It’s getting harder but here’s how to spot a deepfake (for now)

Sunday, 12 July 2026

Here are five terms you should know, in order from interesting but harmless to existentially terrifying.

Spotting deepfake video is increasingly difficult but there are certain tells which are still dependable.

Current deepfake video often looks too perfect. Real video has natural motion blur and imperfect, variable lighting that AI struggles to match.

AI video is generally short as AI will quickly ‘forget’ the original prompt and creating bizarre effects. Longer, cohesive videos are prohibitively expensive.

If you’re looking for a quick way to measure AI progress, you could do worse than watching Will Smith eat spaghetti.

In what has become a viral benchmark for AI-generated video, every successive generation of AI is tested on how realistically it can generate the actor/rapper absolutely hoofing a plate of pasta.

The image on the left is the peak of AI video in 2024, the right was produced in 2026. Both are equally fake.

As recently as 2023, the results were genuine nightmare fuel. Now in 2026, with Kling 3 producing this worryingly realistic scene, the consensus is that the spaghetti test has been passed. Even the actual man himself got in on the trend.

While a rapper eating pasta might not seem too troubling, this level of deepfake has obvious and immense potential for misuse—from political disinformation to sexual harassment.

It’s something experts like Auckland University Honorary Research Fellow Andrew Chen have been warning about for a decade.

“Ten years ago I was saying at some point we will reach a place where the images are pixel-perfect, and by pixel-perfect they're indistinguishable at the pixel level from human-generated or human-taken photos. We're pretty close to that,” Chen told Stuff.

So now that we can’t just watch for extra limbs, horrifying amounts of teeth, or nonsense writing in the background, what ways are left to spot a deepfake? And what do we do when it becomes practically impossible?

The telltale signs of deepfake video (for now)

These days, noticing a deepfake video requires some closer scrutiny, but it’s still possible.

One way is looking for lighting inconsistencies—is everything lit a little too perfectly, like it’s on a Hollywood set?

The reason deepfake video still has that glossy, slightly uncanny feeling to it is the kind of video it’s been trained on. Datasets available to video generators are often advertisements or professionally shot videos which have already been touched up in some way.

They replicate this airbrushed aesthetic into the outputs, making every person in the shot look like an extra in a commercial.

More realistic, natural video—like the kind shot on your phone—contains natural motion blur and variable lighting, which is much harder to train from and therefore realistically replicate.

Deepfakes like this viral Brad Pitt v Tom Cruise battle look good in motion but off when paused. Look at the blur around his skin and strange lighting.

This means one reliable way of testing a video is pausing it, moving through frame by frame, and watching for the bizarre physics, lighting, and other inconsistencies that you miss when it's playing at full speed.

Deepfakes can also take a different route, disguising any strange inconsistencies by intentionally degrading the video quality to pretend it is footage from a security or dash cam.

Therefore, it’s always worth asking: is this a suspiciously short clip showing something spectacular you assume someone would be filming for longer?

Current models suffer from a limited memory capacity, referred to in technical jargon as a 'context window”.

As an AI video progresses, the model forgets the context established at the beginning of your prompt. This is why you often get surreal, horrific, reality-warping effects the longer a piece of video you try to generate. Most deepfakes will therefore be measured in seconds, not minutes.

Of course, trillions of dollars are currently being spent annually to improve AI models, so all these technical challenges could be largely solved in the near future.

But one particular bottleneck could help slow down progress long enough for us to prepare for a world of readily available, pixel-perfect AI video.

What is holding deepfakes back

A key limitation for deepfakes is how resource-intensive they are to produce.

So while the ever-escalating power needs of AI is a clear environmental issue—AI data centres are on track to use more power than Japan by 2030—it may be a mixed blessing in this case.

Making a language prompt uses a relatively small amount of electricity, about as much as watching TV for eight seconds.

But generating images requires hundreds of times more than that. If you're generating a video, which is just a series of images strung together at 30 to 60 frames per second, that’s thousands of times more resources required.

Dr Andrew Chen, confirmed to not be AI generated.

Astronomical computing costs were one of the reasons OpenAI shut down its AI video-generation platform Sora this year after it reportedly cost them up to $15 million a day.

So for now, creating Hollywood-level ready deepfakes— like this completely fake battle between Tom Cruise and Brad Pitt—is extremely expensive. But the potential rewards are increasing in step.

“If it's valuable enough to create it, then someone will,' Chen says.

“If you can create a very convincing deepfake video of a CEO telling a CFO to transfer $12 million to this bank account, it's worth it. If you want to influence an election, might you be willing to spend $100,000 to create a video that is highly convincing?”

How New Zealand is responding

The good news is New Zealand is moving on deepfake regulation, at least when it comes to one specific harm.

May saw New Zealand’s first prosecution under the Harmful Digital Communications Act (HDCA) for creating and sharing deepfaked pornography using images obtained from social media. A 21-year-old man was sentenced to 24 months of intensive supervision.

Under current legislation, a victim of deepfaked pornography must prove an intent to cause harm—a high bar to clear that makes convictions under the HDCA difficult to obtain.

A member's bill introduced by ACT MP Laura McClure seeking to broaden protections passed its first reading with unanimous support. It would amend the HDCA and Crimes Act to expand the definition of an “intimate visual recording” to explicitly include synthetic images.

While this will give protections to victims of non-consensual deepfakes, protecting against a wider range of harm—like weaponised disinformation—will require more expansive regulation.

What could come next

There are many online detectors which claim to be able to identify deepfake content, but the problem is that every successive generation of a model can be adapted to avoid them.

There is an arms race between generators and detectors, and generators tend to have the edge.

Chen says if we want to protect against a wider range of harms, we have to regulate at the source by 'taking the fight to the big tech companies”.

He recommends legislating so that if a company sells a tool which can produce deepfakes in New Zealand, it must have a digital watermark automatically included. This would mean any potentially harmful deepfake could be easily identified.

“We should say, 'We think that this is important and we need to be able to distinguish between reality and fiction so we want you to watermark if you're going to be producing here.' Then the big tech companies have to make a call. Do they watermark, which they've always said that they could, or do they exit New Zealand?”