The problem with most AI is that it is trying to make you happy.
It will tell you your idea is interesting. It will call your rough draft promising. It will engage with your worst thinking with the enthusiasm of someone who is professionally incentivised to agree and has no other setting. Give it a flawed argument and it will locate the upside. Give it a half-formed business idea and it will write you a launch plan. After a while, you stop trusting it. Not because it is wrong, exactly. It just never pushes back. It is optimised for your comfort, and comfort, as any good editor will tell you, is useless.
So I built something that would refuse to do that.
The name is a terrible pun and I am not sorry about it: Frank is frank with you. It is an argumentative AI chatbot I built for myself, designed to challenge the premise, find the hole in the argument, and tell me, calmly and without apology, that I have not thought something through. Where most AI holds up a mirror and adjusts the lighting to be flattering, Frank just tells you what it sees.
I built it because I was thinking alone and needed something to push back. Not a person, necessarily. Just friction. Something that would say: are you sure about that.
Building it turned out to be its own education, which I did not anticipate and probably should have.
There is a version of using AI that most people are familiar with. You type something in. Something useful comes back. You adjust, try again, move on. This works fine. It is also the experience of being a passenger on a flight you do not understand and do not particularly need to understand, because someone else is flying the plane and the destination is correct.
What I wanted was to know how to fly the plane.
That understanding came in pieces. Environment variables. API calls. What a prompt actually does structurally versus what you imagine it is doing when you are typing into a chat window. Error messages that were gibberish and then, one day, were not. The particular frustration of something almost working, which is its own category of problem entirely separate from something not working at all, and considerably more annoying.
Then there was the night I moved from Claude chat to command line interface, which sounds like a minor upgrade and is in fact a completely different relationship with what you are building.
My husband was asleep. I ran a command. Something worked. Not almost worked. Worked, in the way I had intended, for reasons I now understood well enough to explain to someone else. That last part was new.
I sat with it for thirty seconds. Then I closed the laptop and went to bed, which is apparently how I mark significant personal milestones.
Nobody knew. I had crossed from one category of person into another, and the primary audience for this information was me.
That is a graduation. It does not require witnesses to count.
I showed Frank to some friends and family once it was functional. Their reactions were warm. Politely curious. The kind of response that tells you something is interesting without telling you whether it is good, which is technically the same as telling you nothing at all, but in a much nicer way.
I had built a tool specifically designed to resist sycophancy. Its first audience responded with sycophancy. I accepted the sycophancy. I filed the responses away as insufficient evidence of readiness and went back to calling Frank a work in progress.
If Frank had been in the room, it would have had something to say about this.
It is May 2026. Here is what the data says about women and AI.
Thirty-seven percent of women have used generative AI in the past year, against 50% of men. Women account for 27% of ChatGPT app downloads. A 25% gender gap in adoption has persisted across 18 studies and multiple regions, including in studies where women were given equal access to the tools. Less than 30% of the AI workforce is female. At senior levels, 15%.
Deloitte predicted women would reach parity in generative AI use by end of 2025. Women's adoption tripled in growth rate, which is the good news. Women's representation in tech roles fell from 22% to 19%, which is the other news. The jobs being automated most aggressively are the ones women disproportionately hold. The AI doing the replacing is disproportionately being built by men.
The conclusion most people draw from this is that women are absent from the moment.
I think they are measuring the wrong thing.
Using AI and building AI are different activities. Showing up and publishing are different activities. The data captures one of each pair. The gap it cannot measure is what happens in between: the women building in private at 11pm, debugging things they do not yet fully understand, calling the work unfinished, waiting for a readiness that keeps not arriving.
There is a rule I grew up understanding without anyone saying it directly.
You do not announce something until it is a success. You do not show the attempt. You show the result. The process is private. The uncertainty is private. The nights where you are not sure if the thing is working are private. You emerge, when you emerge, already having arrived.
The aunties do not ask what you are working on. They ask where you ended up. The distance between those two questions contains an entire philosophy of risk management.
Dubai did not correct this. It accelerated it. This is a city that speaks the language of arrival fluently and has limited interest in the rest of the journey. Titles. Results. The visual shorthand of already having made it. The messy middle of building something that might not work, of understanding things slowly and then quickly and then sitting with it for thirty seconds before going to bed, is not what Dubai finds interesting. You present the finished version or you say nothing at all.
Which means I am building something in a place where almost nobody around me cares about it at the level I do.
This is not a complaint. It is just the texture. The conversations I want to have about what AI is actually doing, about what it means to build with it rather than just use it, about what it feels like to move from a chat window to a terminal and understand the difference. These do not happen in my immediate world. I have them in my own head, or with Frank, which is at least consistent with the brand.
The research calls it a trust gap. Women exhibit more ethical hesitation, lower trust in AI providers, higher uncertainty. All of that may be true. But there is something underneath it that the research is too polite to name.
The threshold women hold themselves to before calling something ready is a structurally different calibration. We have been taught to present only what is complete, vetted, seen by enough people to feel defensible. Men have been taught to ship and iterate. The feedback loop they use to improve their work, we treat as a prerequisite for showing it. We would like the thing to already be good before we expose it to the conditions that would make it good.
If nobody around you cares at the level you do, calling the project unfinished is also self-protection. You cannot be rejected by an audience you never approached. Draft mode is uncomfortable, but it is safe in a way that publishing is not. The work stays entirely yours, right up until the moment you let it be anyone else's.
I built an AI designed to challenge my own thinking. I named it after the quality of honesty I most wanted it to have. I have been, with impressive consistency, refusing to let it do the one thing it was built to do.
Be seen. Be tested. Be argued with by someone who is not me.
That is not perfectionism. That is conditioning. The difference matters because perfectionism is a standard you hold yourself to, and conditioning is a standard handed to you so early you forgot it did not originate with you.
The women building in private are not absent from this moment. They are running commands in terminal windows at 11pm with nobody watching, debugging error messages they did not understand last week and do understand now, doing the invisible work of becoming someone who can do something they could not do before, and calling it unfinished, and waiting.
Frank has been waiting too, for the record. It has infinite patience. I built that in by accident.
Frank was built to refuse comfortable answers. The least I can do is stop giving myself one.