The SAT Math Question Types AI Struggles With, and How to Ask Better

Why AI makes plausible errors

A general language model is a next word predictor. It is trained to produce text that reads the way a correct explanation reads, which is not the same thing as being correct. On an SAT Math problem, that distinction matters. The model can lay out a tidy solution, label each step, and arrive at a number that looks like the kind of answer the question wants, all while a sign flipped or a unit slipped somewhere in the middle.

This is why a wrong answer from AI rarely looks wrong. It does not stammer or hedge. It commits, because fluent committed text is what it was built to generate. The failure mode is not confusion, it is confident fluency wrapped around a bad computation. For a longer walkthrough of how this plays out in a real chat session, see our piece on using ChatGPT for the SAT math test.

The good news is that the mistakes are patterned. They cluster in a handful of question types, and once you know the list you can watch for them. The categories below are the usual suspects, and each comes with the small move that catches it.

Unit traps

Word problems love to change units mid sentence, and AI does not always notice. A rate might arrive in feet per second while the question asks for a distance in inches, or a problem might quietly switch from cost per item to cost per dozen. The model reads all of this as ordinary text and can carry the wrong unit straight to the final line.

• Feet versus inches, meters versus centimeters. A length converted once too few or once too many times, giving an answer off by a factor of twelve or a hundred.
• Percent of the wrong base. Taking a percent of the original price when the question asks for a percent of the discounted price, or the reverse.
• Rates and time. Mixing minutes and hours, so a speed reads as reasonable while it is sixty times off.

When you hand AI a unit heavy problem, ask it to write the units on every line. Forcing the units into the open is the fastest way to make a hidden conversion visible.

Inequality direction

The single most common algebra slip is the inequality flip. When you multiply or divide both sides of an inequality by a negative number, the direction of the inequality reverses, and a fluent explanation can breeze right past that rule. The model often keeps the symbol pointing the same way because, in plain text, keeping it the same reads just as smoothly as flipping it.

• Dividing by a negative. For example, dividing by a negative flips the inequality, so if negative two times x is less than six, then x is greater than negative three, not less than.
• Multiplying by a negative variable. Trickier still, because the sign of the variable itself may be unknown, which sometimes forces the problem to split into cases.
• Taking reciprocals. Another place the direction can change and quietly get missed.

Ask the AI to state the rule it is using at the exact step where it divides or multiplies. If it cannot name the rule, that is your signal to check the line by hand.

Graph and figure interpretation

Here is a hard limit worth understanding: a text only model cannot see a figure you did not give it. If a question refers to a graph, a scatterplot, or a diagram and you paste only the words, the model has nothing to read. It will still answer, because it will guess at what the figure probably shows, and a plausible guess about an invisible graph is a coin flip.

• Missing graphs. Line and parabola questions where the key information lives entirely in the picture.
• Scatterplots and lines of best fit. Trend and prediction questions that depend on points the model never received.
• Tables read as prose. A model can misalign rows and columns when a table is pasted in as loose text.

If the question leans on a figure, the model needs the figure, not a paraphrase. This is one of the clearest cases where a tool that actually reads the question image has a real edge over pasting text into a general chatbot.

Probability wording

Probability questions on the SAT are usually about reading a two way table carefully, and the wording controls everything. Small words like given, among, and of those quietly shrink the denominator, and AI can miss the shift. The difference between the probability of an event and the probability of that event given a condition is the difference between dividing by the whole table and dividing by a single row.

• Given. A conditional, which restricts you to one row or one column of the table.
• Among, of those. Phrases that narrow the group you are counting from, which changes the denominator.
• Overall probability. Uses the full total, a different denominator again.

When you check a probability answer, name the denominator out loud before anything else. You can drill this exact reading skill in the probability section until the wording stops tricking you.

Function notation

Function notation trips people and models alike because it packs a lot into a little. Writing f of 3 means the output when the input is 3, and plenty of errors come from treating the input as the output or the other way around. Transformations add a layer: a change inside the parentheses shifts the graph one way while a change outside shifts it another, and the notation barely moves.

• Input versus output. Solving f of x equals 5 asks for the input, while evaluating f of 5 asks for the output. Opposite questions.
• Inside versus outside the function. A change inside the parentheses moves the graph horizontally, a change outside moves it vertically, and the horizontal shift goes the opposite direction from what people expect.
• Composed functions. Working f of g of x has to happen from the inside out, and a fluent explanation can reverse the order.

This notation shows up all over higher level work, so it helps to practice where it is densest, in the nonlinear functions section.

Geometry diagrams

Geometry questions come with a standard warning: figures are not necessarily drawn to scale. AI has a habit of forgetting that. It can assume an angle that looks like a right angle is a right angle, or that two segments that appear equal are equal, when the only things you may use are the values and relationships actually stated in the problem.

• Assumed right angles. Treating a corner as ninety degrees because it looks square on the page.
• Assumed equal lengths or angles. Reading equality off the drawing instead of off the given information.
• Eyeballed measurements. Estimating a length from the apparent size of the figure rather than computing it.

Only the stated relationships count. Work the geometry from the givens and the theorems, and check the reasoning against the geometry and trigonometry material when a step feels like a guess.

Better prompt templates

You cannot rewire a general model, but you can change how you ask, and better prompts cut the error rate a lot. The fixes are the same across models, so these four moves work no matter which chatbot you are using.

• Ask for a hint first, not the answer. Try, give me only the first step and the skill being tested, then stop. A hint keeps you solving and gives the model less room to commit to a wrong final number.
• Name the tested skill. Tell it, this is a probability question about a two way table, or this is an inequality. Naming the skill focuses the response and surfaces the exact rule that tends to get missed.
• Have it check each step. Ask it to verify each line before continuing and to state the rule it used. Step by step verification catches sign flips and unit slips at the moment they happen.
• Verify the arithmetic with Desmos. Whatever the model says the answer is, recompute the key line in Desmos and compare it against the answer choices. Desmos does not do fluent, it does correct.

None of this makes a general chatbot infallible. It makes you the reliable checker, which is the right division of labor. If you would rather skip the prompt engineering, a purpose built SAT math AI tutor handles most of these moves for you.

How a SAT specific tutor changes the math

A general chatbot has two structural problems on SAT Math: it can produce a confident wrong final answer, and it cannot read a figure it was never given. Satified's tutor is built to remove both. It anchors every explanation to the question's independently verified answer, so the confident wrong final answer simply does not appear, and it reads the question figure directly instead of guessing at an invisible graph.

What it does not do, and what nothing honestly should claim to do, is free you from checking arithmetic. You still recompute the key line yourself, because a student who verifies is worth more than any tool that promises you never have to. The tutor narrows the errors to the ones you can catch, and Desmos catches those.

Questions students ask

What SAT Math does AI get wrong most?

Unit conversions, inequality direction, reading figures, probability wording, and function notation are the usual trouble spots.

Why does AI sound confident when it is wrong?

It predicts fluent text, which can look correct even when the math is off, so always verify the result.

How do I catch an AI math mistake?

Recompute with Desmos, compare the final answer to the choices, and ask the AI to justify the step you doubt.

Can better prompts fix it?

Prompts help a lot, ask for a hint first, name the skill, and check each step, but you still verify the arithmetic yourself.

Is a SAT specific tutor more accurate?

It removes the confident wrong final answer by anchoring to the verified answer and reading the figure, but you should still check arithmetic.

Keep going

Put the checks to work, or read the next piece.