Spotting AI’s deepfakes

Ged Byrne of Toastmasters International says looking for authenticity is the key to establishing fact from fiction in an increasingly confusing world

Search YouTube and you’ll find scenes of Batman interacting with Wonder Woman where Ben Affleck and Gal Gadot have been replaced by Adam West and Lynda Carter. Just one example of a deepfake. Deepfakes are generated using machine learning employed in artificial intelligence technology; they can be audio, video or individual images. And for many, deepfakes are a bit of fun; perhaps you’ve laughed at the fake photos of Donald Trump in an orange prison jumpsuit. However, it was no joke for the Hong Kong branch of a multinational company when an employee was tricked into transferring more than $25million following video calls that turned out to be deepfakes.

Spotting a deepfake
There’s no guaranteed way to spot a good deep fake, but most scams are not top quality. They’re usually quite cheaply done, unless you are a high value individual and it’s worth investing significantly in a targeted scam.

When trying to spot a deep fake, there are aspects of the human body that it struggles with more than others: eyes; hair; teeth; facial expressions, especially emotion; body posture and movement. Keep a close eye on these, especially when the head is moved. Does the hair move naturally? Do the eyes, nose and mouth move in sync or are they slightly off?

If you are suspicious during a video call, try to get a reaction. Do they respond appropriately with laughter, smiling and eye tracking? Because of these tells, a deepfake will avoid movement. If the speaker is unnaturally still, that is another sign that you are looking at a deep fake.

You versus AI
While most of us are unlikely to be targeted by Mission Impossible level deepfake scam artists, whether we know it or not, AI and its mimicry of humans is creeping into our lives more and more.

ChatGPT was launched in November 2022 with the capability to rapidly generate text on any topic. Suddenly, articles could be written at the push of a button. The result is many content creators generating large amounts of AI created content, making it harder to compete for an audience’s attention. How can humans stand out with so much computer-generated text sloshing about? The answer is to lean into being human.

The good news is that there is a well-established model to help us do this: The SUCCESS model. To stand out as human you need to create: Simple Unexpected Concrete Credible Emotional Stories.

Let’s play the imitation game.

Consider these two reviews for a restaurant. Only one of them was written by a human.

Version 1
“If you visit Riyadh, a good place to eat is the Nadj Village restaurant. Beautiful castle-like building with an open courtyard in the middle with running water and ‘booths’ with rugs and cushions. You leave your shoes outside the booth. We ordered a mix of dishes. Delicious. And very reasonably priced.”

Version 2
“Riyadh in Nadj Village is a hidden gem that transports you to a different world with its castle-like building, open courtyard with running water, and cozy booth seating adorned with rugs and cushions. The fact that it offers all of this at reasonable prices is a delightful surprise. Whether you’re a local looking for a unique dining experience or a visitor wanting to savor Middle Eastern cuisine in an enchanting setting, Riyadh is a must-visit.”

It isn’t difficult to tell the first version is from a human, but why do you find it more engaging? The SUCCESS model makes it clear.

The first review is simple, getting straight to the point, it tells us exactly what we want to know. It is also concrete, painting a mental picture with just a handful of words. The AI’s superfluous ornamentation dilutes the impact. Most importantly, the first review is credible with believable, grounded language. The restaurant is “a good place to eat” that is “reasonably priced.” The AI cannot resist spicing its text with hyperbole, insisting that this “unique dining experience” is a “must-visit” that will “transport you to a different world.” It is all just too much.

ChatGPT struggles with these basics. The paragraph above was taken from 376 words of generated content. The other 281 words were boring, stretching credulity, and adding nothing of value. Perhaps, in time, these issues will be resolved.

The three remaining points of the SUCCESS model will prove much more difficult for AI to overcome: stories containing emotion and the unexpected.

Did you notice that the diners took their shoes off before entering the booths? Of course, because it is unexpected. When was the last time you dined out barefoot? Humans are always attracted to surprising details, but being surprising is difficult for ChatGPT. As a Large Language Model (LLM) ChatGPT works similarly to Google Complete. If you go to Google and type “Dining” it will try to guess what word comes next, suggesting words like “chair”, “table” and “room”.

These suggestions are based on the words that typically follow “dining” in the training data. ChatGPT is always trying to predict what word is most likely to come next. To keep things interesting, it avoids the most obvious word, instead picking at random from a list of alternatives. It has never visited a restaurant, tasted food, or enjoyed a night with friends. It has only read about them. You are a human being, just like your reader, and both of you have shared these experiences. This is how you know the details that are important, interesting, and surprising to your reader.

Human emotions are even more inaccessible to computers, but they are aware of their importance. Like a teenage poet they overcompensate for their lack of experience by striving for heightened emotions that feel inauthentic. It is the prosaic emotions of everyday life that stand out as human. We don’t see it in the text of our review, but it is clear in the accompanying picture showing a group of women who are comfortable in the cushioned booth and each other’s company. They are obviously enjoying their hard-earned rest.

The ultimate weapon in your human arsenal is the story. If you ask ChatGPT to write a story it will create something strange and dreamlike because of its limited memory. Once the context window is full, older details are forgotten. Human lives tell stories that span weeks and years. For example, the review is one of a series of posts journalling a visit to the 2023 Riyadh International Book Fair. A single chapter in an ongoing story.

If you are a small business communicating with your established and prospective customers, you have something very powerful. A shared story that is either in progress or about to begin. You are providing something that they need, joining together in a mutually beneficial relationship. Human relationships are rich and complex, spinning a myriad of narratives.

Now, take a look at what you are writing. Is it simple, credible and real? How can you weave those stories into your text? How can you surprise your readers and make them feel emotion? By applying the SUCCESS model, you will show your readers and potential clients that you are human and connect with them in a way that no machine ever could.