Signal vs. Noise: A Signal-Processing Engineer’s Guide to Information Overload at Work
For a few years, before I ever looked at a brain scan, my job was to remove noise from sound. I worked on speech-enhancement algorithms for hearing aids at Fraunhofer — the unglamorous problem of taking a microphone signal that contains a voice buried in traffic, wind, and the hum of a café, and pulling the voice back out cleanly enough to understand. It is harder than it sounds, and I spent a long time inside it.
I think about that work constantly now, because your inbox is the same problem. So is your calendar, your Slack, your feed, the twelve tabs open behind this one. The sensation people call “information overload” is not really about volume. It is a low signal-to-noise ratio: the thing you need is still in there, drowned in everything that arrived alongside it. And separating the two is a skill — one with a formal, hundred-year-old theory behind it — not a personality trait you either have or lack.
This post is that engineer’s guide to the separation: one old idea, one habit from my current lab, and a handful of filters you can apply this afternoon. Underneath all of it sits a single method — the way of thinking I describe in how to think like a scientist. Filtering information is, at bottom, deciding what deserves your confidence; that is a scientific act, not an administrative one.
The concept that reframes everything: signal-to-noise ratio
Engineers do not talk about “too much information.” We talk about signal-to-noise ratio — the strength of what you want relative to the strength of everything you don’t. It is a ratio, which means there are always two levers, not one. You can raise the signal, or you can lower the noise. Most people, drowning at work, only ever try to process faster. That is neither lever. That is swimming harder in the same water.
The reframe matters because it changes the question. “How do I get through all of this?” has no good answer; the volume is effectively infinite and rising. “What is the signal here, and what is merely loud?” has an answer nearly every time. Your job is not to be faster at reading. It is to be better at attenuating.
EVIDENCE GRADE: STRONG
I grade the framework Strong: it is not a metaphor I am stretching but literally the mathematics your phone uses to hear you in a noisy room. Ratio problems are attacked from both sides. Everything after this is application.
Shannon, 1948: why more information can mean less meaning
The formal basis for all of this arrived in a single paper. In 1948, Claude Shannon published “A Mathematical Theory of Communication” and, in doing so, invented information theory whole. His central move was to define information precisely — and the definition is counterintuitive in a way that is directly useful to you.
For Shannon, the *information content* of a message is tied to how much it reduces your uncertainty. A message that tells you something you already knew, or that you could have predicted, carries almost no information, however many words it contains. A message that genuinely changes what you expected carries a lot. Length and information content are different quantities. This is the whole game.
Hold your inbox up to that definition and it sorts itself with unusual clarity. The “reply-all: thanks!” carries no information — you could have predicted it. The status update identical to last week’s carries none. The one-line message that a launch date moved carries an enormous amount, because it changes your plans. *Reduction in your uncertainty* is the signal, not length. Once you have that ruler, most of what floods you measures as near-zero and can be treated accordingly.
EVIDENCE GRADE: STRONG
Shannon’s theory itself is as settled as science gets. What I grade more carefully, below, is the leap from a theorem about communication channels to advice about your Tuesday. The theorem is Strong; the analogy is a tool, and I will tell you where it bends.
The preprocessing pipeline: what my current work does before it “thinks”
There is a step in my research that almost nobody outside imaging appreciates, and it is the most important there is. Before any brain scan becomes a result, it goes through preprocessing — a pipeline whose entire purpose is to remove noise so the analysis is even possible. We correct for the patient’s head moving, suppress artefacts from breathing and heartbeat, and separate the biological signal we care about from the machine and physiological noise riding on top of it.
The order is the lesson. We do not analyse first and clean up later. We *clean first, then reason* — because reasoning applied to noisy data does not produce a slightly noisy conclusion; it produces a confident wrong one. Garbage in does not give you garbage out. It gives you a beautiful, plausible result about nothing.
Your workday has no preprocessing stage, and that is the core of the overload problem. Information arrives and goes straight into “reasoning” — you read the anxious email and start planning around it before asking whether it was signal at all. The fix is not to think harder about your inputs but to build a cleaning stage in front of your thinking, so your finite mental effort is spent only on data worth reasoning about.
FIELD NOTE — ERLANGEN
The most humbling thing preprocessing teaches you is how convincingly noise imitates signal. Early in my imaging work I watched a promising pattern survive several checks — until we traced it to the fact that the patients, being in more discomfort, moved fractionally more in the scanner than the controls. The “brain difference” was a motion difference wearing the costume of a finding. Nobody had lied and no one had been careless; the noise was simply well shaped enough to pass for the thing we hoped to see. I now assume, by default, that a compelling signal is confounded until it has survived a deliberate attempt to explain it away as noise. That assumption costs a little enthusiasm and saves an enormous amount of wasted work.
AI is now a noise source — a fluent one
There is a new entry in the noise budget, and it deserves naming. Generative AI has made it nearly free to produce large volumes of fluent, plausible text. Some of that raises your signal. A great deal is noise with excellent production values — the summary that is confidently wrong, the ten-paragraph answer to a one-line question, the report padded to look substantial.
This is a genuinely new problem because our instinct for filtering has always leaned on surface cues. We treated fluent, confident, well-formatted writing as a proxy for care and correctness, because historically producing it took effort. That proxy is now broken. Fluency has become cheap, which means it has stopped carrying information about quality. I have written separately about why this raises rather than lowers the value of human judgment, in how AI makes critical thinkers more valuable — and filtering is exactly where that plays out day to day. Asking “is this true, or merely well-phrased?” is the same skill whether the fluent noise came from a colleague, a feed, or a model. AI has simply turned up its volume.
The filters: attenuating noise on purpose
Theory is only worth as much as what it changes on Monday. Here are the filters I actually run, each one an instance of a single move — lower the noise rather than process faster.
The uncertainty filter (email and messages). Before engaging with any message, ask Shannon’s question: *does this change what I expected or intend to do?* If it does not, it is near-zero information regardless of length or tone, and it goes to archive, a batch, or a delegate — not into your active attention. Most of an inbox fails this test.
The bandwidth filter (meetings). Attention is a channel with a hard capacity limit; the evidence that we cannot genuinely parallelise focused cognitive work is robust — what feels like multitasking is rapid switching, and it degrades both speed and accuracy. Treat a meeting as a channel: if it is not transmitting information you need and could be received more efficiently in writing, it is noise occupying scarce bandwidth. Declining it is a filter, not a discourtesy.
The source filter (feeds and reading). Instead of filtering item by item at the point of consumption — exhausting, and too late — filter at the source. Curate the small number of inputs with a consistently high signal-to-noise ratio and cut the rest, the way we choose acquisition settings before a scan rather than rescue bad data afterward. The best noise reduction happens before the signal ever reaches you.
The confound filter (before acting). When something does read as signal, pause exactly where my lab pauses: *what else could produce this, and does my confidence match my evidence?* One alarming data point is usually noise until a denominator and a baseline say otherwise.
| Noise source | The loud, wrong move | The filter (lower the noise) |
|---|---|---|
| Overflowing inbox | Read and reply faster | Uncertainty filter: does it change my plan? |
| Back-to-back meetings | Attend all, multitask through | Bandwidth filter: could this be writing? |
| Endless feeds/reports | Skim everything, item by item | Source filter: curate inputs, cut the rest |
| Fluent AI/colleague output | Trust the polish | Confound filter: true, or just well-phrased? |
Try this today
Open your inbox and, for the next twenty messages, do not reply to any of them. Instead, sort each one with a single question: *did this reduce my uncertainty about something I need to act on?* Yes goes to a short “signal” list; everything else goes to archive or a “no rush” batch, untouched. Do not process the noise — just separate it. When you are done, work only from the signal list. You have just run a preprocessing pass on your own attention, and you will notice that the real work was always a much smaller pile than the volume suggested.
Where the analogy bends — an honest grade
I owe you the limits, because a filter you trust too much is its own noise. The mapping from information theory to your working life is a productive analogy, not a proven equivalence, and I grade it honestly.
EVIDENCE GRADE: MODERATE
Shannon’s mathematics is Strong; its *application* to attention and workplace information is Moderate — well-motivated, consistent with what we know about attention as a capacity-limited resource, but not a quantified law you can compute your Tuesday from. The genuine risk is over-filtering: attenuate too aggressively and you discard weak early signals that mattered, because real signal sometimes arrives quiet before it arrives loud. Noise cancellation has exactly this failure mode — push it too hard and the voice distorts along with the hiss. So calibrate: filter confidently where the cost of a miss is low, keep the gate wider where a missed faint signal would be expensive. No ratio computes that for you.
The signal was always smaller than the noise
The deepest thing my two careers agree on is this: in almost every real system, the signal is a small fraction of the total, and the work is protecting that fraction from everything louder around it. A brain scan is mostly not the finding. An inbox is mostly not the message. A day is mostly not the two decisions that will matter. Overload is what it feels like to have no stage that separates the two — to treat all incoming volume as signal, and then blame yourself for not keeping up.
You do not need to keep up. You need to filter — a skill with real theory behind it and real technique in front of it, and one of a small set of durable thinking tools that outlast whatever software is throwing information at you this year. If separating signal from noise this way is useful, its natural companions are laid out in the five mental models every young professional needs first — the handful I would build before any others, each a way of protecting good judgment from loud, plausible noise.
If this way of thinking is useful to you, the natural next step is the free guide this site is built around: [5 Mental Models to Future-Proof Your Career](/newsletter/) — five models chosen and stress-tested by a brain researcher, each with an honest grade of the evidence behind it. You’ll also get *Signal*, my monthly email: one idea from neuroscience you can use at work. No productivity spam, no AI panic.
*Mageshwar Selvakumar is a doctoral researcher in neuroscience in Erlangen, Germany, studying how chronic pain reshapes the brain using multi-parametric MRI.*