shriker.ca
online
Skip to main content.
People Systems

The Floor is Louder than the Ceiling

The tool multiplies whatever you brought in: competence or a default you never chose. The new risk isn't that it's bad; it's that producing now looks like done, and the expertise is being unmade in rooms that have stopped asking for context.
Jodie Struthers ·Sep 21, 2026·11 min

A multiplier does not ask what it is multiplying. It multiplies the knowing, and it multiplies the not-knowing. Same lever, same pull. For most of the history of making things, the tool was expensive, and that did the checking for you. Then a thing came along that is one click away, and it does not check at all.

It multiplies everything you can do. Including the mistakes.

I am not angry. I am impressed.

The flood is the structure now#

This stopped being a joke.

Low-quality AI output used to be a niche nuisance. Now it is a structural condition of the internet and workplace. It’s been called the microplastics of the web, and… yeah ok. That checks out. Small pieces, everywhere, accumulating in ways nobody can see from the surface until they cannot be removed.

The silly creatures on the feeds have stopped posting. Agentic accounts haven’t. They create the content and post it and engineer the virality. No human in the loop. The human is a ghost with a login and a budget.

Even the platforms that used to be a moat for people who could actually do the thing are getting swamped. By mid-2025, one of the big video platforms was reporting channels drenched in low-effort automated slop, and it did the only thing the economics let it do. It cut ad revenue for channels clearly producing it by volume.

Not for quality. For volume. That is the tell. That is the argument in a single policy change.

The default is doing the damage#

Almost nobody configures the tool. They take the default model. The default tone. The default that sounds confident and means nothing. They hit publish. Millions of times.

We have never held a technology where “just accept the default” produces consequences this wide and this fast. The default is someone else’s compromise. It is not your opinion. It is never your opinion. What happened to the authentic opinions?

A study from last year tested the major systems on a controlled task. When it cited a source, more than sixty percent of the time the citation was wrong, and it almost never flagged its own uncertainty. The more expensive models were more likely to be confidently wrong.

Fascinating.

Here’s the irony. “Convention over configuration” was a good idea. It worked because the default was opinionated: someone who thought hard made a call and baked it in. The default had a point of view. What the AI gave us is the opposite. Not an opinionated default. A default with no opinion and baked-in bias. The mean. The blandest thing that can’t be wrong. We kept the “default” and threw out the “opinion.” And not enough people are yelling about it.

The license#

I get asked for a Claude license. Or a Cursor seat. Or “can we just get them all of Copilot.” Also please enable all the connections for all the systems, and also these MCPs that have varying degrees of access and capabilities. Thanks.

And the pitch they bring me is not “I have a task and I need a tool.” It is “I saw a video.” A video where a person with a very good microphone and a ring light showed the tool doing something. Not their exact something. A demo something. A marketing something. And the take-away, unspoken, is: it does anything. Zero config. No thought. Just type the words and the thing happens.

It is marketed the way a VPN is marketed. You do not know what the routing table looks like. You do not know which ports are open. You do not know what the default cipher is. You just pay and the lock icon appears and you trust the box. And if the box is misconfigured, that is not a problem you have. That is a problem the box has.

I keep telling them: the box does not know what it is for. I do. The tool is a draftsman. I have to tell it where the walls are. If I do not tell it where the walls are, it draws them somewhere.

But they do not want the draftsman. They want the video. They want the ring light. They want the thing that looks like it does anything.

My job is to be the enabler. I make the thing go.

And I enable on a case-by-case basis. Which means I need a case. A use case. A thing. A “here is what I will do with it, here is what it will change, here is what I will stop doing because it now does that.”

And no one has one.

They have a vibe. They have a screenshot. They have “everyone else is getting or doing this and I have FOMO.” I ask “what will you do with it?” and I get “it does a lot of things.” Which is not a use case. “A lot of things” is the default. “A lot of things” is what the marketing said. “A lot of things” is the VPN lock icon.

So I am the hallway, and the hallway is full of people who want the door opened, and none of them can tell me what is on the other side of the door. And I am supposed to open it anyway.

And I have to figure out, one by one, whether the person in the queue is bringing a task or bringing a vibe. (And figure out vendor lock-in, costs, workflows, security, privacy, cross-border implications…)

The thing is in my building#

One thing that gets me: the tools are in my building. Generally free. Provisioned. With onboarding docs. There are Slack channels. The thing is here.

The person in the hallway is not asking “how do I use this.” They are asking “can I get Claude.” Not “I have a task and I need a tool.” They are chasing a brand. A logo. A name they saw in a video. And they are not going to learn the immensely powerful tools we have, or understand the underpinnings are the same. They are going to be loud about a tool they saw.

I gave you the keys and you’re standing in the parking lot.

The document arrived. No one asked for context.#

I have been handed documents recently. Good documents, on the surface. Clean. Structured. Polite. The kind you would print out and put in a binder if you did not know what you were looking at.

They were generated by a model. No one applied business context. No one told it what the constraints were, what the history was, what the room actually agreed to three meetings ago. They just… prompted and exported and sent.

And the room accepted them.

Not because they were good. Because they looked like good work. Because the document existed. Because the act of producing it was visible, and the act of knowing it was invisible, and the visible thing won.

Or.

Someone ran an API spec through a model. Got the full response. Sent the author only the part where the model was critical. Not the part where the model hedged. Not the part where the model was, frankly, wrong about a naming convention this team chose three sprints ago on purpose. Just the red-text bit. The “this looks wrong” bit. Forwarded like a finding.

The spec author is now defending a decision to a ghost. A ghost that also said three things that were best practice, but those do not get forwarded, because they do not serve the point. And the point, apparently, is “I ran a tool and it said your thing is wrong.” I’ve gotten so good at the pattern recognition, that I notice immediately when some of the “default” is missing. (This pattern recognition is how I found someone who had faked their own death, but that’s a story for another time.)

Noted.

And then, because the room needs a reason to stop asking:

“But that’s what I thought the tool was for.”

The tool was for something. The tool was for a task. The task was not “take my half-formed assumption, dress it in confidence, and produce a document I can wave at you and call done.” The tool does not know what it was for. Only the person holding it does. And that person, apparently, was not holding it. The tool was holding them.

Noted again.

Someone generates a thing. Their name is on it. Their login ran it. And now the thing is theirs. Not the tool’s. Not a draft. Not a starting point. Theirs. And when I say “did we talk through the constraint from sprint 3?” I get a wall. Not a “let me think about that.” A wall. Because I didn’t just question the document. I questioned the person, and the person doesn’t know they’re the same thing anymore.

“The AI said so.”

That is the new “because I said so.” Except at least “because I said so” implied the person had formed the opinion. This one implies the opinion formed in a room they weren’t in, and they walked in and claimed the seat.

Fused.

This is what I mean by erosion. It is not a flood you can see coming. It is slower than that. It is the standard quietly moving. One meeting, one document, one “looks good to me.” The next person who does know the context has to decide whether to spend their energy pointing out what is missing, or whether the room has already decided that “produced” and “done” are the same word.

The expertise is still there. But the credit for it is not. The tool got the credit. The tool got the byline. The tool got the “great work.” And the person who actually knew the thing is now the one who has to say “actually, none of this is true for our situation,” which is a harder room to walk into than it used to be.

The tool can raise your ceiling. If you have a ceiling. If you don’t, it just makes the floor louder. And when the floor is loud enough, nobody can hear the ceiling anymore.

What the commons does next#

Trust is a commons. Everything stands on it.

The creatures have noticed. A recent survey says roughly half of them now check across multiple AI sources before they will believe a thing. Only ten percent trust the first answer. (So. Get a second source or do not read it.) When your audience has to verify everything, the shared facts that platforms and journalism and research run on get thinner. And thinner. And thinner.

A report on the information ecosystem out of a university this past spring put it plainly. This is flattening into a monoculture. Homogenized. Commercialized. Anything that does not look like the default gets quietly sidelined.

The board wants a cake. The board also wants to know if the cake is real. 🍰 The board might settle for a nice shrubbery.

The three things#

I am tired, so I will be kind and short.

Treat the output as a multiplier of competence, not a substitute for it. If you handed me a document and did not tell me what I knew about the thing it describes, I cannot tell you it is wrong. I can only tell you it is uninformed. Which is a different, and sadder, word.

Stop accepting the default as your opinion. Configure. Question. Edit. The people who win with this thing are the ones who treat the model as a draftsman, not an oracle. A draftsman draws the lines. You have to tell it where the walls are.

Platforms cannot out-scale the slop. They have to out-incentivize it. The flood is driven by economics that reward volume over quality. So fix the incentives. Provenance. Labels. Demote what is only there to feed the algorithm. Not moderation after the flood. Before.

We did not build a tool that makes everyone smarter. We built one that makes everyone’s existing level of understanding more amplified.

For a while, that was a gift.

I played a game, a few years back, where you strap a rig onto your back so you can carry the weight of the world. The whole point of the game is that the rig does not walk for you. You walk. The rig just takes the weight off.

We are building that rig, for everyone, and we are handing it to people who do not have a map, and calling the straight line into the wall “progress.” Bonk.

The rig is not the problem. The person not knowing where they are walking is.

Observe. Document. Do not feed. 🥩

(I am going to feed it… It is a learning opportunity.)

Further reading, if you want the receipts: