False Alarms: Why We Say “No” When We Should Say “Yes”
Fear
Within numeric wisdom, as we seek to be enlightened,
Lurks circumstance and outcome where one soul will soon be frightened.
And on the field of battle, this poor schmuck will be exhorted,
“The data makes it clear” we cry, and leave them unsupported.
But strategies from leaders, be they CEOs or pharaohs,
Require more than models, but who fields those slings and arrows.
Amidst the idealized notions of a green grass, sunflower seeds, bubble gum, and hot dogs is the reality of little league baseball, as experienced by an anxious child. No, I wasn’t the kid in right field praying that the ball would never find him. That kid is waiting for the orange slices, the snow cone, and the freedom from the pastime he never wished to experience. I was something else entirely, a perfectionist. The diminutive only child, nerdy, studious, and willing to spend hours firing a tennis ball against the wooden paneling of a partially-finished basement for hours on end to refine eye-hand coordination. The coach’s kid who received a steady diet of ground balls in fading sunlight at the local park. Throw to the proper base. “The best throw is no throw.” I loved baseball. I hated the possibility of failure. The thought of striking out struck terror into my soul. Most would have navigated this anxiety by choosing something else to do on sunny spring days.
But not me. Baseball is everything a numerically-inclined, bookish kid could ever hope for. A rich array of statistics-driven narratives spanning a century penning new chapters annually. The trips to the stadium with father and its mixture of ritual and duty in carefully filling the scorecard with the details of every pitch. Baseball was a romantic pursuit. Catching a ball was the gateway to social acceptance as the new kid at school. And still, striking out was horrifying. How could I avoid the shame?
The answer was a shortened swing coupled with an almost pathological fixation with putting a ball in play when a pitch entered the strike zone. Though anxiety produces the type of plate discipline that generates a fair number of walks when opposing pitchers are navigating the awkwardness of their adolescent growth sports, I also generated more harmless ground balls than Mario Mendoza. I avoided a strikeout for literally years. In professional settings, we often cannot choose where and when we enter the batter’s box. The disapprobation of striking out is an unmistakable, career-damaging outcome. However, in far too many cases, the cost of risk-aversion is larger, more difficult to quantify, and a corrosive toxin that prevents taking the swing upside demands.
Data scientists will, without fail, perform the salient expected value calculation and estimate the probabilities of successes and failures acceptably. What neither they, nor the vibe-coding platforms that enable them can account for is the stage on which those failures occur. The world of professional tennis is filled with men and women who deliver their first serves with triple-digit velocity and pinpoint precision. These missiles land within the lines over 70% of the time for some elite servers and in excess 75% of the time for a few folks on the men’s tour with heights that wouldn’t be out of place on an NBA court. When these first serves succeed, the player’s probability of winning the point is extremely high. Alas, tennis offers servers only one opportunity to “miss,” after which the dreaded “second-serve” must be offered, lest that server gift their opponent with a free point. In their hopes of being spared the indignity of double faults, second serves are meeker attempts and the probability of winning the point behind the lesser offering falls significantly.
One particular professional of impressive stature and exemplary talent is Alexander Zverev. Standing 6’6’’, his first serve is formidable. His second serve, and the robustness of his psychological temperament have been questioned throughout his career. Famously1 he allowed a 2-0 lead in a best-of-five US Open final to evaporate, failing to complete his victory as he served for the fifth set.2
During 2020, Zverev landed his first serve at a rate of 68.3%. He won 77.0% of those points. When he was forced to deliver his second (weaker) offering, he won 45.0% of those points3. Worse, not all second serves, despite their comparably conservative approach are successful, and he double-faulted 6.4% of the time. This begets some trivial math, with which your handy excel spreadsheet or your friendly LLM companion will gladly assist you.
P(wins point when attempting 1st serve) = 68.3% × 77.0% = 52.6%
P(wins point when attempting 2nd serve) = (1 - 6.4%) × 45% = 42.1%
Does everyone see the problem? He’d be better off hitting his first serve again! Even though he’d miss nearly a third of those attempts, he’d win the significant majority of the points in which that second (first) serve landed within the box. Moreover, were he to practice only his first serve, one would imagine both parameters in the first equation would improve.
This math is both incontrovertible and extraordinarily simple. Literally, multiplication is sufficient to illustrate the point. And yet. Omitted from this calculation and any other such depiction emerging from an AI is the particular flavor of human suffering the optimal strategy begets. Where, in the trivial mathematical expression lies the emotional wound of standing on the hallowed grounds of Wimbledon, looking up at the royal box with princess Kate and Roger Federer looking on…and striking dozens of double-faults as an overly-polite crowd of British patrons gasps?
Do you imagine this 6’6’’ racquet-wielding virtuoso is similarly adept at navigating the impending press-conference in which his stunningly contrarian strategy will be questioned relentlessly? Is he prepared to field inquiries about whether he will repeat that paradigm in all subsequent matches? Would he, in that fateful final, amidst the throngs of New Yorkers, dare to incur not the potential of defeat, but of a spectacle in which he plays a dubious starring role?
So, dear C{x}O, are you offering not only the data and quantitative expertise required for discovery of the optimal solution, but also the support structures required for those who will execute those plans?
Fear 2.0
When wondering why when adopting AI,
The discourse is crawling with doubters,
Remember the fate of tools driven by data
In novel, peculiar encounters,
There rarely is terror of dull, human error,
We know, all too well, imperfections,
But wand’ring the shallows of optimized algos,
We cringe at their errant selections.
One form of fear lies in our insecurities and our internal fragility. The last essay spoke of the reluctant little leaguer, afraid of the walk back to the dugout following a strikeout or the accomplished professional tennis player who would bear the slings and arrows of the crowd and future podcast punditry as the double-faults accrued. But now, in the age of AI, a second form of insidious concern strangles would-be upside in its crib.
Human beings are profoundly fallible creatures, plagued by their distractibility4, sloth, irritability, irrationality, and general stupidity5. However, we accept these limitations, and hand them the keys to two-ton hunks of steel filled with combustible fuel, and send them off to speed through densely populated urban centers. Predictably, from time to time, this goes horribly wrong. We recognize that the personal agency associated with driving embeds risk and costs for the individual and other passengers and pedestrians.
Because humans are permitted to err.
We allow human beings to move markets with the click of a button, transacting in dizzying sums at the risk of failures of cognition or dexterity.6 Sure, this might catalyze a capital-destroying sequence of reactions, but still place the sack of organic material before the keyboard.
Because human error is acceptable.
But if a driverless vehicle causes a fatality, the reaction to a (sadly) familiar outcome is wholly distinct. If an automated trading algorithm sparks a flash crash, the recriminations are of a different tone. Even if a self-driving car can demonstrate, over hundreds of thousands of miles, that its collision statistics are superior to human drivers, when the artificial intelligence produces a failure a human would have avoided, the discussion is rarely about aggregate and probability distributions. AI’s flavor of “error” often seems silly in the eyes of a “wiser” human being, especially if it fails to discern how to navigate a scenario human beings would consider trivial and obvious.
Among my avocational pursuits is the development of algorithms for the purposes of wagering upon sporting events. Humans who follow their gut transfer massive sums into the coffers of casinos. Algorithms (sometimes) deliver better outcomes. And yet, even an experienced algorithm leaves its operator to shudder and shake his or her head in disbelief at its failure. Perhaps it looks to back a starting quarterback who exited his most recent start with a gimpy ankle and a possibly-concussed cranium. Perhaps it fails to recognize that a major league baseball game played at an elevation in Mexico typically reserved for alpine hiking is likely to produce runs in quantities it lacks the historical comparisons to predict.
Should we eschew the silicon-based for the carbon-based? Of course not!7 Should we continue to allow hunks of gasoline-fueled vehicles to be steered by organic creatures fueled by gas station snacks and caffeine?8
Thus, when the boardroom discourse raises concerns about some AI-driven alternative, is the bearish narrative driven by anecdotal examples of errors a human would not make (if so, remember that self-driving cars don’t drive drunk or text at the wheel)? Or is the argument grounded in quantitatively-defensible reasoning and apt comparisons?
Only one is a reason to say “no.”
Fear and Fear 2.0 are drafted; the remaining sections of this chapter are forthcoming.
Footnotes
Infamously?↩︎
The tennis lexicon for the scenario in which the server can win the set by winning the current game (in which they are the server).↩︎
https://www.tennisabstract.com/cgi-bin/player.cgi?p=AlexanderZverev↩︎
#ScrollingMyLifeAway↩︎
Not that I’m wholly misanthrophic, but I do often wonder how many of the seven deadly sins one commits within a single algorithm-driven rabbit-hole.↩︎
The apt term is a “fat-finger error”↩︎
Though having a human in the loop minding the model ain’t a bad idea.↩︎
I mean, we can still eat the snacks and wash them down with a can of Monster in the passenger seat!↩︎