Einstein's riddle, the 2 percent claim and what the puzzle really measures

0 The short answer

Einstein's riddle, also called the zebra puzzle, is a logic grid puzzle whose earliest known printing is Life International on December 17, 1962, and no source we could open links it to Einstein or gives the claim that only 2 percent of people can solve it a sample. Einstein died in 1955, seven years before that printing. Solving the puzzle draws on deductive reasoning held in working memory, which belongs to fluid reasoning (Gf), but a pass rate nobody measured is not a norm, and a puzzle without a norm group, a time limit or a reliability estimate cannot give anyone an IQ.

1 What Is Einstein's Riddle, and What Does It Ask?

Einstein's riddle is a logic grid puzzle: five houses stand in a row, each has a different color, a different occupant nationality, a different pet, a different drink and a different cigarette brand, and fifteen short clues fix who lives where. The structure is described the same way by the Wikipedia entry on the zebra puzzle and by the Math Lair page on the riddle: five houses, five attributes with five values each, and fifteen clues. The solver is asked two questions, who drinks water and who owns the zebra. Many versions in circulation ask who owns the fish instead, which is why the puzzle goes by both names, and the two pets do not change the logic.

Clues come in a small number of kinds. Some tie two attributes to the same house, as in the first clue of the usual list, which says the Englishman lives in the red house. Some fix a position outright, such as a clue that places one occupant in the first house. Some relate neighboring houses, saying that one lives immediately to the left of another or next to another. Nothing in the puzzle requires knowledge, vocabulary or arithmetic. Every clue is a short sentence in everyday words, and the whole difficulty lies in combining them without losing track.

This page does not reproduce the full clue list. Every solution page already prints it, and the clue list is not what the evidence is about. The questions that readers bring to the puzzle are different: who wrote it, whether it is true that only 2 percent of people can solve it, what it asks of the mind, and whether solving it says anything about IQ. Each of those has an answer that can be checked, and each answer is less flattering to the legend than the legend is to the reader.

A short vocabulary helps. A logic grid puzzle gives a set of categories and a set of clues, and the solver must decide which items go together. A constraint satisfaction problem is the computer science name for the same structure: a set of variables, a set of allowed values and a set of conditions that every assignment must meet. Lin and colleagues, who built a benchmark from such puzzles to test language models, say that logic grid puzzles are derived from constraint satisfaction problems and describe them as Zebra Puzzles or Einstein's Riddle, citing the 1993 paper by Patrick Prosser on algorithms for constraint satisfaction. Deduction means reaching a conclusion that must be true if the premises are true, which is what each step of the grid does.

2 Where Did the Puzzle First Appear?

The earliest known printing of the puzzle is in Life International, dated December 17, 1962, under the title "Who Owns the Zebra?", and the sources that report it do not tie it to Einstein. The Wikipedia entry on the puzzle gives the December 17, 1962 date and adds that the issue of March 25, 1963 printed the solution together with the names of several hundred solvers. The Math Lair page says the puzzle appears to have been first published in the 1960s in Life International under the title "Who Owns the Zebra?", and that the earliest sources its author found, from the 1960s and 1970s, do not mention Einstein.

We want to be exact about how much we verified. We opened the Wikipedia entry and the Math Lair page on October 20, 2026, and both agree on the magazine and the decade. Only Wikipedia gives the full dates. We did not see a scan of the December 17, 1962 issue. A search of the Internet Archive for Life International on October 20, 2026 returned issues from August 27 and September 10, 1962 and from 1963, but not December 17, 1962. The date therefore rests on a reference work that we treat as a pointer, and a reader who needs the original page should look for the scan of that issue. Where the sources disagree, they disagree on detail and not on substance: one says the 1960s, the other gives a day.

Three things follow from a 1962 magazine origin even without the scan. First, the puzzle is a mid twentieth century puzzle, printed in a general interest magazine, and a follow up that printed the names of solvers suggests that readers sent in answers. Second, the printed puzzle named cigarette brands, a detail that places it in its own period; the Wikipedia entry states that the brands did not exist in Einstein's boyhood but tags that statement as needing a citation, so we do not rely on it. Third, a contest printed in a popular magazine is exactly the kind of item that circulates for decades after its source is forgotten, attached to whichever famous name makes it travel.

The name "zebra puzzle" is therefore the older and better documented name. "Einstein's riddle" is the label that later circulation put on it. The ZebraLogic paper, for example, uses the two names together, and Prosser's 1993 paper is the reference it gives for them. The ZebraLogic authors do not credit Einstein with the puzzle, and they name no creator at all.

3 Did Einstein Write the Riddle?

There is no evidence that Einstein wrote the puzzle, and the legend that he made it up as a boy asks us to believe that a puzzle first seen in a 1962 magazine was invented by a schoolboy decades earlier. The Wikipedia entry says the puzzle is often named for Einstein because of an urban legend that he invented it as a boy, that it is sometimes credited to Lewis Carroll instead, and that there is no evidence for either person's authorship. The Math Lair page says the same about Einstein. The Nobel Prize facts page gives his birth as March 14, 1879 and his death as April 18, 1955, so he had been dead for more than seven years when Life International printed the puzzle.

That gap does not prove he never saw it. Proof of a negative is not what this section claims. What the record shows is narrower and checkable. No printing before 1962 has been produced. No letter, notebook, biography or interview has been produced that has Einstein posing, solving or mentioning the puzzle. The earliest sources the Math Lair author found do not name him. Two different famous names, Einstein and Lewis Carroll, are attached to the same puzzle in different tellings, which is what usually happens when a name is added to a story and not drawn from a document. An attribution that cannot be traced to a first source, a date and a witness is a rumor, whoever the name belongs to.

Readers who meet the legend usually meet it in a standard form: Einstein wrote it, and he said that most of the world could not solve it. The two parts are told together because they support each other. A genius named as the author makes the difficulty claim sound credible, and a difficulty claim makes the authorship sound like an achievement. Neither part has a source, so neither can support the other. The same mechanism produces the circulating IQ figure attached to Einstein himself, and the ACIS page on what the Einstein IQ figure rests on takes that number apart in the same way: the figure circulates, and the record holds no test, date or examiner.

The table below separates what the sources document, what they leave open and what circulates without a source.

4 Is It True That Only 2 Percent of People Can Solve It?

No study, test or sample is known behind the claim that only 2 percent of people can solve the puzzle, and the one puzzle page we opened that addresses it calls the claim untrue. The Math Lair page says the figure usually told with the riddle is that 98 percent of people cannot solve it, that this has no truth to it, and that an ordinary person who thinks about the problem logically and has enough patience and time should be able to solve it. We searched for the first appearance of the figure on October 20, 2026 and found pages that repeat it, and none that cite a survey, an experiment or a date.

A pass rate is a measurement, and a measurement has requirements. Someone must define the population whose pass rate is reported. Someone must define the exact task, because a five house puzzle printed with a grid, a five house puzzle read aloud and a six house puzzle are three different tasks. Someone must fix the conditions, including the time allowed, whether paper and pencil are permitted and whether the solver may have seen the method before. Someone must decide what counts as solved, and someone must publish the method so that another researcher can repeat it. The table sets those requirements against what exists.

The only solver count attached to the puzzle's early life is the one Wikipedia reports for the March 25, 1963 issue of Life, which printed the names of several hundred solvers. That is a count of people who chose to write in. It has no denominator, because nobody knows how many readers tried and gave up, and it describes self selected readers of one magazine, not a population. Even a precise count of successful letters would not yield a percentage.

There is a second reason the number deserves suspicion, and it comes from IQ itself. On a scale with a mean of 100 and a standard deviation of 15, which the page on the standard deviation of 15 explains, the top 2 percent of a normal distribution begins at an IQ of about 131. That figure is our arithmetic on the normal curve, using a standard score of about 2.05, and it can be checked with the IQ percentile calculator. The coincidence is worth noticing: "only 2 percent" sounds like a gifted threshold, because it sits near the usual cutoff of 130 (see what an IQ of 130 means). Readers hear a claim about intelligence. But the puzzle was never calibrated against a scale of intelligence, so the resemblance is only a resemblance. A pass rate on a particular puzzle and a percentile on an IQ scale are different kinds of number, and the first cannot be converted into the second. The page on how an IQ score differs from a percentile explains why even real scores cannot be treated loosely.

What the claim does reliably do is invite the reader to rank himself or herself. A reader who solves the puzzle is told, without evidence, that he or she is in a rare group, and a reader who fails is told the same thing in the opposite direction. Both messages are made up. The honest statement is that the proportion of adults who can solve the puzzle depends on how it is posed and on how much time and practice are allowed, and nobody has reported it.

5 How Is the Puzzle Actually Solved?

The puzzle is solved by elimination: each clue removes arrangements that cannot be true, and when enough clues have been applied only one arrangement remains. The method does not need guessing, and it can be shown on a puzzle small enough to read in a minute. The one below is original to this page. It has three houses in a row, numbered 1 to 3 from the left, and three attributes: color (red, green, blue), pet (cat, dog, fish) and drink (tea, milk, juice).

1. The cat lives in the red house.
2. The tea drinker lives in the house immediately to the left of the blue house.
3. The dog owner drinks milk.
4. The green house is at the left end of the row.
5. The dog owner lives in the green house.

The question is who owns the fish and what that person drinks. A solver can start with clues 4 and 5, which together put the green house and the dog in house 1. Clue 3 then puts milk in house 1 as well. Clue 2 says the tea drinker sits immediately left of the blue house. The tea drinker cannot be in house 1, which has milk, so the tea drinker is in house 2 and the blue house is house 3. That leaves red for house 2, and clue 1 puts the cat there. The only pet left for house 3 is the fish, and the only drink left is juice. The fish owner lives in the blue house and drinks juice.

Three habits carry the whole method. The first is to start with the clues that fix a position outright, because they remove the most options at once. The second is to turn every clue into two kinds of facts, what must be true and what cannot be true, and to record both. The third is to carry each result forward before reading the next clue, which is what the chain above does. Pencil and paper do this work with a grid of small marks. Doing it in the head is possible for three houses and demanding for five, a difference the next section explains.

We checked the example by listing every possible arrangement. With three houses and three attributes there are 6 orderings of the colors, 6 of the pets and 6 of the drinks, so 6 by 6 by 6, which is 216 arrangements. We applied the clues in the order above and counted what remained. This count is our own computation on our own puzzle.

The pattern is the point. Each clue divides the space of possibilities, and a puzzle is well made when the last clue leaves exactly one. A puzzle with two surviving arrangements is flawed, and a puzzle in which a clue removes nothing is padded. Computer scientists treat this as a constraint satisfaction problem and solve it with search and propagation, the setting in which Prosser's 1993 paper tested algorithms. The human method is the same idea carried out by hand.

6 Why Does Difficulty Grow So Fast With the Size of the Grid?

The number of possible arrangements grows far faster than the number of houses, which is why a three house puzzle can be done at a glance and the five house version cannot. For a grid with N houses and M attributes there are N factorial orderings of each attribute, so the full space holds N factorial raised to the power M arrangements. That is the definition used in the ZebraLogic paper, which defines the search space as the number of configurations that satisfy the uniqueness constraints and gives it as N factorial raised to the power M. The figures in the table are our arithmetic on that published formula, for square grids.

Nobody solves the puzzle by walking through 24.9 billion arrangements, and nobody needs to. Each clue prunes enormous regions at once, as the example showed when one clue removed two thirds of the space. The real difficulty is different. It lies in the number of facts that must be kept in view at once and in the order in which the clues become usable. Some clues are only useful after others have been applied, and a solver who cannot hold the partial grid steady will lose a deduction already made.

The ZebraLogic authors describe the same two sources of difficulty from the machine side. They say that some puzzles can be solved by direct deduction and others need reasoning that backtracks, in which an assumption is made, followed to a contradiction and withdrawn. Their benchmark spans grid sizes with N and M each from 2 to 6, which gives 25 sizes with 40 puzzles each, 1,000 puzzles in all. Einstein's riddle is one cell in that range, a 5 by 5 grid, with the benchmark extending to 6 by 6.

A different line of research explains why the load matters for people. Halford, Wilson and Phillips proposed in 1998 that working memory limits are best defined by the complexity of the relations that can be processed at the same time, measured as the number of related dimensions. They reported that studies of working memory limits suggest a soft limit corresponding to the parallel processing of one quaternary relation, and that more complex problems are handled by segmentation, in which a task is broken into components that do not exceed capacity and are processed one after another. A grid puzzle invites exactly that strategy. Writing marks in a grid is a way of externalizing the parts so that the head holds only one or two relations at a time, which is why the same puzzle is much harder without paper. That last step is our reading of their framework, not a result they report for this puzzle.

7 What Abilities Does the Puzzle Draw On?

Solving a logic grid draws mainly on deductive reasoning held in working memory, which in the CHC model is fluid reasoning (Gf) supported by working memory (Gwm), with very little demand on stored knowledge (Gc) or arithmetic (Gq). The classification is ours, not a finding of any study, and it follows the way the ACIS battery groups tasks: the Logic Grid subtest sits in the fluid reasoning index. The page on the CHC model defines the abbreviations, and the page on fluid and crystallized intelligence explains the contrast between reasoning with new material and using stored knowledge.

Three demands can be separated. The first is rule following. Each clue is a rule, and every step applies a rule to a partial grid. The second is inference, the movement from two stated facts to an unstated third, which is what Johnson-Laird describes in his account of reasoning. He says that people reason by envisaging the possibilities consistent with what they start from, constructing a mental model of each distinct possibility and drawing a conclusion from them, and that the theory predicts systematic errors in reasoning. A grid is a device for listing the possibilities so that the solver does not need to imagine all of them. The third is maintenance: keeping the results so far available while reading the next clue.

That third demand is where working memory enters, and it is also where a common overstatement begins. It is often said that fluid reasoning and working memory are the same thing. The evidence is more careful. Ackerman, Beier and Boyle reported a meta-analysis of 86 samples and found that the average correlation between true score estimates of working memory and general intelligence was .479, which they describe as substantially less than unity. Working memory and intelligence overlap, then, but they are not identical, and a task that loads on both will not be a pure test of either. The page on working memory tests shows how that ability is measured on its own.

The relational complexity framework has also been used to build tests. Birney, Halford and Andrews described the Latin Square Task, an 18 item test at three levels of relational complexity that they gave to 73 university students aged 16 to 33 and 204 school students aged 8 to 19. They designed it to minimize the role of knowledge and storage capacity, and they reported that it was sensitive to parallel and to serial processing demands. The relevance for the riddle is a design lesson. A task meant to measure reasoning is built so that the difficulty is controlled and graded, while a viral puzzle is built to be shared. Nothing in what we could trace of the 1962 printing suggests that it was designed to grade difficulty or to separate reasoning from the knowledge of a method.

Other abilities play a smaller part. Reading the clues requires comprehension of everyday sentences, but no vocabulary beyond that, so the verbal comprehension index of a battery is barely engaged. Speed matters only if the solver is racing a clock. Attention matters throughout, as it does in any long task, although the specific inhibition effect studied in the Stroop test is not the main demand here. A puzzle that exercises one narrow family of reasoning cannot stand in for the broad profile that a battery measures, and the page on verbal and quantitative ability shows why a single ability family never gives the whole picture.

8 Does Solving It Mean You Have a High IQ?

Nobody has shown that solving this puzzle predicts a measured IQ, so a solve is a pleasant result and not a score. We searched on October 20, 2026 for a peer reviewed study in which adults solved the zebra puzzle and also took an intelligence test or a working memory measure, and we did not find one. That is a statement about what our search turned up and not proof that no such study exists. It does mean that no number linking this puzzle to IQ can be quoted honestly, in either direction.

A link is plausible, and the reasoning is worth stating so that it is not confused with evidence. A puzzle that demands deduction and working memory should correlate with tests of fluid reasoning, because it shares their demands. The correlation would probably be moderate and not strong. The meta-analysis by Ackerman, Beier and Boyle found a correlation of .479 between working memory and general intelligence even for two constructs that are widely discussed as close relatives. A single puzzle solved once is a far noisier measure than either construct. In plain terms, any link between one puzzle and an IQ would plausibly be weaker than the link between two carefully built tests, which is itself only moderate.

Four other things decide whether a given person solves the puzzle, and none of them is intelligence in the sense an IQ scale describes. The first is knowledge of the method. A person who has used grids before solves a new grid far faster than a person who has not, and the knowledge transfers from one puzzle to the next. The second is patience and persistence, which the Math Lair author names as the main requirement along with time. The third is the format: the same clues presented as a printed table, as a spoken list or as a typed paragraph make different demands. The fourth is chance and mood on the day. Tests built to measure ability try to control exactly these factors, which the page on how to prepare for an IQ test discusses for practice effects, and the page on whether you can increase your IQ discusses for what training moves.

There is also a limit that comes from the shape of the data. A single puzzle gives two outcomes, solved or not solved, at one level of difficulty. That separates people only at one point on the ability range. A battery uses many items spread across difficulty, so that people can be separated across the whole range, from low scores to high ones. How well a general factor describes performance at different levels is a separate question, taken up on the page on Spearman's law of diminishing returns. It is another reason to treat one task, however famous, as a poor stand in for a profile.

9 How Do Language Models Do on Puzzles Like It?

Language models lose accuracy sharply as logic grid puzzles grow, and the ZebraLogic study measured that loss on 1,000 puzzles with known complexity, which is the only systematic evidence about this puzzle family that we found. Lin and colleagues built the benchmark from logic grid puzzles derived from constraint satisfaction problems and tested models including Llama, o1 and DeepSeek-R1. They report a marked drop in accuracy as puzzles get harder, which they call the curse of complexity, and they say the limit persists with larger models and more inference time.

The paper's table makes the point with numbers. In the largest search space category, the version dated July 15, 2025 lists 42.5 percent for o1-full, 28.5 percent for DeepSeek-R1 and 17.0 percent for o1-preview. A 5 by 5 grid holds about 24.9 billion arrangements by the formula in the section on difficulty above, which is larger than the thresholds the paper gives for that category in both its text and its table, so Einstein's riddle sits in the hardest category of the benchmark. That placement is our arithmetic on the paper's definitions and not a result the authors report for the riddle itself. The numbers are results for particular models tested in 2025. They are not a ranking of current systems, and newer models may do better.

Three cautions apply before drawing a lesson about intelligence. First, a model solving or failing a puzzle is a statement about that model, with that prompt, on that benchmark. The page on what ChatGPT scores and why it is not an IQ explains why human norms do not carry over to machines, and the page on AI and IQ summarizes the cognitive studies. Second, the benchmark controls complexity on purpose, which is something the 1962 puzzle never did. Third, the result is useful for human readers in one respect: it shows that complexity, defined as the size of the search space, is a measurable property of such puzzles. A claim about how hard a puzzle is can be tested, and the 2 percent claim never was.

10 How Does a Normed Test Measure Deductive Reasoning Differently?

A normed test differs from a viral puzzle in six ways that all matter for what the score means: a reference group, fixed conditions, many items of graded difficulty, reliability evidence, a stated score scale, and documentation a stranger can check. The page on how IQ scores are normed explains the first of these: a score is a position relative to a reference group, usually expressed on a scale with a mean of 100 and a standard deviation of 15. The viral puzzle has none of the six. The table sets the two side by side.

ACIS is a normed, self administered online battery and this page describes it only from the home page as read on October 20, 2026. It has 20 subtests in six CHC domains, and the Logic Grid subtest belongs to the fluid reasoning index, together with Matrix Reasoning, Figure Weights, Visual Number Series and Complex Relations. The report gives a Full Scale IQ and six primary indices on the standard scale with a mean of 100 and a standard deviation of 15, scaled subtest scores from 1 to 19 with a mean of 10 and a standard deviation of 3, percentiles, and a 95 percent confidence interval. Adult norms cover ages 16 to 90. The page on Logic Grid describes that subtest, and the page on fluid reasoning tests places it among others that measure Gf. Readers who want the evidence behind the scoring should read the technical manual, and the page on reliability and validity explains what those words require.

The forms are a one time payment with no subscription, and prices were read on October 20, 2026 and can change. Quick costs 15 dollars for 6 subtests in 3 domains and takes about 45 minutes. Optimized costs 30 dollars for 13 subtests in 5 domains and takes about 110 minutes. Full Scale costs 50 dollars for all 20 subtests in six domains and takes about 175 minutes. Breaks are allowed, and test takers have 30 days to complete a form. There is a free trial that needs no card and a 5 day quality guarantee.

The limits belong beside those strengths. The test is online and unsupervised, which means that conditions are not controlled the way an examiner controls them. It is not a clinical or diagnostic instrument. It is not meant for hiring, for school accommodations or for admission to high IQ societies, and it is offered in English only. A reader who needs a score for one of those purposes should use the route the institution accepts, as the page on what makes a test accurate and the page on who can administer an IQ test lay out.

11 What Should You Do With the Puzzle and With Claims Like It?

Treat the puzzle as practice in deduction that is worth the evening it takes, and treat any claim that only a small share of people can solve something as unsupported until it names a sample. Solving the riddle with paper and pencil is a reasonable way to see for yourself how elimination works, and the example earlier on this page shows the method on a small scale. It is also a good way to notice how much of the work is bookkeeping. Nothing about solving it, or failing to solve it, belongs in a conversation about your intelligence.

When a post says that most people cannot solve a puzzle, five questions test it. Who measured the pass rate, and where is the method published? How many people were tested, and who were they? Under what conditions, with what time limit and with what tools? What counts as solved? When was the measurement made? A claim that cannot answer even the first question is a slogan. A claim that answers all five is rare for viral content, and when one does, it still describes a pass rate on a task and not an IQ.

A second habit concerns difficulty claims in general. Words such as hardest, impossible and genius level on puzzles are not units. A real difficulty statement is a proportion of a defined group that passes, or a parameter estimated from many people, and the page on the hardest IQ test discusses what difficulty buys in a test. The ZebraLogic authors show what a quantified version looks like for this puzzle family: they state complexity as the size of the search space, which can be computed.

A third habit is to ask what the puzzle format is for. Logic grid problems are one family of reasoning item among many, and the page on IQ test questions describes the formats used in published tests without reproducing their items, while the page on the Cognitive Reflection Test covers a different popular quiz. A reader who wants a score rather than a pastime should pick a test with a norm and read the report as an estimate with an interval. The guide to abstract reasoning tests covers a nearby family of tasks, and the pages on free and validated tests and on how IQ is calculated cover the choice and the scoring chain.

12 What Do the Testing Standards Say About a Puzzle That Claims to Rank You?

The Standards for Educational and Psychological Testing (AERA, APA and NCME, 2014) ask test developers to state what a score means and for whom, to describe the norm group and how it was sampled, and to caution users against interpretations the evidence does not support, and the 2 percent claim meets none of those requirements. The Standards are a joint publication of the American Educational Research Association, the American Psychological Association and the National Council on Measurement in Education. They are written for tests used in professional practice, so they are a benchmark and not a law for a puzzle shared online, but they set out what a reader can fairly expect from any claim about a score.

Standard 1.1 says the test developer should set forth clearly how test scores are intended to be interpreted and used, that the population for which a test is intended should be delimited clearly, and that the construct it is intended to assess should be described clearly. Its comment adds that no test permits interpretations that are valid for all purposes or in all situations. The riddle has no developer, no stated population and no construct. Standard 1.2 asks for a rationale and a summary of the evidence for each intended interpretation, and Standard 1.3 says that when validity for a likely interpretation has not been evaluated, users should be strongly cautioned. "Only 2 percent can solve it" implies an interpretation, namely that solving it is a sign of rare ability, and no evidence for that interpretation has been offered.

The norm standards bear most directly on the number. Standard 5.8 says that norms, if used, should refer to clearly described populations. Standard 5.9 says that reports of norming studies should include a precise specification of the population that was sampled, the sampling procedures and participation rates, any weighting of the sample, the dates of testing and descriptive statistics. Standard 5.1 asks that users be given clear explanations of the characteristics, meaning and intended interpretation of scale scores as well as their limitations. A pass rate of 2 percent with no population, no sampling, no participation rate and no date fails every element of Standard 5.9. We say this about the claim and not about whoever repeats it, because most people who pass it on never see the problem.

The American Psychological Association's Ethical Principles of Psychologists and Code of Conduct, as amended effective January 1, 2017, points the same way for psychologists who assess people. Standard 9.02 says psychologists use assessment instruments whose validity and reliability have been established for use with members of the population tested, and that when they have not been established, psychologists describe the strengths and limitations of test results and interpretation. Standard 9.06 asks psychologists, when interpreting results, to take into account the purpose of the assessment, test factors, test taking abilities and other characteristics of the person, and to indicate any significant limitations of their interpretations. The Code governs psychologists, not puzzle authors or readers, but the same logic applies to a reader who wants to know what a result says about him or her.

What the evidence supports, stated narrowly, is this. The puzzle is a real logic grid puzzle with a unique solution, first printed as far as anyone has shown in a 1962 magazine. Its difficulty can be quantified by the size of its search space, and models and people lose accuracy as that space grows. It draws on deduction and working memory. Nothing in the record ties it to Einstein, and nothing supports the claim that 2 percent of people can solve it. It does not measure IQ, and no solving time or outcome can be read as a score. A reader who wants a score that carries a norm, a percentile and a confidence interval should take a test built to give one and read it with the limits that the Standards and the Code describe.

13 Sources Behind This Page

Every figure and attribution above traces to one of the sources below, or is labeled as our own arithmetic. The arithmetic is the number of arrangements for each grid size, the 216 arrangement count and the steps of the worked example, and the IQ of about 131 at the top 2 percent of a normal curve. Web pages were read on October 20, 2026 and can change. The Wikipedia entry is cited as a pointer for the December 17, 1962 and March 25, 1963 dates, which we could not check against a scan of the magazine.

• Ackerman PL, Beier ME and Boyle MO. Working Memory and Intelligence: The Same or Different Constructs? Psychological Bulletin, 2005, volume 131, issue 1, pages 30 to 60.
• American Educational Research Association, American Psychological Association and National Council on Measurement in Education. Standards for Educational and Psychological Testing. American Educational Research Association, 2014.
• American Psychological Association. Ethical Principles of Psychologists and Code of Conduct, adopted 2002 and amended effective June 1, 2010 and January 1, 2017, Standards 9.02 and 9.06.
• Birney DP, Halford GS and Andrews G. Measuring the Influence of Complexity on Relational Reasoning. Educational and Psychological Measurement, 2006, volume 66, issue 1, pages 146 to 171.
• Halford GS, Wilson WH and Phillips S. Processing capacity defined by relational complexity: Implications for comparative, developmental, and cognitive psychology. Behavioral and Brain Sciences, 1998, volume 21, issue 6, pages 803 to 831.
• Johnson-Laird PN. Mental models and human reasoning. Proceedings of the National Academy of Sciences, 2010, volume 107, issue 43, pages 18243 to 18250.
• Lin BY, Le Bras R, Richardson K, Sabharwal A, Poovendran R, Clark P and Choi Y. ZebraLogic: On the Scaling Limits of LLMs for Logical Reasoning. arXiv 2502.01100, accepted to the International Conference on Machine Learning 2025, version 2 of July 15, 2025.
• Math Lair. Einstein's Riddle, page read on October 20, 2026.
• Nobel Prize Outreach. Albert Einstein, facts, page read on October 20, 2026.
• Prosser P. Hybrid algorithms for the constraint satisfaction problem. Computational Intelligence, 1993, volume 9, issue 3, pages 268 to 299.
• Wikipedia contributors. Zebra Puzzle, page read on October 20, 2026, used as a pointer and not as a source of findings.

14 Frequently Asked Questions