Why "Ethical AI" Guidelines Keep Failing the People They're Meant to Protect

In 2018, researchers tested several commercial gender-classification systems, the kind used in everything from photo tagging to security software. The systems worked well for lighter-skinned men. For darker-skinned women, the error rate was dramatically higher. No one had set out to build a biased system. The companies involved had ethics statements. Fairness principles. Good intentions on paper. The gap between what AI ethics guidelines say and what AI systems actually do is the problem I want to talk about in this piece. And I think the fix isn't just better guidelines, it's a different way of building the systems in the first place.

The guidelines aren't the problem.

Over the last decade, governments, companies, and universities have produced no shortage of AI ethics frameworks. Fairness. Accountability. Transparency. Safety. You've seen the words. But researcher Thilo Hagendorff, after reviewing dozens of these guidelines, found a pattern where these guidelines converge on the same abstract principles, and almost none of them come with a way to implement those principles. A company can publish a beautiful ethics statement and still ship a biased hiring algorithm, because nothing in the guidelines requires otherwise.

Bias in AI is usually baked in at the data level, long before a model is ever tested. Data reflects the world it was collected from, including that world's inequalities. When a model is trained to optimise for "success" as historically defined, it can learn to reproduce exactly the disparities we'd hope it would correct. And it's not always obvious where the bias enters. A variable can look neutral, for example, a zip code, a college major, or a gap in someone's résumé and still function as a stand-in for race, gender, or class. You don't need a system to see a protected trait to discriminate against it.

Why can't you just "fix the metric"?

The problem is, even if you aim for a perfectly fair system, it’s often impossible to meet every fairness criterion simultaneously due to mathematical constraints. Researchers have demonstrated that when groups have different base rates, you must prioritise one type of fairness over others (achieving all at once isn’t feasible). This cannot be fixed through engineering alone, values matter in such situations and understanding the context in which the model is to be used, the expected results of the model and its usability. For this reason, fairness in AI therefore extends beyond technical issues to political debates about whose definition of fairness matters the most. Critics of the “ethical AI” movement highlight that framing bias merely as a technical bug to fix can sidestep difficult discussions about the structural inequalities underlying the technology.

What "active inclusion" looks like in practice

So if guidelines alone don't work, what could?

I'd argue for something more demanding: active inclusion - building equity into every stage of an AI system's life, not checking for it at the end. In practice, that means:

• Auditing data sources before modelling begins, not after a system has already been deployed - checking who's represented, who's missing, and why.

• Documenting datasets and models transparently, using tools like datasheets (which explain a dataset's origin and intended use) and model cards (which report how a model performs across different demographic groups).

• Bringing affected communities into the room early - not to review a finished product, but to help define what "success" even means before the first line of code is written.

• Monitoring after deployment, especially in high-stakes areas like hiring, where vendors' fairness claims are notoriously hard to verify from the outside.

None of this is new, versions of it already exist in pockets of the AI research community. What's missing is treating it as the default, not the exception.

The real question isn't "is this AI fair?"

It's: who got to decide what fair means, and when in the process were they asked?

Guidelines that arrive after a system is built are a compliance exercise. Inclusion that starts before the first dataset is collected is a design principle. The second one is harder, but it's also the only one that takes into consideration lived experience and centres the voices of users, especially those already experiencing the extremes of being marginalised by these systems.

This piece draws on research into AI fairness, sociotechnical systems, and participatory design, including work by Buolamwini & Gebru (2018) on algorithmic bias, Hagendorff (2020) on AI ethics guidelines, Selbst et al. (2019) on fairness in sociotechnical systems, and Gebru et al. (2021) on dataset documentation.

Contact

I’m always open to collaboration and conversations about inclusive AI and accessibility.

Email

tmakati04@gmail.com

© 2026. All rights reserved.