How accurate are meat thermometers: 1 to 2 degrees and 2 to 4 degrees are both correct, and neither source says why
Google's own AI Overview cites the USDA at two to four degrees and three other sources at one to two degrees, then leads with the smaller number without noticing they disagree. "Accurate" is standing in for four different claims: resolution, factory tolerance at one point, calibration drift, and how long you waited. A spec sheet only ever answers one of them.
Position two on this search is a Reddit thread titled by its own frustration: five thermometers, and not one of them reliably agrees with itself. Underneath it, a Facebook group for home smokers asks the plainer version of the same question: is a twenty dollar thermometer trustworthy, or are you paying the extra thirty dollars for something real. Google answers first, with a full AI Overview, and the overview answers by contradicting its own citations. One of the five sources it quotes, the USDA, says two to four degrees. Three others it quotes say one to two degrees. The summary leads with the smaller number and never mentions that its own sources disagree.
That is not rounding. Two to four degrees is double the width of one to two degrees, and a page that opens by quietly picking the smaller of two disagreeing numbers has already shown you that “accurate” is doing more than one job. Nobody on the page, including the summary, says which job.
This site has already worked through three of the ways a correct-looking thermometer produces a wrong-looking number: a dial that averages its whole immersed length, a probe whose stem steals or donates heat depending on what it is sitting in, and a reading taken before the number has finished moving. What has not been done yet is put those three in one place next to a fourth nobody here has named, and ask which one a single spec sheet is actually describing when it prints one number on the box.
What is tested here, and what is not
Analyzed. Sensor tolerance, display resolution, calibration drift and settling time are four separate engineering concepts with four separate causes. Keeping them apart is arithmetic, not opinion, and the reasoning below is worked from published tolerance concepts rather than assumed.
Tested. Not yet, and it rides on the same bench session already planned for calibration: a bag of ice, a pan of boiling water, every thermometer in the house. That session was already going to record what each one reads at two fixed points. This page adds one more figure at no extra cost: how long each one takes to stop moving. Method at How We Test.
Two claims wearing one word
The first claim is resolution: how finely the display divides its scale. A screen that shows 154.3°F is not more accurate than one that shows 154°F, it is only more precise about the number it is wrong by. One of the sources sitting in Google’s own summary makes this point directly, buried in a video description rather than in the text anyone actually reads: resolution is how many digits it shows you, accuracy is whether that number is true. A five dollar thermometer and a hundred dollar thermometer can share the same decimal point and disagree by ten degrees underneath it.
The second claim is accuracy proper: how close the display sits to the real temperature, and that is the number USDA and the university source are both trying to report. They disagree because a figure like “plus or minus two degrees” is a spec, and a spec is only meaningful attached to a stated condition. Which sensor. Checked at which point on the scale. Fresh from the factory, or a year and forty cooks later. Neither citation says, so both numbers are honest descriptions of something, and neither tells you which something.
Why one number checked at one point is not the same claim as accuracy across a range
Every thermometer on this site’s bench is to be checked at two points, not one: ice water at 32°F and boiling water at 212°F, corrected for altitude. That is not caution for its own sake. A sensor’s output is a curve, not a fixed offset, and a device that is trimmed or checked at a single point can be dead on there and several degrees off forty degrees away, because nobody looked at the slope. Two points fix two things: where the curve sits, and how steep it is. One point only fixes the first, and a spec sheet that quotes a single tolerance figure without saying at which temperature it was measured is describing the intercept and staying silent about the slope.
This is the likeliest reason the two numbers in Google’s own summary do not match. “One to two degrees” and “two to four degrees” can each be a perfectly good description of a real device measured a real way, and there is no way to tell from either citation which point they checked, how many they checked, or what population of thermometer sat behind the number: a five dollar dial, a fifty dollar digital probe, or a cabled probe built for a smoker rather than a kitchen. A figure that does not name its own conditions is not lying. It is answering a narrower question than the one being asked of it, which is the same complaint this site has already made about a lid dial, a leave-in probe, and now the box the instrument came in.
The three other ways a correct instrument can show you a wrong number
Sensor tolerance is only the first of four. A dial thermometer does not read one point, it reads the average temperature along its whole immersed stem, which is why the lid dial on a smoker can disagree with a grate-level probe by forty degrees without either one being defective. That is not a tolerance problem. It is an averaging problem wearing a tolerance-shaped complaint.
A probe left in place has its stem sitting in whatever surrounds it, and heat moves along that stem toward whichever end is colder. The same probe reads low in your hand and reads high in a smoker, for the same reason a poker left half in a fire feels hot at the end you are not holding. That is not a tolerance problem either. It is a conduction problem, and it does not appear on any spec sheet.
And a reading taken before the display has settled is neither of those. It is simply early. The calibration page already says to wait for the number to stop moving before trusting it. This page is about how long that actually takes, and whether it has anything to do with how many decimal places the screen bothered to print.
| The complaint | What is actually wrong | What fixes it |
|---|---|---|
| Screen shows three decimal places | Resolution, not accuracy: a precise display on an unverified sensor | Check it against a known point, not the digit count |
| Spec sheet says ±2°F | True at one checked point, unknown slope elsewhere on the scale | A two-point check: ice and boiling, not one |
| Dial disagrees with a probe by 40°F | Averaging along the immersed stem, not sensor error | Know which instrument is reading a length and which reads a point |
| Reading climbs for ten seconds, or runs high all day | Settling time or stem conduction, not sensor error | Wait it out, or know which way the stem is leaking heat |
None of these four is fixed by buying a more expensive thermometer with more decimal places. Three of the four are fixed by knowing what the instrument is doing. The fourth, sensor tolerance, is the only one a spec sheet describes, which is presumably why it is the only one anybody advertises.
What we are measuring
The size of each of these four is what decides whether any of it matters at the grate. A one degree resolution complaint is trivia. A ninety second settling time on an instrument read four seconds after insertion is a brisket pulled at the wrong number.
The prediction goes down before the ice melts, because a prediction recorded afterward is worth nothing.
Every thermometer in the house, two fixed points, one stopwatch: planned, not yet recorded
Protocol. The same ice-water and boiling-water bench already planned for the calibration page, extended by one instrument: a stopwatch. Every thermometer in the house goes into the ice bath together, and the time from insertion to a settled reading is logged for each, alongside the final value calibration already asks for. Repeated in the boiling water, altitude noted. Where a device’s display resolves finer than one degree, that resolution is written down too, next to its own settling time and its own final error, so the three can be compared to each other instead of assumed to travel together.
The figures and the photograph are published once this protocol has been run on the bench. Until then this page carries the method, not the numbers.
What to actually do
Ignore the decimal places. A display that shows tenths of a degree is a marketing choice, not a claim about the sensor behind it, and the only way to know what that sensor is actually doing is to check it, the same way the calibration page already describes.
Check it at two points if you can manage both, not one. A single ice-water check catches a sensor that has drifted badly and misses one that is fine at 32°F and several degrees off at 200°F, which is exactly the range that matters for meat.
If two checked thermometers disagree, do not default to trusting the pricier one. Ask instead which stem is doing something unusual: buried the wrong depth, sitting half in the fire, read four seconds after insertion instead of thirty. Most disagreements between two already-checked instruments are not a sensor problem at all. They are one of the other three.
Where it fits
This page is about what “accurate” is quietly standing in for. The bench procedure that actually checks it is how to calibrate a meat thermometer. The averaging error specific to a lid-mounted dial is lid thermometer against grate probe. The conduction error specific to a probe left in the meat is do you leave the meat thermometer in while cooking. Why the chamber number will not hold still in the first place is offset smoker fire management. The burn that happens before any food does, where a chamber’s own gradient shows up on the metal, is how to season an offset smoker. What is actually going into all of this measuring is brisket. How this site keeps measurement apart from reasoning is on How We Test, and who is running the pit is on About.
