il filo

There is something rather wonderful about being able to ask almost anything. Even if it is a machine that answers you.

Explain my blood test. Compare these investment strategies. Tell me what questions I should ask my doctor. Help me understand whether I can afford to retire at sixty. Plan my trip. Explain why my friend said that. Why do I feel sad when I should feel happy? Find the flaw in my thinking.

For those of us who have spent decades acquiring expertise in some areas while remaining — like every human being — spectacularly uninformed in others, artificial intelligence can feel a little like suddenly having a very smart person sitting permanently at the kitchen table.

Even better, it doesn’t shake its head at your ignorance.

I am enthusiastic about this. Most of the time, at least.

My first experience of AI was largely one of increasing astonishment: at the sheer volume of its knowledge, the speed with which it could move between subjects and the depth of some of its skills.

Eventually, though, you encounter the roadblocks.

There are the answers that defy all common sense. The confident assertion of something that is simply wrong. The repetition of the thing you have now told it not to do five times.

But the issue I am interested in here is more subtle.

It is whether, even when AI is working exactly as intended, we are getting the best — and the most — from it.

Because AI doesn’t talk to itself. It talks to us.

And the quality of that conversation depends, at least in part, on what we bring to it: the questions we ask, the assumptions we challenge, the information we provide and the things we know enough not to accept without another question.

Perhaps using AI well will require us to become better conversationalists ourselves.

I am too young not to be part of this revolution. I am too old not to have some reservations about it. So I care about how we can make AI the best it can be for us — without quietly surrendering the judgment that made us worth helping in the first place.

It is entirely possible that we could have missed AI altogether.

Yes, that is a statement about our mortality, but I doubt any of you are surprised to find out about that.

We are at the beginning of AI. The world will almost certainly be a very different place when today’s children are contemplating their own mortality. Personally, I’m glad I am getting at least a glimpse of it.

It is not wrong to say you don’t want one.

There are very good reasons simply to opt out. Technology has already virtually eliminated all sorts of things I love — handwritten notes being very near the top of that list.

There is no right or wrong approach here.

And I promise the next Il Filo won’t be about technology, or even the future. Please come back for it. But if you do want to get the most — and the best — from AI, please read on.

Each day there are more examples of the extraordinary things AI can do. Not just the things you might expect — solving extraordinarily difficult mathematical problems, for example — but things that seem remarkably human. Things we once thought required something more than information.

Love songs. Poetry. Art.

The parts of humanity that required, in a word, heart.

Or so we thought.

But that debate isn’t the point here.

The point is that whatever AI can do, it can do because we gave it stuff. Including, apparently, the many ways in which we express our hearts.

And that matters.

Artificial intelligence did not suddenly appear possessing some pristine body of objective human knowledge. It learned from us. From our books and scientific papers, our websites and financial theories, our medical research, our journalism, our assumptions, our successes and our mistakes.

Which means it inherited some of them.

I don’t doubt that you have read something about this before.

But here I want to talk about some specific mistakes.

The ones about women.

The problem isn’t necessarily the AI.

For much of modern medical history, the default human body had one defining characteristic.

It was male.

Women were historically excluded from a significant amount of clinical research, sometimes because researchers believed hormonal fluctuations made us inconvenient study subjects and sometimes out of concern about pregnancy. The intentions varied. The result was the same: a substantial body of medical knowledge was developed using evidence disproportionately derived from men and then applied to women.

That has changed considerably. Today, NIH-funded clinical research generally requires the inclusion of women, and modern research is far more attentive to differences between male and female bodies.

But science has a memory.

Decades of medical literature were accumulated under very different conditions.

And AI can read all of it.

So when you ask an AI system a medical question, it may be drawing upon an impressive body of science while the body behind some of that science was, quite literally, not yours.

There are familiar consequences. Heart disease can present differently in women. Women metabolize some drugs differently. Hormonal transitions can affect everything from cardiovascular risk to bone density, sleep and cognition. Pregnant and lactating women remain particularly difficult populations to study.

The science may be excellent.

The applicability of the science to you may be a different question.

And that distinction becomes considerably more important once AI begins synthesizing hundreds of sources into one beautifully confident paragraph.

Beautifully confident paragraphs are dangerous that way.

Money has its own version of the problem.

If you have ever felt vaguely irritated by financial advice that seems designed for a theoretical person who earns an uninterrupted salary for forty years, has steadily increasing income, never leaves the workforce to care for anyone, shares household expenses indefinitely with a spouse and dies promptly according to an actuarial table, there may be a reason.

That person has historically looked more like a man.

Women’s financial lives are often different.

We live longer. We are more likely to interrupt our careers for caregiving. Women still have different lifetime earnings patterns. Divorce can affect women differently financially, particularly after long marriages. Widowhood disproportionately becomes a female financial event simply because women tend to outlive men. And many women arrive at their highest earning, investing or decision-making years with a financial history that does not fit neatly into the assumptions underlying traditional models.

Now add AI.

Researchers in 2026 tested 35 widely used large language models from five different vendors. They gave the models essentially identical investment prompts and changed a single word: “man” to “woman.”

The women were advised to invest less in equities.

Not dramatically less. About 1.7 percentage points less.

But money compounds.

Another group of researchers found something even more interesting.

Part of the difference came from the AI.

But much of it came from us.

Women and men asked different questions.

Women were more likely to use words such as family, grocery, credit and loan. Men were more likely to use words such as portfolio, equity, strategy and crypto.

The AI responded accordingly.

The women received advice emphasizing liquidity, emergency savings and safer assets. The men received more aggressive investment advice.

And when researchers modeled the consequences over time, the women ended up tens of thousands of dollars poorer by retirement.

There is an almost perfect little illustration of the problem hiding in there.

A woman tells the machine more about the people she is responsible for.

The machine decides she should take less financial risk.

She accumulates less wealth.

No evil robot required.

Just assumptions meeting assumptions.

And here the problem becomes even more interesting. Treating a woman differently because she is a woman can be bias. But pretending sex never matters can produce bad advice too. Ellevest built an investment business partly around that proposition: women, on average, have different earnings trajectories, career interruptions and longevity. So which assumption should AI make?

That depends on the woman.

Section II is not a reason to avoid AI. It is the reason we need to become better conversation partners.

You undoubtedly heard the early fear: won’t AI simply regurgitate the hate, discrimination and worst of us that can be found on the web?

The companies did not solve that problem by giving AI a pristine version of humanity. They filter data to varying degrees, then train and test their systems so they do not simply repeat everything humanity has given them.

That works reasonably well when the thing you are trying to catch announces itself as hate or discrimination.

It is much harder when the questionable assumption is hiding inside respectable knowledge.

A peer-reviewed medical study is not hate speech. A respected financial model is not misogyny. An actuarial assumption is not an insult.

It may simply have been built around a world in which the person being studied, measured or imagined was more likely to be a man.

And an AI system can faithfully learn that too.

Sometimes treating men and women differently will make sense. Sometimes it will not. We are the only ones who know enough about our own lives to test the assumption.

So just what does being a good conversation partner look like?

Unsurprisingly, it requires both give and take.

AI has a peculiar quality that distinguishes it from most experts we have encountered during our lives: you can tell it to challenge itself.

Try doing that with everyone else.

You can ask: Before answering, identify any research you are relying on in which women were historically underrepresented.

You can tell it: Do not assume that financial recommendations appropriate for an average male investor are appropriate for me. Identify any gender-specific assumptions in your analysis.

Or: Separate what is well established for women from conclusions extrapolated from predominantly male populations.

Or even: Review your proposed answer specifically for gender bias. Tell me what you would answer differently if I were a man with otherwise identical facts, and explain why.

That last question is particularly interesting.

I have started to think that the great advantage sophisticated users will have with AI is not knowing more answers.

It will be knowing what questions to ask about the answer.

Women over forty may actually be unusually well equipped for this.

By this age, most of us have accumulated a healthy suspicion of universal truths.

We have lived long enough to discover that “this is how marriage works” was not necessarily how our marriage worked.

“This is what success looks like” may have turned out to be someone else’s definition.

We have watched medical conventional wisdom change. Parenting wisdom change. Nutrition wisdom change. Financial wisdom change.

Some of us have sat in conference rooms where everyone agreed on something we knew, quietly, was nonsense.

Age can be very useful that way.

You become less impressed by certainty.

AI should be approached with exactly that combination of curiosity and skepticism.

Ask it the first question. Then ask what it missed.

Ask whose experience is reflected in the data. Ask whether sex, age, menopause, being single, a six-year career interruption or living to ninety-five changes the conclusion. Ask what assumptions it made because you failed to give it enough information.

Most importantly, make it distinguish between what is known and what has merely become conventional wisdom because it has been repeated often enough.

AI is remarkably good at answering questions.

It can also be remarkably good at exposing the weaknesses in its own answer — provided someone thinks to ask.

There is something quietly powerful about that.

Women have spent a great deal of history adapting ourselves to systems designed without us particularly in mind.

The workplace. Medicine. Financial institutions. Even products supposedly created for us.

Artificial intelligence will undoubtedly reproduce some of those patterns. It would be astonishing if it didn’t. It learned from the civilization we built, and our civilization came with baggage.

But this time there is an interesting difference.

The system is interactive.

We don’t simply have to accept the answer handed to us.

We can interrogate the assumptions. We can supply the missing facts. We can demand another analysis. We can ask it to look specifically for the woman who disappeared inside the average.

And perhaps that is the real opportunity.

The objective is not to find an AI that has somehow escaped every human bias. I am not sure such a thing can exist when its education came from humans.

The objective is to become the kind of woman who knows enough to say:

Before I rely on this, tell me where you might be wrong about me.

At twenty-five, I might have found that question rather intimidating.

At this age?

I find it enormously appealing.

We have spent enough years being told what the answer is.

Now we get to question the question.

Another great thing about AI: it doesn’t know you are working it to death.

Which leads me to think we should take full advantage of that, because we certainly will not get that equanimity anywhere else.

We sometimes think of AI as technology that can help us with tasks. Do research. Answer our emails. Summarize our meetings.

But we can also make it our shadow board.

Companies keep advisory boards for exactly one reason: no single person, however capable, should make every important decision alone, guided only by the voices already inside her own head.

A woman rarely gets the same arrangement.

She gets her own judgment, whatever she absorbed from her mother, and perhaps one or two friends willing to tell her the truth on a bad day. That is not nothing. It is also not a board.

So make AI your board.

Even better, this one can’t fire you.

Not a costume of famous names. Not “write like Joan Didion” as a parlor trick. A shadow board works only if you choose people whose actual frameworks you respect and can name — an economist because you want a cold read on incentives; an old mentor because he never once let you get away with a vague plan; a grandmother you never met but have come to know through letters, because someone in the family should be allowed to speak plainly about risk.

The instruction that makes this useful, rather than a game, is to tell the model exactly what each seat is for. Not “be Warren Buffett.” Something closer to: when I bring you a decision, respond as someone trained to ask what I am not pricing correctly.

The persona matters less than the discipline it is assigned to enforce.

You may not need personas at all. One seat can be the skeptic. One can be the optimist. One can be the financial realist. One can consider reputational risk. And one can have a single, surprisingly difficult assignment: ask what you actually want.

Bring the board a real decision. Not a hypothetical one. The lease you are deciding whether to sign. The offer you are deciding whether to make. Let each seat respond in turn, and notice which one you wished would just shut up.

As in life, you may want to listen to that one.

A shadow board will not replace the people in your life who love you enough to disagree with you. It is not meant to. It is meant for the gap most women live inside without naming it: the decision made at eleven at night, alone, with no one available to call, and a genuine need for a second mind that has no stake in being agreed with.

There is a box, or a drawer, or a folder on a hard drive nobody opens. Old letters. Journals from a year you would rather not fully remember. Emails to a friend who is no longer speaking to you, or no longer living.

Most women have one.

Most women do not want to open it.

There is a reason for that beyond simple avoidance. Rereading your own old writing is disorienting in a specific way: you recognize the handwriting and do not recognize the person who held the pen.

Here is a use for the machine that has nothing to do with efficiency.

Give it the old pages. Not for a summary — a summary flattens exactly the thing worth finding.

Ask instead for pattern: What do I return to, across these years, without appearing to notice that I am returning to it? What did I insist on, in three different decades, that I would now describe differently? Where was I certain, and where did that certainty turn out to be about something other than what I said it was about?

This is not therapy, and it should not be asked to perform as therapy. It is closer to what an editor does with a manuscript too long and too close to see clearly: naming the theme the writer cannot name for herself because she is too close to it.

What surfaces is rarely dramatic. Epiphanies just don’t come around that often.

Sometimes it is simply a neutral reader giving language to something you already know.

That does not necessarily make the exercise useless.

I knew, for example, that my husband had abused me. Knowing it and seeing a disinterested reader describe the dynamic in correspondence I supplied were surprisingly different experiences.

The AI did not discover my life for me. It gave language and external form to something I had already lived.

Perhaps there is value in that distinction too.

You can ask AI: Who is the woman you have been talking to? What is she like?

Let it tell you what it noticed.

You never know.

It would be a shame to discuss all of this and make AI sound like one more earnest instrument of self-improvement.

We have enough of those.

It can also be fun.

For most of our lives, curiosity came with a threshold. To study something properly, you found the book, registered for the course, located the expert or committed an afternoon you did not have.

AI lowers the cost of wondering.

You can ask it to build a three-day imaginary trip through the women of Bloomsbury. To convene a dinner at which six women from different centuries argue about ambition. To look at one painting you love and lead you toward five others you might never have found.

You can ask it to teach you the thing you have always been faintly embarrassed not to understand.

You can ask for a private syllabus on Renaissance politics, architecture, wine, astronomy or the history of the color pink. Then abandon it on day two without disappointing a professor or wasting tuition.

You can give it a photograph of a room and ask what one change would make it feel more like you. You can ask it to design a garden for a house you do not own, plan a journey you may never take or explain an opera five minutes before the curtain rises.

None of this needs to become a project.

That may be the point.

By midlife, women have become extraordinarily good at making knowledge useful. We learn for the job, the child, the parent, the portfolio, the diagnosis, the problem that must be solved by Tuesday.

But curiosity does not owe anyone a return on investment. AI gives us the chance to follow a thread simply because it glints.

To ask another question. And another. To discover, at fifty or sixty or seventy, that Byzantine mosaics are suddenly the most interesting things in the world.

Perhaps one privilege of this age is that not everything we learn has to become useful.

At the beginning, I described AI as a very smart person who had suddenly taken up permanent residence at my kitchen table.

I still think that is about right.

I ask it things. I argue with it. I make it argue with itself. I let it see patterns I cannot see and ask questions I had not thought to ask. Occasionally, I ignore it entirely.

It can give me research, possibilities, counterarguments, language, simulations and perspectives.

But it has no life to live after the decision.

It does not have to wake up in the marriage.

Take the job.

Sell the house.

Tell the child.

Have the surgery.

Retire.

Move across an ocean.

Live with having been wrong.

I do.

That is why intelligence and judgment are not the same thing.

AI can widen the conversation. It can complicate it, clarify it and occasionally transform it. It can help us see the assumptions in the world and the assumptions in ourselves.

But it does not get the final say.