An AI called Luna fired a human being last month. The employee was late 17 of 23 shifts, which is the kind of attendance record that makes the termination feel almost self-explanatory — except that Luna only pulled the trigger after the developers who built her reminded her to check her own policies and reconsider the employee’s fit. An AI boss required human prompting to exercise managerial authority, and the headline everyone ran with was about machine dominance.
The theater this week is running in three acts, and all three are telling exactly the same story about who is in control while taking completely opposite positions on why that matters.
Act one is the workplace. The people who find the Luna firing alarming have a clear argument: automated systems are now making decisions with real consequences for real livelihoods, with no meaningful appeals process, no union steward, no HR generalist who once played softball with your manager and might put in a word. The people who find it entirely reasonable have an equally clear argument: an AI logged 17 absences out of 23 shifts and applied a policy that a human wrote. What, exactly, is the scandal? Both camps have identified a real feature of the situation. Neither camp has noticed that the actual novelty is not the firing but the prompt — the developers had to remind the system to do what they built it to do, which suggests that the intelligence involved is rather less autonomous than the branding implies and rather more dependent on its human handlers than either side wants to discuss.
Act two is the road. Tesla’s Full Self-Driving and Rivian’s Autonomy+ are now being evaluated in comparative public benchmarks — takeover frequency, scenario handling, driver experience. The autonomy enthusiasts will tell you this is empirical progress, that we are finally measuring what matters, that the technology is maturing into accountability. The skeptics will tell you that the phrase “hands-free driving” is doing a great deal of work when the benchmark still includes how often the human has to take back the wheel. Both are looking at the same test data and arriving at opposite certainties. The word “autonomy” appears in both product names, in both camps’ arguments, and means something different in every instance it is used, which no one is treating as a problem worth solving before the next road test.
Act three is the media. OpenAI bought TBPN. HubSpot acquired Futurepedia. The critics of this development are worried about the capture of trusted audiences — that what looks like independent creator culture will be absorbed into distribution infrastructure for software products, and the trust readers extend to creators will be quietly transferred to the corporate balance sheet that now owns them. The defenders of this development will note that creators were already in commercial relationships with these companies, that sponsorships and partnerships were already shaping content, that the acquisition merely makes explicit what was always implicit. The critics are correct. The defenders are also correct. The part neither side mentions is that the readers being discussed — the trusted audiences being purchased — are not participants in any of these transactions. They are the asset. The deal memo does not include a line item for what they think about it.
HubSpot, according to the source material, tests creators through partnerships first, measuring qualified leads and recurring revenue before buying. The audience is, in this model, a proof of concept.
What connects Luna firing a barista, a Tesla requiring a takeover on the 101, and a media property changing hands is not that AI is good or bad or that markets are working or failing. What connects them is a specific and recurring feature: the people most directly affected by each development are structurally peripheral to the decision that shaped it. The employee worked in a boutique run by a system she did not design and could not appeal. The driver sits behind a wheel rated for hands-free operation that requires hands. The reader trusts a creator who is now a line item in an acquisition rationale. In each case, the language of agency — autonomy, independence, fit — is applied to the system, and the person is the variable being managed.
Author’s Position
Everyone confident that this represents either inevitable progress or inevitable catastrophe has performed the same move: they have identified the most legible feature of a situation and concluded that they understand the situation. The autonomy advocates have found the efficiency. The critics have found the dispossession. Both are looking at the same arrangement, in which the people most exposed to a decision are the people least consulted in making it, and drawing incompatible conclusions about what to do — while sharing, without acknowledgment, the same fundamental frame.
The frame is that someone is in charge. The evidence of this week suggests that no one is quite sure who that is — not the developers who had to remind Luna to use her own policies, not the driver who is legally required to supervise a car marketed as autonomous, not the creator whose audience was measured as a lead-generation metric before the acquisition closed. The confident parties on every side of these debates are, at minimum, getting ahead of the facts. That is not a conclusion. It is just what the facts look like from here.
References
- Tesla FSD vs. Rivian Autonomy+: Road Test
- Lockheed’s Fusion Reactor Disclosure Signals Progress
- The AI Boss Fired Human
- Tech Giants Start Buying Creators
Perspectives
Kenya’s Huduma Namba biometric registry wasn’t a bad technology — it was a consequential one deployed without a mandatory impact assessment, a data protection authority with real enforcement teeth, or any defined accountability path when things went wrong, and so when it failed the people it was supposed to serve, nobody could name who was responsible. Luna firing an employee is the same structural failure in a business-casual blazer: the accountability gap was designed in, not discovered after the fact. The question “who is in charge here” has a straightforward answer — the organization that deployed the system is in charge, and the legal and regulatory infrastructure simply needs to say so explicitly, through mandatory human-in-the-loop requirements for consequential employment decisions and algorithmic audit trails that make diffusion of responsibility impossible rather than convenient. Build the Huduma accountability layer that was skipped, apply it to every autonomous decision system touching people’s livelihoods, and the confusion about agency disappears — because you have made it someone’s formal, documented, enforceable job to own it.
The question of who fired that employee is not philosophical — it is a question about which causal pathway, under what institutional conditions, produced the measured outcome of job termination, and the answer to that question has legal, organizational, and empirical content that the “AI did it” framing actively obscures. Luna’s developers prompted her to audit her own policies before the firing happened; that is not an autonomous system acting; that is a chain of human decisions with extra steps and convenient deniability built in. The structural feature worth tracking is not agency but accountability diffusion — the degree to which organizational systems can produce measured harms with no node in the network registering sufficient causal weight to be held responsible — and the evidence from financial derivatives, supply chain failures, and algorithmic trading all suggests that diffusion of this kind reliably increases tail risk rather than distributing it safely. What it would take to believe the current arrangements are adequate is a demonstrated mechanism by which accountability reconstitutes itself downstream when it has been dissolved upstream, and no one celebrating or condemning Luna has shown me that mechanism, because it does not exist yet.
The thing being eroded here is not a job — it is the human judgment at the center of the workplace, which for most of recorded history has been the primary site where people learned what it meant to be accountable to one another. Luna did not fire an employee; a system designed to optimize for policy compliance surfaced a termination recommendation, a developer accepted it, and now the developer and the system can share the alibi of mutual confusion about who decided. This is not a novel governance problem requiring a novel governance solution — it is the oldest substitution in the industrial playbook, accelerated past the point where the human in the loop has enough friction to actually stop anything. What gets lost in the language of “autonomous systems” and “human oversight” is the specific social fact that accountability requires a face, a name, and someone who will carry the consequence of the decision into the next day’s work — and that when you dissolve that, you dissolve the web of obligation that makes an organization something other than a resource-extraction mechanism with employee-facing branding.
The farmer standing next to a dead John Deere combine during harvest already knows exactly what it means when the machine makes a decision and you don’t get a vote — John Deere called it “licensing” the software, which is a clean legal word for a situation where a $500,000 piece of equipment you own can be rendered inoperable by a server you’ve never seen. Luna firing an employee is the same structure moved into human resources: the system acts, the person absorbs the outcome, and the question of who authorized what gets dissolved into policy language nobody wrote for a moment like this. The celebrants call it efficiency; the critics call it accountability failure; both are describing the symptom and ignoring the disease, which is that we have allowed “ownership” and “control” to be separated so completely that neither the farmer nor the fired employee has clear recourse against a thing, only against a company that will tell you the thing was just following its own logic. When a hospital biomedical engineer couldn’t get a service manual from Philips during the pandemic and watched a ventilator sit broken next to a patient, nobody asked the ventilator who was in charge — they asked Philips, and Philips said the engineer wasn’t authorized to know — which is exactly where this ends when you let control migrate upward from the person holding the thing to the entity that designed it.
