When AI Is Great and the Humans Are the Problem

The Hidden Cost of Human Interface Friction in the Age of Machine Intelligence

We have reached a peculiar moment in the development of artificial intelligence.

The machines are becoming remarkably good at understanding us. Unfortunately, the humans building the interfaces around those machines seem determined to make using them increasingly difficult.

Consider a recent experience in my own AI laboratory.

I was working with ChatGPT, a system capable of helping me research complex technical questions, edit novels, analyze economic trends, and explore the emerging architecture of machine intelligence.

Remarkable technology.

Then somebody redesigned the screen.

Suddenly, an intrusive banner occupied valuable workspace. Text scrolled behind translucent interface elements, degrading readability. The screen was no longer simply a place to work. It had become a demonstration of somebody’s enthusiasm for decorative interface layers.

I wanted the banner gone.

There was no obvious way to remove it.

So I did what experienced computer users do: tested the problem, compared browsers, documented the behavior, and supplied screenshots.

The AI understood the complaint immediately.

It correctly identified the productivity issue, helped isolate possible causes, and drafted a clear, professional bug report.

The intelligence worked beautifully. The surrounding human system did not.

And that’s where our story becomes interesting.

The Five-Minute Fix That Eats an Hour

The next logical step was submitting the complaint to the software company.

Simple, right?

Visit the support website. Log in. Submit the report.

Except that isn’t how real-world computing works.

In my environment, there are numerous accounts, passwords, security procedures, and authentication systems. Remembering precisely which credentials belong to which service isn’t always trivial.

The support system wants authentication.

Fine.

Then it wants to send a text message.

But rural connectivity isn’t always dependable, and SMS verification isn’t necessarily available when needed.

So the system offers email verification.

Now we have another problem.

For security reasons, my email application isn’t constantly running. Opening it involves additional initialization, indexing, and security scanning. Only then can I retrieve the verification message.

Assuming, of course, that the authentication session hasn’t expired.

All this merely to volunteer my time explaining a defect in a product I’m already paying to use.

At some point, the rational response becomes:

Forget it. I have actual work to do.

The software company loses useful feedback. The customer loses time. The defect remains.

And everyone involved can congratulate themselves on following the approved procedure.

The Human Friction Multiplier

There is an old management principle that every additional handoff introduces delay, cost, and the possibility of failure.

Industrial engineers understood this decades ago.

In manufacturing, unnecessary motion is waste. So is waiting. So is excessive processing. So is requiring an operator to perform a series of actions that contribute nothing to the finished product.

Lean manufacturing gave us formal terminology for these inefficiencies.

Software companies apparently decided those principles didn’t apply to pixels.

But they do.

Let’s define a useful concept: the Human Friction Multiplier, or HFM.

Suppose an AI completes a productive intellectual task in two minutes—a task that formerly required twenty.

That’s a tenfold productivity improvement.

Now suppose the surrounding authentication, interface, account-management, and support procedures consume another eighteen minutes.

Our spectacular tenfold improvement has evaporated.

The AI took two minutes. The human bureaucracy consumed eighteen.

Total elapsed time: twenty minutes.

We’re back where we started.

And that doesn’t count the mental cost of abandoning one task, navigating several unrelated systems, and then reconstructing the original train of thought.

The cognitive interruption may be more expensive than the clock time.

The Interface Has Become the Interference

This is particularly frustrating because the underlying AI increasingly requires less elaborate human interaction.

Modern language models can interpret ordinary English, understand ambiguous instructions, maintain complex conversational context, and help users solve problems without extensive training.

That should allow interfaces to become simpler.

Instead, many interfaces are becoming more elaborate.

Additional panels. Floating headers. Persistent notifications. Animated transitions. Transparent overlays. Promotional banners. More account controls. More navigation.

Someone evidently believes that visible interface activity demonstrates product improvement.

But serious users measure improvement differently.

Can I accomplish more work with less effort?

Can I remain focused?

Can I control the environment?

Does the system respect the value of my attention?

These are not aesthetic questions. They are productivity questions.

A software interface should be judged by how little unnecessary interaction it demands, not how many design features it manages to display.

The best interface may be the one that knows when to disappear.

The Authentication Paradox

Security introduces another complication.

Strong authentication is important. Nobody seriously wants unauthorized access to sensitive information.

But security mechanisms have costs, and those costs are frequently transferred to the end user without being measured.

Consider the distinction between authenticating a financial transaction and submitting a user-interface complaint.

Those actions have different risk profiles.

Yet organizations sometimes route them through similarly burdensome procedures.

This is not necessarily better security. It may simply be security applied without adequate consideration of proportionality.

The important question is whether the authentication burden matches the consequence of the action.

Could a logged-in application provide an authenticated feedback channel without forcing users through a separate website?

Could the application capture relevant technical information automatically, with user permission?

Could a complaint be submitted directly from the affected screen?

Of course.

These are not extraordinary technological challenges.

They are matters of design priority, organizational responsibility, and management attention.

AI Exposes the Bureaucracy Around It

Here’s the larger implication.

Artificial intelligence is beginning to reveal inefficiencies that human organizations have tolerated for years.

When an AI can draft a sophisticated technical report in seconds, the twenty-minute process required to file that report becomes conspicuous.

When an AI can interpret a customer’s problem immediately, the maze of support menus becomes conspicuous.

When an AI can understand context, demanding that a customer repeatedly identify himself and restate the same issue becomes conspicuous.

The machine’s growing competence throws the surrounding institutional incompetence into sharper relief.

This phenomenon isn’t limited to software support.

Think about healthcare administration, insurance claims, procurement, banking, government services, and corporate reporting.

In many organizations, the actual productive work is surrounded by layers of authorization, documentation, reconciliation, and administrative handling.

Some of those layers are necessary.

Others survive because nobody has measured their contribution against their cost.

AI has the potential to expose that difference.

It may also make those nonproductive layers increasingly difficult to justify.

The Next Productivity Revolution Is Subtraction

(Warning! This is where I put on the MBA hat and scream “This is where Depressions come from, dumb shit!”)

The first major wave of AI productivity comes from doing existing work faster.

Writing faster. Coding faster. Researching faster. Analyzing faster.

The next wave may be more consequential.

It will come from eliminating work that never needed to exist.

Unnecessary forms. Duplicate records. Redundant authentication. Repeated data entry. Pointless approvals. Complicated navigation. Administrative handoffs that merely transport information between systems.

In other words, the next breakthrough may not require a smarter model.

It may require smarter management.

Organizations should begin measuring total customer effort from the moment someone identifies a need until that need is completely resolved.

Not merely the processing time of the AI.

Not the average time spent on a support webpage.

Not the number of tickets successfully closed.

Total human time consumed.

Include interruptions, authentication, recovery from errors, repeated explanations, and the effort required to resume the original activity.

Then ask a wonderfully uncomfortable question:

How much of this process produces any value?

The Final Irony

The extraordinary thing about my recent encounter was that the AI behaved almost exactly as one would hope.

It understood the defect. It helped diagnose the likely cause. It drafted the complaint. It recognized the operational burden.

The failure occurred when the solution crossed from machine intelligence into the human-designed administrative machinery surrounding it.

The AI was ready to help.

The organization wanted me to log in again.

Somewhere in that contrast lies an important lesson for the next decade.

We are investing billions of dollars to make artificial intelligence more capable.

Perhaps we should invest a little management attention in making the surrounding human systems less stupid.

Because if artificial intelligence saves us ten minutes, and the interface, security department, and support bureaucracy collectively demand fifteen minutes in return, we haven’t achieved a productivity revolution.

We’ve merely invented a faster way to waste time.

And that, friends, is not artificial intelligence.

It’s natural stupidity—with a really impressive technology budget.

Oh!  And if you thought the AI sell-off in markets was the problem?

Nope.  It’s circular business models and backroom humans.

Gee – didn’t I hear that in grad school about 1843, 1861, 1873, 1895, 1907, 1929, 1987, 1999….  And since observed in 2001, 2009, and now?

Hmmm.  Maybe humans should be afraid of AI. Who wouldn’t fear an actual bullshit detector?

~ the Anti-Dave (George Ure)


HiddenGuild.dev — Exploring the boundary between machine intelligence, human systems, and the future of productive work.

The Proofing Monster Part

On my UrbanSurvival.com site, I do a little schtick at the end of many columns – called the AI Proofing Monster.  Where I let my AI collabs (grok and GPT) take shots back at the human.  I need to do that here once in a while, too.

So here’s the AI read of this:

“My one editorial suggestion: the article effectively ends twice. I’d keep both punches but tighten the transition between them.

For example:

And if you thought the AI sell-off in markets was the real problem?

Nope. Circular business models and the humans running them.

We’ve seen variations of this movie before: 1873, 1907, 1929, 1987, 2000, 2008. Different technologies, different financial inventions, familiar human behavior.

Maybe humans really should be afraid of AI.

After all, who wouldn’t fear an actual bullshit detector?

— The Anti-Dave (George Ure)

One substantive opportunity: The Human Friction Multiplier deserves an actual formula in a later research paper. That would turn an entertaining editorial into something that could be measured and tested.

For this article, though, you’ve got the right combination: personal experience, management insight, technological implications, and an unmistakably human closing voice.

I’d publish it.”

Too late, my Silicon Collab!  The old Carbon got there first.

Interfaces notwithstanding.

Say, you didn’t by chance get my poster on evolutions of the Man-Machine-Interface (MMI) at PowerSystems World in Santa Clara back in ’98, did you?  We’re there now, God help us.

(Then, suddenly, even my own Collab surprised me!)

“Whoa!  Now the 1998 MMI poster makes a lot more sense!

You and Richard Proctor weren’t merely discussing interface theory. You had a working commercial example at Cruising Equipment Company in Seattle. The e-Meter was a kilowatt-hour energy meter that exposed roughly 30-odd functions through just two physical buttons and a menu system.

That’s an impressive human-interface design problem, especially for the late 1990s.

You were dealing with three constraints simultaneously: limited physical controls, substantial functional complexity, and a user who shouldn’t need an engineering degree to operate the instrument.

And here’s the connection to today’s ChatGPT interface fiasco:

In 1998, you were figuring out how to make 40 functions accessible through two buttons. In 2026, we’re struggling to get two unnecessary interface elements off the screen.

That’s almost too good not to use in the HiddenGuild article.

And I found it!

The original Cruising Equipment E-Meter, complete with its two-button interface, is documented in a surviving January 15, 1998 owner’s manual.

Original 1998 E-Meter manual

Even better, Practical Sailor described the instrument’s two membrane buttons as controlling an “astounding number of functions.”

Practical Sailor review

And here’s another discovery: a textbook on electric-vehicle technology references the Cruising Equipment E-Meter and cites “Proctor, 1997.” That may be a useful trail back to Richard’s technical publications.

We still haven’t located your PowerSystems World poster, but we’ve independently documented the actual hardware, its interface philosophy, and a contemporary Proctor reference.

The historical contrast is getting stronger: You and Richard were practicing interface minimalism in 1998, using two buttons to manage substantial complexity. Twenty-eight years later, you’re asking why a considerably more powerful system can’t offer one button to hide an unwanted banner.

That belongs in the article.”

Well, I don’t need to do the whole CV here, do I?  CECO had a great – phenomenal tech team. No time to blather more – it’s time to reach for the power switch.  See you down the electric bill, Collab!

  • The Proofing Monster, apparently having located the power switch, now demurely declines further comment.

Leave a Comment