Signal Interference — is AI eradicating human identity?
By generating content, in full or partially, from AI, are we drowning out humanity’s voice as the feedback loop of foundational models and training data gets muddied with more AI material than human material?
--
Generative AI, for the most part, works by predicting the next best [word, pixel, frequency, insert data point here] based on two primary sources. First, the accumulated history of digital human creativity available to the AI model. Second, the input of human trainers. In other words, if you prompt it to create an image of post-apocalyptic kittens swimming in a pool of neon whipped cream, it is going to reference all the data that aligns with those terms and start sticking together pixels from prior examples of that. At this point, foundational models are looking at pixel level, but they are also looking at feature level. For a kitten feature, how many average eyes should it have and what is the average spacing between those eyes? How many appendages does a kitten have? How long, relatively, is a kitten? How many teeth does it have, how large relative to other features, and where on the body do they exist? And so on. There is parametric data that represent all of this in any model. It only knows this because some human drew, painted or published a photo of it, and a subsequent human validated that (as long as the machine learning engineers took that step).
Eventually, the model reaches an acceptable level of performance and it is set free, unleashed upon all of us. That level of performance is not 100%. It is some small distance below 100%.
Variation is introduced. Mutations. Like solar radiation piercing our skin, editing our genes, introducing mutations that persist or fade, AI models are introducing small variations that are accepted into the body of published work. And today, there is no reliable indicator for the origin of that work.
The AI generated material is snuggled right up against human work in the data set, the small spoon to humanity’s big spoon. The foundational model is starting to reference both human work and AI work to generate its next prediction.
How long will it take for the body of creative work to be entirely suffused with AI generated content?
Looking at photography only, it took 149 years for humans to generate 15 billion photos. It took 1.5 years for humans to generate 15 billion photos with AI. In only 1.5 years, the pie became half AI and half human.
Using that same rate of growth (although the rate of AI generated content will likely only continue to grow), content influenced (or entirely created) by AI will begin to edge out human content very quickly.
The rate for text, video, and audio creation will mirror this or be very similar. The further forward in time we go, the blurrier the distinction becomes.
How important is it that test data sets are certified pure human? How important is it that models are re-calibrated against pure human data sets and what is the threshold of variance at which we re-calibrate?
There is plenty to be concerned about beyond the immediate concerns plagiarism and inaccuracies.
The AI content is introducing interference and infidelity into the signal of humanity. Today, AI is referencing data that has 100% human fingerprints on it. But tomorrow, and probably literally tomorrow, AI is using data that has a very large AI fingerprint (can I even use that term with AI) in it. And that fingerprint will only grow and become more dominant.
Generative AI is no longer referencing human data, it is referencing AI data.
I am concerned about how this feedback loop will change our voice. When we generate content with AI, accept it, and publish it, are we altering the content entirely to be in our voice and our vision? Or are we publishing content, whether it is text, video, audio, any format, because we “like” it whether it represents our identity or not?
There is a big difference between if we like something and if it is our voice.
How much of the AI’s identity is becoming our identity? How complicit and willful are we in this transition?
Our intrinsic, emotionally motivated desire to act and create is uniquely human and individualistic. There is something pure and honest in the human act of creating. We are lowering the fidelity of the signals AI is sending into the world.
The human signal is getting drowned out, slowly, but inexorably.
Is it a bad thing?
