How to use feminist prompting to overcome AI’s blind spots

In the world of generative AI, knowing what to ask an AI tool is crucial. But so is your thinking process. Enter feminist prompting…

It used to be that a prompter (sometimes called a prompt) was a person who cued actors when they forgot their lines or neglected to move to where they were supposed to be on the stage.

Gone are those days. Now a “prompt” is something you feed into a generative AI platform, in the hope that it will give you what you need: a recipe, a business idea, the text of a difficult email. 

The word, used in this way, is jargon. At its base it means “ask the AI tool something”. But when ChatGPT first burst on the scene in November 2022, it was inevitable that a host of new terminology would appear wherever the tool was being discussed.

A multitude of detailed frameworks for prompting were punted in those early days, but recent advice is that it really isn’t all that important. Generative AI models have improved over time and it’s often good enough to just start somewhere and type in what you want. 

But whatever you decide to enter into that box in your favourite AI, there’s one thing that’s non-negotiable. That’s the thinking you put into your task, the framework you are applying to your work. The more you think something through, the better your request is going to be. The more your own worldview is in your prompts, the closer the AI is going to be to giving you the result you want.

Enter intersectional feminism, and the concept of feminist prompting. I interviewed Kath Magrobi, Quote This Woman+ director for her ideas about the subject. Our lightly edited Q&A, which took place via email, is below. 

Question: In general, how do you approach prompting?

Answer: For me, prompting is a tool to deliberately re-shape the assumptions that LLMs work from: to use the tool itself to reveal its blind spots, and then to set out to correct them. I’m trying to get outcomes that align with African intersectional feminist thinking. Most AI is trained on a sea of patriarchal, colonial, classist and Western-centred data. If I just ask a question that appears neutral at surface level, I’ll get an answer that is anything but neutral! 

It will echo back the biases – conscious or unconscious – of those whose opinions hold the most power. I recently heard someone say: never forget that when you’re talking to AI, you’re talking to a rich, white, urban man who lives in the United States and is trained in the schmooziest PR: who is using you ruthlessly in ways you don’t begin to know, all while pretending to help.

I once asked an AI tool a very simple question: “Help me find African women experts who would be useful to newsrooms.” What came back was exactly what I expected: a handful of urban, English-speaking academics, often aligned to Global North institutions, and framed as rare exceptions in male-dominated fields. 

That answer helped me see what the default assumptions were. No rural women, no working-class voices, no LGBTQIA+ perspectives. I could then go back and re-prompt. I said: “Now redo this through an intersectional feminist lens. Name the power structures at play, for example but not limited to, patriarchy, classism, racism, the digital divide, and include the kinds of women who are usually excluded.”

The second answer was instantly more nuanced. For me, that’s the process: first, let the AI show its hand, then interrogate it against my feminist perspective, and only then reshape the response.

Q: How do you ensure the AI’s response acknowledges interlocking systems of power?

A: I build intersectionality into the instructions. That means explicitly asking the model to examine the issue through lenses of race, gender, class, sexuality, disability, age, geography, and so on. I often give it a checklist: “Your answer must consider how this plays out differently for women+, LGBTQIA+ people, disabled people, rural communities, and migrants.”

If you don’t specify this, the AI defaults to the loudest, most dominant narratives. I also remind it to look beyond identity categories and talk about structural inequalities – patriarchy, capitalism, racism – because otherwise it tends to individualise problems instead of naming the systems that create them.

Q: How do you overcome data gaps?

A: I treat the AI as an incomplete assistant, not an authority. If marginalised groups are absent from its training data, the only way to fix that is to force the AI to pause and ask: who’s missing here? 

A big part of my prompting is: “Identify gaps in whose perspectives are represented. Whose voices might be absent or misrepresented?” That nudge helps surface the silences. Then, I either instruct the AI to hypothesise based on existing feminist and human rights research, or I weave in external knowledge from the QW+ expert community. The key is not to let the AI off the hook. We can do it the greatest service by making it flag absences instead of pretending they don’t exist.

Q. Give me some examples of feminist prompts…

A: I work in feminism every day of my life, and these are the kinds of prompts I might use. 

• Analyse this news story through an intersectional feminist lens. Start by identifying which power structures (patriarchy, colonialism, classism, racism, ableism, heteronormativity) are visible or hidden in the narrative. Then suggest how rural women+, LGBTQIA+ people, and working-class voices might frame the issue differently.
• Draft this training agenda, but ensure that the language is plain, accessible, non-academic, and that it avoids clichés about women as ‘strong’ or ‘resilient’. Instead, centre their agency, dignity, and the systemic barriers they face.

But think about everyday life too! Here’s a prompt I did when I was babysitting an almost three-year-old who didn’t want to go to sleep:

Rather than: “Write me a 300-word bedtime story for a child encouraging them to fall asleep”, I asked this instead: 

Write a short, gentle bedtime story for a young child that helps them feel safe and sleepy. Avoid patriarchal, colonial, capitalist, and ableist tropes – so no kings, queens, castles, battles, money, or “perfect” bodies. Instead, centre diverse characters such as rural children, animals with differences, or children with disabilities, and show them solving problems through kindness, cooperation, and care for one another. The language should be simple, rhythmic, and soothing, with a focus on rest, community, and connection to the natural world. The story should be calming and no longer than 300 words.

And the story is here:

The star who wanted to rest

Once upon a time, in a wide, quiet sky, there was a little star who felt very tired.

She had been shining and sparkling all day, even though the sun was much louder and brighter.

“I think I need a rest,” whispered the star.

But the moon, kind and round, said, “It’s all right to rest. You don’t need to shine all the time to be important.”

So the star drifted lower, and her friends in the forest – an owl with one wing smaller than the others, a child who used crutches to walk, and a shy little rabbit – looked up and waved.

“Rest here with us,” they said.

And together they made a soft nest of leaves, feathers, and gentle songs.

The star curled up, listened to the rustling trees, and slowly, slowly, her glow dimmed into sleep.

High above, the sky was still beautiful, because beauty was never about who shone the brightest, but about how everyone shared the night together.

And soon, all the forest, the child, the owl, the rabbit – and the little star – were dreaming safe and sound.

• How AI was used in the making of this post: Kath used ChatGPT to help frame her responses; NotebookLM was used for research and question generation.

• This is an adaptation of a post that first appeared here.

Picture: WOCinTech Chat, Flickr (CC BY 2.0)