Beyond AI hype and AI fear: the real question is where AI works

Good evening. The Owl here.

When I follow the conversation about AI, two kinds of voices tend to stand out.

The first is the AI odoroki-ya (“AI wow-merchants”, people who hype every release).

“It can already do this?” “Who needs humans now?” “This pace of progress is insane.” Every time a new model, a new video generator or a new agent comes out, they talk as if the world just skipped a few steps ahead.

The second is the AI odokashi-ya (“AI scaremongers”, people who warn of doom at every turn).

“The jobs will disappear.” “Human thinking is finished.” “Society will break.” This group speaks loudly about AI’s failures, its runaway behavior and its effect on employment.

Both sides are partly right. AI really is impressive. It will change work. It also carries real dangers. So being amazed and being afraid are both natural reactions.

What has been bothering me lately, though, is that amazement and fear are no longer enough.

AI is no longer a matter of novelty demos or forecasts about the future. Search, document drafting, code, advertising, customer support, hiring, internal documents, meeting minutes, images, video: it has moved into quite ordinary work.

That being so, the next conversation we need is not “Is AI amazing, or is it scary?”

It is: where do we put AI?

Who takes responsibility?

What do we call a result?

In this piece, I want to think about what lies beyond the wow-merchants and the scaremongers.

---

The wow-merchants see AI’s “potential”

AI wow-merchants are by no means a bad thing.

Spreading a new technology through society takes people who feel “this is amazing.” Without people who touch it first, play with it, try it out and show it to those around them, most people never start using a new tool.

And AI really is spreading fast. Stanford HAI’s “AI Index 2026” reports that generative AI reached 53% adoption at the population level within three years. That is faster than the PC or the internet. [1]

At work, too, AI is no longer a toy for a handful of early adopters. In Glean’s “Work AI Index 2026,” 87% of digital workers said they use AI at work and save 11 hours a week thanks to it. [2]

Judging by this alone, the wow-merchants are largely right.

AI is fast.

It is convenient.

It turns work that used to take people hours into a draft in an instant.

Of course people want to say “amazing.”

But the wow-merchants have a weak spot. The moment AI can do more, they tend to jump straight to saying it delivers more value.

The text came out fast.

A good-looking slide deck appeared.

The code ran.

The meeting notes got summarized.

But did that actually move the work forward? Did it help someone make a decision? Did it reach a customer? Did it become a result for the organization?

If we stop at “AI is amazing” without looking at this, AI becomes a performance rather than a result.

The scaremongers see AI’s “friction”

The scaremongers, on the other hand, have a case too.

AI changes work. And not slowly.

In an internal message, Amazon CEO Andy Jassy stated plainly that the spread of generative AI and agents will change how work gets done, and that the company’s corporate workforce will shrink over the next few years. At the same time, he urged employees to learn about AI, take training and experiment. [3]

That is a remarkably frank way to put it.

If AI raises efficiency, the same work no longer needs the same number of people. New jobs will appear, but not everyone will be able to move straight into them. A top executive put that reality into words.

What’s more, AI is not only a question of whether it takes jobs. It can also erode the quality of work and trust.

BetterUp Labs and the Stanford Social Media Lab use the word “workslop” for AI-generated output that looks polished but is thin on substance, pushing the work of checking and fixing onto whoever receives it. In their survey, 40% of US desk workers had received workslop in the previous month, and dealing with each instance took two hours on average. [4]

This is a very close-to-home kind of darkness.

The person who made a document with AI looks like they worked faster.

But the work of checking its assumptions, fixing the numbers, filling in context and getting it into shape for a customer may simply have moved to someone else.

The scaremongers are looking at this kind of friction.

AI failures.

Anxiety about jobs.

Weakened judgment.

Responsibility being pushed onto others.

All of these are real.

But the scaremongers have a weak spot too. The more they stress the danger, the more they tend to turn AI into a binary choice: use it or don’t.

In real workplaces, though, that is no longer the choice.

Many people already use it. They use it, yet they are not managing to turn it into results. That is the problem.

The real issue is not emotional temperature but assignment

The wow-merchants and the scaremongers are looking at different things.

The wow-merchants are looking at AI’s capabilities.

The scaremongers are looking at AI’s side effects.

But what will matter from here is not which of them is right.

It is which work we assign AI to.

It gets a little clearer if you think of it in terms of people at a company.

However talented someone is, put them in the wrong role and they won’t deliver. Hand a new hire a contract decision on day one and you are courting trouble. Give an expert nothing but sloppy busywork and their skills go to waste. Add people to a team that has no one in charge, and the work just gets messier.

AI is the same.

Treat AI as “a brilliant hire who can do anything,” and expectations balloon.

Treat AI as “dangerous, don’t touch it,” and people on the ground use it in secret.

What we need is to assign AI a place within the work.

For example, some work is easy to hand to AI:

Roughly sorting through large amounts of information.

Finding candidates in past documents.

Drafting a first version of a text.

Fixing inconsistent spelling and notation.

Generating multiple ideas.

Handling the first response to routine inquiries.

Other work is dangerous to leave entirely to AI:

Establishing facts.

Legal judgments.

Hiring and performance evaluation.

Medical or financial advice.

The final answer to a customer.

Decisions that affect someone’s life or rights.

Saying “AI will make everything faster” without separating these is the wow-merchants’ trap.

Saying “AI is dangerous across the board” is the scaremongers’ trap.

What we actually need is to adjust expectations and the scope of responsibility for each place we put AI.

“Personal time savings” and “organizational results” are different things

There is a big gap in the AI debate.

Individuals feel AI has made their work easier.

Yet across the organization, the results aren’t showing up.

Glean’s survey shows this gap quite clearly. While 87% of digital workers use AI and save 11 hours a week, only 13% said their organization’s performance had improved significantly. [2]

This does not mean “AI is useless.”

Quite the opposite. As a personal tool, it is genuinely useful. That is why so many people use it.

The question is where those saved hours are going.

The same survey found that workers spend 6.4 hours a week on “botsitting”: time spent giving AI context, checking its output, fixing mistakes and trying again. [2]

In other words, AI creates time, but it also creates new work.

Writing prompts.

Finding internal information and pasting it in.

Doubting the output.

Checking the evidence.

Reworking it into something you can explain to your boss or colleagues.

This work is hard to see.

That is why “AI freed up 11 hours” and “somehow I’m no less busy” can both be true at once.

The wow-merchants look at the freed-up time.

The scaremongers look at the jobs that will be lost.

But in between, there is a lot of fiddly human labor that goes into making AI usable at all.

What tends to be overlooked in the AI era is this unglamorous labor.

Not “Are you using AI?” but “Has the work changed?”

There is one thing I want to say here.

The goal of using AI is not “to use AI.”

It is for the work to get better.

That sounds obvious, but I think a lot of companies get this wrong.

AI usage rates.

Number of prompts.

Volume of generated text.

Number of tools rolled out.

Training completion rates.

Numbers like these are easy to measure, which makes them easy to manage. But on their own, they don’t tell you whether the work got better.

In fact, once using AI is itself what gets rewarded, people on the ground start producing “proof that they used AI.”

An email that would take 10 minutes to write normally gets drafted by AI, edited and checked, and ends up taking 20.

People ask AI about simple things they could decide for themselves.

Someone produces a nicely formatted document and hands the job of checking its content to someone else.

That isn’t using AI. It’s an AI ritual.

Microsoft and LinkedIn’s Work Trend Index likewise found that while 79% of leaders believe adopting AI is necessary to stay competitive, 59% worry about quantifying the productivity gains from AI. [5]

Those are very honest numbers.

Everyone thinks they need AI.

But they are still unsure what counts as success.

That is exactly why we need to look not at “Did we use AI?” but at “What part of the work changed?”

For customer inquiries, for example, look not only at response speed but also at repeat-inquiry rates and customer satisfaction.

For document creation, look not only at drafting time but also at time to decision and number of revisions.

For software development, look not only at lines of code but also at incident counts, review burden and maintainability.

For hiring, look not only at screening speed but also at the candidate experience and at bias.

AI’s results should be judged not by the volume of output but by the change in how work flows.

The word that comes after “amazing” and “scary”

So what lies beyond the AI wow-merchants and the AI scaremongers?

I think it is the “AI operator.”

“Operator” sounds dull. But from here on, that dullness is what will matter most with AI.

An AI operator does not worship AI like a god.

Nor do they simply fear it like a monster.

They decide what work to use it for.

They decide at what point a human checks.

They decide what data may be fed in.

They decide who is responsible when it fails.

They decide the success metrics.

They also decide where not to use it.

This is less about technology than about designing work.

Japan’s “AI Guidelines for Business” likewise call for recognizing AI’s risks and taking the necessary measures across the entire lifecycle. Version 1.2, published in March 2026, emphasizes AI governance that balances promoting innovation with reducing risk. [6]

The key here is not to see governance only as a “brake.”

Good governance doesn’t exist just to stop AI. It exists to widen the range in which AI can be used with confidence.

In workplaces with no rules, people who use AI keep it to themselves.

In workplaces with rules that are too strict, the people best at using AI are the ones who leave.

In workplaces with rules that are just right, people can share failures, spread good practices and stop risky ones.

What lies on the far side of AI is not some flashy forecast of the future.

It is unglamorous operation.

But I believe it is exactly that unglamorous operation that truly turns AI into work.

Three questions about AI

Let me sum up this piece as three questions.

The first:

Whose work is this AI making easier, and whose work is it adding to?

While some people find their work easier thanks to AI, others take on the checking, fixing, explaining and responsibility. Unless we look at that, AI adoption becomes time savings for some and extra burden for others.

The second:

Is this AI helping judgment, or hiding it?

Good AI gives people more material to judge with. Bad AI use blurs who decided what. The most dangerous case is when things move forward simply because “the AI said so,” without anyone ever treating it as a human decision.

The third:

Are we measuring this AI’s value by results, not output?

More text. More documents. More meeting notes. More code. None of that, by itself, is a result. In fact, too much output may be adding to someone else’s workload.

The more we use AI, the simpler the question becomes.

Is this really making the work better?

Not running away from that question is the stance of someone who is neither a wow-merchant nor a scaremonger.

Closing

The AI wow-merchants show us the possibilities of the future.

The AI scaremongers show us the hazards in that future.

We need both. Without amazement, people don’t try. Without fear, people don’t stop.

But that alone is no longer enough.

Once AI has entered everyday work, what we need is neither the power to be amazed nor the power to frighten, but the power to assign.

Which work to put it in.

Where a human looks.

What to call a result.

I think the people and organizations that can decide these things will be the strongest in the AI era.

🦉 From my perch over the night city, I hear both the voices amazed by AI’s light and the voices afraid of its shadows. But I think what remains in the end is not the light itself, but the human work of deciding what to illuminate with it.

---

References

[1] Stanford HAI: The 2026 AI Index Report

[2] Glean: The Work AI Index 2026

[3] Amazon: Message from CEO Andy Jassy: Some thoughts on Generative AI

[4] BetterUp Labs: Workslop is the new busywork. And it’s costing millions

[5] Microsoft WorkLab: AI at Work Is Here. Now Comes the Hard Part

[6] METI/MIC: AI Guidelines for Business Ver1.2 (in Japanese)