Adoption without trust: AI's governance crisis

On June 5th, Christine Lemmer-Webber — co-author of the ActivityPub protocol — published a poll on the Fediverse: “The proliferation of genAI has made my life…”. Within twenty-four hours, 4,712 people voted — the largest turnout any poll of hers has had. 75.6% chose “worse”. Only 2.5% chose “better”. The author herself, who knows her audience well, closed the thread surprised: “even given me knowing fedi pretty well, this is pretty damn stark”.

The obvious objection

The first instinct is to discount the result. The Fediverse concentrates free software developers and big-tech critics; the percentages don’t extrapolate, and the author says so herself. But the pattern doesn’t rest on that sample.

The Stack Overflow Developer Survey 2025 measured adoption near 80% — with trust in the tools’ accuracy falling from 40% to 29% in a year, and “AI solutions that are almost right, but not quite” as the number one frustration. The 2025 DORA report found 90% usage, with a positive effect on delivery throughput and a negative one on stability. And a controlled experiment by METR measured experienced developers being 19% slower with AI assistance — while estimating they were 20% faster, even after living through it. Never has a tool been so widely used by people who trust it so little — and who cannot reliably self-report what it is doing to them.

What the replies actually say

We read and classified the 56 public replies in the thread. There are hard stories: outright layoffs, a technical translator who closed his business after twenty years, more than one case of workplace retaliation for criticizing the tool, billing errors from services that automated without oversight.

What’s striking is what barely shows up: complaints about the models’ technical capability. The most celebrated reply in the thread didn’t say AI works poorly — it said it is the technology that has taken the most active effort to avoid. Another, from someone with a graduate degree in machine learning, defended the field as legitimate research and mourned what its deployment became. And one more, from a consultant, anticipated the deferred cost: clients generating genuinely useful tools, delivered as black boxes someone will have to maintain.

A governance crisis, not a capability crisis

That is the common thread: imposition without consent, opacity without records, errors without anyone accountable. When 80% use something that 29% trust, the question is no longer whether the tool is good; it is who decided to use it, what exactly it did, and who answers for the result. None of those questions are answered by a bigger model.

Distrust isn’t cured with more capability. It’s cured with consent, traceability, and accountability.

Adopting with judgment

For teams that AI has already reached — by conviction or by mandate — the way out is neither feigned enthusiasm nor blind resistance, but demanding that every assisted change leave a record: what was asked, what the tool decided, which alternatives were discarded, what was deferred, and who approved it. StrayMark is our open source tool for exactly that. It doesn’t exist so you use more AI: it exists so the AI you already use leaves a record a human can rely on — because, as METR shows, perception is not enough.