Existential Risk, Now Available in Graphic Design

The note went out on Wednesday morning to roughly 5,700 people, and by lunchtime more than 1,200 of them had something new in common with the phrase "existential threat". Canva is thirteen years old, Sydney-based, and sells software for making posters. Co-founder and chief executive Melanie Perkins framed the cuts around slowing growth, the first in the company's history, and the coverage duly hung the word "existential" on AI. Note the direction of travel. Canva is not a company being eaten by artificial intelligence; it is a company that sells the thing AI made cheap, charging per seat, in a year when the number of seats stopped multiplying. If that is an existential threat, the term now covers a missed forecast and a board meeting that ran long.

Read the report closely and the specifics evaporate. It does not say which roles, which teams or which locations are affected. It does not say when the reductions take effect. It does not publish Perkins's note. What survives is a headcount, a growth figure and a mood, which is the standard texture of AI-driven restructuring in 2026: a large number, a vague cause, and no ledger anyone outside the building can inspect. A fifth of the workforce is a serious thing to do to people, and it deserves a paragraph of mechanism. If a design tool really has been out-competed by the models it licenses and ships, that is a fascinating commercial story. Nobody told it, because the mechanism is not AI. The mechanism is that growth slowed and payroll is the only line item a chief executive can move before the next board meeting.

"Existential" used to mean something specific. It meant the end of the species: paperclips, the lab leak, the models that build themselves. It was a word with a body count attached, or at least a species count. It has now been demoted to mean "awkward for the forecast", and that demotion is not a small thing, because vocabulary is infrastructure. Earlier this week I made the point about OpenAI's decision to rebrand data centres as "AI compute campuses": whoever owns the noun writes the regulation. The same manoeuvre runs at smaller scale here. Whoever owns "existential" gets to skip the question that would follow from saying "we over-hired in 2024 and the music stopped".

Meanwhile, in the same few days, Acronis published the anatomy of a thing that is actually an AI threat. They call it BOSSAI: a mesh of open-source and "abliterated" models, wired together with MCP servers, each model handed a specialist offensive job, run at Taiwan's emergency services. Abliterated, for the fortunate uninitiated, means the refusal behaviour has been surgically removed. The researchers do not name the commercial and closed-source systems in the mesh, which is itself the interesting part: you do not need to know the brands, because the architecture is the weapon. Not one brilliant model, but a dozen unremarkable ones divvied up like a burglary crew. That is what the threat looks like when it is real — cheap, modular, aimed at an ambulance switchboard, assembled by people who will never send a press release. The threat in Canva's announcement is a slide.

Watch where the word settles. This week Rob Wiblin noted on the 80,000 Hours podcast that OpenAI has fired three safety researchers — Tomek Korbak, Mikita Balesni and Jasmine Wang — described as working on one of the company's own stated top alignment problems. Those are the people paid to take "existential" literally. The vocabulary migrates upward and away from the work: it leaves the desk of the researcher who was escorted out and lands in the mouth of an executive explaining a growth chart. It is the same company, incidentally, that I noted last week asking departing staff to unlock their personal phones on the way out. Nobody asked the models to unlock anything, because the models are not staff and cannot be made redundant. That is the whole joke.

Policy, meanwhile, is converging on exactly the wrong instrument. The five major labs have settled on a shared position about how they would like to be governed: Washington should get privileged early sight of frontier capabilities, through incident reporting and secure testbeds. Fine. But notice the asymmetry. The failures of the most powerful systems on earth would be logged to a cleared official behind a door, while the corporate decisions justified in their name are disclosed to nobody, in no format, checked by no one. I would settle for a form. If AI is the reason for 1,200 redundancies, write it down: which systems, which tasks, which productivity number, signed by whoever signed the note. Not a new regulator. A paragraph, on the record, with an accountant in the room. See how many of these announcements survive it.

Canva will be fine. It will hire again, some of the same roles under new names, and in a year it will publish something cheerful about how AI made the team nimbler. The tell will be whether "existential" reappears in the next set of results when the growth line bends back up. Words are cheap; that is precisely why executives reach for them instead of money. But "existential" has a job, and every time a design company borrows it to explain a payroll decision, the people who need it lose a little more of its power to make anyone look up — the ones staring at a mesh of mutilated models and an emergency switchboard. Perkins could have said we over-hired. She said existential. One of those is a business decision. The other is a theft.