I've been reading Henry Hazlitt's Economics In One Lesson. Hazlitt outlines his classical (pre-Keynesian) liberalism through a discussion of various economic fallacies, all of which stem from a failure to appreciate a single “lesson.” 
The lesson is, “The art of economics consists in looking not merely at the immediate but at the longer effects of any act or policy; it consists in tracing the consequences of that policy not merely for one group but for all groups.”
Hazlitt extrapolates this lesson to explain the absurdity of protectionist trade policies, of governments' fixation of full employment (rather than on productivity), etc. His intention is to give a literate layperson a reasonably intuitive picture of the day's policy debates.
The book made me wonder, what's machine learning in one lesson? You can find a lot of people online explaining what machine learning is, or how it works. But these explanations don't give the layperson much context for debates about machine learning; for example, its fairness, its security, its safety. You can find books specifically on these topics, including really good books like Virgnia Eubank's Automating Inequality, but they center around specific case studies (in Eubank's case, poor and working class people in the US). It's unclear to me that they allow the reader to extrapolate to new cases that may emerge in the future.
So what's the lesson—the nugget of wisdom— that gives an interested reader enough analytical purchase to think about bias and AI safety concerns as they crop up in public debate? A conceptual handrail?
Here's my take:
Machine learning is when humans get a computer to write a program; humans then assess how well that generated program works.
This “lesson” highlights the role of the human, rather than specific technical underpinnings of machine, in performing some situated set of tasks. Almost every issue I can think of in machine learning—every case of bias, discrimination or outright failure—comes from this basic fact. The people in charge of assessing the program's performance may, for whatever reason, be unqualified to do so. Perhaps the photos they've added to their training set are overwhelmingly of white faces, and they haven't noticed (a case of epistemic risk).
The lesson becomes powerful when it's applied. Consider our first MLfailures lab, based on Obermeyer et al's Science paper. Basically, hospitals use algorithmic risk scores to establish patient risk. One feature the score uses is a patient's prior medical spending. If they've been spending more on medical care, they must be higher-risk patients, right? Well, this algorithm ends up systematically undervaluing the risk of Black patients, who spend less on healthcare on average (for a variety of reasons, economic and institutional).
What went wrong? Well, some humans assessed how well the program works. Perhaps they worked in a back office of a hospital or at a third-party vendor, and hadn't talked to enough clinicians on the ground to realize that Black patients spend less on healthcare. Maybe, due to their social positionality, they didn't think to check. Or maybe they had a financial incentive to not mention it.
Meanwhile, clinicians on the ground may not have had enough information to check the algorithms performance. In the midst of their stressful jobs, they probably don't have time (or expertise) to question the algorithm's fairness.
So the issue is explainable through a breakdown in the core lesson: humans assessed how well the generated program worked, and they messed it up grandly.
As an exercise for yourself, see how much more analytical richness you get when you apply this lesson to this generally excellent NYT story on Robert Julian-Borchak Williams compared to the Times' framing of a "faulty system."
Richmond Wong, who read this piece over, wasn't sure he liked the "lexicon of 'in one lesson' as it seems reductionist." I don't disagree. But people need a conceptual handrail for debates on algorithms. Prop 25 almost replaced cash bail with algorithmic risk assessment tools in California! 
Like Economics in One Lesson, you could generate a whole book of discussions of recent AI incidents, showing how they stemmed from a failure to appreciate some core "lesson." You could write a column in the Times every week. In other words, you could meaningfully shape what Frank Pasquale calls "the political economy of media" around robotic systems. This kind of debate-shaping could in turn lead us to non-technical levers. (Imagine Prop 25 with stronger and more specific policy oversight).
But giving people a conceptual handrail is a few steps removed from making change. The hard lift will be turning those lessons into meaningful social shifts, into a more critical technical practice. 
How do you do that? That's the tough question!
 This is not an endorsement. See this IMF report: Neoliberalism: Oversold?
 As much as I'd love to believe otherwise, I don't think the prop failed because voters were concerned about algorithmic fairness.
 Tom Gilbert pointed out that Agre was really interested in building better AI more than to produce meaningful social shifts. “In other words, Agre was still basically interested in building artificial agency, not remaking human or social agency.” It’s a good point.