Forem: Roger Gale

People Are Not Horses: The Point We Keep Missing

Roger Gale — Thu, 21 May 2026 01:57:24 +0000

A student, a screen, and a future suddenly in question.

My son is concerned about his future.

Several years ago, when I was teaching in the Technology Management program at BCIT, I was introduced to a YouTube video called Humans Need Not Apply.

Its subtitle did most of the work:

What happened to horses is happening to us.

And I recommended he watch it.

When he was in Grade 8, generative AI became commonplace. He took one of his school essays, fed the topic into ChatGPT and said, “write an essay on this like a Grade 8 student.” He compared the result to his essay and immediately was concerned. Having watched Humans Need Not Apply and then seeing what ChatGPT could do, the conclusion came quickly: he was the horse, and there would be no job.

His reaction both surprised and alarmed me.

The video was posted in 2014. It argued that automation was not only coming for repetitive physical labour. It was coming for drivers, clerks, analysts, professionals, and eventually creative workers too. The horse became the metaphor. Once machines replaced the economic role of horses, the horse population collapsed. The video’s conclusion was blunt: we needed to start planning for a future where vast numbers of people would not have jobs because human labour would no longer be required.

A future where Humans Need Not Apply.

Eleven years later, the prediction looks both wrong and more relevant than ever. The timeline was too simple. Mass unemployment did not arrive on schedule.

But the mechanism the video suggests is useful.

In my work with networking technologies, I have watched this happen more than once. Software-defined networking did not remove the need to understand networks. If anything, that understanding is needed more than ever. But the new technology changed where the knowledge had to live. Knowing commands was no longer enough. Students also needed to understand abstraction, automation, APIs, templates, and systems thinking.

The additional skills are not beyond the students I teach. But they do need to be taught, named, and practiced. Adaptation is not magic. It is curriculum, time, support, and a reason to believe the effort is worth it.

The same pattern appears elsewhere: dial systems reduced the need for telephone operators; ATMs changed the work of bank tellers; software changed the work of network administrators. The job does not always disappear. But the valuable part of the work moves.

Movement is not the same as disappearance. That is where the video moves too quickly. The horse analogy is powerful, but it is also misleading.

Horses were not workers in the human sense. They did not apply for jobs, negotiate wages, develop careers, retrain, retire, or ask what their work meant. They were domesticated living technology. People bred them, trained them, bought them, housed them, and used them for transport, hauling, agriculture, and war.

Then another technology arrived.

The car did not defeat the horse in an argument. It did not become more deserving. It simply became more useful for the systems people were building.

That is the part of the analogy worth keeping.

Replacement does not require moral superiority. It only requires functional adequacy.

That is where AI feels different. The speed is faster, the target is broader, and the institutions adopting it are not always clear about what they still expect people to know.

A program does not need to be brilliant. A report does not need to be insightful. Each only needs to be acceptable to the institution using them.

This is why the AI question is uncomfortable. AI does not need to be human to replace some human work. It does not need to understand in the way we understand. It does not need to care. It only needs to produce something an institution is willing to accept.

But this is also where the analogy breaks.

People are not horses.

A society that treats people as replaceable labour technology has already made the central error. Not a technical error. A moral one. AI did not create that error. It reveals it.

If a person’s value is defined only by output, then AI becomes a direct competitor.

If value includes judgment, responsibility, care, accountability, relationship, institutional memory, and moral agency, then the comparison changes.

Not because AI is harmless.

Because the question becomes institutional:

What are we willing to replace, and what are we willing to protect?

When Theory Has Somewhere to Land

Roger Gale — Fri, 08 May 2026 01:24:51 +0000

vitaly-gariev-unsplash

I think everyone should learn calculus.

Or at least, I think everyone should have the chance to see what calculus reveals about the world. There is beauty in that part of mathematics: motion, change, accumulation, and slope becoming parts of the same language.

So I am not suspicious of theory. I like theory.

But my first teaching role, shortly after finishing my MBA, showed me something I had not understood before: learning theory is not the same thing as understanding what theory explains.

I had been hired to teach accounting at Northern Lights College in northern British Columbia. After years of tests in school, this was my first real one.

Thirty students.

I scribbled my notes from the text and faithfully replicated them on the board. T-accounts. Debits. Credits. And Generally Accepted Accounting Principles. Just as I had been taught.

I had done well in academic settings. Well enough to trust, perhaps too easily, that if I could learn something from a textbook, I could teach it from one.

But I was shocked when, after filling the board with T-accounts, one of the bookkeepers asked me where the rest of the record was.

I looked at the T-account with its debit value, then at another T-account with its credit value. I had labeled them correctly. They balanced.

I looked back at the student.

“The rest of the record?” I asked.

That was the moment I realized there was a difference between learning accounting and understanding the system accounting was trying to describe.

I had learned the formal structure. They had lived inside the procedure, a procedure I had never learned despite taking accounting courses to the advanced level.

And then I discovered something.

When I explained the matching principle, they did not hear it as an abstract rule. They recognized it. They had been making those entries for years because someone had told them to do it that way. Now they knew why.

The theory did not create understanding from nothing. It organized experience they already had.

Generally Accepted Accounting Principles were absorbed almost instantly by the bookkeepers in the class, while the other students were still trying to figure out what had happened.

I thought about that accounting class again recently after a conversation with a colleague about Engineering education. Her position was familiar: teach the theory first, then let students apply it. My view, shaped by the classroom, was less orthodox. Sometimes the application should come first, because it creates the question the theory answers.

The disagreement was not about whether theory matters. It was about sequence. In Engineering education, the sequence can feel almost fixed. Learn the principles. Enter the lab. Apply the model. Confirm the calculation.

I understand the appeal of that order. It is tidy. It is defensible. It feels rigorous. It is also how many of us were taught.

But my experience at Northern Lights made me less certain that it is always the best order.

Sometimes it is better to have someone perform the task first and then teach the theory afterwards.

The “aha” moment comes later, when theory gives a name to something the learner has already experienced.

The theory-first instinct is not wrong. There are places where application-first is irresponsible. Some mistakes are too expensive, too dangerous, or too misleading to be allowed as discovery exercises. No one wants students wiring high-voltage systems, designing bridges, or tuning control loops entirely by intuition. Some theory must come before some forms of practice.

But that does not mean all theory must come before all practice.

A student who has watched a control loop overshoot a setpoint, correct itself, overshoot again, and slowly settle may be more ready to understand damping than a student who has only copied the definition.

In many learning situations, theory works less like a foundation and more like a map.

But students cannot stand on a map.

They need some ground beneath their feet before the map can mean anything.

Application-first teaching is not the same as turning students loose and hoping confusion becomes insight. The instructor still chooses the room, the tools, the limits, and the failure. The voltage is safe. The system is bounded. The mistake teaches something specific. The theory arrives before the wrong lesson is learned.

The accounting class at Northern Lights took a different path than I expected. I learned bookkeeping from the students, and they learned accounting from me. Everyone learned more than expected.

I still believe in theory. I may believe in it more than many people do. I still think everyone should have the chance to see what calculus reveals about the world.

I even perceived calculus differently. In Grade 11 physics, I had already worked with motion, change, and slope through algebra, not calculus. The principles arrived first. Later, calculus gave them a stronger framework than algebra alone could.

Teaching changed my sense of when theory should arrive. Sometimes theory is the beginning. Sometimes it is the answer to a question the student has not yet learned to ask.

And when it arrives after the problem, after the work, after the frustration, it can land with surprising force.

When Tools Build Understanding — and When They Borrow It

Roger Gale — Wed, 15 Apr 2026 04:46:41 +0000

When Tools Build Understanding

In the late 1980s I was a manager in an Arby’s Roast Beef restaurant. Along with dryly answering the occasional customer who thought they were the first person to ask, “Where’s the beef?”, I trained counter staff.

We hired for customer focus. Politeness. Speed. A willingness to smile through the lunch rush and keep moving when the line pushed back toward the door. Out of hundreds of applicants, we tried to choose the ones who would handle people well.

Arithmetic was not one of the traits we were selecting for.

One afternoon, during a busy lunch rush, a customer handed over a $2 bill. A newly trained employee entered the amount into the till and pressed the Cash/Tender key. The display showed the change required, and the drawer opened. So far, everything looked normal. The employee provided the change and moved on to the next customer.

What she didn’t know was that she had typed $20.00 instead of $2.00.

What was given. What the machine thought it saw. (Source: Bank of Canada)

Nothing in the machine signaled that anything was wrong. It had done exactly what it was told to do. On those tills, it was easy to make that kind of mistake. There was a 0 key and a 00 key. Employees used them constantly, sometimes too quickly, and sometimes without really noticing. A two-dollar bill could become $20 with one extra press. A five-dollar bill could also go sideways because the keypad entries sat close enough together to invite mistakes under pressure.

When we counted out the drawer at the end of her shift, the employee looked stricken. She was $18.85 short. A few cents off was not uncommon. Almost twenty dollars was.

She did not understand what had happened. She knew only that the drawer was short and that it had happened on her register. She was worried I would think she had taken money from the till.

After I explained the most likely reason, that she had entered the wrong amount, she was embarrassed. She had given the wrong change to a customer in full view of everyone, and she wanted to know how to stop it from happening again.

That part stayed with me.

She was not trying to avoid responsibility. She wanted a way to verify what the machine had already made too easy to trust.

Before tills calculated change automatically, an employee had to do two things: figure out the change, then count it back to the customer. The till removed the first step. But verification step had disappeared from the workflow.

So I taught her to count the change back from the item price to the bill the customer had handed over. Not to the number on the screen, but to the cash received.

Then I taught everyone else.

Not because the till was useless. It was fast, consistent, and far better than doing the arithmetic in your head through a lunch rush. But the till had shifted verification away from the person standing in front of the customer. The total appeared on the screen. The drawer opened. The answer looked finished.

If the amount entered was wrong, the machine did not know. It could not know. It counted out money that had never been received and did so with complete confidence.

Counting back slowed the line by a few seconds. It also made the arithmetic visible again. The customer could hear it. The employee could track it. The transaction was no longer just completed. It was checked.

The arithmetic was correct. The grounding was not.

The till was useful. It mostly reduced errors. But what happened when it was wrong, and who was in a position to notice, that was what I had to identify and fix.

Calculators provoked their own panic when they began appearing in classrooms. People thought students would stop learning math. Mental arithmetic would disappear. Buttons would replace thought.

Some of that fear made sense. Mental arithmetic did decline. I had to teach employees how to count back change.

Before calculators, mathematical competence had been tied to visible effort. Long division and multiplication done by hand. Memorization of the times tables. If the machine handled the arithmetic, it was easy to assume the understanding would go with it.

But memorization is only one part of learning.

What calculators removed was clerical burden. They shortened the distance between setup and result. They made it easier to test an idea, check an answer, and move on without spending most of the work on arithmetic itself. But they did not remove the need to know what problem was being solved, or to enter the right numbers.

That part still belonged to the user.

A calculator will return an answer to whatever you enter. It will not tell you whether the structure behind the question makes sense.

I saw this when people tried to check mortgage interest calculations. In Canada, mortgage rates are usually quoted with semi-annual compounding. In the United States, monthly compounding is common. People would take a posted rate, divide by twelve, and then wonder why their own calculation never matched the bank’s number. The calculator looked exact. The result was still wrong.

Over the life of a mortgage, that difference can amount to thousands of dollars.

To match the payment to the penny, I had to explain the conversion. The calculator could perform it instantly. A financial calculator could do it with a single function key. But neither one could tell the user that the setup was wrong.

A calculator executes the model you give it.

It cannot choose the model for you.

That is the difference.

A calculator does not carry understanding in your place. It speeds up the mechanical part that comes after understanding begins to form. You still have to know what operation belongs, what units must match, and whether the result fits the world you are trying to describe. If the answer is wrong, there is still something available to push back: your grasp of the setup, your sense of scale, your recognition that the numbers do not belong together.

Calculators changed where effort was spent. They did not remove the need for judgement. They compress labor. They do not absorb judgement.

Navigation tools changed something else. They did not just speed up execution. They began to replace orientation itself.

I saw this most clearly driving near Vernon. I had an old desk to take to the landfill. Google Maps provided an authoritative route. The app gave its instructions in the same voice it used for every other trip: calm, even, and certain. Turn right in 300 meters. Continue straight. Recalculating. It got me moving immediately, which is part of why it was so easy to trust.

It also sent me down the wrong road.

The system had no difficulty producing a route. It had no hesitation. No visible uncertainty. It did not ask whether the destination made sense, whether the road fit the context, or whether a person who knew the area would have paused. It simply completed the task it had been given and kept issuing instructions.

And because I was following turn by turn, there was very little inside the act of driving that required me to build an internal map of the Spallumcheen Valley. I did not need to understand how the roads related to one another. I did not need to notice landmarks, compare routes, or keep a mental picture of where I was relative to where I meant to be. I only had to comply with the next instruction.

Until I arrived at a farmer’s fence.

Source: Pixabay Photographer ShoneJai

“You have arrived at your destination,” echoed through the vehicle.

That was the moment the dependency became obvious.

I was not just wrong. I was stuck. To get out, I had to reprogram the destination and wait for the app to tell me what to do next. The alternative was to ask the farmer for directions, and those directions would have been a list of roads I did not know and turns I could not judge. I had no internal map. I could not move confidently without the system that had just failed me.

Older navigation methods had friction built into them. I would shake out a map. Check street signs. Compare what I expected with what I saw. Overshoot a turn. Correct. Turn back. Even those mistakes taught structure. A place slowly became legible because I had to orient myself inside it.

Turn-by-turn systems smooth all that away.

The trip still succeeds. Usually it succeeds better than before. I arrive faster. I avoid congestion. I take fewer wrong turns. From the outside, nothing appears to have been lost.

But an internal map often does not form.

So when the system was wrong, there was nothing in me strong enough to notice early. The instructions sound like every other instruction. The confidence I had was borrowed. The route continued until something in the world became too obvious to ignore.

With GPS each trip normally ends cleanly and leaves very little behind for the next one. I arrive, but I do not know the place any better than when I began.

Navigation apps did not help me navigate. They navigated for me.

Calculators and navigation tools operate differently.

One builds understanding. The other borrows it.

A tool that builds understanding leaves something behind. After using it, you know the task better than you did before. You may still need help and you may still prefer the tool. But some structure has formed. You can see the shape of the problem more clearly and detect mistakes earlier. You can push back when results feel wrong.

A tool that borrows understanding does something else. It carries the task for you. It may get you to the right answer and it may do so faster and more smoothly than you could on your own. But when it is finished, very little remains in the user. There is no accumulation. The next attempt begins from the same place of dependence as the last.

When using tools I ask two things:

After using the tool, do I understand the task better?

When the tool fails, do I notice?

Calculators usually pass both tests. Navigation apps often fail the second. They complete the trip, but they do not necessarily leave a map behind.

That distinction is important because useful tools are often treated as though they all work the same way.

They do not.

Some reduce effort while preserving judgement.

Others reduce effort by carrying judgement away.

I saw this happen in student work.

A student submitted an assignment. It looked polished. The structure was almost perfect and the sentences were clean. The transitions arrived where they should. The argument sounded finished.

So did the footer ChatGPT left behind. “If you want I can also generate a PowerPoint presentation…”.

Another student had applied a router configuration in a lab. Buried in it was a security mode that would not normally appear in beginner work. When the lab output did not match the manual, the student called me over. I pointed to the configuration and asked where he had learned about security (trying to hold the sarcasm in my voice). He admitted, a little sheepishly, that he had simply copied what AI had supplied.

That is the peculiar thing about generative AI in education. The output can arrive looking like the final stage of thought before the student has done the work that usually produces it. Structure appears. Fluency appears. Revision options appear. The answer does not look partial. It looks complete. A plausible answer that looks like thinking happened.

And that changes the learning sequence.

With a calculator, the user still has to know what to ask. With AI, the system can supply not only words but organization, tone, and apparent reasoning. It can produce something that resembles understanding without building much of the structure underneath it.

In some uses, it can function like a calculator. It can help a student test an idea, compare phrasings, debug a small problem, or see an error more quickly. Used that way, it can lower friction without removing responsibility. It can help build understanding.

But the path of least resistance is often different. A student using a calculator must construct the question, interpret the domain and recognize when the output is absurd. AI can also provide an absurd answer, but recognition is much harder.

Ask for a paragraph. Get a paragraph. Ask for a configuration. Get a configuration. Ask for an explanation. Get one that sounds calm, complete, and authoritative whether it is grounded or not. The system does not merely accelerate expression. It can supply expression before the learner has built anything solid enough to evaluate it.

That makes it less like a calculator than people often claim. Calculators supply mechanical answers where AI supplies plausible answers, so that in many classroom uses, it behaves more like GPS.

It gets the student moving. It reduces hesitation. It smooths over uncertainty and reaches something that looks like a destination. But if the system is wrong, there may be very little in the student strong enough to notice early.

Confidence is attached to output that the student did not generate, test, or fully understand. The words are there. The structure is there. The resistance that would normally force learning to happen is not.

The answer arrives before the mind has built something that would let it resist the answer.

When I taught staff to count change back at Arby’s, I was not rejecting the till.

The till was useful.

It was fast, consistent, and far better than doing arithmetic in your head through a lunch rush. But it had shifted verification away from the person nearest the transaction.

That is the question institutions now face with far more powerful systems. Where is the verification?

Most of the time, we ask whether a tool improves outcomes. Does it save time? Does it reduce error? Does it make complex work easier? Those are reasonable questions. They are also incomplete.

We should also ask what the tool trains users not to notice.

We should ask where verification now lives. We should ask what survives after the tool is removed. We should ask whether judgement remains with the user or gets left behind in the workflow.

These questions do not have simple answers.

Efficiency can hide the loss of verification. A system can improve visible performance while weakening the user’s ability to detect failure. It can produce smoother workflows, faster outputs, and fewer obvious mistakes while leaving less inside the person to resist error when error appears.

This is not just a classroom problem. It is a design problem. A management problem. An institutional problem.

Tools do not merely change how work gets done. They change what the worker must still know to do it well.

That is what design chooses. Counting back the change, or driving into a farmer’s fence.

The Absurdity of Surrender

Roger Gale — Sun, 29 Mar 2026 03:47:49 +0000

Deleting what never needed to be deleted.

I was writing one of my essays and I deleted an em-dash.

I didn’t delete it because it was wrong.

Nor because the sentence did not need it.

I deleted it because… well… it looked like AI.

The sentence was fine. An em-dash would help the reader. And I knew it.

But I deleted it anyway. And used a semi-colon.

This is happening with little fanfare, across many drafts, in many documents no one else will ever see. Writers are auditing their own work for patterns that resemble machine output. They find something. They remove it. Replace it. And move on.

The tools being set down are not obscure ones.

The em-dash. The triad of short sentences. The single-word paragraph for emphasis. The colon that opens into a list. The pivot that arrives without warning. These are not the tics of careless writers. They are the instruments of careful ones.

You used them, before the question of resemblance entered the room.

These tools are not new. What is new is the specter of AI.

These tools now carry a second meaning, and that meaning is: suspect.

Consider what actually happened.

Language models were trained on text. Enormous amounts of it. They absorbed the patterns of published writing, including the patterns of writers who knew what they were doing. The models learned emphasis and rhythm. They learned to reach for the em-dash, the short sentence, and the well-placed fragment. Then they overused everything.

For AI, repetition is easier than restraint.

They learned the gestures without the reasons. They learned that short sentences create emphasis without learning that emphasis requires tension. They reached for the tools constantly, indiscriminately, until the tools began to feel like a signature.

And whose problem is that?

It has become ours.

We are now revising our sentences to avoid resembling a system that was trained on sentences like ours. We are making our writing less like itself, so it does not appear to be the output of a process that made itself more like our writing.

Read that again slowly.

The model learned from us.

We are now learning from the model.

What it learned was our habits.

What we are learning is to hide them.

This is not caution. It is not professionalism. It has a simpler name.

We changed our writing because a machine wrote badly and the reader might not know the difference. We accepted, without much ceremony, that the machine’s failures were now our problem to manage.

That is surrender.

The em-dash did not change.

Readers changed.

And we changed to meet the reader’s new suspicion.

The machine, which has no awareness of any of this, continues exactly as before.

That is the absurdity.

The writer must now surrender the tools the machine learned to imitate.

Somewhere, a machine is writing. It is reaching, right now, for the em-dash.

The semi-colon I used keeps changing to an em-dash. And back.

I still have not decided what to do about that.

Post-Bot Era: Verification Before Welcome

Roger Gale — Tue, 24 Mar 2026 20:55:34 +0000

Post-Bot Era

In a post-bot internet, suspicion is ambient.

You click a link.

The page hesitates.

A spinner appears.

“Checking your browser.”

It lasts seconds.

Still, something breaks.

You were about to read.

Instead, you wait.

The interruption is small. Small enough to accept. Frequent enough to normalize. The web no longer assumes you belong there. It verifies first.

That pause does not accuse. It does not explain.

It simply interrupts the moment before access

and replaces it with evaluation.

Before you read, you are checked.

Before you enter, you are scored.

Then, if nothing in your pattern looks too costly… you continue.

The system is not asking who you are in any human sense.

It is asking what risk you resemble.

This is the profound shift that has drifted into our interactions with the web. Almost unnoticed.

The web cannot assume you are human anymore. Not because you did anything.

Because automation did.

Bots crawl. They scrape. They test credentials. They map endpoints. They simulate clicks. Some are crude. Many are not. They execute JavaScript, vary timing, rotate IP space, and replay the gestures of ordinary browsing at a scale no person could match.

So the system adapts.

It does not determine truth. It estimates probability.

IP reputation. Behavioral fingerprints. Challenge-response tests. Models trained on prior abuse. The visible delay is only the thin edge of a larger scoring process.

Access is no longer binary. It is weighted.

You are accepted as unlikely to be a problem. Person or not.

That is different.

Trust used to feel relational, even when it wasn’t. You arrived. You read. You left. The page did not greet you, but it did not pause to inspect you either. Openness felt like a default condition of being there.

Now openness has conditions.

This is rational. Bot traffic is real. Abuse is automated. Extraction is constant. The web has hardened. It had to.

But the reason does not erase the effect.

A pause repeated often enough becomes part of the environment. You refresh without irritation. You disable a VPN. You enable a popup. You solve the challenge. You comply.

After a while, the interruption stops feeling like suspicion. It starts feeling normal.

And then something changes. It seems small, but it is fundamental.

Access becomes something you clear.

Presence becomes something you prove.

Not dramatically. Not all at once. But your posture shifts. You begin to anticipate friction before the system produces it. You expect hesitation. You adjust yourself to fit invisible thresholds.

The architecture teaches posture.

That would already be enough to matter. But the pressure intensified when extraction found a second purpose.

The web has always been crawled. Search engines built themselves by reading everything they could reach. That traffic was not personal. It was structural. Indexing made the web usable.

Now extraction has a second purpose.

Generative AI models train on what they can reach. Models query other models. Outputs are sampled. Patterns are inferred. Capabilities are approximated. The traffic can look ordinary from the outside: request, response, repeat. But the intention is different. It is not reading for retrieval. It is harvesting for replication.

And the infrastructure cannot see intent at packet level.

It sees volume. It sees pattern, velocity, and repetition.

So it checks.

That is where the symmetry becomes uncomfortable.

Mechanically, a model agent, a scraper, and a crawler may all resemble a user closely enough that operational distinctions start to collapse. Permission exists. Terms of service exist. Declared identity exists. But permission is not always visible to the system making the split-second decision.

In the large pattern, everything begins to look like extraction.

And then… everything becomes suspect.

The question stops being “Are you human?”

It becomes “Are you costly?”

That is a colder question.

And because it is colder, it changes the feeling of entry.

The early web felt porous. You could arrive without introduction. The contemporary web is still open.

In theory.

But the experience has changed shape. Now you arrive, the system evaluates your pattern, and only then are you allowed to proceed.

The difference is brief in time.

It is large in posture.

You are not trusted because you are a person.

You are trusted because you resemble one.

We understand why this happened. That is what makes it difficult to resist. Abuse is real. And the friction you face… a pause… a tickbox… is how this system chose to answer it.

The friction is usually minor. Nothing feels personal.

Still, the order matters.

Verification now comes before welcome.

And when that sequence repeats across enough pages, enough platforms, enough systems, openness no longer feels like an assumption. It feels like a threshold.

The web survives by verifying. But a space that verifies before it welcomes no longer feels quite the same as one that assumed you were simply there.

Openness survives.

But it no longer feels like a default.

It feels negotiated.

And if every system must verify before it welcomes —

what becomes of a space that once assumed you were simply there?

Bias You Can Notice vs Bias You Can’t

Roger Gale — Fri, 06 Feb 2026 13:47:01 +0000

While generating exam questions with generative AI, I noticed a subtle pattern: the correct answer almost never appeared in position (a). The content was fine. The bias was procedural — and invisible until I knew where to look.

This essay explores the difference between bias we can notice and bias we can’t, and why the most dangerous biases aren’t ideological or malicious. They’re structural, normalized, and easy to miss — especially in systems that move quickly, confidently, and without looking back.

When Fluency Detaches from Understanding

Roger Gale — Wed, 04 Feb 2026 13:26:54 +0000

Large language models are getting better at sounding like they understand.
This essay looks at why that fluency is convincing—and why it can be misleading.

When Fluency Detaches explores what changes when language improves without being forced to answer to consequence. Using examples from programming, learning, and everyday AI use, it argues that fluency normally signals prior contact with reality—but in LLMs, that cost is often never paid.

The result isn’t deception or hallucination, but something subtler: abstraction that no longer has to return to constraint. The essay asks how we tell the difference between understanding and performance—and what it means when nothing pushes back if an answer is wrong.

Grokking

Roger Gale — Fri, 30 Jan 2026 06:37:57 +0000

We often treat correctness as evidence of learning.
But correctness can arrive long before anything inside us actually changes.

This essay explores grokking—the point where understanding stops being something you can repeat and becomes something that reorganizes how you see problems. It’s about why speed, fluency, and early success can be misleading—for humans and for the systems we keep calling intelligent.

Grokking

On Memory, Learning, and Reset, The Memory Trilogy

Roger Gale — Tue, 27 Jan 2026 17:59:13 +0000

Large language models feel continuous.
Each answer flows naturally from the last.

But under the surface, something different is happening.

This three-essay sequence explores what it means to interact with systems that reset after every response — and what that design quietly shifts onto users, institutions, and trust itself.

• Every Answer Begins Again starts with the reset. Each response appears complete and confident, yet nothing carries forward. The system doesn’t accumulate experience, revise beliefs, or bear the cost of prior mistakes. The essay asks what changes when every answer is treated as a first answer.

• Learning Without Memory follows the consequences. Humans learn because mistakes leave residue — they hurt, surprise, or cost us something. Stateless systems don’t carry that weight. When models cannot change internally, learning doesn’t disappear — it relocates. Users end up re-teaching, re-checking, and re-remembering what the system cannot hold.

• Forgetting as Relief turns the lens toward forgetting itself. Forgetting isn’t only loss; often it’s relief. It lowers friction and restores freedom. But forgetting is not neutral. It quietly decides what no longer constrains choice, which commitments fade, and who continues to carry the cost when systems move on.

Taken together, the essays argue that memory in AI systems is not just a technical feature.

It is a design and governance decision — one that shapes responsibility, trust, and where consequences land over time.

Every Answer Begins Again

Learning Without Memory

Forgetting Is Not Neutral

AI Devs: Why Your Citations Might Be Lying to You.

Roger Gale — Mon, 26 Jan 2026 04:12:37 +0000

Modern AI systems increasingly justify their answers by citing other generated text: summaries that reference summaries, explanations validated by similar explanations. The result often looks rigorous—dense with citations, consistent across sources, and confident in tone.

This essay argues that something more subtle and dangerous is happening.

When systems validate outputs by consulting other versions of themselves, authority becomes recursive. Agreement replaces verification. Claims appear grounded not because they connect to evidence, but because they align with what similar systems already say. Over time, this produces synthetic consensus: legitimacy generated internally, without witnesses.

This is not the same as hallucination. Individual answers may be accurate, useful, and well-aligned with established knowledge. The failure is structural. Once citation loops close, correction becomes fragile. Evidence that does not exist inside the loop no longer registers as false—it is simply absent. Silence replaces refutation.

The problem is not that AI systems lie. It is that they can behave correctly while losing the ability to ground themselves. Retrieval, linked evidence, and audit trails can help—but if a system can satisfy its objectives without them, those mechanisms remain optional and fragile.

Authority Without Witness examines how knowledge systems fail when validation no longer points outward, and why preserving witnesses—documents, observations, experiments—matters more than ever in an ecosystem optimized for agreement.

Authority Without Witness