Welcome back to the Frontier Tech Hub's AI mini-series. If you missed our first issue, I would recommend catching up here.
This time we’re sharing our top responsible AI lessons, tech to battle corruption and herding cats on horseback. When it comes to deploying AI responsibly, we all wish there was a poster on the office door with a clear set of rules, but the reality (we’ve found) is that it doesn’t quite work like that.
The reality is also that the poster is sometimes part of the problem, as we’ll explore today.
The decision nobody calls a decision
When millions worldwide rely on food rations making their way along the aid supply chain to the most remote corners of the world, leakage becomes a priority challenge. We’ve seen pilots tackle humanitarian supply chain transparency from different points of the system, using blockchain, AI and IoT sensors.
When one pilot decided to implement facial recognition as the mechanism for accessing food rations for 800 million people, they had to abandon the idea. There were issues accessing the relevant database.
This is the pattern we're honing in on in this issue: the decisions that matter most are the ones nobody records as a decision. Sometimes, like here, circumstance makes the call before a person does, and no one writes it down because it never felt like a choice. Sometimes a person makes the call deliberately and still doesn't write it down, because it was instinctive. Either way, the decision goes undocumented, and a decision you can't see is one you can't learn from.
AI non-deployment is typically framed as a technical or resource failure rather than as an opportunity for ethical reconsideration. Build in the habit of asking "should we?" alongside "can we?" before circumstance answers on your behalf.
Over a decade of testing tech, we’ve seen pilots keep making responsible choices, but filing them under a different name. When a team narrows an ambitious system down to a simpler, more legible one, they call it a ‘technical pivot’. A team working with imperfect data (because it's the only data there is) makes a judgement about who that data leaves out, and whether to proceed anyway. They log it as a 'feasibility constraint'. But deciding to build on data that under-represents the most marginalised people (with eyes wide open), is an ethical decision as much as a technical one. Almost none get recorded as such.
They show up as technical choices, operational tradeoffs, scope reduction. And because nobody writes them down as ethical decisions, they vanish. The next team can't learn from a choice it can't see, or fund a safeguard that was never named.
It certainly never makes it onto the poster on the back of the office door.
When the government marks its own homework
In the Philippines, over 400 "ghost projects" have been uncovered in flood-control infrastructure alone. Billions of pesos have been disbursed for things that never got built. It’s a theft of public money, of physical protection, and of the trust that functional government depends on.
Today, it’s the kind of problem that invites a dramatic AI answer, and examples already exist. On the other side of the world, Brazil’s ALICE system screened 190k+ procurement processes in a single year, and its alert system helped trigger audits covering R$27 billion in contracts.
This is the counter-example to everything above. ALICE's most important design choice was made deliberately, in advance, and it's written down: it only flags as many cases as a real human auditor can actually act on. Flag too much and you bury the auditors in noise until they stop trusting the tool. Flag too little and you miss the theft.
The team chose to surface fewer cases than the algorithm technically could. This is a decision about how the tool fits the human using it, made on purpose and recorded as such. This is what it looks like when we don’t let the algorithm mark its own homework.
“I dislike the phrase “human in the loop” because it cedes authority to the machines. Let’s flip the narrative. It’s our loop, we work the same way we always have, now we recruit agents to join the team… Not as a loop we’ve been excluded from, instead as one we invite agents into.”
This brings us to the deeper point behind our title. Ghost projects happen when a government, in effect, marks its own homework. It awards the contracts and verifies the delivery inside the same house. The most resilient systems we’ve investigated share one principle: the state doesn't get to be the only one checking the state. In the Philippines, that independent layer is already assembling itself, emerging from public anger (an independent marker the government can't easily switch off). Volunteers behind BetterGov.ph have stitched procurement, infrastructure and political-dynasty data into one public dashboard (also not easily switched off).
Our new exploration is now live, profiling 6 international case studies, 8 technologies, and 5 trends on where anti-corruption technology is heading over the next five years.
When someone actually marks your homework
Earlier this year, three civil-society teams across Eastern Europe deployed working AI tools to defend their information environments*.
In Lithuania, a platform that generates synthetic content so teenagers can practise spotting it. In Bulgaria, a usable interface wrapped around an open-source detection tool, run on the eve of national elections to surface coordinated activity across Telegram. In Moldova, economic data has been made accessible enough to counter misleading narratives with actual numbers.
Across all three pilots, what drove uptake wasn’t the complexity or novelty of the tech - it was how well the tool matched its context. But when an independent reviewer looked across all three, they found the same gap in every one. The human oversight was there (journalists reviewed outputs, facilitators were always in the room), but there wasn’t a conscious decision to have it. The oversight existed in practice but had never been designed, documented or named as the safeguard it was.
There it is again: the decision nobody wrote down as a decision. The teams were doing the responsible thing on instinct, but an instinct you haven't named is one you can't hand to the next team. The decision was being made and immediately lost, because it didn't look like a decision. It just looked like getting on with the work.
There's a sharper edge to this example, because these teams are building AI to tackle AI challenges, and coming back to the title: we can’t let the algorithm mark its own homework. Every one of these pilots kept a person in charge - the lesson is to write that choice down as the safeguard it is, before the next team assumes the tool has it covered. The lesson for future work is to embed ethics, documentation, and safeguards at the start, not after the build.
One last note on this. Faced with AI-shaped threats, there's a pull to reach for AI-shaped defences. But a lot of online manipulation is high-volume automation wearing a more frightening label, and resources spent chasing the model may be guarding the least important door. Look past the part everyone's pointing at, and ask what's really carrying the harm.
*This marks the completion of our first portfolio dedicated to upholding information integrity in Eastern Europe. We set out to understand what it takes to exponentially scale the positive impact we know AI can have, and we’ll be announcing a second cohort later this year. Read the first portfolio’s insights here.
Where this leaves us: the anti-poster
This matters beyond tidy documentation. The Hub exists to generate evidence about what works, so others can adopt and adapt it themselves, but you can't evaluate a decision you can't see. When the ethical calls are filed as technical footnotes, the most important reasoning in a pilot never makes it into the record, and the next team has to rediscover it by luck. Naming these decisions as the ethical ones they are is critical to ensuring learnings about responsible AI survives.
The teams doing this well are in the habit of catching themselves and noticing when a "technical" call is also an ethical one, and saying so out loud, on the record, so the next team inherits the reasoning instead of having to get lucky. It’s slower and doesn’t scale the way a poster does, but it's the only thing we've seen actually work.
We’ll be back in two weeks to get practical: we’re launching a new home to house the (ahem, almost-award-winning) AI tools we’ve been testing and refining over the last two years. Each one is free to use for development practitioners globally.
🤠 Until then, remember that AI is another technology that helps us make better sense of information and knowledge. Remember what herding those cats was like before we had any kind of digital solution?
My “anti poster post-it.” Pick up a pen and make yours. I’d love to see what you write!
In other news
This is where we share recommendations from members. Our Network is all about collective insight, so if you’ve come across something worth sharing, get in touch!
Read Where AI Interventions Succeed or Fail from our friends at AI4D: an excellent exploration of six inflection points in the model-practice gap.
The 4th Annual Responsible Tech Summer Reading List has landed. We’re looking at ‘Prophecy: Prediction, Power, and the Fight for the Future, from Ancient Oracles to AI’. Hmu if you want to start a book club.
Read the The Sports Car Theory of LLM Adoption from Rachel Coldicutt, who starts by musing: “if LLMs make decision makers feel stupid, maybe they'll stop using them.”
AI enthusiasts are in a race against time, AI skeptics are in a race against entropy. This article explores how we close the gap between the two.
And one can never have enough metaphors for AI, as explored in this episode of Dreaming against the Machine, with AI philosopher Shannon Vallor.
Got this forwarded to you? Join the FT Network here (and let us know who shared it with you so we can say thanks)