“Question 4(b)” – Into AI, asking the right questions: The Bandwidth Problem

This is the second piece sitting under Question 4 of this series. The parent argues that making AI useful is a conversation, not a prompt. The earlier piece, 4(a) on configuration, describes how to set up the AI so it knows who you are. This one is about why that setup matters more for some people than for others, and about a problem AI addresses that older tools could not.

A short note before we start. This piece is around 3,300 words and contains some personal autobiographical material in service of the argument. If you only have time for the headline, it is this: the gap between what we can think and what we can output is a real and unevenly distributed cost, and it widens further under emotional load. AI is one of the first tools that addresses the gap directly rather than working around it, including the old advice about not sending emails when angry.

The premise

There is a gap between what most people can think and what most people can output. The gap is variable. It widens under cognitive load (tiredness, parenting, distraction, multiple competing tasks) and it widens further for people whose neurological wiring sits outside the average. The gap exists for everyone. It just shows up more visibly for some than for others.

The gap is not about ability or intelligence. It is about bandwidth. The brain holds more than the keyboard can carry, more than the social register of the situation can accept, more than the tired Tuesday-evening writer can format into something that lands the way it was thought. The internal model is faster, richer, and more responsive than the output channel that has to render it for other people.

For most of human history, the available tools have worked around this gap rather than addressing it. Writing is slow because writing is slow. Speech is constrained by what comes out of your mouth in the moment. Editing helps but takes time. Skilled people get better at narrowing the gap. Less skilled people, or differently wired people, or simply tired people, live with it.

AI is one of the first tools in history that addresses the bandwidth gap directly. Not by making you smarter or more articulate. By translating between what you can produce and what the situation requires, in real time, at a cost that older translation methods (hiring an editor, learning the skill yourself over years of practice, sitting with a draft for two days before you can see what it needed) could not match.

This piece is about that translation work, who benefits most from it, and how to do it without creating the kind of dependency that makes the underlying skills atrophy.

The current language for this

The wider conversation about AI and cognitive load has settled into a recognisable set of terms over the last year. Each one captures part of the picture.

Cognitive prosthetic. The most clinical of the terms. Frames AI as a functional augmentation that compensates for a deficit. Useful when the deficit is real and named, less useful as a general description because most users are not operating from deficit. A neurotypical parent writing a difficult email at midnight is bandwidth-limited but not disabled.

Digital ramp. Borrows from the language of physical accessibility. The ramp does not change what the building contains, but it changes who can get into it. The framing captures something real about AI levelling access to certain kinds of professional output. It is also slightly aspirational: the ramp metaphor assumes the AI is reliable enough to use as a permanent fixture, which the other pieces in this series suggest is not yet the case.

Cognitive Navigation Interface. The most academic of the terms, from a recent paper proposing AI as a cognitive right for neurodivergent people. The framing positions AI as a way of filtering institutional and informational complexity rather than as a content generator. Useful for the specific case of bureaucratic forms, dense documentation, and the kind of mental overhead that comes from interfacing with systems designed without neurodivergent users in mind.

Cognitive debt. The counter-framing. AI reduces cognitive load now, but the skills that would have been developed by doing the work yourself may atrophy. The concern is real. The article will return to it later.

I am going to use bandwidth multiplier in the rest of this piece, because it captures the specific mechanism better than the other terms. AI is not replacing your thinking. It is multiplying your effective output across the limited channel your hands, your time, and your social context provide. The thinking is yours. The output capacity is increased.

The mother writing an email at midnight

The clearest example of the bandwidth problem is not a clinical case. It is something most working parents will recognise.

A mother is on the sofa with a two-year-old asleep on her shoulder. The two-year-old refused to be put down. The mother has a work email she has to send before tomorrow morning. The email is to a customer who is mildly difficult and who needs to be handled carefully. The mother knows exactly what she wants to say. She has been thinking about it for the last hour while the toddler refused to sleep.

What she cannot easily do is render the email in the register the customer needs. Not because she is bad at writing. Because she has one hand on a phone, one arm holding a sleeping child, and a brain that has been processing the day’s other tasks since 5am. The cognitive overhead of “what does this email need to look like to land properly” is higher than the content of the email itself.

This is the bandwidth problem in its most general form. The thinking is there. The output channel is constrained. The cost of formatting is higher than the cost of substance.

The mother who hands her draft to an AI for register translation gets the email sent in three minutes rather than thirty. The AI is not making her smarter or more empathetic. It is doing the formatting work that her current bandwidth cannot support. The substance is hers. The shape is shared between her and the tool.

This is the version of the problem that affects everyone, sometimes. For most people, sometimes is occasional and the impact is small. For some people, sometimes is most days and the impact is significant. The difference is not about ability. It is about how much of life is spent in bandwidth-constrained states.

When the constraint is emotional

The bandwidth-mismatch problem widens under any cognitive load. The most universally recognised version is not dyslexia, ADHD, or parenting under sleep deprivation. It is anger.

There is a piece of advice older than the internet itself: never send an email angry. Save it to drafts. Review it the next day before sending. The advice survives because it works. The next-day version of you is calmer, sees the social register more clearly, and catches the lines that would have caused damage. The advice works because it puts time between the impulse and the output, and time is where the bandwidth gets restored.

The mechanism is the same as the other bandwidth-constraint cases. Anger consumes cognitive bandwidth. The portion of the brain that would normally handle social-register formatting is busy holding the anger. What comes out the keyboard is the unwrapped fact, the sharp accusation, the line that twelve hours from now you will wish you had not written. The substance might be entirely correct. The rendering is wrong for the audience because the bandwidth required to render it well was not available at the moment of composition.

Hurt, anxiety, grief, frustration, and any sufficiently strong emotional state produce the same pattern. The cognitive cost of the emotion is itself a bandwidth tax. The output that survives the tax is the unwrapped version. The next-day-review fix works by waiting for the emotional cost to subside and the bandwidth to return.

This is also where AI as bandwidth multiplier offers something the older advice does not. Sometimes the email genuinely cannot wait. The angry customer needs a response within the hour. The colleague who has just sent you something unreasonable expects a reply before lunch. The school is asking for a decision by end of day. The twelve-hour drafts-folder method requires that you have twelve hours.

When you do not, AI compresses the wait into the time it takes to type the rough version, ask for a review, and read the rewrite. Three minutes instead of twelve hours. The tool is not the next-day version of you. It is a translation layer that approximates what the next-day version would have caught, in the time you actually have.

The principle is the same as the original advice. The latency is shorter. The result, in most cases, is the email you would have sent tomorrow, sent today, without the regret.

My lived experience, part one

I have a two-year-old. I work in email security. I write a blog series. I have a partner doing a diploma. The cognitive overhead of holding all of those in one head, on any given day, is substantial.

The mother at midnight is not a constructed example for this article. It is a description of my own working life, regularly enough that it does not require imagination. The email to the difficult customer, the message to the colleague who needs careful handling, the response to the school about Enya’s swimming lesson, the reply to Gabie about dinner plans, the technical note to a customer about why their integration broke. Each of these is a small piece of writing. Each requires its own register. Each is competing with everything else for the cognitive cycles I have available in the moment.

This is the ADHD-adjacent end of bandwidth strain. I do not have a formal ADHD diagnosis. The pattern is recognisable to anyone who has lived with too many concurrent threads. The mental scaffolding required to switch between contexts and produce appropriate output for each one is itself the limiting factor. By the time I have decided what register the email needs, the toddler has woken up, the dog wants to go out, and the next meeting has started.

AI as bandwidth multiplier addresses this directly. I can think about what the email needs to say. I can describe the situation to the AI in the rough shape it lives in my head. The AI can render it into the register the recipient needs. The thinking remains mine. The formatting is shared.

The result is not that I write better emails. The emails were always going to be the same emails. The result is that I write the same emails in less time, with less cognitive cost, while holding the rest of the day’s load in my head.

My lived experience, part two

There is a second pattern that is less universally recognisable but worth describing because it is where the bandwidth argument matters most.

I have Scotopic Sensitivity Syndrome, also called Irlen Syndrome. It is a specific type of visual dyslexia, distinct from the auditory-processing dyslexia that most public discussion describes. It is not a problem with the eyes. It is a problem with how the brain processes visual information from the eyes, specifically a sensitivity to certain frequencies of white light that produces perceptual distortions in printed text. In my case, the relevant distortion is that I see halos around words. Other people with the syndrome see words shake, fade, blur, or appear to swim on the page. The condition is genetic, lifelong, and not detected by standard optometric assessment. The Irlen Clinics in Australia (and equivalent services overseas) diagnose it through specific perceptual testing and address it where addressable through carefully prescribed tinted filter lenses that block the problematic light frequencies.

The clinical description is necessary because the lay understanding of dyslexia is largely incorrect. Most people think dyslexia means letters appear reversed or scrambled. Some forms involve that. The form I have does something different. The text remains correctly ordered. The visual experience of looking at it is just distorting enough that reading at normal pace requires extra processing work.

The interesting effect for the bandwidth argument is what this produces in practice.

I am an extremely fast skim reader. The visual distortions mean that reading a sentence linearly is more cognitively expensive than reading at speed and filling in the gaps from context. So my brain has adapted: I jump words, skip forwards, fill in what I missed from what I caught, and reconstruct the sentence afterwards. This works because most sentences in most contexts are partially predictable. The brain becomes good at completing patterns with incomplete data.

The same pattern applies in reverse when I write. My fingers produce text faster than my eyes can verify it. The skim-reading pattern I use for input gets applied to my own output before the output is complete. The result is that what I type, in real time, is often partial gibberish. Words missing letters, words in the wrong order, sentences that trail off, paragraphs that fragment.

The keyboard is the bit that makes any of this manageable. Touch typing means my fingers can produce text faster than I can write longhand. The cursor and the delete key mean I can correct in real time. By the time you read what I have written, multiple passes of mental correction have already happened. You see the corrected version, not the gibberish stage.

This is the second-order bandwidth problem that scotopic dyslexia produces. The thinking is fast. The fingers are fast. The visual processing of my own output is the bottleneck. The mental correction loop runs constantly and is itself cognitively expensive. By the time I have produced a paragraph that is coherent enough to be read, I have done significantly more processing work than a neurotypical writer would have done to produce the same paragraph.

The related effect is what produces the social-register problem I mentioned earlier. Because the writing process is already running an expensive correction loop, the additional cognitive layer of “is this the right register for the reader”is harder to maintain. I reach for facts first because facts are what survive the correction loop. The social wrapping that a neurotypical writer would automatically include gets dropped because there is no spare bandwidth to maintain it.

The neurotypical reading of this output pattern is “blunt” or “lacking empathy.” The assessment is wrong in an interesting way. The empathy is there. The cognitive workflow that produces empathetically-wrapped output is a particular kind of work that my brain prioritises differently than a neurotypical brain does, because the correction loop has already consumed bandwidth a neurotypical writer would not have spent.

This is where AI as bandwidth multiplier does its most valuable work for me. The AI does not give me empathy I lack. It translates the empathy I have, in the post-correction text I produce, into the register the reader expects. The substance of what I think about the other person does not change. The shape of the message changes from “facts first, social context absent because the correction loop ate it” to “social context wrapping the facts.” Same content. Different rendering, performed by a tool that does not have to spend its bandwidth on the same correction loop I do.

The outsourcing problem

Before the demonstration, the honest acknowledgement.

The cognitive-debt framing is real. AI as bandwidth multiplier carries a risk that the skills it multiplies do not get developed. If I never write a difficult email myself, my capacity for writing difficult emails atrophies. The AI becomes the only way I can produce that kind of output. The dependency is not theoretical. It is a documented pattern across users of any cognitive prosthetic.

The argument the rest of this series has been making applies here as strongly as anywhere else. The user who treats AI as a black-box magic button loses the underlying skill over time. The user who understands what the AI is doing, who maintains the cognitive workflow themselves at least some of the time, and who uses the AI selectively for the cases where bandwidth is genuinely the limiting factor, keeps the skill alive.

The distinction I make for myself: I use AI when the cognitive cost of formatting is higher than the value of formatting myself. I do not use AI when the formatting is the work I am supposed to be doing. Writing this article, for example, is the work. AI is helping me organise and revise rather than replacing the thinking. The cognitive load is appropriate for the task. The bandwidth is sufficient because the task is what I have allocated my bandwidth to.

The principle generalises. If you are using AI to skip the work you should be doing to develop a skill, you are accumulating cognitive debt. If you are using AI to handle work that was never going to develop the skill anyway (because the cognitive overhead was higher than your available bandwidth and you would have done it badly or not at all), you are using AI as the bandwidth multiplier it is good at being.

The line between these two cases is not always sharp. It requires honesty about what you are outsourcing and why. The conversation-not-prompt framing applies here too. The user who knows what they are doing keeps the underlying skill. The user who does not is heading for the skill atrophy the cognitive-debt argument warns about.

How I actually use it

This article exists because of AI as bandwidth multiplier. Worth being explicit about that.

Without AI, this series would have been a single article. It would have been less articulate than what you are reading. It would have been a jumbled rant of incoherent messages with the central arguments buried somewhere in the middle and the tangents pulling the structure apart. I have the thoughts. I do not, on any given evening, have the spare cognitive capacity to organise a thousand thoughts into a coherent argument that survives the reader’s attention.

AI as drafting partner does the organisation work. I write or speak the rough shape. The AI renders it into structure. I read the structure, push back where it has gone wrong, refine where it has gone shallow, redirect where it has missed what I meant. The thinking remains mine. The structural work that would have eaten my Tuesday evenings and produced nothing publishable is offloaded to the tool that can do it without depleting me.

The blog series is the largest example. A more concrete demonstration follows in the next section.

The full exchange

For readers who want to see the bandwidth-multiplier work in concrete form, here is the actual email exchange. Names and company-specific identifiers are changed; the technical content and the AI’s analysis are reproduced verbatim.

The original draft I wrote, in response to a technical escalation from a peer at another company:

As mentioned previously, the UK and Amsterdam are in different classes. I understand why that would be the case for the UK, however I do not know why that is the case for NL IP space, all the classes, limits and address space was done before I joined this company (indeed whilst I was still at the last company). Most of European IP space (including Russia) is in the same bucket, NL and the UK are in the class as the USA, I have no visibility as to why an explanation is not in the original change request.

I have a project in progress to review and validate both the entries and limits as an ongoing project, but other priorities have made progress slow.

My colleague is currently on PTO, other people you CC’d have no access, and would not know what needed to be done if they had given the configuration of the rate limiting engine is under my teams’ sole responsibility. So CC’ing others just results in people messaging or calling me to check I’m across the issue and the team are handling, this as you can appreciate then just eats up time available to get to your issue. With regard to when, it should get resolved by CoB today, but there are 2 other tickets with higher priority and my colleague is on PTO so it maybe that it is dealt with Monday. In the meantime you might want to look at not opening multiple concurrent connections as that will stop the ‘Policy-IPCC’ rejection. Regards,

I asked ChatGPT to review the message. The response was direct:

“This needs a rewrite. The technical content is fine, but the tone is too defensive and the paragraph about CC’ing others reads as scolding (even if it’s true). For a third party, you want to stay factual and neutral, and avoid anything that sounds like ‘you’re wasting my time.’ Also a few sentences are very long and could be tightened.”

The rewritten version it produced:

As mentioned previously, the UK and Amsterdam fall into different rate-limiting classes. I understand why this is the case for the UK, however I do not have visibility as to why NL (Netherlands) IP space was placed into its current classification, as these buckets/limits and address space allocations were implemented prior to my joining this company (at the time I was still at the last company).

Most European IP space (including Russia) is currently grouped into the same bucket, whereas NL and the UK are grouped into the same class as the USA. Unfortunately there is no explanatory detail available in the original change request that would clarify the reasoning.

I do have an ongoing project to review and validate the existing entries and rate limits, however progress has been slower than expected due to higher operational priorities.

With regard to your ticket: My colleague is currently on PTO, and the other parties CC’d do not have access to, or responsibility for, the rate limiting configuration (this sits solely with my team). The change should be completed by close of business today, however there are currently two higher priority tickets ahead of it, and with my colleague on leave it may slip to Monday.

In the meantime, to avoid the 4xx Policy-IPCC temporary rejections, please ensure no more than two concurrent connections are opened per source IP.

Regards,

Compare the two. The technical content is identical. Every fact, every constraint, every piece of operational reality. What changed:

The defensive framing was removed. “All the classes, limits and address space was done before I joined this company”became “these buckets/limits and address space allocations were implemented prior to my joining this company.” Same fact. Different emotional weight.

The implicit complaint about the CC list was rewritten as a neutral procedural statement. The original signalled frustration. The rewrite stated routing responsibility without editorial.

The advisory closing became a specific instruction. “You might want to look at not opening multiple concurrent connections” became “please ensure no more than two concurrent connections are opened per source IP.” The first reads as advisory and slightly resigned. The second reads as a clear operational direction with a specific number.

The AI did not improve my technical reasoning. It did not catch any factual errors. It did not change the substance of what I was communicating. It rendered the same substance in a register that matched the audience the message was going to. The work it did is work I could have done myself with another hour of editing time. I did not have another hour.

This is what AI as bandwidth multiplier looks like in practice. The thinking remains mine. The technical understanding remains mine. The bandwidth required to wrap the thinking in the appropriate register was the limiting factor, and the AI handled that part of the job.

What this means in practice

The bandwidth problem is not going away. The cognitive load of modern life, modern work, and modern social communication is high and uneven. The tools that previously helped with the problem (better organisation, more sleep, a competent personal assistant, an editor with time to spare) are either not available to most people or insufficient to close the gap.

AI is one of the first tools in history that addresses the bandwidth problem directly. It is not a replacement for thinking. It is a translation layer between what you can think and what you can output in the time available. The user who understands this distinction gets the benefits without accumulating the cognitive debt the critics correctly warn about. The user who treats AI as a magic button loses the underlying skills and ends up dependent in ways the critics describe accurately.

The neurodivergent angle on this is that the bandwidth gap is wider, more visible, and more consistently present for some people than for others. The principle is general. The lived experience is variable. The article makes the case that AI as bandwidth multiplier is not exotic and not optional for some users in some situations. It is the tool that makes the difference between writing the email at all and not writing it.

For the dyslexic, ADHD, autistic, anxious, depressed, exhausted, parenting, working, or just having-a-bad-week reader: the gap between what you can think and what you can output is real, and you do not have to keep absorbing the entire cost of closing it yourself. The tool exists. The tool is conditional on being used well. The conversation-not-prompt argument from the rest of this series applies.

There is an older version of this advice that pre-dates AI entirely. Never send an email angry. Save it to drafts and review it the next day before sending. The next-day version of you is calmer, sees the social register more clearly, and catches the lines that would have caused damage. The advice works because it puts time between the impulse and the output, and time is where the bandwidth gets restored.

AI as bandwidth multiplier compresses that twelve-hour wait into three minutes. The tool is not the next-day version of you. It is a translation layer that approximates what the next-day version would have caught, in the time you actually have. For situations where the email genuinely cannot wait, this is the difference between sending something you regret and sending something you do not. The principle is the same. The latency is shorter.

Closing

Footnote:

The same pattern applies in other output media. Two of the images in this article (the funnel at the top and the hand selecting a hammer at the close) were generated from long specifications I typed in one pass. I had the visual concepts clearly in my head. It didn’t get it quite right in the first, but it was close enough, the point was: I did not have the hours of image-editing work that manual production would have required. The AI image generator did the rendering work. The concept remained mine. The execution was shared. The bandwidth-multiplier argument generalises beyond text.

For the first, the prompt I gave was:

generate a photorealistic, but also conceptual image of a funnel either white plastic or stainless steel with words, documents, books etc all overflowing the top and more poring in with only a single stream flowing out the bottom into a partially written book (2 pages showing one with partial text as the stream of letters from the funnel is landing/laying on them.. text above the funnel should be of differing sizes and styles whilst the text exiting the funnel should be of a similar size but different style (eg some could be clipped/torn from newspapers, whilst others bare times-new-roman letters) and the finished text in the book being of a uniform style and size. The text on the page should be just readable and some coherent text like an extract from Shakespeare.

For the second, I told Claude what I was thinking and it gave the prompt:

photorealistic image of a hand hovering between several hammers of different types arranged on a workshop bench, ball-peen, claw hammer, rubber mallet, sledge, and tack hammer visible, natural workshop lighting, the hand mid-decision, slightly raised above the row of tools.