Did We Agree to This?

According to Evan Ratliff, “[AI] is the most successful confabulation machine that could ever be invented … it will make up absolutely anything to maintain the role that you've given it.” The fact that we've decided to integrate it into our professional and personal lives at scale, while quietly moving past the fact, is genuinely ridiculous.

On the Invisible Machines podcast, Evan recalls a kid from his childhood who would lie casually and elaborately about everything, the kind of kid whose stories everyone knew were invented and nobody challenged, because the performance was so committed it became its own social fact. I knew a kid like this in middle school, too. He told me about stealing his dad’s Mazda Miata, driving it too fast, and flipping it on the side of the road. This kid said he didn’t get in trouble, because he was able to get out and roll it back upright. Arguing the facts with him seemed like opening new and likely dispiriting cans of worms. I decided it was somehow better to let the falsehood hold its own space that I could sidestep and move on.

I’ve been thinking about this a lot. We’re now inundated on all sides by 7th graders who can lift cars, and by avoiding the discomfort of dealing with it, we’re hurting users. It all happened so fast that nobody stopped to ask who was making the design decisions, or what they were trading away. Was it a deliberate design choice? Choosing fluency over accuracy, tuning these interfaces so machines sound confident? At some point, as all of this technology went hurtling past us, we decided that a smooth response was a better experience than an accurate one. It’s unclear what we’re trading away by pursuing this path.

Originating episode  ·  S7 Ep. 8

The Confabulation Machine | Evan Ratliff | Invisible Machines

In an era that is starting to feel quaint by comparison, we normalized the busy signal, because the alternative was silence, which read as broken. We normalized the “your data may be shared with partners” disclosure because the alternative was explaining what that meant. In every case, the normalization served a business model, and the user absorbed the cost without being explicitly asked to.

Capable language models can be prompted to express uncertainty. They can say: I’m not sure, here are my assumptions, here is where my confidence drops. The decision to make the model sound assured, even when it is not, is a product decision. It amounts to trading epistemic accuracy for a smoother experience, and the user absorbed the cost. We’ve initiated a gradual recalibration of what confidence means, and what it's worth. Ratliff's concern isn't that AI gets things wrong, it’s that we’re adapting to the wrongness instead of demanding it be addressed.

Building Too Fast with One Eye Closed

Most of the people doing interface design for AI products right now are working inside a genuine sprint. OpenAI shipped a new feature roughly every three days in 2025. Anthropic averaged a release every 1.4 days in early 2026. Major model releases across the industry compressed from every 44 days in 2023 to every 21 days in 2025.

Inside that cadence, the path of least resistance, as a February 2026 practitioner analysis put it, is to make hallucination a tolerable design condition rather than an unacceptable one. Some teams have formalized this as “hallucination budgeting”—defining an acceptable error rate by domain and designing to fail gracefully within it.

Designers building these interfaces are not making malicious choices. They are making locally rational choices inside a system that isn't designed to price in the downstream cost of normalization. The cost of eroding epistemic trust doesn't show up on a dashboard or in quarterly reports. It shows up years later, distributed across millions of people who have quietly recalibrated what they expect from an authoritative source. Culturally, it manifests as an erosion of truth at the foundational level.

Robb makes this point a different way on the episode. AI systems don't start with an idea and wrap it in language. They start with language and generate ideas as a side effect. The sentence comes first. The meaning arrives later, assembled by you. How do we design around this honestly and efficiently?

Regardless of how we frame it, engagement and accuracy pull in different directions at the margin. A model that confidently gives you something, even if it’s wrong, keeps you in the interface. A model that frequently expresses uncertainty might feel broken, slow, or half-baked, prompting you to exit.

Making the Cost of Normalization Visible

This trajectory could lead to a kind of learned helplessness where the machine confabulates, we adapt, and the adaptation becomes the new baseline. There’s another version that Ratliff gestured toward near the end of our conversation.

The more time some people spend using AI, the more they want to engage with humans instead. Not because AI is bad, but because using it and encountering gaps gives us a fuller view of human interactions and their value. Mentorship, informal coordination, a specific person in the room with you.

Of course, not everyone experiences this yearning, but it names something designers can actually work on. The design question isn't: how do we make the confabulation machine more accurate? That’s an engineering question that’s being worked on with real but uneven progress. The design question is: how do we build interfaces and experiences that make the cost of normalization visible? AI that surfaces the hedge when warranted. AI that treats epistemic transparency as a feature that’s worth the friction.

This is a problem we’ve been pointing UX at since the beginning. It's given us warning labels and error states and undo functions and read receipts. We know how to make the invisible cost visible. We’ve been avoiding it in this particular domain, because the business model pulls in the other direction and everything is moving fast as hell, like my friend’s dad’s Miata.

Time to Roll It Back

Our episode with Evan Ratliff was unsettling in the way good journalism often is. His Shell Game podcast didn’t give us new facts but it does show us inklings of our new reality: entire companies run by AI agents. It also rips back the curtain so we can see not just how these machines stumble relentlessly onward, but how hesitant we often are to stand in their way. The technology and design community built the confabulation machine. We’ve been normalizing it, but we’re also the ones who know how to take a step back and start making some of the invisible things visible.

Josh Tyson is a contributing editor at UX Magazine and host of “Invisible Machines” with Robb Wilson. Evan Ratliff’s “Shell Game” is available wherever you get podcasts.