Understanding Human Intent
08.25.2026
In my previous article, I argued that commands aren't enough for embodied AI.
Here is why: in the physical world, user intent is rarely complete. Even when a person gives a clear instruction, the system still has to interpret what the user meant, what is happening around them, and how that situation changes the meaning of the instruction.
The design challenge is not only helping the robot form an appropriate interpretation of the user's intent, but also making that interpretation legible enough for the user to understand and correct if they wanted.
What can actually be missing from an apparently clear instruction?
The Gap Beneath the Signal
An embodied AI system should treat a user's declaration as a signal to weigh — not a command to execute immediately.
Imagine your child calls on the way to school and says, "Can you bring the box on my desk?" The instruction sounds clear until you enter the room and see five boxes on her desk.
“
Which one did she mean? Do you guess based on what you know about her class, her habits, and what she usually carries? Or do you call to verify?
When a robot receives the same instruction, it faces the same ambiguity — maybe even more. When the system hears "Bring the box on the desk to me," like you standing in your daughter's doorway, it's facing multiple unresolved questions at once.
“
Which of the five boxes? Does "to me" mean wherever the speaker is standing right now, or wherever they end up? And how is each box currently sitting — two are next to the monitor, three are below two books.
The system should form a probabilistic interpretation from available cues, including the user's behavior, interaction history, and the state of the environment.
In this case, what would you do in that doorway — pick one box based on your inference, bring all five and let her decide, or call to clarify which one she meant? Each choice resolves uncertainty differently — and places a different amount of effort, interruption, and risk on the human. That ambiguity shouldn’t remain an invisible system problem. I believe an important part of HRI design is shaping both the interaction logic around uncertainty and how that uncertainty is communicated: what evidence the robot is relying on, how confident its interpretation is, and when that uncertainty becomes significant enough to seek more context or clarification.
Then, Context Narrows the Ambiguity
If evidence is incomplete, the system should gather more context, form a tentative inference, or ask for clarification. Embodied AI needs both — the ability to infer, and the option to ask when inference alone isn't enough. The equation below can form more complete intent:
Declared intent (e.g., "bring me the box") + context → possible intent
Context may include:
Temporal: time of day, sequence of events
Spatial: where things are, where the user is
Behavioral: what the user was just doing
Historical: habits, preferences
Social: presence of others, norms
Go back to the box example.
If you notice that your daughter has science class that morning and remember that she built a small rocket from a plastic bottle the night before, the same instruction suddenly becomes easier to interpret.
The signal did not change. Your interpretation did.
Context is not simply more data for the robot. It’s part of the shared situation that helps human and robot arrive at a more grounded understanding. The challenge for design is to determine which context is relevant, how it should influence the robot’s interpretation, and how that interpretation can remain correctable when its assumptions are wrong.
Intent Appears in Layers
Context can narrow ambiguity, but it doesn't always provide the same degree of clarity. Some of what you know comes directly from what your daughter said. Some comes from what you can observe in the room right now. And some comes from what you remember about her recent behavior and routines. I find it useful to think about these as three layers of intent: declared, implied, and latent.
Declared is what she explicitly said: “Bring the box on my desk.” It provides the strongest direct evidence, but it still doesn't tell you which of the five boxes she means.
Implied intent comes from what you can infer from the situation in front of you. Perhaps one box is open on the floor, with a cut plastic bottle and tools beside it, suggesting that she was using it recently, while three others appear untouched. This evidence is not something she stated directly; its meaning depends on how you interpret the current context.
Latent intent draws on broader patterns and history. You remember that she built a bottle rocket the night before and that she has a science class today. Another box contains that project. Nothing in her instruction explicitly refers to it, but those contextual clues make it a plausible candidate.
Now the evidence conflicts. Her declaration points only to a location. The current scene suggests one box, while broader context suggests another.
Which evidence should the robot trust—and how much should it infer from each?
As intent moves from declared to implied to latent, explicit user evidence generally decreases while the amount of interpretation required increases. Action risk, however, depends on the action and situation — not on the intent layers alone. Bringing the wrong box to a kid may be a minor inconvenience. Bringing the wrong equipment to someone repairing a water pipe might not be.
Why You Can't Just Rely on Latent Intent
So, back to the question from the last section: which evidence should the robot trust—and how much should it infer from each? When declared, implied, and latent evidence point in different directions, the system should not silently collapse them into a single fact. The disagreement itself is evidence of uncertainty.
Contextual patterns can strengthen a latent inference, but they should not turn an assumption into a fact about the user. As repeated interaction provides more evidence — including the user’s corrections — those inferences may become more reliable, but they should still remain correctable by the user.
Understanding Is Not Permission to Act
Users will not always give embodied AI a fully specified task. The system must form an interpretation from incomplete signals, contextual evidence, and uncertainty. But a plausible interpretation does not automatically justify action.
Understanding is not permission to act.
Once the system thinks it understands what the user may want, the next question becomes: How much confidence is enough for how much action?
That's where the next chapter begins: decision.
Disclaimer: The daughter who called her mom in this story was me. And yes—I had pointed her to the completely wrong box.