by JL Ecochard
You know the VHS clock. Blinking 12:00. Always blinking 12:00.
Not because the feature did not work. Because it was built by engineers for engineers. Nobody asked whether the people who would actually live with the machine could program it. Eventually nobody bothered, because the tape still played.
That is not a technology failure. It is a design philosophy failure. Built by people who would never have to live with its limitations. Deployed on people who were never consulted. Abandoned when it proved too complicated.
AI in the humanitarian sector follows the same pattern. A model fine-tuned for Amharic or Somali by a team in San Francisco. Tested on datasets that do not reflect the community it will serve. Deployed by an organization that won a grant to pilot it. Celebrated when the pass rate reaches 77%.
Nobody asks who bears the 23%.
I said something at the Good Tech Summit that stopped the room.
“When my work with AI fails horribly, I go to jail. The AI doesn’t.”
I facilitated a session with 50 humanitarian practitioners — five tables, all field experience, no observers. We looked at Signpost, the IRC-built information service for displaced people operating across 17 countries. Six-month pilot. Staff reported a 70% efficiency gain. They published it.
They also published the fail rate. 23%. One in four responses not good enough to give to someone in crisis.
That transparency is what makes Signpost a good case study. Not the 77%. The 23%. The honest number everyone learns from.
The room’s reaction was immediate. People wanted the 77%. They could not accept the 23% — not for what it means statistically, but for what it means in practice. Someone displaced, frightened, trying to find food or shelter or legal help, getting a wrong answer from a system that sounded confident. That is not an error rate. That is harm.
Then the room named something harder. The acceptable error rate should not be set by engineers, funders, or media. It must be set by the community bearing the consequence. And in most humanitarian AI deployments, that community has no seat at the table where the threshold is decided.
When frontline staff resist trusting AI predictions, that resistance is sometimes framed as a deployment barrier. A trust gap to close through training.
The room disagreed. Frontline staff questioning opaque predictions may be the healthiest institutional immune response we have. They and the people they serve bear the consequences of errors. The question “should I trust this?” is not a gap to close. It is a signal to respect.
A human in the loop who is afraid to push back is not accountability. It is decoration.
The VHS clock had a human in the loop too. Standing in front of the blinking 12:00, manual in hand, eventually giving up. Present. Technically in control. Accountable for nothing. Shadow IT became Shadow AI.
The room pushed further. When an AI system causes harm — a wrong displacement prediction, dangerous information to someone in crisis, a population mapping tool that misidentifies a community — who owns it?
The NGO that deployed it? The tech company that built the model? The funder who structured the grant? The consultant whose liability contract already released them? The community-based organization at the last mile that had no say in the design?
This is not theoretical. In Moffatt v. Air Canada (2024), a British Columbia tribunal ruled that companies are liable for what their chatbots say, regardless of which component generated it. A chatbot is not a separate legal entity. The organization owns all information it provides, regardless of source.
That ruling applies to a humanitarian chatbot giving wrong information to a displaced person in exactly the same way it applied to a customer service chatbot selling flights.
Now the practical question. When did you last read your D&O policy? Stop reading this. Go find it. Look at the renewal date. That date tells you how much time you have.
The Insurance Services Office, whose standard forms underpin approximately 82% of US property and casualty policies, introduced generative AI exclusions effective January 2026. The era of silent AI coverage — where ambiguity gave organizations inadvertent protection — is over. The humanitarian sector is not Oil and Gas. It will get the commercial exclusions by default and the legal precedents by proximity. Your board may already be uninsured for the AI your organization is already using. AI-related board liability claims doubled in 2024.
Average settlement values rose 27% to approximately $56 million. Plaintiffs do not need to prove the AI system failed. They need to prove the board failed to govern it.
Before Sarbanes-Oxley, financial reporting was an accounting department problem. After, the CEO, CFO, and Treasurer signed their name and faced personal criminal liability if the statements were false. The loop became a jail cell. Shared. Named.
BJ Fogg’s behavior model explains why it worked. Motivation was always present — nobody wanted to go to prison. What SOX added was the trigger: a signature with personal criminal consequence. Behavior changed overnight.
Something else happened. CEOs and CFOs had to develop financial literacy. Not become accountants. Become competent enough to understand what they were signing.
AI will require the same. The mistake I see repeatedly: “it is a technology problem, give it to the CTO.” The same mistake made before SOX — “it is a finance problem, give it to the CFO.” We know how that ended.
AI is the new air. Every organization is already breathing it whether they know it or not. Fish that do not grow lungs do not become slower swimmers. They become extinct. We are out of the water world. Time to crawl and learn to breathe.
Name the last AI-assisted decision your organization made that affected the people you serve. Now name the person who owned that decision and its consequences. Not the team. Not the working group. Not the policy document. The individual person.
If you cannot name them, you do not have human accountability. You have human presence.
Francesco Schettino was on the bridge of the Costa Concordia. He had every instrument available. Thirty-two people died. He went to jail.
The tech did not.
JL Ecochard is Chief Innovation Officer at NetHope and co-founder of its Center for the Digital Nonprofit.
This is Part 2 of a four-part series developed from his session “When Every Hour Counts” started at the Good Tech Summit, April 2026. Part 1: Fund it like roads, not meals. Part 3: Local. Part 4: The Brain Drug Argument.
