Your transparency toggle is a default, and the default is off

Somewhere in your product there’s a control that says “show reasoning,” or “see sources,” or “why this?” It’s collapsed by default. It went in because a trust problem turned up in testing and giving people the option to look was the fix everyone agreed on. You almost certainly don’t know how often anyone opens it.

Two things landed this month that make that number worth having.

The first is a price tag on explanation. On August 20 the design tool company Figma published the second issue of Sightlines, its newsletter for design and product leaders, and took a hard line on agents that summarize things: when an agent reads across a pile of interview transcripts and hands back findings, every claim has to link to the specific transcript line it came from. That citation discipline, Figma argues, is what keeps the output trustworthy once the volume gets large. Meanwhile the research literature on AI transparency has been measuring the other side. Detailed explanations slow decisions when time is short. And observed usage patterns show people preferring speed over explanation when finishing the task is the goal, which raises the risk that they take the recommendation without checking it. Both of those are reported findings rather than settled ones, but the second is drawn from what people actually do when the explanation is sitting right there and can be skipped.

Given a measured cost, the sensible move is to spend it where it earns something. Show the reasoning for high-stakes, hard-to-reverse decisions. Let people move fast on the cheap ones. Scale the explanation to the consequence. That’s careful reasoning from the evidence, and it’s the conclusion I’d have written too.

The second thing that landed this month is why it doesn’t hold.

Product teams are rebuilding their back ends for agents. The vehicle is MCP, the Model Context Protocol, which Anthropic open-sourced in late 2024 and OpenAI, Google DeepMind, Microsoft and Cursor have since adopted. It’s a common way for an AI agent to reach your tools and your data, and it collapses integration work from every-agent-times-every-tool down to every-agent-plus-every-tool. The architectural advice arriving with it is consistent. Human workflow furniture doesn’t transfer. Step-by-step wizards, progress indicators and approval gates were built for human cognitive load, and agent workflows want intent-level abstractions and structured result schemas instead. Hand the agent the outcome and let the path go.

Scaling explanation to stakes assumes the reasoning is still sitting there to show, and the layer you’re rebuilding underneath is the one deciding whether to keep it.

Run the two decisions in the order a shipped product runs them. The architecture layer replaces step-wise state with structured results. The interface layer then decides, per stakes, whether to surface the reasoning. That second decision is a policy only if the first one preserved something to surface.

Look at what got discarded as human furniture. A progress indicator is transparency: it shows what the agent is doing at each step. An approval gate is override: it’s where a human redirects. Step-by-step is the intermediate state that makes both legible. A stakes rule is a throttle, and a throttle only controls something that’s flowing. Below a certain line in the stack it was never retained, and the interface can’t reinstate it when a decision turns out to matter after all. The same gap hits escalation. Knowing when to hand a decision to a human is one of the trust behaviors everyone lists, and escalating requires having the intermediate state to escalate with.

Neither call is wrong on its own. Both carry the same justification, that the human-facing step costs time. What nobody decided is what happens when you make both, and the two calls get made by different teams in different weeks.

Now the population problem, which is worse. A 2026 study by Fernandes and colleagues in Computers in Human Behavior, with 246 participants and a 452-participant replication, found the metacognitive risk concentrating somewhere counterintuitive: in AI-literate users specifically, who grow more confident in their self-assessments while getting less accurate. The efficiency preference this month’s research measured lives with users whose goal is immediate task completion. Read those two descriptions together and they’re the same person. The design fix is least likely to be voluntarily engaged by exactly the people it was built to protect, and once explanation becomes contextual, they’re the ones holding the control.

That deserves a fair hearing on the other side, because the case for correctable-by-design is real. Transparency genuinely does improve calibration; users who see the reasoning become appropriately trusting or appropriately skeptical rather than indiscriminate, and that one is verified. There’s a well-known list of interventions from a Harvard Business Review piece on July 20 that work on the same principle: reverse prompting, AI-free stages, parallel human and AI analysis, multi-interpretation interfaces. Each has a plausible mechanism. Each also adds time. Reverse prompting adds a turn, an AI-free stage adds a stage, parallel analysis runs the work twice. Nobody has priced any of them, and every one is a step a human has to agree to take, every single time.

Figma’s claim deserves a fair hearing too, and it’s narrower than the general debate it fed. In a product whose output is synthesis, the citation is the artifact. Strip the transcript line off the claim and you have an assertion nobody can check, which is a different output entirely. Some products sit in that class, where the dial doesn’t exist at all. That’s worth separating from a high-stakes setting on a dial that does.

So the question people are now asking, where exactly citation should switch from optional to mandatory by task risk and reversibility, is the second question. The first one is about the dial itself. When reasoning is available and optional, what share of users open it, and does that share move with the stakes of the decision or with how much time the person has? This month’s evidence hints at the answer and it isn’t stakes. The preference sharpens when immediate task completion is the goal, which is a property of the user’s clock. A designer’s stakes model and a user’s schedule are two different variables, and only one of them is holding the switch.

Both of those are cheap to answer for your own product, well before anyone settles them in general. Pull the open rate on your disclosure control this quarter and cut it by how fast the session was moving. Ask your platform team whether the agent layer still emits a trace or only a result. Nobody has published either number for any product, and you can have both for yours inside a month. Until then, what you’ve shipped is a default, and the default is off.