictive effects of its algorithms and infinite scrolling. Those insidious features are bad enough, but my biggest fear is that these platforms will rob us of our grip on reality.
It remains to be seen whether the young can be protected by National’s 11th hour bid to introduce legislation banning under-16s from social platforms (ACT says not, and does not support the banning Bill introduced yesterday to Parliament). But let’s not stop at children in assessing the grip on reality: The potential to lose the ability to determine what is real is a danger facing all of us.
Once upon a time, losing touch with reality was a quick way of being declared insane. Today, it can afflict all of us, and I have to admit that it is driving me mad.
Every time I addictively scroll down Facebook, I am confronted by images that look real, but which are the product of artificial intelligence. I see the reimagining of historical events, bikini-clad maidens with near-impossible torsos, idyllic scenes, horrific scenes, and cats and dogs that are endearingly cute.
Those are the ones that disclose their origins or which I can readily identify as AI-generated, but I have been caught out with images I took to be real.
In fact, I was caught out yesterday. I shared a Facebook post relating to a small piece of nostalgia from the time I was a child, but did not notice the image was a concoction and had been marked as AI-generated. It appeared to be a pleasant reminder of my childhood. Nostalgia is a strong emotional response and, as you will see later in the commentary, it is a believability trigger.
Some AI-generated images, like the piece of nostalgia, are innocuous. A rather splendid art deco building left me disappointed when I found it did not exist, but no real harm was done. The same cannot be said of a fake image of body bags being loaded into an ambulance following a triple homicide in Hastings in April. It looked real. That incident led Netsafe to issue a statement about its concern over the proliferation of pages on Facebook that are now regularly posting AI imagery that purports to represent real scenes of tragedy.
I don’t beat myself up when I fail to recognise a fake. A study last year, carried out by British biometric technology company iProov, tested 2000 people in the United Kingdom and United States and found that only 0.1 per cent correctly identified all fake and real images and video. The ‘fail’ rate was higher on video than on still images.
Those findings worried me. Yes, I knew that breakdancing tabbies and talking labradors had a zero per cent chance of being real, but the statistics meant there was an extremely high probability that I had been duped by other images and videos. Equally disturbing was the finding that – despite the failure rate – people remain overly confident in their fake-detection skills. I’m afraid I would put myself in that category in spite of the finding that over-confidence is highest among 18-34 year olds.
Over-confidence could cost you money. A study released earlier this year by the Commercial Bank of Australia found that nearly nine out of ten Australians believed they could spot an AI-generated scam, but only 42 per cent correctly identified the fake images.
That is only one of the concerns about the negative impact of such fakes. Three out of four people in the UK/US survey were worried about their social impact but more than 80 per cent of those over 55 expressed such anxieties.
Concerns about social impacts were obviously worrying Meta’s Oversight Board (an appointed body of experts) when it told the company it needed to do something about the proliferation of AI-generated images and videos on Facebook and Instagram. So the juggernaut created META.O, a detection tool that could assess images on its platforms to identify those that were AI generated. You might have praised Mark Zuckerberg for moving with commendable speed, were it not for the fact that the implementation of the tool coincided with the launch of Meta’s own AI assistant to create images online.
The image you see at the head of today’s commentary was created with that Meta AI tool. Yes, of course it’s a fake: I have not been inside the White House and Donald Trump would not shake my hand unless his was encased in one of those newfangled ICE shock gloves. I think you’ll agree the tool can produce convincing images. However, when I posted the image on Facebook as a test, nothing flagged it as AI-generated. To determine its provenance, I had to proactively copy and load the image onto Meta AI and ask for verification. When I did so, Meta confirmed it was AI-generated. However, how many people would routinely go to that much trouble while addictively scrolling through Facebook?
I share all of the anxieties people have over the spread of fake images and videos on social media, but my greatest fear is for reality itself.
Let’s not become bogged down in postmodernist arguments about a fragmented world and the truths we create for ourselves. I’m not talking about grand explanations of society and its foundations but something much more basic: Our ability to believe what we see with our own eyes.
Susan Sontag, in her acclaimed essay On Photography defined the photographic image and made what was a helpful distinction between the written word and photographs.
Photographs really are experience captured, and the camera is the ideal arm of consciousness in its acquisitive mood…What is written about a person or an event is frankly an interpretation, as are handmade visual statements like paintings and drawings. Photographed images do not seem to be statements about the world so much as pieces of it, miniatures of reality that anyone can make or acquire.
That essay was written almost 50 years ago and, even then, veracity was never absolute. Photographs could be manipulated but, given the abundance of uncorrupted images, the public could generally be satisfied that they were looking at ‘miniatures of reality’.
Artificial intelligence is fast robbing us of that ability. I despair of a time when reality can no longer be taken for granted but becomes a mixture of distorting mirrors and works of self-serving fiction. The consequences of that possibility are the destruction of trust and a validation of falsification that strikes at the heart of everything from social structures to personal relationships.
Some jurisdictions are moving to head off that dystopian possibility. Australia recommends AI developers mark generated content to help people identify it as such. The European Union has gone further and earlier this month implemented a law that places obligations on companies using AI systems. They must, for example, disclose AI-generated or manipulated images, audio, or videos which resemble existing persons, places, or events.
The accompanying codes are explicit about what needs to be labelled and how but it is a fair bet that transnationals like Meta and TikTok will meet only minimum obligations – just enough to keep them on the right side of the law.
I know what that looks like. I have lost count of the number of times I have looked at images, doubted their veracity, and then searched until I found a miniscule AI symbol sitting where it was least likely to be seen.
In any case, such a ‘Where’s Wally’ approach may be too late. Research by Dr Márk Miskolczi of Corvinus University of Budapest has found there are much stronger drivers than authenticity. He found that AI-generated images (AIGIs) often elicited strong responses:
“The analysis identified five main thematic categories framing AIGIs: emotion and nostalgia, arousing empathy and compassion, extraordinary achievements and hobbies, birthdays and social ills, and religion and spirituality. Among these, the most prevalent were the emotionally nostalgic and empathy-inducing themes, often centred on elderly individuals, rural life, and intergenerational bonds.”
He found AI-generated images deceived users by simultaneously triggering multiple cognitive biases such as confirmation bias, anchoring, and familiarity effect, which together reduce critical thinking. The study showed such images used several biases at once, creating a more powerful and harder-to-resist persuasive effect than typical misinformation.
The cute kitten fakes have seduced us into accepting a level of AI-generated images that is hazardous now and prejudicial to the future of civil society. Prominent disclosure is part of the solution, but ways must also be found to limit the nature and volume of AI-generated images and videos. They form an integral component of the harms caused by social media. Miskolczi, and other researchers, have found that emotionally-charged images are more likely to be perceived as credible when they are accompanied by what he terms “social proof” – likes and supportive comments on social media.
Regulatory and legislative measures to limit social media harms need to include specific provision for the harmful effects of AI-generated images – beyond the sexualised imagery that ACT MP Laura McClure’s member’s bill seeks to prohibit. One of those harms is the dangerous distorting of reality by creating ‘events’ that did not occur and misrepresenting the actions of real people.
We can expect European nations to enact tougher laws if the platform owners do not show greater responsibility and willingness to follow not only the letter but the spirit of the recently enacted laws. What can we expect of New Zealand’s Parliament? For now, we are likely to have only limited solutions that owe more to electoral opportunism than a genuine desire to bring socially irresponsible platforms to heel.