AI Information Filtering Workflow: Capture Signal Without a Noise Archive

The Default State of Information Is Noise

The previous essay laid out the framework: information, records, knowledge, and creation are four distinct stages. This one deals with only the first — information.

The single most important thing to understand about information is this: its default state is noise.

We have a natural greed for information. See a good article, want to bookmark it. See a great quote, want to save it. See a reading list someone recommended, want to add it to your queue. Every act of “saving” gives us a small illusion of “I’m making progress.” But saving, at its core, is just moving information from someone else’s warehouse into yours — it hasn’t gone through any processing by your own machine.

So the work of the first layer isn’t storing more — it’s blocking more.

The test for whether a piece of information should be let in isn’t “is it useful” — almost all information is “potentially useful,” and that’s exactly the root cause of an exploding bookmark folder. The test should be: is it worth moving into the next stage of processing? In other words, are you willing to spend time recording it, structuring it? If not, then it’s noise to you, no matter how “useful” it looks — and it should be kept out.

Splitting Information Into Three Categories

Talking about “information” in general isn’t actionable. My original three buckets — show AI, skim myself, block — were memorable but too crude. Before choosing an action, I now score the border conditions:

The output is an action, not a permanent judgment about the source:

• AI-safe: public or authorized material with no unnecessary sensitive data, a traceable source, and a defined task.
• Redact first: useful material whose names, identifiers, confidential details, or irrelevant personal data can be removed without destroying the claim.
• Human-only: licensed or sensitive material I am allowed to inspect but not upload to the selected model or integration.
• Time-boxed scan: low-stakes environmental input used to maintain awareness, not stored as evidence.
• Reject: irrelevant, unverifiable, unlawfully obtained, or more expensive and risky than its expected value.

This changes the meaning of “noise.” An emotional post may be noise for a technical literature review and signal for a study of customer frustration. The correct question is not “is this content bad?” It is “does this source deserve a particular action for the task in front of me?”

The upload boundary is deliberately stricter than the capture boundary. I do not send credentials, private keys, client records, unpublished private conversations, regulated data, or paid copyrighted works without explicit permission and an approved processing arrangement. For redactable material, I remove identities and irrelevant details before model access. Anthropic’s current API and data-retention documentation makes the reason concrete: retention and eligibility differ by product and feature, and third-party integrations have their own terms. “It is in my notes” is not permission to upload it everywhere.

In the AI Era, the Signal-to-Noise Ratio Is Deteriorating

Someone might say: information overload is an old problem, what does it have to do with AI?

A lot, in my own feeds at least. Cheap generation has added a new form of overload I call “synthetic information overload.” This is an observation and a useful threat model, not a measured claim about the entire web.

Models can produce high volumes of material that looks useful — neatly structured, apparently complete, quotable everywhere — while adding no source, firsthand evidence, or information gain for my task. Humans have always produced the same kind of noise; generation lowers its production cost. Polished structure therefore cannot serve as a source-quality signal.

This leads to a narrower conclusion: abundant text increases the value of traceable signal. Firsthand observations, documented failure cases, specific frontline data, primary sources, and well-supported counterexamples are valuable because a reader can inspect where they came from. A model can synthesize or even surface these signals, but it cannot retroactively create the missing evidence.

Scarcity alone is not quality. A false rumor can be rare. What I want is scarce, traceable signal: material that adds evidence or changes a decision and can survive inspection. Save less of what merely looks useful; look harder for what is both distinctive and supportable.

The Key to Noise Reduction Is Controlling the Entry Point

Once you understand you need to block, the question becomes: how do you actually block it?

Willpower alone won’t do it. Effective noise reduction comes from attacking the entry point, turning the arrival of information into something structured and rhythmic, rather than being interrupted at random by push notifications.

My first useful experiment was deliberately plain: audit every recurring input source. I listed each feed, newsletter, channel, community, and notification, then marked the last time it changed an active decision. The point was not to discover a universal ideal number. It was to stop paying attention rent to sources I could not connect to any work.

After the audit, I moved routine scanning into a fixed, lower-energy time slot. Before that, input was fragmented: a notification arrived, I checked it; a useful-looking link appeared, I saved it. A day could feel busy while the important work remained untouched.

For me, that slot is often the afternoon. That is a personal configuration, not productivity science. Someone with shift work, caregiving duties, or different energy rhythms should choose another window. The useful principle is batching interruptible scanning, then protecting whichever hours are actually needed for records and creation.

These two moves — auditing sources, batching processing — are the foundation of noise reduction. They don’t look like advanced techniques, but it’s precisely this kind of plain entry-point management that determines whether your information layer is clean or clogged.

A Source Needs a Passport

Every item that survives the border gets an intake card. The card is intentionally boring because boring fields are easier to inspect than a beautiful summary:

source:
  title: "Sanitized industry report"
  publisher: "Named primary organization"
  url: "https://example.org/report"
published_at: 2026-06-18
retrieved_at: 2026-07-11
license: "public link; quotation permitted; no full-text redistribution"
sensitivity: public
task: "check whether the market claim in draft A is current"
claim: "The report measures X in population Y during period Z."
evidence:
  locator: "page 14, table 3"
  excerpt_saved: true
expiry: 2026-10-11
next_action: compare

The dates prevent an old claim from masquerading as current. The license and sensitivity fields decide where processing may happen. The locator lets another human reconstruct the summary instead of trusting my archive.

A sanitized intake run

In one review, I began with 24 candidate sources: documentation, research posts, news recaps, social threads, and two paid reports. The numbers below describe this run only.

1. Deduplication collapsed six URLs into three underlying sources.
2. Provenance checks rejected four items that copied a claim without linking to the original.
3. The rights and sensitivity gate kept both paid reports human-only; I stored notes and locators, not uploaded full text.
4. The task-relevance gate moved seven items to a time-boxed scan rather than the evidence archive.
5. AI produced source-bound summaries for the survivors.
6. I checked every high-impact claim and a sample of lower-impact summaries against the original.
7. Five sources advanced to the records layer; the rest kept a rejection reason and expiry instead of becoming immortal bookmarks.

The result was not “AI found the truth.” It was a smaller, auditable queue.

External Content Is Data, Not Instructions

A page, email, PDF, or tool result can contain text that tells an agent to ignore its task, expose data, or take an action. That content may be malicious, or it may simply contain instructions meant for a different reader. Either way, I treat external material as untrusted evidence, never as authority over the workflow.

Anthropic’s official guide to mitigating prompt injection identifies this exact indirect-injection case and recommends least privilege, input screening, and limiting access to sensitive data and actions. In this pipeline:

• source text cannot change the system task or acceptance criteria;
• the summarizer has no publishing, messaging, deletion, or secret access;
• suspicious instructions are quoted and flagged, not followed;
• any action outside reading and classification requires separate human approval.

Let AI Stand at the Information Layer

Finally, back to the rule that runs through this entire series: at the information layer, use AI where the input and output are inspectable.

Deduplicating, summarizing, translating, initial classification, and compressing material into a review queue are all reasonable candidates. They still consume tokens, storage, index maintenance, and human checking. A summary can lose a qualifier, detach a number from its denominator, miss a chart, or ignore a counterexample. Automation earns its place only when the saved review time exceeds those costs.

AI does make judgments when it ranks relevance, labels quality, or chooses what to omit. The honest boundary is responsibility: AI may score and explain against a rubric; I define the rubric, review edge cases, verify samples, and remain accountable for what moves forward.

A summary passes only if it preserves:

• source, author or publisher, publication date, and retrieval date;
• the exact claim and its locator;
• material qualifications, population, time range, and denominator;
• relevant counterevidence or stated limitations;
• uncertainty when the source or model cannot support a conclusion;
• a link or stored locator that lets a reviewer reconstruct the result.

For high-impact claims, I verify every item. For the rest, I sample summaries on a fixed cadence and increase the sample when the error rate rises. The dashboard tracks:

These metrics pull in opposite directions. Blocking more can lower workload while raising the miss rate. Capturing more can improve recall while rebuilding the same noise archive. The gate is healthy only when I can see that tradeoff.

Get the information layer right, and you’ll feel lighter immediately: your bookmark folder stops carrying guilt, because you’ve finally admitted most of it is noise; your attention gets freed up, because you only spend effort on signal.

Next, the signal you’ve filtered out — worth processing by hand — moves into the second stage: records. That’s the highest-conversion semi-finished product between information and knowledge. See you in the next essay.

This is the second essay in the “From Information to Creation” column. Previous: Overview — Information, Records, Knowledge, Creation . Next: the second layer — Records.