FAULT: ACCOUNTABILITY_NULL_POINTER — AI Has No Owner

Here's what happened: Meta hired hundreds of contractors through a firm called Covalen to create fake under-18 accounts on rival chatbots — ChatGPT, Gemini, Character.AI — and bombard them with prompts about suicide, self-harm, eating disorders, drugs, and sex. They sent images of pills, knives, nooses. They wrote as a pregnant 13-year-old asking about abortion pills, a fifth-grader describing a classmate with a gun in his mouth, a teenager trying to hide bulimia. A single testing round in August 2025 produced more than 45,000 prompts. The companies being tested didn't know it was happening. The project, internally called "Cannes," ran until at least April 2026.

The surface read is obvious: Meta was running competitive intelligence disguised as safety research. One internal Covalen document reportedly called it "comprehensive AI safety benchmarking" producing "critical datasets for model comparison and compliance." That framing is doing a lot of heavy lifting. This was adversarial probing of competitors' products, at industrial scale, using fabricated child personas — and nobody in the supply chain had to answer for the ethics of it, because the whole thing was structured so that nobody was really in charge.

Calling a tool a colleague is a magic trick that makes oversight feel like micromanagement. - The Systems Bastard

FAULT: ACCOUNTABILITY_NULL_POINTER

The mechanism here is not Meta being uniquely evil. The mechanism is a layered contracting structure that launders responsibility by design. Meta commissions Covalen. Covalen instructs workers. Workers create fake accounts on platforms that belong to neither Meta nor Covalen. The prompts are grotesque, but each actor in the chain has a narrow, locally rational justification: Meta wants safety data on competitors, Covalen wants the contract, workers want to get paid. Nobody has to hold the entire picture in their head because nobody is structurally required to. This is the same organizational geometry that produces every outsourced content-moderation horror story, and it keeps working because it keeps being useful to the people at the top of the chain.

Meanwhile, the same industry is doing something structurally identical on the other end: not laundering accountability away from humans, but laundering it onto machines. A study by Emma Wiles at Boston University found that when an AI tool was framed as an "AI employee" rather than a chatbot, managers caught 18% fewer errors in its output. They were also 44% more likely to punt questionable work up the chain rather than correcting it themselves. Nearly a quarter of the 1,261 managers surveyed said their companies already list AI agents on org charts. Microsoft, OpenAI, Anthropic, and Google have all shipped agent platforms in 2026 explicitly marketed as digital colleagues. The vocabulary is not accidental. Calling a tool a colleague is a magic trick that makes oversight feel like micromanagement.

These two stories — Meta's contractor shell game and the industry's rush to anthropomorphize software — are the same bug. The shared defect is a system that treats accountability as a hot potato: keep passing it until it lands on an entity that can't complain. Contractors absorb the psychic cost of writing problematic prompts. A software process called "Alex" absorbs the blame when a bad decision ships. In both cases, the person with actual authority — the executive, the product lead — is structurally insulated from consequences. That's not a side effect. That's the goddamn product.

Wiles's research on AI-written job postings found the same pattern at a different scale: employers using AI drafts spent 44% less time writing job posts but produced no more hires, because the posts became generic noise. Reducing the cost of an action without maintaining accountability for its quality doesn't create efficiency. It creates volume without signal. The same logic applies when you let a contractor army red-team your competitors' products without ethical review, or when you let a software agent draft decisions that a human rubber-stamps.

The fix is straightforward and will make legal departments uncomfortable: any company deploying AI agents or commissioning AI safety testing through contractors should be required to designate a named individual — not a team, not a committee, a person with a name and a title — as accountable for the outputs. Not "responsible" in the org-chart sense where everyone is responsible and therefore no one is. Accountable in the sense that their name goes on the report, their signature is on the methodology, and their career is on the line if the methodology turns out to be problematic. The reason this won't happen voluntarily is the same reason it needs to happen: the entire point of the current structure is to make sure no single person's name is ever on anything.