There is a sentence written in 1923 that ought to haunt everyone who has ever published an AI ethics statement.
It appears in a British Colonial Office whitepaper, formally titled Indians in Kenya, issued that July under the Duke of Devonshire. The document was Britain's official ruling on a bitter political fight in its Kenya Colony, and buried inside its careful prose is a line of startling moral clarity:
"Primarily Kenya is an African territory... the interests of the African natives must be paramount, and if and when those interests and the interests of the immigrant races should conflict, the former should prevail."
Read that today and it sounds like justice. It sounds like a government naming the right beneficiary and committing itself, on the record, to protect them. It is the kind of sentence you would be proud to put at the top of a policy page.
It was also, in practice, almost entirely hollow. And the reason it was hollow is the same reason so much of today's "responsible AI" is hollow. We write this from Nairobi, in the country that whitepaper was written about, because the lesson did not stay in 1923. It simply changed vocabulary.
What the paper actually did
To understand why "paramountcy" rang so false, you have to see the fight it was written to settle.
Colonial Kenya in the 1920s held three groups locked in an unequal contest. White settlers, few in number but politically dominant, wanted to entrench their control: an elected majority reserved for Europeans, exclusive ownership of the fertile "White Highlands," and a path to self-government. Indians, more numerous than the settlers and many descended from the laborers who built the railway, demanded equality: a common electoral roll, the right to own land in the Highlands, an end to segregation. And Africans, the overwhelming majority of the population, were barely admitted to the conversation at all.
So the whitepaper had to rule on a dispute. And here is the move worth studying. It resolved the settler-versus-Indian conflict largely in the settlers' favor: it rejected the Indians' demand for equal franchise, handed them a separate communal roll with only a few reserved seats, and preserved the White Highlands for Europeans. Then, to justify overriding the Indian claim to equality, it reached for the loftiest possible principle. It declared that neither immigrant group could dominate, because Africans came first. African paramountcy.
The principle that sounded like a commitment to the majority was, in the machinery of the document, a rhetorical instrument for settling a quarrel between two minorities. Africans were named as paramount and given, in that same moment, no meaningful vote, no restored land, no seat at the table where the decision was made. The declaration cost its authors nothing because nothing in the enforcement changed. For decades afterward, settler interests continued, in practice, to prevail over the very people the paper called paramount.
That is the anatomy of a hollow principle: a stated commitment, elevated precisely because it is expensive to contradict in words and cheap to ignore in practice.
The same sentence, a new century
Now open almost any frontier AI lab's ethics page.
"We put people first." "Safety is our highest priority." "We are committed to fairness, to the communities most affected, to the interests of those who cannot speak for themselves." The interests of the affected must be paramount.
The words are sincere, often written by people who genuinely mean them. That is exactly what makes them dangerous. Because sincerity is not enforcement, and a principle that carries no cost and no consequence is not a control. It is a declaration. And declarations, we learned in 1923, have a way of describing a world that does not exist.
Look for the tell. It is the same tell in both documents:
- Who is named as paramount, and who actually holds the vote? The Devonshire paper named Africans and empowered settlers. An AI charter names "affected communities" and empowers the shipping deadline. In both, the beneficiary in the prose is not the decision-maker in the room.
- What happens when interests conflict? The whitepaper had beautiful language for that moment — "the former should prevail" — and no mechanism to make it so. When your model's safety interest conflicts with a launch date, what actually prevails? If the answer lives only in a values statement and not in a blocking gate, you already know.
- Is the principle measured, or merely proclaimed? "African interests are paramount" was never audited, never scored, never tied to a consequence. "Our AI is fair" is, far too often, exactly the same: an assertion with no test attached, no threshold that fails a release, no evidence anyone could check.
A principle you cannot fail is a principle you are not really holding. It is decoration.
Why we anchor in standards, not sentences
At Reseni Labs we are, frankly, suspicious of high-minded language about AI, including our own. Not because the values are wrong, but because we have seen, in our own history, how a noble sentence can become a substitute for the thing it describes.
This is why one of our founding principles is deliberately unromantic: standards before opinions. We anchor recommendations in NIST, ISO, OWASP, and the AI Act before invoking taste. A standard is the opposite of the Devonshire sentence. It is a principle with teeth — a specific, testable, falsifiable claim that something either passes or fails. "We value privacy" is a declaration. "This system was assessed against a data protection impact assessment, and here is the artifact" is a commitment. One can be ignored at no cost. The other leaves evidence when you betray it.
And it is why we insist on being adversarial by default: assume your system will be attacked, and prove otherwise with evidence. Evidence is what the 1923 whitepaper never had to produce. Its paramountcy was self-certified. Nobody made the Empire demonstrate, with data, that African interests had in fact prevailed — because if they had been made to show the evidence, the whole hollow structure would have collapsed on contact.
The lesson transfers cleanly. Do not let an AI system self-certify its ethics. Make it produce the artifact. Make the principle fail loudly when it is violated, or admit it was never a principle at all.
The test the paper failed, and the one to apply
Here is the single question that separates a governing principle from a decorative one, and you can apply it to a colonial whitepaper and a model card with equal force:
What breaks when we violate this?
For African paramountcy in 1923, the honest answer was: nothing. No election was voided, no minister resigned, no land returned. The principle violated itself daily and the machinery ran on undisturbed. That silence is the whole story.
Ask it of your own AI governance. When your fairness commitment is breached, does a release actually block? When your data-minimisation pledge is violated, does a pipeline actually fail? When the interests you named as paramount conflict with the ones holding the schedule, does the document have any power to make "the former prevail" — or does it only have the words?
If nothing breaks, you have not written a policy. You have written a 1923 whitepaper. Beautifully worded, morally confident, and describing a world that does not exist.
The Devonshire paper is worth remembering not because its authors were uniquely cynical, but because they were, in a sense, ordinary. They reached for the highest principle available to justify a decision already made, and they never built the mechanism that would have forced them to keep it. The AI industry does this now, at scale, in fluent and sincere prose. The vocabulary is new. The gap between the sentence and the system is exactly one hundred years old.
Close the gap. Or at least have the honesty to know which side of it your ethics page is written on.
Reseni Labs is an independent research lab working at the intersection of privacy engineering, security research, and AI governance, founded in Nairobi and working globally. We help teams prove their AI systems are trustworthy with evidence, not assurances. If your principles need mechanisms behind them, get in touch.