Latency thresholds describe perception, not permission

Year 2 / Module 15 Interaction techniques and usability evaluation / Unit 6 / Latency thresholds describe perception, not permission

Latency thresholds describe perception, not permission

HCI2107 · Unit 6 of 10

The published response-time limits describe what users notice rather than what a system is allowed to do, and because the tolerance users bring is set by the fastest systems they already use, a delay figure is meaningful only against the interaction it disrupts and the baseline the user arrived with; the measured detection thresholds are far lower than the folk limits and depend on the task and form factor.

Latency thresholds describe perception, not permission

Three numbers govern how the field talks about responsiveness: 0.1 second for a system to feel instantaneous, 1.0 second for the flow of thought to stay uninterrupted, and 10 seconds for attention to remain on the dialogue (Nielsen, 1993). They are read as a budget a system may spend, and this sixth of ten units argues that the reading is wrong. The limits describe the user's state, not a system's entitlement, and the measured thresholds for noticing and for damage sit below the folk limits and move with the task, the form factor and the strategy the user adopts. What a delay costs depends on the interaction it interrupts and on what the user has already adapted to, which on a poor or intermittent connection is a great deal.

The folk limits come from a paper that refused to set them

Miller (1968) surveyed the controversy over response-time limits and concluded that it could not be resolved as posed, because two semantic issues stood in the way, the first being the question "response time to what?", since different purposes and actions have different acceptable response times. Nielsen (1993) cites Miller when presenting the three limits as advice unchanged for decades: at 0.1 second the system feels instantaneous, at about 1.0 second the delay is noticed and the feeling of operating directly on the data is lost, and at 10 seconds attention is at risk. His advice is about what to do while slow, which presupposes that slowness is expected.

The measured threshold is not one number, and it moves with the task

Ng and colleagues (2012) built a touch demonstrator delivering latencies from the commercial range of 50 to 200 milliseconds down to about 1 millisecond. Users greatly preferred the lower latencies, and noticeable improvement continued well below 10 milliseconds, a level difficult to reach commercially. The authors therefore propose immediate low-fidelity feedback followed by high-fidelity visuals. That perceptual threshold is two orders of magnitude below the 1 second limit.

Jota and colleagues (2013) studied dragging with a direct-touch device, noting that common commercial devices present up to 125 milliseconds of latency. Performance fell as latency increased, more severely when targets were smaller or farther away. Deber and colleagues (2015) found significant effects of both form factor and task, and they found that improvements which reduced but did not eliminate lag still had value. Long and Gutwin (2019) found differences between a touchscreen, a mouse, a controller and a drawing tablet in throughput and pointer paths. Targeting has been shown to be affected by latencies as low as 41 milliseconds. The pattern is a dependence on form factor, task and target geometry, so a latency figure is a measurement with conditions attached.

Yang and colleagues (2025) measured the threshold directly in virtual reality and put it between 130 and 170 milliseconds, above the tenth of a second Nielsen attached to feeling instantaneous. It fell when the interaction strategy emphasised speed or the task grew harder, the effect of latency on performance was nonlinear, and electroencephalogram signals registered latency that participants did not report. A threshold quoted as a property of a person is therefore a property of a task, a strategy and a measurement instrument at once.

What users notice and what harms them are two different thresholds

The cleanest evidence comes from an unusual task. Martens and colleagues (2018) had participants balance a ball on a beam under second-order control with added latencies of 14 to 198 milliseconds. Latency affected performance at the smallest added value of 14 milliseconds (d = 0.61). A larger controller gain increased the effect on control indicators (all d at or above 0.68) and on perceived control difficulty (d = 0.85). Most importantly, performance and perceived difficulty were affected before participants were aware of the latency. Nielsen's limits are defined by what the user feels, so the felt threshold is a late indicator. A test that asks only whether the user noticed measures the most forgiving threshold.

Responsiveness is more than one number

Pavlovych and Stuerzlinger (2009) characterised the latency and jitter of several common input devices, and crossed the two properties in an ISO 9241-9 pointing experiment. Latency had the stronger effect on performance, but jitter sharply increased the error rate, roughly in inverse proportion to target size, and filtering jitter adds lag, so a single responsiveness figure conflates properties that cannot be tuned independently. Warburton and colleagues (2022) show the same for the measurement: across four headsets their camera-based method found mean latency at the onset of a sudden movement of 21 to 42 milliseconds, functionally 2 to 13 milliseconds once motion prediction was accounted for, with sudden accelerations lowering spatial accuracy.

Delay belongs inside the difficulty model, not beside it

Delay does not cost a fixed amount, which is the formal reason the folk limits cannot be permissions. Hoffmann (1992) showed that a modified Fitts's law applies when a transmission delay separates control movement from feedback. Movement time at a constant index of difficulty rises linearly with delay, and the relation is multiplicative. MacKenzie and Ware (1993) measured target acquisition under four lag conditions. At 225 milliseconds of lag, movement times rose by 64 per cent and error rates by 214 per cent against zero lag, and a model in which lag multiplies the index of difficulty accounted for 94 per cent of the variance. If delay multiplies difficulty, the same delay costs more on a small distant target than on a large near one, and no single budget fits the whole interface.

Contested question: are the thresholds facts about people or budgets for systems?

The literature disagrees about what kind of object a threshold is, not about the measurements. The folk reading of Nielsen (1993) treats the limits as facts about perceptual ability and therefore as budgets: the user notices at a tenth of a second, tolerates a second and stays for ten, so a system may spend up to them. The measurement literature places detection between ten milliseconds in a direct-touch demonstrator and 130 to 170 milliseconds in a virtual reality task, and shows it shifting with form factor, task and strategy (Ng and colleagues, 2012; Jota and colleagues, 2013; Yang and colleagues, 2025). The folk numbers are about willingness to wait, and a willingness is not a perceptual fact.

A third position locates the binding threshold in the task. Beigbeder and colleagues (2004) monitored operational Unreal Tournament 2003 servers to deduce the latency players actually met, then ran controlled studies on maps isolating movement and shooting. Typical packet loss had no impact on performance or play quality. Typical server latency, while sometimes unpleasant, did not significantly affect the game's outcome. That regime sits far above the laboratory thresholds, and Martens and colleagues (2018) supply the mechanism in their evidence of adaptation.

What would settle the dispute is a design none of these papers runs: one task family at two levels of error tolerance, three latency levels from effectively zero to clearly perceptible, and the baseline experience manipulated between groups, with noticing and performance measured separately in the same participants. Martens and colleagues show damage below awareness but vary neither baseline nor tolerance; Beigbeder and colleagues vary the task but measure only the game's outcome; Deber and colleagues measure noticing but not damage.

The sides do not carry equal evidence. The folk reading rests on Nielsen's (1993) advice, which cites Miller (1968) for a controversy he left unresolved, while the measurement side rests on controlled experiments, Ng and colleagues (2012) on a touch demonstrator, Martens and colleagues (2018) on a second-order control task with effects from d = 0.61 upward, and MacKenzie and Ware (1993) on four lag conditions. The strongest objection is that the unit's own thesis cannot yet set a number, because a threshold that moves with task, path and adaptation yields no budget a designer can apply. A designer should measure noticing and performance on the actual task while the dispute is open.

Where the argument is heading

The direction is away from a single number and towards a profile with a measurement statement attached. Warburton and colleagues (2022) show motion prediction changing a headset's functional latency to between 2 and 13 milliseconds, so a report naming only a mean is incomplete by construction. Martens and colleagues (2018) make the user's history a variable, and Deber and colleagues (2015) show that improvements below detection retain value. The question is not how fast is fast enough, but fast enough for which task, on which input path, compared with what the user has adapted to.

The sources this unit rests on

Beigbeder, T., Coughlan, R., Lusher, C., Plunkett, J., Agu, E. and Claypool, M. (2004). The effects of loss and latency on user performance in Unreal Tournament 2003. Proceedings of NetGames 2004. doi:10.1145/1016540.1016556 Monitored operational game servers to deduce the loss and latency players really met, then ran controlled studies on maps isolating movement and shooting. Typical packet loss had no impact on performance or play quality, and typical server latency, while sometimes unpleasant, did not significantly affect the game's outcome.
Deber, J., Jota, R., Forlines, C. and Wigdor, D. (2015). How much faster is fast enough? User perception of latency and latency improvements in direct and indirect touch. Proceedings of CHI 2015. doi:10.1145/2702123.2702300 Compared direct and indirect touch for tapping and dragging across two experiments and measured the ability to perceive small latency changes. Perception showed significant effects of both form factor and task, and improvements that reduced but did not eliminate lag retained value.
Hoffmann, E. R. (1992). Fitts' law with transmission delay. Ergonomics. doi:10.1080/00140139208967796 Shows a modified Fitts's law applies when a transmission delay separates control movement from feedback, with movement time at constant index of difficulty rising linearly with delay. Movement time, difficulty and delay are related multiplicatively, so delay is not an additive tax on a fixed budget.
Jota, R., Ng, A., Dietz, P. and Wigdor, D. (2013). How fast is fast enough? A study of the effects of latency in direct-touch pointing tasks. Proceedings of CHI 2013. doi:10.1145/2470654.2481317 Studied dragging with a direct-touch device, noting that common commercial devices present up to 125 milliseconds of latency. Performance fell as latency increased, more severely for smaller or more distant targets, and a follow-up showed perception of latency in the land-on phase of the drag.
Long, M. and Gutwin, C. (2019). Effects of local latency on game pointing devices and game pointing tasks. Proceedings of CHI 2019. doi:10.1145/3290605.3300438 Compared a touchscreen, a mouse, a controller and a drawing tablet on targeting and interception under latency, analysing throughput and pointer paths. Targeting has been shown to be affected by latencies as low as 41 milliseconds, and one figure cannot describe four input paths.
MacKenzie, I. S. and Ware, C. (1993). Lag as a determinant of human performance in interactive systems. Proceedings of CHI 1993. doi:10.1145/169059.169431 Measured target acquisition under four lag conditions: at the highest lag tested, 225 milliseconds, movement times rose by 64 per cent and error rates by 214 per cent against zero lag. A model in which lag multiplies the index of difficulty accounted for 94 per cent of the variance and beat additive models.
Martens, J., Franke, T., Rauh, N. and Krems, J. F. (2018). Effects of low-range latency on performance and perception in a virtual, unstable second-order control task. Quality and User Experience. doi:10.1007/s41233-018-0023-z Balanced a ball on a beam with added latencies of 14 to 198 milliseconds. Latency affected performance at the smallest added value of 14 milliseconds (d = 0.61), a larger controller gain increased the effects on control (all d at or above 0.68) and perceived difficulty (d = 0.85), and performance fell before participants were aware of the latency.
Miller, R. B. (1968). Response time in man-computer conversational transactions. Proceedings of AFIPS 1968, fall joint computer conference, part I. doi:10.1145/1476589.1476628 Surveys the controversy about response-time limits and argues that two semantic issues prevent resolving it, the first being the question "response time to what?", since different purposes and actions have different acceptable or useful response times. The thesis of this unit is in that paper, decades before the numbers it is usually used to justify.
Ng, A., Lepinski, J., Wigdor, D., Sanders, S. and Dietz, P. (2012). Designing for low-latency direct-touch input. Proceedings of UIST 2012. doi:10.1145/2380116.2380174 Built a high-performance touch demonstrator delivering latencies from the commercial range of 50 to 200 milliseconds down to about 1 millisecond. Users greatly preferred lower latencies and noticed improvements well below 10 milliseconds, so they propose immediate low-fidelity feedback and high-fidelity visuals later.
Nielsen, J. (1993). Response times: the three important limits. In Usability Engineering. Morgan Kaufmann. https://www.nngroup.com/articles/response-times-3-important-limits/ States the three limits: 0.1 second for the user to feel the system is reacting instantaneously, 1.0 second for the flow of thought to stay uninterrupted, and 10 seconds for attention to stay on the dialogue. The page presents them as limits set by perceptual ability and recommends progress indicators past about 10 seconds.
Pavlovych, A. and Stuerzlinger, W. (2009). The tradeoff between spatial jitter and latency in pointing tasks. Proceedings of EICS 2009. doi:10.1145/1570433.1570469 Characterised the latency and jitter of several common input devices and crossed the two properties in an ISO 9241-9 pointing experiment. Latency had the stronger effect on performance, but jitter sharply increased the error rate, roughly in inverse proportion to target size, and filtering jitter adds lag.
Warburton, M., Mon-Williams, M., Mushtaq, F. and Morehead, J. R. (2022). Measuring motion-to-photon latency for sensorimotor experiments with virtual reality systems. Behav. Res. Methods. doi:10.3758/s13428-022-01983-5 Presents a camera-based, system-independent method that co-registers real and virtual controller movement, applied across four headsets. Mean latency at the onset of a sudden movement was 21 to 42 milliseconds and functionally 2 to 13 milliseconds once motion prediction was accounted for, and sudden accelerations lower spatial accuracy.
Yang, S., Yue, K., Gao, H., Guo, M. L., Liu, Y., Zhang, D. and Liu, Y. (2025). Dynamic Changes of Latency Perception Threshold in Virtual Reality. IEEE Transactions on Visualization and Computer Graphics. doi:10.1109/tvcg.2025.3596919 A Fitts's law task placed the latency perception threshold at 130 to 170 milliseconds, lowered when the interaction strategy emphasised speed or the task grew harder, and refined the model to include latency. Electroencephalogram signals registered latency that participants did not report.

Check your understanding & apply it · 4 questions, 1 essay

1A team signs off a fixed budget of 50 milliseconds for every operation, citing limits of 0.1, 1.0 and 10 seconds. Which diagnosis fits the unit?

2A study adds the same 100 milliseconds of lag to two pointing tasks, one on a large near target and one on a small distant target. What does the unit predict?

3Two papers report mean headset latency of 21 to 42 milliseconds and of 2 to 13 milliseconds. A buyer wants one comparable figure for the two devices. Which judgement fits the unit?

4A game studio measures the latency its players actually meet on live servers, finds no significant effect on match outcome, and concludes that latency does not matter to its users. Which diagnosis fits the unit?

5Essay A product team is deciding how much latency a drawing application may carry on a tablet, and proposes to sign off on a fixed budget of 50 milliseconds for every operation, on the grounds that users notice a delay somewhere between 0.1 and 1 second and that 50 milliseconds is well inside the published limits. Write a critique and an evaluation design. Set out what the team's reasoning assumes about the three limits in Nielsen (1993), and why Miller (1968) would reject the reasoning as posed. Say what the measurement literature establishes about detection thresholds, using Ng and colleagues (2012) for the commercial range and the sub-10 millisecond improvement, Jota and colleagues (2013) for the effect of target size and distance in a direct-touch dragging task, Deber and colleagues (2015) for the role of form factor and task, and Long and Gutwin (2019) for device differences. Use Martens and colleagues (2018) to argue the difference between what a user notices and what harms performance, and MacKenzie and Ware (1993) with Hoffmann (1992) to argue why a fixed budget is wrong when delay acts on the index of difficulty. Use Pavlovych and Stuerzlinger (2009) and Warburton and colleagues (2022) to say why a single figure is underspecified, and Beigbeder and colleagues (2004) to say what kind of task would tolerate far more. Design the study that would settle the question for this product, naming the conditions, the measures and the baseline you would manipulate, and state what the result would and would not license.

The team's reasoning fails at its first premise. Nielsen (1993) states the three limits as limits on the user's state: 0.1 second for the system to feel as though it is reacting instantaneously, 1.0 second for the flow of thought to stay uninterrupted, and 10 seconds for attention to remain on the dialogue. Those describe when a person notices and when they begin to disengage, not how much delay an operation may carry, and Miller (1968) refused exactly this move when he argued that the controversy cannot be resolved without asking response time to what, because different purposes and actions have different acceptable or useful response times. A budget of 50 milliseconds is therefore an answer to a question the team has not asked. The measurement literature shows what the numbers do and do not cover. Ng and colleagues (2012) found users greatly preferred lower latencies in a touch demonstrator, with noticeable improvement continuing well below 10 milliseconds against a commercial baseline of 50 to 200 milliseconds, so detection is far below the folk limits, while Yang and colleagues (2025) place the perceptual threshold in virtual reality at 130 to 170 milliseconds, moving with the strategy and the difficulty of the task. Jota and colleagues (2013) found performance in direct-touch dragging decreasing as latency increased, more severely for smaller and more distant targets, which means the same millisecond figure costs different amounts in different parts of a canvas. Deber and colleagues (2015) showed detection depends on form factor and task, and Long and Gutwin (2019) found different input paths behaving differently under the same latency. A single figure cannot describe that spread. Martens and colleagues (2018) matter most for the sign-off decision. In a second-order control task, latency of 14 milliseconds already affected performance, and performance and perceived control difficulty were affected before participants were aware of the latency. A test that asks users whether the drawing feels delayed measures the most forgiving threshold available. MacKenzie and Ware (1993) show why a fixed budget also fails arithmetically: at 225 milliseconds of lag, movement times rose 64 per cent and error rates 214 per cent, and a model in which lag multiplies the index of difficulty explained 94 per cent of the variance, which Hoffmann (1992) had predicted from a modified Fitts's law. Delay multiplies the cost of a difficult target rather than adding a constant, so a budget that is generous for a large brush stroke is tight for a small precise one. Pavlovych and Stuerzlinger (2009) add that latency and jitter have different effects, the second raising errors roughly in inverse proportion to target size while the filtering that removes it adds lag, and Warburton and colleagues (2022) show a latency figure is a property of device, movement phase and method, with motion prediction reducing a measured 21 to 42 milliseconds to a functional 2 to 13 milliseconds. Beigbeder and colleagues (2004) set the outer bound on tolerance: in a networked game, typical latency, while sometimes unpleasant, did not significantly affect the outcome, so tolerance belongs to the task. The study for this product should therefore be run inside the product's own task family. I would take three operations with different error tolerance, a precise tracing operation, ink stroke entry and canvas panning, and cross them with four latency levels: near zero, below the detection threshold, above it but within the commercial range, and clearly perceptible. Each participant would complete both a detection measure, comparing pairs of delays, and a performance measure, completion time and error, and I would manipulate the baseline experience between groups so that one group works with a near-zero-latency build before the test and another group arrives from a consistently slow build, which turns adaptation from an inference into a comparison. Reporting would name the device, the movement phase, the jitter as well as the mean, and the target sizes used. The result would license a latency profile for these three operations on this tablet with these targets and this baseline, and it would not license a single budget for the product, a general claim about perception, or any statement about other input paths such as a stylus with a different sampling rate. The sign-off artefact should be that profile, with the conditions printed beside it.