← Processing speed

Are online reaction time tests accurate?

Published 6 September 2026 · Last reviewed 6 September 2026 · Next review due 6 September 2027

Short answer: the number you are shown is real, and a large part of it is not you. Somebody measured exactly how much — by taking the human out of the experiment altogether — and the figure is bigger than almost anyone assumes.

The experiment worth knowing about

Preprint · not peer reviewed

In 2019 a group of researchers did something unusual. Instead of asking people to take reaction time tests and arguing about the results, they removed the person. A light sensor watched the screen. A mechanical actuator pressed the button. The machine was built to "respond" at a fixed, known 105 milliseconds every single time.

Then they ran that same rig against four different setups, 100 trials each, on one computer. Any difference in the recorded times could not be the participant, because there was no participant. Whatever came out other than 105 ms was the equipment.

Read the first row again. A phone touchscreen added 400 milliseconds to a response that actually took 105. The equipment contributed almost four times as much as the thing being measured. And that protocol was not only the least accurate — it was also the most variable.

The touchscreen is the setup a phone gives you by default.

Where the time actually goes

None of this is a defect in any particular test. It is what general-purpose consumer hardware does, and the same paper enumerates it:

• The screen. At 60 Hz, a flash can appear up to 17 ms before or after the moment the computer records as "stimulus shown".
• The operating system. Typically 2–20 ms of latency just from task scheduling, and more when something else is running.
• The input device. Drivers for identical hardware can differ by tens of milliseconds. Touchscreen-to-display lag ranges from 50 ms to 200 ms.
• All of it together. Hardware and software variation can introduce timing delays of up to 100 ms.

Why you cannot just subtract it

Preprint · not peer reviewed

The obvious response is to measure the lag once and take it off the top. That does not work, for two reasons the paper is explicit about.

The error is not a constant. It varies between devices, between drivers, and from moment to moment on the same machine depending on what else is happening. A single correction cannot remove something that keeps changing.

And more trials do not help. Averaging fixes random noise, not systematic bias. The authors put it bluntly: with a systematic offset, taking more samples only produces a result that is "more precisely inaccurate".

What this does and does not mean

It does not mean online reaction time tests are worthless. That would be the overcorrection, and it is not what the evidence supports. If you take the same test, on the same device, twice in a week, most of the equipment lag is common to both readings — so the comparison between them survives even though neither absolute number does.

What it does mean is that the absolute figure travels badly. Your 240 ms and someone else's 210 ms, measured on different devices, may differ entirely because one of you used a phone. Comparing your result to a published clinical figure is worse still, because those were collected on equipment built for the job.

There is a specific casualty worth naming. Researchers increasingly regard reaction time variability — how much your responses scatter — as more informative than the average. Its normal range is around 10 to 20 ms. The authors state plainly that this "simply cannot be measured by systems that have variable latency ranges of 84 msec". The most useful part of the measure is the part consumer hardware destroys first.

What we did about it

We build cognitive tests, so this was not an academic question for us. It decided what we built.

Our Processing Speed instrument is not a reaction time test. It is a symbol-matching task scored over ninety seconds. You match symbols to digits using a key that stays on screen, and the score is how many you complete correctly.

The reason is arithmetic. Ninety seconds is 90,000 milliseconds. A 40 ms input lag is 0.04% of that window — it moves the count by a fraction of one item. The same 40 ms sitting inside a single 240 ms reading is a sixth of the measurement.

The task is less dramatic than a reaction test. It is measurable on the hardware people actually own, which the reaction test is not.

If you want to use an online reaction test anyway

Plenty of reasons to, and the evidence suggests how to get something real out of it:

• Use the same device every time. Most of the lag then cancels between your own readings.
• Prefer a physical mouse or keyboard to a touchscreen — 80 ms of added latency against 400 ms, on the figures above.
• Compare yourself to yourself, and only that. Not to a friend on a different phone, and not to a number from a study.
• Ignore the last digit. Anything claiming single-millisecond meaning on consumer hardware is reporting precision it does not have.

Sources

• Holden, J. K., Francisco, E. M., Tommerdahl, A., Lensch, R., Kirsch, B., Zai, L., Pearce, A. J., Favorov, O. V., Dennis, R. G., & Tommerdahl, M. (2019). Accuracy of different modalities of reaction time testing: Implications for online cognitive assessment tools. bioRxiv 726364. biorxiv.org/content/10.1101/726364

Every figure on this page was read in the primary source, not in reporting about it, and is logged in our evidence ledger with the section it came from. The 400 ms, 80 ms and 30 ms figures are the four-protocol results; the 17 ms, 2–20 ms, 50–200 ms and 100 ms figures are from the same paper's account of where latency originates. The paper is a preprint on bioRxiv, and we say so rather than presenting it as peer-reviewed literature. If you think something here is wrong, tell us at corrections@brainscoredaily.com — we correct publicly and date the correction.