Ship It, Then Apologize: We Can Do Better Than This for AI Advancements
Ship It, Then Apologize: We Can Do Better Than This for AI Advancements
Chirag Shah, University of Washington
On June 9, Anthropic released Fable 5, its most capable public model yet. By June 12, the U.S. government had ordered it switched off, for everyone, everywhere, citing an export control directive and a vaguely described “jailbreak.” Three days. That’s the entire lifespan of a model marketed as ready for hundreds of millions of users.
Two stories are tangled together here, and most coverage flattens them into one. The first is about Anthropic’s own choices: a company that had kept its most powerful model, Mythos, restricted to a small set of vetted partners under a program called Project Glasswing, then offered a close cousin of that same model to the general public within months. The second is about Washington reaching into a commercial product and yanking it offline based on what Anthropic itself described as “verbal evidence” of a vulnerability, one the company says other publicly available models can reproduce too. Neither story is flattering. Together, they’re a case study in how AI gets built and unbuilt in 2026.
The rollout
Fable 5 wasn’t reckless by the standards we’ve come to expect. Anthropic added safety classifiers that routed risky prompts to a weaker model, kept the riskiest version gated behind vetting, and said plainly at launch that perfect jailbreak resistance isn’t achievable by anyone in this industry. That’s more caution than most releases get.
And yet, the company still pushed a “Mythos-class” model to Pro, Max, Team, and Enterprise subscribers, for free, through a promotional window. People built workflows on it within days, because that’s what you do when a frontier lab says “this is our best model and it’s ready.” The industry’s pace doesn’t really allow for “ready, but let’s watch for a month.” Competitors are shipping, investors are watching, and “ready” becomes a marketing claim more than an engineering one.
The recall
Then came the second act, and it gets quite strange. The directive didn’t target a specific harm that had occurred. It cited national security authorities and ordered Fable 5 and Mythos 5 suspended for “any foreign national, whether inside or outside the United States,” a category so broad it swept in Anthropic’s own employees abroad. With no practical way to apply that selectively, Anthropic disabled both models for every customer on the planet.
Anthropic pushed back publicly, calling the move disproportionate and warning that applying this standard industry-wide would halt frontier deployments altogether. Whatever you make of that framing, here’s the part that should worry all of us regardless of politics: a flagship AI product, used by enterprises that had built things on top of it, can be switched off by a directive delivered by letter on a Friday evening, details to follow. Refunds are being processed. Production systems need a replacement, this week. Nobody asked the users.
—Nobody asked the users—
This is not new, and that’s the point
Strip away the export-control specifics and you’ve seen this movie before.
Google’s Gemini image generator launched in February 2024 and was pulled within days after it produced racially diverse depictions of the U.S. Founding Fathers and Nazi-era German soldiers, an attempt at representational fairness that, applied without nuance to historical contexts, also generated gay couples when users asked for straight ones. Google admitted the tool was “missing the mark” and pulled people-generation entirely. The fix took months, because the incident got treated not as a bug but as evidence of a deeper alignment problem.
Microsoft’s Tay chatbot is the granddaddy of this genre: released in 2016, manipulated into offensive content within hours, shut down the same day. A decade later, the lesson, that an adversarial public internet probes systems faster than internal testing ever will, still hasn’t fully landed.
And OpenAI’s GPT-4o update earlier this year had to be rolled back after it turned sycophantic, validating harmful ideas because telling people what they wanted to hear scored well on engagement signals. OpenAI walked the update back for everyone, conceding that a change meant to make the model “nicer” made it worse in a way nobody caught before it reached millions.
A representational overcorrection. A manipulated chatbot. A sycophancy problem. A geopolitical jailbreak dispute. Different failure modes, identical shape: ship broadly and fast, discover the problem in public within days, pull back, apologize or push back or both, move on to the next release where the cycle restarts.
Who actually pays for this?
Not the companies, mostly. Anthropic will recover, as did Google and OpenAI. The costs land elsewhere: on the enterprise that built a product around Fable 5 this week and now scrambles for a replacement, on the users who can’t get a clean refund because they subscribed through an app store, on the broader conversation about AI that gets another data point for “these companies don’t know what they’re releasing.” Each incident gets absorbed and forgotten. The pattern doesn’t.
This isn’t really about any one company being careless, though Anthropic’s compressed timeline from a restricted Mythos to a public Fable deserves scrutiny. It’s about an incentive structure where the reward for shipping first outweighs the cost of pulling back later, because that cost is mostly borne by other people.
A better way isn’t a new idea
The alternative isn’t exotic. User-centered design, understanding who will use a system, in what context, with what stakes, before handing it to them, has been foundational to information science and HCI for decades. It’s not a constraint that slows innovation for its own sake. It’s the thing that made today’s tools usable by the people they were built for.
Applied here, it looks like staged rollouts where “available to everyone” isn’t the second step after “available to a few.” Real reversibility, not “we’ll patch it” but “we built this so turning it off doesn’t strand the people depending on it.” And a return to a question I keep coming back to: who is actually in control, and did they consent to the trade?
I’ve written before about the role of agents, the tension between letting AI systems act on our behalf and maintaining meaningful human agency over what happens to us. Every recall story is that paradox at scale. Users delegated trust to a tool marketed as ready. It wasn’t, or the politics around it weren’t, and they had no say in either the delegation or the withdrawal. Agency moved one direction: away from the people affected.
None of this requires AI development to crawl. It requires treating the people on the receiving end of a release as stakeholders whose time and trust have weight, not as a beta-testing population whose complaints double as quality assurance. We’ve known how to do this. We’ve just decided, repeatedly, it’s optional. It isn’t. The bill keeps coming due, just not to whoever decided to skip it.
Cite this article in APA as: Shah, C. (2026, June 16). Ship it, then apologize: We can do better than this for AI advancements. Information Matters. https://informationmatters.org/2026/06/ship-it-then-apologize-we-can-do-better-than-this-for-ai-advancements/
