How to tell if an image or video has been created by AI - and if we still can

Explainer - Love it or hate it, artificial intelligence is everywhere these days.

But for every new technology, there are always people who will exploit it.

AI-generated images are unavoidable online, with many intentionally being used to mislead, scam or monetise outrage.

If you are not a computer or AI expert, how do you even know what's what anymore?

The good news: You can often still tell the real from the AI. The bad news: It's getting harder all the time.

False images and videos are frequently being churned out in response to real-life events - many regarding the Israel-Iran conflict were just the latest example of propaganda hitting social media.

RNZ spoke to AI and fact-checking experts about how the average internet scroller can figure out if that amazing viral image of Abraham Lincoln with his iPhone might just possibly be fake.

How can I even tell if an image is fake now?

"It's really hard," Victoria University senior lecturer in AI Andrew Lensen admitted.

"Generative AI models, where users provide a description (caption) of what they want and the model tries to create it, have come a long way over the past few years."

Andrea Leask, deputy CEO of the online safety non-profit organisation Netsafe, said the easy availability of apps and technology using AI had played a big part.

"Anyone can generate AI content," she said. "A child can create digitally altered images."

Ben James is the editor of AAP FactCheck, a division of the Australian Associated Press. (Full disclosure: This reporter also writes for AAP FactCheck.) James said the flaws and mistakes in AI imagery were becoming much more subtle.

"AI fakes are far more sophisticated than they were six months ago, and they will continue to evolve. Gone are the days of six-fingered people."

Lensen agreed, noting the early rubbery clunkiness of older AI images was rapidly fading away.

"We all laughed at Will Smith eating spaghetti or people having seven fingers in 2023, but now the AI models are sufficiently advanced that it basically no longer happens.

"I think for the general public, reliably recognising AI images is no longer possible, and for AI/media experts, there's not much time left."

James said that AAP FactCheck asked three questions at the beginning of every check it does: Who is making the claim, what is the evidence, and what do trusted sources say.

"Whether we are talking to journalists, schoolchildren, or seniors, we always return to that three-question process.

"It doesn't have to be overly burdensome, but it is a way of prompting those key critical thinking skills and will, nine times out of ten, keep you out of trouble."

So what can the average person do to detect an AI image?

Looking closely at images can reveal clues they are not what they appear to be.

"AI still struggles with textures and shadows," James said.

"It also struggles with small details, particularly writing. Look for name badges, logos, road signs, and similar elements; often, the words will be garbled.

"Be wary of perfection. Often, AI images are just a little too perfect: airbrushed skin, background details all perfectly framed."

With video, there could also be tell-tale signs.

"Look for distortions, particularly around the face, hair and hands," Netsafe's Leask said.

"For example, flickering around the face and hair, inconsistencies in skin texture, unnatural eye movements or finger placements."

Another simple solution is to use what's called a reverse image detector - a search engine that looks for other instances of a photo to determine where it first came from.

It can be useful for tracking down the original version of an altered photo or a photo of a past event being presented as a current one.

Popular reverse image searchers included TinEye and Google Images' own search.

"Various AI-detection tools, while not perfect, can offer further indications as to whether an image or video is genuine," James said.

However, Lensen was sceptical about their long-term usefulness.

"There is a real 'arms race' of AI generation vs AI detection, and I don't see a way for the detectors to win," Lensen said.

"After all, we train these AI models to generate the most realistic content possible, which means as the generator gets better, the detector has an increasingly harder task.

"In fact, many of these models will be trained by having an AI detector that the generator has to 'fool' during the training process!"

Lensen said in his university work, he does not support using AI detectors to check over student work as the consequences of "false positives" - being accused of using AI when you aren't - can be quite harmful.

Look for trusted sources - who might not be your cousin sharing posts on Facebook

When it comes to detecting AI, your own brainpower and detective skills may be the most important tool.

That means cultivating a wider base of sources - and yes, responsible media plays a big part.

"You need to have your trusted sources," James said.

"Despite all the talk of deception, reputable media organisations do a pretty good job of separating fact from fiction. Therefore, you need reliable sources you can count on."

"My advice in 2025 is to look at the provenance of the image," Lensen said.

"On social media, is it a profile with a history of legitimate posts (around a common theme) or is it a strange profile who seems to post very regularly on different topics?

"This is, really, all those 'critical thinking'/source checking that we used to do pre-internet days.

"It is, unfortunately, more work for/onus on the person consuming the content, but I think being a sceptic is a really important skill in the age of mis/disinformation."

Video scams aren't just of celebrities anymore

False videos could be trickier to debunk, but one of the key things to look out for was whether it was an unlikely subject for the person in the video to be talking about - celebrities were often "duplicated" for crypto and investment scams.

For instance, last year, a fake video featuring Prime Minister Christopher Luxon backing an online trading company did the rounds.

Sometimes, victims were not even celebrities. Sir Jim Mann, a leading New Zealand endocrinologist, was shocked recently to discover his face and voice were being used to scam patients with type 2 diabetes.

"The AI was so effective, it looked like I was actually saying those words," he told RNZ.

With video, a key thing to look at is whether the person's face movements match what they were saying, or whether the audio sounded a bit distorted or robotic.

It used to be more difficult to make convincing "deepfakes", Leask said.

"Now, a single image or a handful of words are all you need to create a very convincing deepfake. And ordinary people are being targeted."

Can anything be done about malicious use of AI deepfakes?

ACT Party MP Laura McClure recently put forward a bill in the House to restrict the generation and sharing of sexually explicit deepfakes.

In Parliament when discussing the bill in May, she held up a faked nude photo of herself that she created, saying "this image is a naked image of me but it's not real."

McClure argued the sharing of explicit deepfakes could ruin a person's life.

"For the victims, it is degrading and it is devastating," she said.

The bill was lodged in Parliament's members ballot, where it could get pulled at random, but it was still a long way from becoming law.

Netsafe said they received a significant increase in reports of harm relating to digitally altered images in the past year, Leask said.

"We have found that where the producer of the content is a young person, typically, the digitally altered content will have been created for fun, to ridicule or bully someone.

"In contrast, where the producer of the content is an adult, they are more likely to be motivated by sexual gratification, abuse or harassment."

But it may be difficult to ban such images outright.

"Outlawing it is good in principle, but how will it ever be enforced?" Lensen asked.

"There is a pretty high burden of proof to show that someone produced a deepfake, and that gets even more complex when it could be done cross-border."

He said the government needed to provide more detail of implementation and enforcement to make it a substantive effort to actually solve the problem.

As for legally requiring all AI-generated content to be labelled, it might be a good idea in theory but Lensen said it was not very workable.

"I think that ship sailed a long time ago. Even if we brought in such legislation now in NZ, there is no way it would be adopted worldwide, and so the 'bad actors' could just be located in another jurisdiction.

"And that's not to mention enforcement: even if we could detect AI content (we can't), who is going to police that and take the content down?"

What if you've been targeted by false images or video?

"If the content is digitally altered and abusive, threatening, harassing, or includes intimate material (such as nudity, sexual activity, undressing, or toileting), and it has been shared or threatened to be shared without consent, the Harmful Digital Communications Act may apply," Leask said.

"This includes deepfake intimate images or videos."

"We can often help you get the online content removed and explain the options available under the law."

You could report online harm at Netsafe's website, text "Netsafe" to 4282, email help@netsafe.org.nz or call 0508 638 723.

Media literacy is key, say experts

Then there was the broader problem of media literacy.

A survey last week showed that for the first time the majority of Americans are getting their news from social media and similar results were seen in Australia too.

In New Zealand, the Trust In News Aotearoa New Zealand report released earlier this year found only 32 percent of respondents trust the news.

"Media literacy needs to start from an early age," James said.

"There really needs to be a co-ordinated effort if we are to have functioning democracies making big decisions based on fact."

Lensen criticised the increasing use of AI in some newsrooms in New Zealand.

"We need the media to be a trusted source now more than ever, and using AI really makes that social license harder to maintain."

RNZ has laid out a series of Artificial Intelligence principles which state it "will generally not publish, broadcast or otherwise knowingly disseminate work created by generative AI" and that any use of AI, generative or otherwise, should be done in consultation with senior managers.

Is social media being ruined by the flood of AI slop?

"Maybe?" Lensen said.

"I think it undermines the value of social media for many," he said.

"The big appeal of social media originally was being able to connect with friends/family (e.g. OG Facebook) and likeminded humans (e.g. Reddit, FB groups) to share human experiences and have social connections that don't rely on physicality.

"Injecting AI into that inherently removes that human-ness."

A frequent issue seen by fact checkers is almost nonsensical AI 'slop' being posted simply to harvest engagement.

"A lot of social media accounts have gained followers by posting AI media (without disclosure) because it allows them to draw clicks/reactions," Lensen said.

News organisations are also still being caught out by AI falsehoods despite their best efforts.

A story earlier this month that was picked up by media worldwide featured a manipulated video of a Chinese paraglider covered in ice after supposedly being sucked into the upper atmosphere.

RNZ partner the Australian Broadcasting Corporation was one of those and posted a note explaining why they removed the story.

"It is difficult" to catch some things now, James said.

"We live in a time of instant news.

"Journalists are no longer just competing with other journalists but also with influencers, agitators, and commentators on social media. They can afford to be wrong, but journalists have to be more careful. Our reputation is everything."

Sign up for Ngā Pitopito Kōrero, a daily newsletter curated by our editors and delivered straight to your inbox every weekday.