Deepfakes, disinformation and doghead-men: sorting truth from lies in the age of AI

AI-generated misinformation is evolving rapidly, from deepfake audio in elections to disinformation campaigns that erode public trust. For local government, the challenge is twofold: recognising misinformation early and responding effectively. The good news? AI isn’t just a threat—it’s also a tool. Councils can leverage AI to monitor emerging false narratives, fact-check claims, and craft rapid, targeted responses to counter misinformation.

In this think piece, LGIU’s Ingrid Koehler employs humour, history, real-world examples of disinformation deepfakes and her own hoaxing to highlight the dissolving boundaries between truth and lies.

Wild lies of the Wandle

I have to admit it, I’m prone to an occasional fib. One of my little parenting pleasures was creating fantastical yarns to beguile children. I like to think I was helping them develop skills of discernment, judgment and evidence evaluation, but mainly I just found it funny.

I once convinced a gullible intern, (at least for a few minutes), that there were several colonies of feral marmosets along the River Wandle which snakes through South London. Pleased by my success, I wanted to see if I could pull this over on the kids. My son, who’d been dragged along on too many Wandle walks, wasn’t buying it. But maybe my stepson would? My son didn’t think so, citing his stepbrother’s nature knowledge. But I offered him ten pounds sterling if he’d be my accomplice in the ruse. And it worked. At the end of the weekend, my boy demanded his fee and I paid up.

I was proud to discover that I got boomeranged. My stepson delivered a stellar performance of belief, and those two scamps split the tenner between them. But perhaps if I’d had access to a convincing AI image of marmosets scampering in South London, I’d be ten pounds richer today.

AI-generated feral marmosets on the Wandle River.

Breeding disbelief: the parents who cried wolf or maybe dog

My partner was equally invested in helping our blended family offspring build a sense of healthy scepticism. Some time after the marmosets, my stepson overheard us talking about the doghead people and immediately accused us of spinning another tall tale. No matter how much we protested, that yes, we were talking about people with human bodies and dog heads, and no, the doghead people weren’t real, but once people had believed that they were. These Cynocephali inhabited lands just beyond where the furthest travelled person known of had ventured. But he insisted we made it up out of whole furry cloth.

Certainly, not everyone believed in the doghead people. But they were well documented. They’re in the pages of Pliny’s Natural History. Herodotus reported that someone told him they were in eastern Libya. Much medieval ink was blotted on the notion that St. Christopher was one of these creatures. Generally, in the West, people believed that they lived far to the East, but in the East, that they might be from the West. The ancient Welsh said that dog-headed men resided in Edinburgh.

Of course, there were no photos in those times, but there were plenty of drawings. Presumably based on third-hand eyewitness accounts.

By Hartmann Schedel (1440-1514), – Nuremberg Chronical (Schedel’sche Weltchronik), page XIIr — digital source: Beloit College, Public Domain

The doghead people of today

I’m not aware anyone believes in them now, and they seem to be fading from cultural prominence. Perhaps they could make a comeback as shadowy disinformation workers? But they have their modern equivalents in Little Green Men and Bigfoot, both of which have had their share of drawings, faked photos and true believers.

AI-generated disinformation factory.

I recalled the marmosets and man-beasts and how we learned to have appropriate trust and doubt after attending an excellent LGIU training on dealing with misinformation. Although I’ve written about misinformation extensively, I learned many new things. Rachel Eden, who is also a councillor, led a thought-provoking but pragmatic session, enriched by contributions from my fellow delegates. I was curious if anyone had dealt with misinformation augmented by artificial intelligence. Conversations on the course are confidential, but I can say that none of the examples of misinformation, on the face of it, seemed to be about AI. But it won’t be long before this will become common.

Curious about what other training courses we have coming up? Check out our full calendar here 📅

Is seeing believing?

Much coverage of AI and misinformation is particularly focused on ‘deepfakes’. Deepfakes are video or audio content that use powerful AI learning and computation to mimic real people or events to a sometimes shocking level of realism. It’s quick and easy to make a deepfake, but it’s not yet super simple to make convincing ones.

There have been several examples of deepfakes generated to influence electoral outcomes. One was a ‘deepfake’ audio of the incumbent Sadiq Khan making disparaging remarks about Remembrance Day honouring those lost in war ahead of the 2024 London mayoral election.

In early 2024, thousands of voters received automated ‘robocalls’ featuring an AI audio deepfake of then-President Biden encouraging them not to vote in their state’s primary elections. The fake was identified right away, and steep fines were issued. Back in November, I wrote about election misinformation in the US or – in old-fashioned language – lies. I looked at the lies spread about the disaster response to Hurricane Helene, but I didn’t even mention the AI-generated image of a crying little girl clutching her puppy in a boat floating on rising flood waters. To me, it hardly seemed worth mentioning, it was an obvious fake. But that obvious fake was widely circulated and given that many social media algorithms and humans favour a message with a compelling image, it paved the way for the lies to spread further.

Confirming your bias

In each of these examples, people checked against their pre-existing biases before taking action. Joe Biden wouldn’t tell people not to vote, and the call is reported. The photo of the little girl? The floodwaters devastated North Carolina, so of course there were scared, upset kids and probably puppies, why not share it? The faked ‘secret audio recording’ of London Mayor Khan was first posted by someone who said “It’s what we all know Sadiq thinks,” according to the BBC. Even the person who apologised for helping the clip go viral holds onto the notion that even if the audio wasn’t real, it represented a bigger ‘truth’.

The version of AI imaging that I’ve used below doesn’t fulfil requests for images of public figures. So I couldn’t make an image of Sadiq Khan rescuing the little girl from the flood. Others could, but they shouldn’t.

Mixed memes: AI-generated image

For now, deepfakes are a lot like drawings of dog-headed men. They help people visualise what they already believe. And if they don’t believe it, then it’s easy to disregard it. Human nature means we tend to value facts that confirm our pre-existing biases. AI means that it’s easier to manufacture that ‘evidence’.

Are deepfakes even the issue?

For now, I think there are two present dangers with deepfakes. Some of the audio is pretty convincing. But what’s more common is deepfake pornography. This is targeted at people in all walks of life, but almost always at women or girls. In the context of political disinformation, it’s used against women politicians across the political spectrum. It’s pervasive because the software is widely available. It can look convincing because it uses well-tested face-swapping technology to put real images of the target woman snatched from innocent social media posts, official portraits or news reports onto the bodies of other women filmed in real pornography.

The sense of realism isn’t entirely the point. This may not even be disinformation in the strictest sense. The goal seems less to convince the electorate and more to harass and threaten the individual politician and attack the credibility of women in politics overall. Beyond the horror and disgust experienced by the target, its most devastating impact could be to dissuade women and girls from entering public life. It’s another degrading layer to the usual slurs and rape threats that female politicians endure so often that it’s wrongly seen as part of the rough and tumble that goes with the job (read more about the impact from Vandinika Shukla at Tech Policy Press).

People have been able to manipulate images for a long time and mimic voices for even longer. It takes skill to manipulate images without leaving tell-tale traces, and it takes a lot of effort to find someone who not only can impersonate a well-known figure but will. The key challenges to creating fakes in a high-churn news cycle are speed, cost and efficiency. This is a problem AI is designed to solve.

The proliferation of detectable deepfakes erodes trust, too. Knowing that deepfakes are swirling around means that we have to expend energy to mentally fact-check everything we see and read. Seeing is no longer believing. Checking if content is real will increasingly require more skill than most of us have. This puts every claim on shaky ground, and it’s exhausting. The greater temptation will be to check out and withdraw further into our online bubbles of confirmation bias, creating further polarisation.

The 2024 London mayoral election was impacted by other instances of generative AI misinformation. One disinformation campaign used AI to quickly and cheaply generate social media content published by bots that were ‘hashtagged’ or labelled with voter fraud messages, making it a trending topic. While individual messages weren’t impactful or even about elections, the sheer volume was designed to undermine confidence in the election. In other words, most of the AI-generated content didn’t matter as individual messages, but AI helped malicious actors ‘flood the zone’ with messages that looked like they could have been written by real people.

When ChatGPT first hit the scene, it was easy for me to tell which text was written by a person and which had been written by AI. I’d find it much harder today with software trained on individual writing styles. It’s also easy to generate text which is persuasive, authoritative and free from the kind of grammatical errors or gaffes that might trigger scepticism. One person’s rage over a planning proposal can now become a social media campaign from a whole fictitious community with the click of a few buttons. Dr Rumman Chowdhury, whose work focuses on ethical AI, has suggested that even more convincing compositional AI, building whole backstories about politicians’ lives or their political records including real links to fake articles, will take online smearing to a far more persuasive new level (UNESCO report). The rising volume of content, which may or may not be written by real people or contain real facts, remains the biggest danger… for now.

Using AI to fight misinformation

While bad actors are using new tools to spread old lies, local governments can use AI to fight back.

• AI monitoring tools can be used to detect emerging misinformation campaigns and pick out anomalies in community sentiments.
• Automated fact-checking tools are already on the market, and AI can also be harnessed to check claims against verified sources such as trusted media outlets, scientific papers or the council’s own published policies and statements.
• AI can be used to analyse images and audio to detect deepfakes.
• Of course, local government communicators can use AI to help craft messages that counter misinformation to a variety of audiences at a previously impossible speed.

But we have to embrace a culture of experimentation and the use of AI for good ends. This means providing training, a framework of ethical guidance and trusting individuals to use their judgment when they use AI.

Navigating the waters of truth and lies in the age of artificial intelligence

The spread of misinformation has never been easier, faster, or more pervasive than in the era of AI. While deepfakes and AI-generated content may still function like modern-day drawings of dog-headed men—reinforcing existing beliefs rather than radically shifting them—the sheer volume and sophistication of AI-assisted misinformation pose an unprecedented challenge.

The real issue isn’t just the existence of deepfakes but the broader erosion of trust. As AI tools become more adept at generating convincing text, images, and audio, distinguishing between reality and fabrication demands more cognitive effort than ever before. This constant vigilance risks not only exhausting individuals but also further entrenching ideological echo chambers. If seeing is no longer believing, then what do we trust?

The answer may lie not in retreating from technology but in reshaping how we engage with information. AI may be a tool for misinformation, but it can also be harnessed to detect and counteract it.