In the 1950s, Claude Shannon, often regarded as the father of information theory, developed a mechanical device designed to predict human choices in a simple game of selecting between two levers. Shannon’s machine, inspired by an earlier model with a modest success rate, was able to guess correctly up to 65 percent of the time by recording and analyzing a person’s previous choices, exploiting predictable patterns in human behavior.
The principle behind Shannon’s device and similar experiments is that humans tend not to generate truly random sequences. People often fall into subconscious habits or predictable behaviors when asked to produce random outcomes. For example, studies as far back as the late 1930s have shown participants favor certain sequences when asked to mimic coin tosses, such as starting with "heads" more frequently than "tails," or choosing certain patterns far more often than others, despite all outcomes being equally likely in a genuinely random process.
Ergonomics pioneer Alphonse Chapanis, in the 1950s, extended this research by asking volunteers to complete grids of random digits. The resulting data revealed strong preferences for descending digit pairs and an aversion to repeated digits, highlighting common cognitive biases in generating random-like sequences. Such predictable tendencies have practical implications, including vulnerabilities in security systems and strategies for guessing answers on multiple-choice examinations.
These human limitations have cascading effects in areas where randomness is critical. Attempts to use machines for randomness have sometimes been compromised, as demonstrated in the 1980 Pennsylvania lottery scandal of 1980, where weighted lottery balls skewed expected outcomes. Similarly, digital random number generators can be vulnerable; for instance, an IT professional admitted to embedding a backdoor into lottery software that enabled certain numbers to be predicted on specific dates.
In the realm of computing, generating unpredictable random numbers is essential for cryptography and security. However, even sophisticated devices have been compromised because the underlying random processes were insufficiently random or predictable.
Coin tossing, often regarded as the gold standard for fairness, is also subject to subtle biases. Mathematician and magician Persi Diaconis has shown that a coin is more likely—about 51 percent of the time—to land on the same face it started on. Diaconis even created a mechanical coin tosser capable of consistently replicating this effect.
More recently, mathematics communicator Matt Parker highlighted these quirks in a public experiment at the International Congress of Mathematicians. Parker challenged attendees to flip coins and achieve ten heads in a row, promising a monetary prize. Surprisingly, several participants succeeded quickly, a result attributed to biased flipping techniques, such as insufficient coin rotation. When Parker switched to dice the following day, results aligned more closely with expected random distributions.
Together, these findings underscore the inherent difficulties humans and machines face in producing truly random outcomes, influencing fields ranging from gambling and cryptography to educational testing and recreational math.