<\/p>\n
For as good as generative AI has gotten, simple visuals are what it recreates best. \u201cIt\u2019s not that it struggles with tattoos, per se,\u201d Farid says. \u201cIt\u2019s that it struggles with details.\u201d In other words, a wardrobe full of Dan Flashes shirts would arguably keep you just as protected as a tattoo sleeve.<\/p>\n
There are also fewer people with visible ink than without\u2014only 22% of people in the U.S. have more than one tattoo. That means the model training data set of people with tattoos is much smaller than the set of people without them, says Saetbyeol Leeyouk, a critical technologist and graduate student at MIT Media Lab whose work focuses on the intersection of ethics and AI.<\/p>\n
But both Farid and Leeyouk say there is a big caveat here: We are talking about the capabilities these programs have right now in August 2026. If you\u2019re going to say \u201ctattoos make it harder for generative AI to recreate your likeness,\u201d you have to end that sentence with \u201ctoday,\u201d Farid says. \u201cTomorrow, next week, next month, it\u2019s going to get better. We know this.\u201d Leeyouk agrees, saying \u201cEventually, [the tattoo issue] will be solved.\u201d<\/p>\n
For proof of this acceleration, Farid invites us all to remember the 2023 Will Smith eating spaghetti video, a nightmarish montage of vaguely Will Smith-shaped figures shoveling handfuls of pasta-like blobs into what could potentially pass as a human mouth if you were unfamiliar with the species. Less than three years later, a Reddit user generated an updated version that, as long as you aren\u2019t looking too closely for a human soul behind the eyes, could definitely be a spaghetti commercial starring Will Smith.<\/p>\n
These AI-generated videos are getting better because the programs are getting better at temporal coherence, which is a visual consistency between frames. \u201cTemporal coherence has come from the fact that we\u2019re not generating frame after frame after frame [to create a video,]\u201d Farid explains. Rather than generating videos one still image at a time\u2014Will Smith eating spaghetti is Eadweard Muybridge\u2019s horse for the internet age\u2014these newer programs can generate a short video clip all at once.<\/p>\n
There are limits on how long those clips can be\u2014Farid says about 10 to 15 seconds\u2014which is why, for right now, your tattoos probably would get messed up in any longer generated video, which is just a bunch of short clips back to back. \u201cIf you have a fair number of [tattoos] and I made a video of you, they\u2019d probably move a little bit from clip to clip,\u201d he says.<\/p>\n
But how much will that little bit of movement actually matter? Leeyouk points out that, particularly in the case of non-consensual sexual deepfakes (an area of research focus for her), \u201cthe purpose of that is not just the itself, but to humiliate the victim,\u201d she says. Whether they got the shading on your neotraditional snake exactly right probably won\u2019t matter much to you if your likeness is being stolen for a nefarious purpose.<\/p>\n<\/div>\n