Facts as feedstock, noise as feature
a plausible trajectory for generative AIs of the future?
I am not an expert in either AI or the media despite the occasional dip into both fields, for example here on AI and corruption or here on media freedoms. So perhaps slightly unhinged from reality here comes a bit of a weird speculation.
Mirroring quaint speculations during the toddler years of the internet that digital information wants to be free, a similar set of pronouncements seems to have hit the road and the conference circus with regard to AI. The story goes like this: Yes, it is uber-expensive to train frontier AI systems (even if you allow for DeepSeek type of efficiency gains) but sunk costs are at the end of the day, well, sunk and then AI supply gets commodified, becomes ultra-affordable as everyone is trying to blitz-scale their low-variable-cost user bases. So in a sense we are back to “AI wants to be (almost free)” as a principal dynamic and mantra.
Yet, this is where my idea of information positionality comes into play and plausibly spoils it all. As I have developed elsewhere, information goods, no matter if analog, digital or AI-generated have a peculiar utility function. In many applications the value of a piece of information for me depends on others not having access to this information, at least not at the same time or at the same quality. This gives me an edge in acting upon this information be it in markets, in political competition, in research, in the judicial arena etc. Information exclusivity adds additional value or creates this value in the first place. In short: information is a positional good, it’s utility for a user in many contexts depends on how many other users also consume it.
Information service providers have long embraced — consciously or not — this tenet of positional information and introduced a host of mechanisms to create (temporary) exclusivity via insanely high subscription rates for legal databases, market and political intelligence, cutting-edge research journals. The ones who can afford access are willing to pay an exclusivity premium.
So far so good but how does this play out with AI? I suspect in very similar terms: Once the market is maturing consumer-facing generative AI services will segment into two stylized antipodes:
- very expensive, high-end, highly factual and instantaneously up-to-date AI models that give an information edge to a small band of deep-pocketed subscribers; and,
- low-end cheapo models for the masses, embedded in advertising, subtly skewed by paid promotions and geared towards emotion-heavy and attention-grabbing answers either of the amusing or enraging type — and most likely available in different shades of political tribalism for extra identity-affirming creature comfort
So if you want to work with an AI that has been fed on high-end information and loops in real-time updates from leading intelligence providers you will pay dearly for this privilege. Facts are feedstock for these models. Click-through to sources is always included, hallucination rates are carefully monitored and minimized, pros-and-cons in arguments, confidence levels of forward looking statements, all inclusive.
In contrast, if you are an everyday user that appreciates a bit emotional color, tribal traction and a low level of ambiguity you will plug straight into an advertising sponsored apparatus that has ingested all the trashy cheap noise that came before and that can be hovered up across the social media sphere at relatively low costs. These low-end AIs will convert all this chatter into fact-adjacent, under-complex but easily digestible insights, and generate answer blurbs tailored to your type of humor, anger and outrage level. Hallucination will matter less than vibe compatibility. All the noise these systems feed on and generate will be a feature, not a bug as good hallucinations generate better stories than bad reality and noisy answers ensure the low-information utility that drives high-end users away and towards expensive high-end models.
So in sum, the dynamics of positional information turn facts into feedstock and noise into a feature not a bug for generative AIs of the future.
Too unrealistic a scenario? Of course I exaggerate a bit and paint a condescending caricature of the normcore user. But the general dynamics are all there. The higher end of the market is already moving into this direction: the facts-as-feedstock wheel is already spinning: close to 100 partnerships between AI developers and publishers / platforms have already been struck. Open AI has even announced to fund local newsrooms to get its hands on more high-quality local media content. So the variable costs on keeping training data and prompt replies fresh are on the rise. For how much longer these services will be available to the low end of the market in subsidized form is therefore far from clear. Meanwhile affective polarization and political tribalization seem to all but deepen in many countries, creating fertile ground for AI offerings that embellish and market noisier types of generative outputs.
