You are nine minutes into an agent run, and your database migration transcript has reached a few hundred lines of schema reads, generated SQL, and test output, all in the same gray monospace. Near the middle sits the one line that matters: "Note: a recent backup could not be confirmed for this table. Proceeding with the migration script." It has the same color, indentation, and typeface as everything around it. At the bottom, a green summary says the tests pass, and that is the only thing your eye catches before you approve.
The postmortem will say the warning was right there, but a warning that is displayed and a warning that is detected are different things, and only the second one protects anyone. Before we explain why, try the demo.
If one target reached you instantly while the other made you hunt, you have felt the mechanism this chapter rests on. Our essay The Human Factors turns that mechanism into the practice's first recommendation: make the warning impossible to miss.
What a single glance can and cannot detect
Before you consciously read anything, your visual system makes one fast parallel pass over the whole display and decides what your attention gets offered. This pass is called pre-attentive processing, and it follows strict rules.
- It detects a short list of features. Roughly a dozen attributes can pop out, including color, size, motion, and flicker. Everything else, including the meaning of words, waits for slow serial reading.
- It detects presence, not absence. A signal that appears can pop out. A missing element never does, because noticing an absence requires item-by-item search.
- It cannot combine features. A target defined by a combination, like the one red square among red circles and blue squares, shares every feature with its neighbors and will not pop. You have to hunt for it.
Heavy visual load suppresses detection. When visual short-term memory is loaded, people become measurably worse at detecting a stimulus in plain sight. A long agent transcript or a dense generated dashboard is exactly that kind of load, so the wall of output around your warning lowers the odds it is ever seen.
The research papers behind this part documented these rules failing in real products.
- A dense e-commerce homepage ran five competing grouping schemes that acted as mutual distractors, and wasted its one highlight color on elements that varied randomly in size, shape, and placement. No grouping ever formed, and spotting a deal collapsed into a search for absence.
- A stock-trading interface showed why detection matters: it is the first event in the cognitive chain, and whatever a user fails to detect they cannot reason about, verify, or act on.
Our opener, Why AI products need human factors, walks through the rest of that chain, and this chapter covers the first link.
Reserve a visual feature for each critical warning
The recommendation follows from those rules.
Decide the few signals a user must never miss, give each one its own visual feature, and let nothing else in your product use that feature.
For most AI products the must-never-miss list is short and stable: uncertainty in the model's output, the irreversible action, the boundary between generated and verified content, and the moment the system is about to act outside the user's view. Each signal needs three properties.
- Presence. The signal is a thing that appears, never a thing that is subtly missing.
- A single feature. One attribute different from everything around it, not a combination the parallel pass cannot compute.
- A monopoly. The feature means one thing, everywhere in the product.
Making something stand out only helps when you highlight what the user actually needs to act on, not what is easiest for you to compute. In a study of AI code completions, highlighting the tokens most likely to need editing made programmers faster and their edits more targeted, while highlighting the lowest-probability tokens gave no benefit over no highlighting at all.
How shipped products apply this
- Diff highlighting is the oldest example. In any code review tool, deletions render red and additions render green, so you register the shape and size of a change before you read a word of it. That is single-feature color pop-out working on a boundary that matters.
- Grammarly's tone detector makes a risk visible while you can still fix it. Once a draft passes about 150 characters, an emoji indicator appears in the corner of the text field showing how the message will likely sound, such as Confident, Formal, or Angry. A risky tone becomes a distinct visual object before you hit send instead of a regret that arrives after.
- Apple shipped the violation and then the fix. At launch, AI-generated notification summaries looked identical to real app notifications, so errors carried the source's authority. In December 2024 a summary of BBC News alerts falsely claimed a murder suspect had shot himself, and the BBC filed a formal complaint. In iOS 18.3, Apple paused summaries for news and entertainment apps, italicized every summary to set it apart from a standard alert, and added a Settings warning that the feature may contain errors. The fix was to give machine-generated text a feature of its own, the very thing it had launched without.
Build the reserved channel into your product
- Write the must-never-miss list. Walk every flow and record each caveat, uncertainty, irreversible action, and permission grant a user can encounter. If the list runs past a handful, you are recording nice-to-notice, not must-never-miss.
- Reserve one feature per signal. One color, one block style, or one icon, used for that signal and nowhere else. A color that also decorates promotional moments is not reserved.
- Add redundancy on the critical object. Pair the color with a position or a shape so the signal survives a degraded glance on a small screen, with split attention, in the sixth hour of the workday. The homepage analysis above failed here, with a highlight color that carried no second attribute.
- Pull warnings out of uniform output. Never inline a caveat in a wall of matching text. Give it whitespace, a border, or an interruption that stops the flow.
- Verify the variable before you highlight it. Confirm that whatever you are encoding predicts what the user must do next, the way edit-likelihood highlighting did and probability coloring did not.
Test your riskiest screen with a glance
This drill takes about 15 minutes. Screenshot the riskiest output your product can show, such as the agent about to act or the answer carrying a caveat. Flash it for about a fifth of a second at a colleague who has not seen it, or have them squint until the text dissolves, and ask them to write down what they noticed. Compare their notes with your must-never-miss list. Anything that matters but did not pop is your next design ticket, and anything that popped but does not matter is stealing from the budget. Rerun the test after the fix. It needs only a glance from the viewer, and it pairs well with the verify-first rule, which trains the human half of the same habit.
Chapter Summary
- A warning the product shows and a warning the user actually notices are two different things, and only the one the user notices protects anyone.
- In a fast glance, the eye catches a short list of features like color, size, motion, and flicker, and nothing else. The meaning of words waits for slow reading.
- A glance catches something that appears, not something that is missing, and it cannot pick out a target defined by a mix of features.
- A busy, crowded screen makes people worse at noticing a signal in plain sight, and a long agent transcript is exactly that kind of screen.
- List the few signals a user must never miss: uncertainty in the output, an irreversible action, the line between generated and verified content, and the moment the system acts out of view.
- Give each of those signals one visual feature of its own, something that appears rather than something missing, and let that feature mean one thing everywhere in the product.
- Pair the color with a position or shape so the signal still reads on a small screen, with split attention, late in a long day.
- Pull warnings out of uniform output with whitespace, a border, or a break in the flow, and only highlight something once you have confirmed it predicts what the user must do next.
- Attention is a fixed budget, so guard the reserved feature against emphasis inflation: give it an owner, and make every new claim on it replace an old one.
- Detection is only the first step. Once the signal lands the user still has to track the whole session, which is where Working memory: keep the session on the screen picks up, and the glance test also opens the checklist when you run the human factors audit.
Sources
- Treisman, A. (1986). Features and objects in visual processing. Scientific American, 255(5).
- Healey, C. G., Booth, K. S., & Enns, J. T. (1996). High-speed visual estimation using preattentive processing. ACM Transactions on Computer-Human Interaction, 3(2).
- Konstantinou, N., Bahrami, B., Rees, G., & Lavie, N. (2012). Visual short-term memory load reduces retinotopic cortex response to contrast. Journal of Cognitive Neuroscience, 24(11).
- Konstantinou, N., & Lavie, N. (2013). Dissociable roles of different types of working memory load in visual detection. Journal of Experimental Psychology: Human Perception and Performance, 39(4).
- Vasconcelos, H., et al. (2024). Generation probabilities are not enough: Uncertainty highlighting in AI code completions. ACM Transactions on Computer-Human Interaction.
- The graduate research papers behind this part (2013): a pre-attentive processing analysis of a dense e-commerce homepage and a visual sensory analysis of a stock-trading interface.
- Grammarly support documentation on the tone detector.
- Apple iOS 18.3 release notes and BBC reporting on AI notification summaries (2024 to 2025).