The Amplitude Problem

TAM-076 · The Approximate Mind

Priya keeps a small cactus on her desk that she has not watered in three weeks. It seems fine. She cannot tell whether this means cacti are resilient or whether this particular one is dying in a way she has not learned to read. She has the same relationship with the forty-seven papers open in browser tabs on her laptop, each one about maternal health interventions in districts like hers.

Priya is a district health officer in a state that appears frequently in development reports and rarely in the experience of the people writing them. Her job, on paper, is to allocate limited resources to programs that work. Her job, in practice, is to figure out what “works” means when everyone with a keyboard and a model can produce a document that says their thing works.

She is not struggling with misinformation. That would be simpler. Misinformation has a shape you can learn to recognize: missing sources, implausible claims, obvious bias. What fills Priya’s screen is worse than misinformation. It is information. Properly sourced, internally consistent, published in places that used to mean something. Forty-seven papers, and she cannot find the signal.

This is the amplitude problem.

What the Kitchen Knows

#

There are approximately 1.2 million recipes for chicken biryani on the English-language internet. Nobody finds this particularly distressing.

The reason is not that biryani is simple. A good biryani is genuinely difficult, and the difference between a great recipe and a mediocre one matters to anyone who has tasted both. The reason nobody panics about a million biryani recipes is that the filtering model is built into the domain itself. You cook it. You taste it. You know.

Popularity helps. Reviews help. Your aunt’s opinion helps. None of these filters are perfect, but they do not need to be perfect. They need to be functional. The feedback loop between a recipe and its consequences is short enough that you can navigate abundance without drowning in it.

This is what filtering looks like when it works: imperfect mechanisms that are adequate to the domain. The volume is high but the signal is recoverable because reality provides feedback on a human timescale.

Most of what we learned about managing information abundance, we learned in domains like this. Search engines rank by relevance and popularity. Review systems aggregate preferences. Social networks surface what your circle values. These are all biryani-tier solutions. They assume that somewhere in the chain, someone touches reality and reports back.

The assumption held for a long time. It is breaking now, and the break is not where most people are looking for it.

The Old Filter Was Effort

#

Before AI writing tools, the global corpus of serious-looking research on any given topic was constrained by something that had nothing to do with truth and everything to do with production cost. Writing a paper was hard. Getting it reviewed was slow. Publishing it required navigating institutional gatekeepers who were inconsistent, biased, and sometimes corrupt, but who nonetheless kept the volume at a level where a thoughtful reader could survey the field.

The friction was the filter.

Not a good filter. Not a fair one. Entire perspectives were excluded because the people who held them lacked institutional access. The old system’s gatekeeping was a legitimate target of criticism for decades, and much of that criticism was earned.

But the friction did something that was easy to overlook while it was operating and impossible to ignore now that it is gone: it kept the ratio of signal to noise within the range of human cognitive capacity. A researcher could read the major papers in a subfield. A policy maker could survey the evidence on a question. A district health officer could, with effort, distinguish the interventions that had been tested from the ones that had merely been described.

That world is over. Not because AI introduced falsehood into the research ecosystem, but because AI removed the production cost that kept volume proportional to effort. The floor rose. Every paper now looks competent. Every abstract is well-structured. Every citation list is plausible. The markers that used to correlate, however imperfectly, with someone having actually done the work no longer correlate with anything at all.

Priya’s forty-seven tabs are not a personal failing. They are a structural condition. The tools she was trained to use for evaluating evidence, source credibility, methodological rigor, institutional reputation, all of these assume a world where production cost serves as a first-pass filter. In a world where production cost approaches zero, she is navigating with instruments calibrated for a different atmosphere.

Three Domains, Three Failures

#

The amplitude problem does not break the same way everywhere. It breaks differently depending on how far the domain sits from direct human experience, and understanding this unevenness is the first step toward knowing what to do about it.

The kitchen tier. Domains where you can taste the result. Cooking, fitness routines, basic home repair, language learning. The feedback loop is personal, physical, and fast. A million AI-generated recipes are annoying but survivable because your tongue still works. Popularity-based filtering remains imperfect but functional. The old internet was built for this tier, and it still mostly serves it.

The clinic tier. Domains where reality provides feedback, but slowly, distantly, and through institutions. Healthcare interventions, educational policy, development programs, urban planning. Someone, somewhere, can eventually determine whether the intervention worked. But the distance between a paper’s claims and their real-world consequences is measured in years, thousands of miles, and layers of institutional interpretation. This is Priya’s tier. AI did not create bad research here. It made the volume of adequate-looking research exceed the capacity of any individual, or any reasonable team, to evaluate. The old filter, effort as proxy for seriousness, is gone. Nothing has replaced it.

The abstract tier. Domains where there is no ground truth to touch. Theoretical physics, pure mathematics, formal philosophy, parts of economics and social theory. The old filter here was never effort alone. It was comprehension scarcity. Only a small community of people on earth could evaluate a paper on quantum chromodynamics or higher-dimensional topology. That community’s smallness was the membrane. Not effort, not popularity, not institutional prestige, but the simple fact that the pool of qualified evaluators was tiny enough to function as a quality filter.

AI breaks this tier differently than it breaks the others. It does not fool the experts. It buries them. When the volume of plausible-looking theoretical work exceeds the expert community’s capacity to evaluate it, the membrane does not get penetrated. It gets overwhelmed. The signal is still there, somewhere, but the people who could identify it no longer have enough hours in their careers to find it.

What Amplitude Actually Means

#

The word matters. This is not a volume problem, though volume is part of it. Volume is about quantity: there is more of everything. Amplitude is about intensity: each individual piece is louder than it used to be.

AI amplifies the signal and the noise with equal fidelity. Your genuine insight and your motivated reasoning both get the same quality of expression. A careful observation drawn from years of clinical work and a speculative framework assembled from pattern-matching both emerge from the tool looking equally polished, equally cited, equally authoritative.

The system cannot distinguish between what you know and what you merely believe. It treats both as input and produces output at the same quality level. This is not a bug in the technology. It is a faithful reflection of what the technology does: it models language, not truth.

In the old world, the difference between knowledge and belief was partially, imperfectly legible in the effort required to express each one. A person who had actually done the fieldwork could write about it with a specificity that a person theorizing from a distance could not easily match. The writing itself carried traces of contact with reality. Those traces were never reliable enough to serve as proof, but they were often enough to serve as signal.

AI erases the traces. Not by fabricating them, but by making the surface quality of all writing converge. When everything reads like it was written by someone who knows what they are talking about, the reader’s ability to distinguish expertise from fluency collapses.

This is the real amplitude problem. Not that there is more noise, but that the noise and the signal have become the same volume.

The Honest Inventory

#

Before reaching for solutions, it is worth being honest about what we have lost and what we have not.

We have not lost truth. The papers that describe interventions that actually work still exist. The theorems that are actually proven are still proven. The district where a specific maternal health program reduced mortality by a specific percentage: that district is still there, and the data is still real.

What we have lost is the set of ambient, imperfect, socially constructed mechanisms by which a thoughtful person could find truth without already knowing it. Institutional reputation, publication venue, citation patterns, writing quality, methodological signaling: these were never truth-detectors. They were heuristics. And they worked well enough, in a world where production cost kept volume manageable, that we could operate as though they were truth-detectors without too much damage.

The heuristics have not been disproven. They have been rendered inoperative by a change in the production environment. It is as though a city’s entire wayfinding system, street signs, landmarks, local knowledge, was designed for a town of fifty thousand people, and overnight the population became five million. The signs are not wrong. They are just no longer sufficient for the navigation task.

I wonder whether we have ever faced a transition quite like this: not the introduction of false information, but the collapse of the conditions under which true information could be recognized.

Toward a New Noise Cancellation

#

In audio engineering, noise cancellation works by generating a signal that is the inverse of the noise, so that the two cancel each other out. It requires knowing what the noise sounds like. This is the central difficulty of the amplitude problem: in most domains that matter, we cannot identify the noise without already knowing the truth. If we knew the truth, we would not need the filter.

So the new noise cancellation cannot work like the old one. It cannot be a single mechanism applied uniformly. It has to be domain-aware, sensitive to the specific way each tier of knowledge breaks under amplification.

In the kitchen tier, the existing filters mostly hold, but they need reinforcement. Provenance tagging, marking content with its generation method, helps a reader know whether a recipe was developed through testing or assembled through prediction. This is the easiest tier to address because the feedback loop does most of the work. The real risk here is not that people will be deceived but that they will be exhausted by volume. The intervention is curation, not filtration.

In the clinic tier, the most promising direction is what might be called chain-of-contact verification. How many steps between this claim and someone who stood in the room where the thing happened? AI can make any paper sound proximate to lived experience. It cannot fabricate the chain of institutional, geographic, and temporal contact that connects a claim to an observation. The new filter in this tier is not “is this well-written” or “is this well-cited” but “how close to the ground is this, and can I trace the path?”

This is not easy. It requires infrastructure that does not yet exist: registries of fieldwork, tagged datasets, provenance metadata for observations as well as publications. But the shape of the solution is at least visible.

There is also a different relationship with the amplifying tool itself. Priya does not need AI to find her more papers. She needs AI that will stress-test the ones she has. Not “help me summarize this research” but “show me where this paper’s reasoning is weakest, where its evidence is thinnest, where its conclusions outrun its data.” Adversarial AI, not assistive AI. The same tool that made everything louder, turned around and used to interrogate its own output.

In the abstract tier, honesty requires admitting that the solution is least clear. When the knowledge has no ground truth to anchor it and the filtering mechanism was always the scarcity of qualified evaluators, increasing volume without increasing the evaluator pool creates a problem that no tagging or provenance system can address.

The best available direction, and it is genuinely uncertain, is AI as pre-filter for expert communities. Not replacing expert judgment but performing triage: identifying which of the thousand new papers in a subfield contain genuinely novel arguments versus which are recombinations of existing ones, flagging logical gaps, checking proofs mechanically where mechanical checking is possible. This preserves the human membrane while giving it a chance to function at higher volume.

But there will be a gap. There will be a period, and we may already be in it, where the abstract tier operates without adequate filtering. Where plausible-sounding theoretical work circulates and is cited and builds reputation before the expert community can evaluate it. This is not a problem to be solved by cleverness. It is a condition to be named and navigated with humility.

What She Needs

#

Priya closes twelve of her forty-seven tabs. Not because she has evaluated them, but because she has been staring at her screen for two hours and her eyes are tired. She will pick different ones tomorrow, or the same ones, or new ones that appeared overnight. The workflow feels like bailing water from a boat that is filling faster than she can empty it.

What she needs is not a better search engine. Not a smarter ranking algorithm. Not an AI assistant that summarizes papers for her, because a summary of noise is still noise, just shorter.

What she needs is a way to ask a different question of the material in front of her. Not “which of these is right” but “which of these was here?” Which one carries the traces of someone who stood in a clinic like hers, counted patients like hers, watched what happened over months and years in conditions she would recognize? The answer might still be wrong. Proximity to reality is not proof. But it is a filter that works in her tier of the problem, and right now she does not have it.

She picks up the small plastic watering can she keeps behind her monitor and gives the cactus a little water. Not much. She is not sure how much it needs, and overwatering, she has read, kills more cacti than neglect.

Somewhere in her tabs, there is probably a paper about that too.

This is Part 76 of The Approximate Mind, a series exploring how artificial intelligence reshapes what it means to think, create, work, and live. Part 26 examined the promise of democratized cognition; this essay asks what happens when the same tools that democratize cognitive power also democratize cognitive amplification. Part 50 explored the monoculture that emerges when AI-mediated curation narrows economic diversity; here the convergence is epistemic rather than economic. Part 47’s three delegations assumed that what we delegate retains its quality; this essay argues that cognitive delegation without noise filtering is delegation without quality control. Part 74 proposed the interrogator, an AI that questions objective functions; the adversarial AI described here is a narrower version of that idea, applied to the specific problem of evaluating evidence rather than evaluating purpose. The companion essay, Part 77, examines what happens when the amplitude problem is not accidental but engineered.

References

#

Information Overload and Epistemic Quality

Benkler, Yochai. The Wealth of Networks: How Social Production Transforms Markets and Freedom. Yale University Press, 2006.

Blair, Ann M. Too Much to Know: Managing Scholarly Information Before the Modern Age. Yale University Press, 2010.

Nichols, Tom. The Death of Expertise: The Campaign Against Established Knowledge and Why It Matters. Oxford University Press, 2017.

Filtering, Gatekeeping, and the Production of Knowledge

Collins, Harry, and Robert Evans. Rethinking Expertise. University of Chicago Press, 2007.

Merton, Robert K. The Sociology of Science: Theoretical and Empirical Investigations. University of Chicago Press, 1973.

Ziman, John. Real Science: What It Is, and What It Means. Cambridge University Press, 2000.

AI, Synthesis, and Trust

Floridi, Luciano. The Ethics of Artificial Intelligence: Principles, Challenges, and Opportunities. Oxford University Press, 2023.

Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021.

Global Health Evidence and Decision-Making

Chambers, Robert. Whose Reality Counts? Putting the First Last. Intermediate Technology Publications, 1997.

Greenhalgh, Trisha. How to Read a Paper: The Basics of Evidence-Based Medicine and Healthcare. 6th ed., Wiley-Blackwell, 2019.