What the examiners said about IB Maths AI HL — and exactly what to do about it
Built from five of the IB's own AI HL subject reports (May 2024 TZ2, November 2024, May 2025 TZ1 and TZ2, November 2025), read in full, covering Papers 1, 2 and 3 and the exploration. They were supplied by the site owner; the IB makes subject reports available to IB World Schools through its Programme Resource Centre. We have not seen every timezone's report for May 2024. Everything here is in our own words: we do not reproduce the reports or any exam question, and every worked example is our own. The question topics are our own matching from the reports' comments (parts whose topic we could not tell, or that fall outside our topic list, are marked "not mapped"); we show no marks per question because we did not use the markschemes.
Summaries are in our own words with links to the originals. Predictions are informed guesses, not a guarantee. You'll meet these notes as you revise too: under practice questions, inside mark schemes, in your AI marking and in “What to do next”.
Ranked for May 2027
What to focus on for May 2027
November 2026 candidates can use the same list. Ranked by how often each area appeared and how weakly it was answered.
Informed prediction, not a guarantee: any topic on the specification can be examined.
1
Statistics: hypothesis tests and combined distributions
Hypothesis testing appears in every session (often on Paper 3) and the reports repeatedly criticise hypotheses, test choice and conclusions; combining normal variables and binomial set-up lose marks every time.
Calculus: differential equations, volumes, rates in context
Euler's method, coupled systems, separable equations and volumes about the y-axis are named as difficult in four or five sessions, and the reports call for more algebraic skill in calculus.
Markov chains, finance, complex numbers and matrices
Transition matrices, TVM finance and complex/matrix work recur each session, with marks lost on set-up (columns, signs, periods) rather than the calculation.
The 10 mistakes that cost the most marks (and 9 more)
Ordered by how often and how widely they were reported. The first 5 are free in full; Pro unlocks the fix and a worked example for every other one.
#1Accuracy: premature rounding, 2 s.f. answers and ignored instructions
What examiners saw: Every AI HL report. Unsupported 2 s.f. answers; intermediate values rounded and reused (the May 2025 TZ2 Paper 2 report says this is now the biggest accuracy problem, especially in probability chains); answers not given to the 2 decimal places, nearest dollar or whole years asked for; and, at the other extreme, every calculator digit copied into a final answer. When a question gives an answer, show your value to more figures than the printed one.
Fix: Use the full GDC value in every later step, round once at the end to 3 s.f. (or as instructed), and write the unrounded value first. Round, don't truncate. Give a sensible level of accuracy for the context.
Worked example (our own): Heights are N(170, 8²). Find p = P(height > 185), then the probability that at least 2 of 20 people are taller than 185 cm.
#2Hypothesis tests: hypotheses in words, the wrong test, and conclusions that don't follow
What examiners saw: All five sessions. Hypotheses written in words that mix up sample and population (or omit the population mean); a two-tailed alternative for a one-tailed test; a two-sample test where a paired test was needed (and its assumptions not stated); a z-test instead of a t-test; χ² goodness-of-fit degrees of freedom not reduced for an estimated parameter; cells with expected frequency below 5 not combined; the binomial test for a proportion hardly known; a correct comparison followed by the opposite conclusion; and confidence intervals without inequality notation.
Fix: Write H₀ and H₁ with symbols for population parameters (μ, ρ, p). Choose the test from the design (paired data → paired t; σ unknown → t). State the p-value, compare it with the stated significance level, then conclude in context. For goodness of fit: df = cells − 1 − parameters estimated.
Worked example (our own): A Poisson model is fitted to count data, with the mean estimated from the data; after combining small cells there are 5 categories. State H₀ and the degrees of freedom.
Loses marks
H₀: the mean is 2.1
df = 5 − 1 = 4
Earns the marks
H₀: the data can be modelled by a Poisson distribution
df = 5 − 1 − 1 = 3 (one parameter estimated)
Why: Goodness of fit tests the whole distribution, not the mean, and each estimated parameter removes a degree of freedom.
#3HL algebra: product rule, brackets, logs and exact solutions
What examiners saw: Every HL report stresses that the course is not only calculator work. Candidates could not apply the product rule, expand brackets or find a common denominator, confused (−x)² with −x², forgot ± when square-rooting, could not turn ln w = a ln d + b into a formula for w, failed to simplify the reciprocal of a sum of fractions, and did not use the quadratic formula when exact values were asked for.
Fix: Practise the core manipulations without the GDC: product and chain rules, brackets, log laws, exponentiating both sides, exact quadratic solutions. Then use the GDC to check.
Worked example (our own): A log–log regression gives ln w = 0.4 ln d + 1.2. Write w in terms of d.
Loses marks
w = 0.4 ln d + 1.2
or w = e^(0.4) d + e^(1.2)
Earns the marks
w = e^(0.4 ln d + 1.2) = e^(1.2) × e^(ln d^0.4)
w = e^(1.2) d^0.4 ≈ 3.32 d^0.4
Why: Exponentiate the whole right-hand side: a sum in the exponent becomes a product, and 0.4 ln d = ln d^0.4.
#4IB notation: AB as a length, vectors in columns, f⁻¹ vs f′, the variables asked for
What examiners saw: A "perennial problem" in the reports' words: in May 2025 at least a third of candidates read AB (a length) as the vector from A to B. Vectors were written as coordinates, position vectors used where displacement vectors were needed, the inverse-function symbol used for a derivative (and f⁻¹ confused with 1/f), answers given in x and y when the question used other variables, and "h =" left off an equation.
Fix: AB (no arrow) is a length; AB with an arrow is a vector, written as a column. Displacement AB = b − a. f⁻¹ is the inverse, f′ the derivative. Answer in the question's variables.
Worked example (our own): A = (1, 2, 0) and B = (4, 6, 0). Find AB and the vector from A to B.
Loses marks
AB = (3, 4, 0)
AB→ = 5
Earns the marks
AB = √(3² + 4² + 0²) = 5 (a length)
AB→ = b − a, the column vector with components 3, 4, 0
Why: Without an arrow, AB means the length of the segment. Vectors are columns, not coordinates.
#5Finance on the GDC: signs, months vs years, P/Y and C/Y, and the rounding asked for
What examiners saw: Three sessions: PV and PMT with the same sign, N = 5 (years) for 60 monthly payments, the compound-interest formula used where regular payments made it invalid, P/Y and C/Y confused (including "quarterly"), not all TVM entries shown, and answers not given to the nearest dollar or 2 decimal places as required.
Fix: List every TVM entry (N, I%, PV, PMT, FV, P/Y, C/Y). Money out and money in have opposite signs. N = years × payments per year. Round the final answer as the question says.
Worked example (our own): A $24 000 car loan is repaid monthly over 5 years at 4.8% per year, compounded monthly. Find the monthly payment to the nearest cent.
#6Graph algorithms: Prim vs nearest neighbour, no method shown, bounds mixed up
What examiners saw: Prim's algorithm and the nearest-neighbour algorithm were confused; Prim's was done "by eye" with no order of edges (the vertex order does not show the algorithm); the stopping rule was not stated; the cycle did not return to the start or repeated vertices; a spanning tree was treated as a path; the best upper bound was not the smallest, and the deleted-vertex lower bound was not understood. The reports call method "all important" in algorithms.
Pro: the fix and a worked example (wrong vs right method) for this mistake are in Pro. See plans
#7Markov chains: columns, the initial state, the right power of T
What examiners saw: Candidates built transition matrices whose columns did not sum to 1 or whose entries were inconsistent, did not know the initial state vector, used the wrong power of T (counting days wrongly), multiplied two powers of T when a two-stage path was needed, left a steady state as a matrix instead of a vector, and used a transition matrix where simple probability was enough. "To which percentage" was read as a change in percentage points.
Pro: the fix and a worked example (wrong vs right method) for this mistake are in Pro. See plans
#8Volumes of revolution: about the y-axis, π, squaring, limits
What examiners saw: Four sessions: the function integrated without squaring (an area, not a volume), π omitted, y² integrated with respect to x when rotating about the y-axis, limits taken from the wrong axis or not changed after a transformation, an upper limit wrongly assumed, and a composite volume taken as a simple sum.
Pro: the fix and a worked example (wrong vs right method) for this mistake are in Pro. See plans
#9Sketches from the GDC: domain, endpoints, turning points, angle mode
What examiners saw: Sketches extended beyond the domain, did not start at the right point, missed a turning point hidden by the window, put endpoints away from the axis as the screen showed, and were drawn in degree mode (so a trig graph came out almost straight). Complex-number and calculus answers were given in degrees, and zero acceleration was confused with where a velocity graph crosses the axis.
Pro: the fix and a worked example (wrong vs right method) for this mistake are in Pro. See plans
#10Distributions: variances add, the normal is continuous, the binomial coefficient
What examiners saw: Candidates added standard deviations (or subtracted variances) when combining normal variables, halved a standard deviation instead of scaling the variance, rewrote a strict inequality for a normal variable using the next whole number, as if the variable were discrete, left out the binomial coefficient, forgot to square a probability for two independent events, and used binomial and Poisson models the wrong way round. When variances are given, they are simply added.
Pro: the fix and a worked example (wrong vs right method) for this mistake are in Pro. See plans
#11Derivatives in context: not differentiating, gradient and angle, rates
What examiners saw: Candidates did not differentiate to find a normal (and worked back from the given answer), described a derivative as an average instead of an instantaneous rate, found dy/dt and dx/dt and stopped in a related-rates part, used f′ = 0 for a point of inflection, solved dy/dx = 0 when the question asked for something else, and rarely used "gradient = tan of the angle with the x-axis" to find an angle between lines.
Pro: the fix and a worked example (wrong vs right method) for this mistake are in Pro. See plans
#12Euler's method: formula copied, not applied, and GDC variables
What examiners saw: Copying the booklet formula earns nothing; many wrote little except a final answer (full marks if right, none if wrong), used calculator notation, or wrote the iteration in the GDC's default variables. For systems and second-order equations the reports ask for the iteration in the question's variables, at least the first step shown, and the answer.
Pro: the fix and a worked example (wrong vs right method) for this mistake are in Pro. See plans
#13Coupled systems and phase portraits: equilibria, eigenvectors, trajectories
What examiners saw: Candidates thought equilibrium meant the two derivatives were equal (it means both are zero), could not write a system in matrix form, made errors finding eigenvectors, confused the vector solution with a single function, and drew spirals or saddles where the eigenvalues said otherwise. Phase portraits were named as difficult in two sessions.
Pro: the fix and a worked example (wrong vs right method) for this mistake are in Pro. See plans
#14Separable differential equations: the constant, the sign, and simplifying
What examiners saw: Candidates wrote the separated equation without integral signs, integrated 1/(2x) as ln 2x, reused one constant for two different integrations, omitted a minus sign, found the constant from the wrong condition (L = 0 instead of L = L₀), left C unevaluated, and could not simplify the reciprocal of a sum of fractions to finish.
Pro: the fix and a worked example (wrong vs right method) for this mistake are in Pro. See plans
#15Modelling: regression on the GDC, r² for non-linear models, context of the variable
What examiners saw: Candidates fitted curves through two chosen points instead of using regression, forgot to subtract a background value before fitting, compared non-linear models with r instead of R² (and without quoting values), gave a model in x and y instead of the question's variables, misread what t measured (years after a start date), and compared models on accuracy when the question asked something else.
Pro: the fix and a worked example (wrong vs right method) for this mistake are in Pro. See plans
What examiners saw: Candidates multiplied complex numbers without the middle terms, did not equate real and imaginary parts, worked in degrees, did not use the GDC for modulus-argument sums, confused a transformation with its matrix, and did not recognise a product of matrices as a reflection. Some centres had not taught reflection and rotation matrices at all.
Pro: the fix and a worked example (wrong vs right method) for this mistake are in Pro. See plans
What examiners saw: In every session the reports describe whole questions left blank or topics not taught (graph theory, differential equations by the matrix method, the test for a proportion, reflection and rotation matrices), and too many candidates scored very low marks. The reports conclude that teaching SL plus a few extras is not enough preparation for an HL paper.
Pro: the fix and a worked example (wrong vs right method) for this mistake are in Pro. See plans
#18Generic statements on Paper 3, ignoring the help given, giving up
What examiners saw: Paper 3 is not a longer Paper 2: it tests links between parts and interpretation. Candidates wrote general statements that did not engage with the scenario (which score nothing), ignored the guidance lines before parts, did not reuse information given earlier, and stopped after one stuck part instead of re-entering with the given value.
Pro: the fix and a worked example (wrong vs right method) for this mistake are in Pro. See plans
#19Exploration (IA): templates, invalid tests and maths that doesn't serve the aim
What examiners saw: All five IA sections: templated explorations (a Voronoi diagram followed by a travelling salesman route) rarely score well and risk malpractice; statistical tests are applied without checking their conditions, and χ² is used on quantities rather than frequencies or to test a curve's fit; higher-order polynomials are chosen for R² alone; mathematics that does not serve the aim earns no credit, however correct; and reflection is descriptive, with AI-generated lists of limitations adding nothing.
Pro: the fix and a worked example (wrong vs right method) for this mistake are in Pro. See plans
How often each topic appeared, and for how many marks
Every question in the papers we read, matched to a topic. Columns marked * count questions (q) because the marks could not be verified from the mark scheme. Hover or tap a cell to see the questions.
7q (Paper 1 (TZ2) Q7, Paper 1 (TZ2) Q10, Paper 1 (TZ2) Q11, Paper 1 (TZ2) Q13, Paper 2 (TZ2) Q2, Paper 2 (TZ2) Q3, Paper 3 (TZ2) Q2)
6q (Paper 1 Q14, Paper 1 Q16, Paper 1 Q17, Paper 2 Q2, Paper 2 Q7, Paper 3 Q2)
14q (Paper 1 (TZ1) Q7, Paper 1 (TZ1) Q8, Paper 1 (TZ1) Q12, Paper 1 (TZ1) Q15, Paper 2 (TZ1) Q3, Paper 2 (TZ1) Q6, Paper 2 (TZ1) Q7, Paper 1 (TZ2) Q7, Paper 1 (TZ2) Q9, Paper 1 (TZ2) Q12, Paper 1 (TZ2) Q15, Paper 2 (TZ2) Q2, Paper 2 (TZ2) Q3, Paper 2 (TZ2) Q6)
8q (Paper 1 Q7, Paper 1 Q10, Paper 1 Q14, Paper 1 Q15, Paper 2 Q2, Paper 2 Q3, Paper 2 Q6, Paper 3 Q2)
5q (Paper 1 (TZ2) Q8, Paper 1 (TZ2) Q12, Paper 2 (TZ2) Q3, Paper 2 (TZ2) Q4, Paper 3 (TZ2) Q2)
7q (Paper 1 Q1, Paper 1 Q5, Paper 1 Q11, Paper 2 Q2, Paper 2 Q5, Paper 2 Q7, Paper 3 Q1)
11q (Paper 1 (TZ1) Q5, Paper 2 (TZ1) Q2, Paper 2 (TZ1) Q4, Paper 2 (TZ1) Q5, Paper 2 (TZ1) Q7, Paper 1 (TZ2) Q4, Paper 1 (TZ2) Q8, Paper 2 (TZ2) Q3, Paper 2 (TZ2) Q4, Paper 2 (TZ2) Q6, Paper 3 (TZ2) Q2)
5q (Paper 1 Q4, Paper 1 Q12, Paper 1 Q13, Paper 2 Q7, Paper 3 Q1)
5q (Paper 1 (TZ2) Q4, Paper 1 (TZ2) Q9, Paper 2 (TZ2) Q6, Paper 2 (TZ2) Q7, Paper 3 (TZ2) Q1)
4q (Paper 1 Q3, Paper 1 Q15, Paper 2 Q1, Paper 3 Q2)
10q (Paper 1 (TZ1) Q4, Paper 1 (TZ1) Q6, Paper 1 (TZ1) Q9, Paper 2 (TZ1) Q1, Paper 3 (TZ1) Q2, Paper 1 (TZ2) Q3, Paper 1 (TZ2) Q6, Paper 1 (TZ2) Q13, Paper 2 (TZ2) Q5, Paper 3 (TZ2) Q1)
8q (Paper 1 Q1, Paper 1 Q3, Paper 1 Q9, Paper 1 Q16, Paper 2 Q1, Paper 2 Q3, Paper 2 Q5, Paper 3 Q1)
6q (Paper 1 (TZ2) Q1, Paper 1 (TZ2) Q3, Paper 1 (TZ2) Q5, Paper 1 (TZ2) Q15, Paper 2 (TZ2) Q5, Paper 2 (TZ2) Q6)
6q (Paper 1 Q4, Paper 1 Q7, Paper 1 Q9, Paper 2 Q3, Paper 2 Q6, Paper 3 Q1)
9q (Paper 1 (TZ1) Q11, Paper 1 (TZ1) Q13, Paper 1 (TZ1) Q14, Paper 3 (TZ1) Q1, Paper 1 (TZ2) Q1, Paper 1 (TZ2) Q5, Paper 1 (TZ2) Q16, Paper 2 (TZ2) Q1, Paper 2 (TZ2) Q7)
4q (Paper 1 Q2, Paper 1 Q5, Paper 1 Q16, Paper 2 Q7)
3q (Paper 1 (TZ2) Q14, Paper 2 (TZ2) Q1, Paper 2 (TZ2) Q6)
4q (Paper 1 Q2, Paper 1 Q10, Paper 1 Q12, Paper 1 Q13)
8q (Paper 1 (TZ1) Q2, Paper 2 (TZ1) Q2, Paper 2 (TZ1) Q3, Paper 1 (TZ2) Q2, Paper 1 (TZ2) Q4, Paper 1 (TZ2) Q11, Paper 1 (TZ2) Q14, Paper 1 (TZ2) Q15)
4q (Paper 1 Q8, Paper 1 Q11, Paper 2 Q3, Paper 3 Q2)
4/4
Confidence: high for the series shown — every question was read and mapped by hand. "Series" = how many of the 4 series examined the topic.
Paper by paper
Every paper we analysed, question by question
Topics are ours (matched from each report's question-by-question comments). We don't reproduce the questions: open the official paper from our past-paper page.
Reading trajectories on a slope field; Eigenvalues & PageRank-style centrality
Pro
Your mistakes vs the examiners' hot-spots
We set the topics in My mistakes and your practice scores against this page's mistakes library and focus list, and rank where they overlap, with a practice link for each.
Free account: how many of your notebook entries match, and the top one. Pro: the full ranked list, your focus areas and your risk list.
A one-page PDF of the 10 most common mark-losing mistakes and how to avoid them. Print it and stick it inside your revision folder.
Sources
Where this comes from
Built from five of the IB's own AI HL subject reports (May 2024 TZ2, November 2024, May 2025 TZ1 and TZ2, November 2025), read in full, covering Papers 1, 2 and 3 and the exploration. They were supplied by the site owner; the IB makes subject reports available to IB World Schools through its Programme Resource Centre. We have not seen every timezone's report for May 2024. Everything here is in our own words: we do not reproduce the reports or any exam question, and every worked example is our own. The question topics are our own matching from the reports' comments (parts whose topic we could not tell, or that fall outside our topic list, are marked "not mapped"); we show no marks per question because we did not use the markschemes.
What do examiners say students get wrong in IB Maths AI HL?
The most repeated points are: Accuracy; Hypothesis tests; HL algebra. Each one on this page has the fix and, where free, a worked example showing the wrong and right method.
Which IB Maths AI HL topics come up most?
In the material we analysed, the biggest areas were Calculus, Geometry and trigonometry, Statistics and probability.
Is this a prediction of the May 2027 paper?
No one outside the exam board knows what will be on the next paper. The "focus" list is an informed prediction from how often topics appeared and how weakly they were answered; it is not a guarantee, so revise the whole specification.
Where does this information come from?
We read 5 International Baccalaureate examiner reports (May 2024 TZ2, November 2024, May 2025 TZ1, May 2025 TZ2, November 2025) in full, summarised them in our own words and linked each point to the original.
IB Math Revision