1962: Mariner 1 destroyed by a software error

On 22 July 1962, NASA launched Mariner 1 from Cape Canaveral, hoping to send the first successful probe past Venus. Just 293 seconds after liftoff the Atlas-Agena rocket veered dangerously off course. Range-safety officers had no choice but to destroy it over the Atlantic, turning an ambitious planetary mission into an expensive fireball.

Investigators quickly traced the failure to the guidance software. In the original handwritten equations an overbar marked a smoothed average of velocity measurements. When the formula was transcribed into code the overbar was omitted. Without that smoothing, ordinary noise was treated as a real deviation; the computer issued ever-larger corrective commands until the vehicle tumbled out of control. The episode entered computing lore as “the most expensive hyphen in history.”

In 1962 digital computers were still rare and code was written by hand on coding sheets. Mariner 1 demonstrated how a single notational slip could cascade into catastrophe. The lesson spurred stricter practices—independent verification, configuration control, and clearer mathematical notation—that became standard in aerospace software. Five weeks later the identical twin Mariner 2, flying corrected code, completed the first successful planetary flyby.

The same chain of reasoning that links a missing overbar to a destroyed rocket now links specification errors, numerical instability, or silent data-corruption to the unexpected behaviour of large-scale AI systems. Modern machine-learning pipelines contain millions of parameters and countless preprocessing steps; a subtle mismatch between the intended mathematical model and its executable form can produce outputs that are confidently wrong. Techniques descended from those early aerospace safeguards—formal verification, differential testing, and rigorous audit trails—remain essential if artificial intelligence is to be trusted in safety-critical settings.

Thus a brief, well-documented mishap on a summer morning in 1962 still speaks directly to anyone building intelligent machines today: correctness is not an optional extra, and the gap between human intention and machine execution can be bridged only by relentless attention to detail.