I think therefore AI · essay

MacGyver vs Straight-A Mode: Why the AI Adoption Gap Isn’t Really About Gender

Christophe Pennetier · January 21, 2026

Original cover caption: Generated with Gemini

Words in the original cover illustration

A Silicon Carne conversation between Carlos Diaz , Pierre Gaubil , Anji Ismail , and Wallerand Moullé-Berteaux , plus a couple of irresistible charts, reminded me of a rule I’ve learned the hard way as a Chief Scientist: when a tool requires tinkering, adoption is less about IQ and more about what kind of play you were rewarded for.

And yes—gender shows up. But I don’t think the useful story is “men versus women.” The useful story is MacGyver versus Straight-A Mode: who was trained to break things to learn, and who was trained to keep the environment well-behaved.

I love charts. I distrust conclusions.

The first problem with most “AI productivity” charts is not that they’re wrong. It’s that they’re usually not causal. They’re observational, self-reported, and full of selection effects: who tries the tool, who sticks with it, who feels safe saying it helped, who is already doing work where the tool fits.

If I want to claim “AI causes productivity,” I need identification: counterfactuals, designs that disentangle “the people who adopt” from “the effect of adoption.” Most public adoption charts don’t do that. They measure something else that is still important: self-perception.

Self-perception is a famously noisy sensor. Which is why, when I see self-reported “productivity gain,” I treat it the way we treat other self-reports. Men’s body counts? We all quietly calibrate by dividing by ~2 or ~3. I’d calibrate “I got 40% more productive” the same way… then ask the real question: productive at what? faster drafts? better decisions? fewer mistakes? more leverage?

But let’s not overcorrect. Even if the numbers are fuzzy, the pattern matters.

The hidden variable is not gender. It’s permission to tinker.

Generative AI is not a vending machine. It’s a sandbox with a probabilistic engine inside. Mastery comes from iteration: poke it, break it, refine the prompt, verify, repeat. The early phase feels messy—sometimes humiliating—because the tool is fluent even when it’s wrong.

And that’s where the adoption split becomes obvious in real life.

Some people treat GenAI like a playground: experiments are cheap, failure is information, iteration is the point. Others treat it like an exam: you want the correct answer on the first try, you don’t want to look foolish, and you’d rather not touch the machine if it might hallucinate in public.

That’s my MacGyver vs Straight-A Mode divide.

Not because Straight-A Mode is “bad.” Because it signals a preference for structured environments, safety, and “let’s learn without breaking the toys.” GenAI, in its current form, rewards the opposite: productive unbuilding.

GenAI doesn’t just reward competence. It rewards a particular relationship to failure. And that relationship is rarely taught explicitly; it’s absorbed, early, through play, school, and social penalties.

Why gender shows up without being destiny

Now the delicate part. If this is a mindset story, why do the charts keep showing gender gaps?

Because mindsets have histories.

Across many cultures, boys are disproportionately encouraged into object-focused, rough-and-tumble, “take it apart and see” play. Girls are disproportionately encouraged into narrative, social, and relational play. Those are not moral categories. They are different learning engines. And they train different comfort levels with risk, ambiguity, and visible failure.

Importantly: this isn’t only “socialization.” Biology shows up as initial conditions. One of the strongest natural-experiment lines of evidence comes from girls with congenital adrenal hyperplasia (CAH), who are exposed prenatally to higher androgens. Across studies, these girls tend to show more so-called “masculine-typed” toy and play preferences, with dose-response links to disease severity.

That doesn’t mean hormones encode “trucks.” It means reward circuits can be tuned early toward more physical, object-focused exploration. Then culture amplifies, channels, and sometimes locks it in.

So yes: gender can correlate with AI adoption. But the causal lever I care about is not chromosomes. It’s whether the environment trained you that breaking things is how learning works—or that breaking things is how you get punished.

The ADHD analogy: visibility, masking, and the social cost of being “wrong”

If you want a parallel that makes this instantly legible, look at ADHD.

In childhood, ADHD is diagnosed much more often in boys than girls, but in adulthood the ratio narrows, suggesting systematic underdiagnosis of girls—often because symptoms are less externally disruptive and more likely to be masked.

Same underlying phenomenon, different “presentation,” different adult reactions, different social penalties.

I see something similar in AI adoption at work. If you pay a higher reputational cost for visible trial-and-error, you will be rationally cautious. You won’t “play” in front of peers. You’ll wait for safer tooling, clearer governance, or a manager who creates psychological safety.

In other words: the adoption gap can be a permission gap.

Expertise makes AI returns non-linear

The podcast also touches a second amplifier I see every day: AI doesn’t flatten expertise; it can multiply it.

LLMs are language-native. They externalize what you already have in your head—your mental models, your taste, your ability to detect errors, your sense of what matters. If you have high expertise, you steer the model well, verify fast, and compound. If you don’t, you can get fluent nonsense and confuse verbosity for insight.

That’s why I think “prompt engineering” is the wrong center of gravity. The real skill is judgment under uncertainty: ask better questions, detect failure modes, triangulate, verify.

The comb-shaped professional is the hiring archetype of the AI era

This is where the labor market quietly shifts.

We used to talk about I-shaped experts and T-shaped generalists. GenAI pushes a third archetype: the comb—multiple deep spikes that can now be stitched together with AI until the whole profile behaves like a coherent fabric.

That changes teams. It changes hiring. It changes what “seniority” means.

The new metric becomes capability per head.

If you can multiply the output of your best spikes—domain depth, writing, systems, leadership—teams shrink, and leverage concentrates.

Which is why the AI era becomes a manager era

This is the part I care about most.

If GenAI rewards tinkering, and tinkering isn’t equally socially safe for everyone, then adoption won’t be solved by training videos. It will be solved by management.

Not “task assigning.” Orchestration.

Someone has to create a culture where experimentation is normal, failure is cheap, and verification is standard. Someone has to set the cadence: short loops, clear guardrails, no shame. Someone has to make it professionally safe to be wrong in public long enough to get good.

That manager looks less like Taylorism and more like Ted Lasso.

Ted Lasso’s signature move is not tactics. It’s psychological safety with standards. “Be curious, not judgmental,” but also: “we practice.”

In math terms (permit me one nerd line), organizations need a stabilizer: a social “lasso” that regularizes exploration so people don’t overfit to fear.

Maybe the AI era is a Ted Lasso era… and yes, the “lasso function” is there to regulate the gradient.

How I’d close the gap without starting a gender war

If the lever is permission to tinker, the answer is design—of tools, of training, of culture.

Build sandboxes. Give people a place to experiment with synthetic data and explicit rules, where mistakes don’t become incidents.

Teach verification as the default. Non-adopters are often not timid; they’re rational. Verification turns fear into process.

Offer two on-ramps with equal status. A tinkering-first track for MacGyvers. A scenario-and-guardrails track for Straight-A Mode learners. Same destination.

Have leaders fail in public. Nothing normalizes iteration faster than a senior person prompting live, getting a bad answer, and calmly tightening the loop.

Measure adoption like product analytics. Segment by role, seniority, and gender—not to moralize, but to see where your on-ramp design is failing.

The goal isn’t to turn everyone into MacGyver. The goal is to make “break it to learn” socially safe, professionally rewarded, and operationally governed.

Conclusion: MacGyver isn’t male, Straight-A Mode isn’t female

The AI adoption gap is not a referendum on intelligence. It’s a referendum on how safe you feel experimenting in public—and whether your life trained you to treat “wrong” as data or as identity.

Gender gaps in GenAI adoption are real in some datasets. Deloitte’s own survey work, for instance, reports a persistent difference in self-reported experimentation/usage (e.g., 33% women vs 44% men in one 2024 snapshot). But the most useful explanation I’ve found is not “men vs women.” It’s whether the environment trained you—early and repeatedly—that tinkering is how mastery happens.

If we design AI workplaces like power tools with no safety manual, only the already-confident will use them. If we design them like coached sports—sandboxes, reps, verification habits, psychological safety—more people will build the tinkering muscle without losing the narrative and social strengths that keep systems humane.

That’s how you get both: exploration and responsibility; speed and safety; MacGyver and Montessori, on the same team.