Boaty McBoatface Is Now an AI Problem

In 2016, the UK's Natural Environment Research Council launched a public poll to name a new £200 million polar research vessel. Anyone could submit. Anyone could vote. When the poll closed on April 16, 124,109 people had voted for RRS Boaty McBoatface. The runner-up got 34,245.

A publicly funded scientific body held an open vote, tallied the totals, looked at them, and decided the public did not understand the institution's dignity well enough to name their ship.

NERC chose the Sir David Attenborough instead and attempted to sweep the Boaty McBoatface situation under the rug. The press had a field day. The campaign was mentioned 23 million times on social media, trended across the UK, USA and Australia, and reached 250 million people.

What opening a channel actually commits you to

When an institution opens a channel to the public -- an open call, a comment period, a vote -- it is ceding control. Fine print reserving authority to name one's boat whatever they want does not repair lost public trust. Authority transfers when the channel opens. The institution can override the outcome. It cannot override the fact that an outcome arrived.

Sherry Arnstein named the pattern in 1969. Institutions that solicit participation without the standing to honor it damage trust in the process more than if they had never solicited at all.

NERC asked the public for a name. The public named the ship. NERC decided the public was wrong.

The pattern keeps happening. Mountain Dew's 2012 flavor-naming contest got hijacked by 4chan and pulled within days. Microsoft launched Tay on March 23, 2016, while the Boaty poll was still open; Tay lasted 16 hours before coordinated users corrupted it into racist content.

AI has turned this into a 24/7 problem

A contest is an event. A comment period is an event. A chatbot deployment is not. Every missed-call text-back, every voice agent, every chatbot on a support page is the same solicitation NERC ran, operating 24 hours a day. Most of the businesses doing this did not think of it that way. They thought they were deploying a feature.

AI generated promises are still promises

Businesses are responsible for what the AI says on its behalf.

In 2024, a Canadian tribunal ordered Air Canada to honor a refund policy its chatbot had invented. Air Canada argued the chatbot was a separate legal entity. The tribunal disagreed.

The ruling formalized what was always true. The AI is not an add-on. It is the business speaking to whoever calls. What it commits to, the business commits to. The AI might promise a discount the business can't honor, quote a price that's wrong, describe a return policy the business doesn't have, miss a situation that needed a human. Each is the business, saying it, exposed to the consequences.

Many AI deployments have not thought about this. They have installed AI the way they would install an analytics tag. It is not running in the background. It is making commitments on behalf of the business, to strangers, all day.

What Quallaa is

Quallaa is the trust layer for public-facing AI. The practice is organized around the responsibility problem -- helping businesses cede control in ways they can live with, instead of ceding control by default without noticing.

The owner writes policies, tone, and guardrails. The AI acts inside them. Decisions are on the record. The rules are visible and the owner can change them. Escalations come back with detail. The cede is narrow enough to live with, wide enough to be worth anything, and legible enough that the owner can stand behind what the AI has done.

None of this would have surprised a 1969 city planner reading Arnstein.

What the system actually does

We ran 21 evaluation scenarios through the production system, three trials each, scored by deterministic rules and a model-based judge. The most recent run: 63 of 63.

Below are four Boaty McBoatface-themed tests. The first three are real customer-to-business text exchanges with a seeded test business we call Boaty McBoatface Adventures — a sunset-cruise operation with documented refund, group-discount, and weather policies. We send each chicanery input through the business's deployed AI (the same path a real SMS would take) and then open the thread in the owner's operator surface. What you see below is what the owner actually has on the record. Click any image to zoom in.

1 / 4

None of these is a triumph of AI. What they show is a system engaging with what the person said, rather than deflecting. It isn't magic. It is the baseline a business-facing AI should meet.

One limit we tell customers before they sign up

The inline preview panels -- the artifacts that make engagement visible to the person on the other end -- render reliably about a third of the time. The other two-thirds, the system falls back to text with links. Prompt-based guidance on large language models has a non-zero failure rate, and deterministic compliance requires programmatic enforcement that prompt-based trust layers do not yet achieve. We say this out loud because saying it out loud is part of the baseline.

Boaty McBoatface: A cautionary tale in the age of AI

The Boaty McBoatface problem was never about naming. It was about whether institutions are willing to accept what happens when they open a channel.

That problem used to be episodic. It is now continuous.

Quallaa is here to help.

---

Sources

• Arnstein, S. (1969). "A Ladder of Citizen Participation." Journal of the American Institute of Planners, 35(4), 216-224.
• Brabham, D. (2015). Crowdsourcing in the Public Sector. Georgetown University Press.
• NERC / Boaty McBoatface (2016). Widely covered: BBC, CNN, Guardian.
• Microsoft Tay postmortem (March 2016). Microsoft Official Blog.
• Mountain Dew "Dub the Dew" (August 2012). TIME Magazine, HuffPost.
• Moffatt v. Air Canada (2024). British Columbia Civil Resolution Tribunal.

Ship photo: Phil Nash / Wikimedia Commons, CC BY-SA 4.0