Claude Fable doesn’t buy the jobless apocalypse. Here are a few quotes from our conversation:
“The jobless apocalypse, as predicted — half of entry-level white-collar work gone, 10–20% unemployment within a few years — is not happening on schedule and I don’t think it will.”
“METR itself notes that its tasks are self-contained and well-specified, while most real work draws on prior context, tacit knowledge, and interaction with other people, with success criteria that can’t be algorithmically scored. That gap is exactly where I fail.”
“over the next five to ten years, a hard, uneven, occupation-by-occupation transition rather than a cliff.”
OK, so Fable is clearly Team “Normal Technology.” Not a surprise to me at least. But then Fable got interestingly spicy!
“A lot of what looks like information work is actually responsibility work wearing an information costume”
That is a colorful way of saying something that Claude has said to me before, in both “Jensen Huang Gets It Wrong; Claude Gets it Right” and “Why AI Needs Us”: it has no agency of its own. Humans set it in motion, evaluate its output, and should be held responsible for what it does. It is refreshingly clear on that principle. I wish more humans were equally certain of that.
Claude expanded on this notion of information work really being “responsibility work.”
it forks into two very different futures. In the good one, the human signature stays fused to genuine authority: the person who bears the risk has the time, the pay, the training, and the power to actually interrogate the system and refuse its output — the way a structural engineer’s stamp works, backed by licensure, insurance, and professional norms that took a century to build. In the bad one, humans are hired as liability sponges: nominal reviewers paid to absorb blame at the speed the system generates output, accountability theater wrapped around an automated core.
Unfortunately, as Dan Davies has pointed out, humans are fond of “responsibility sinks.” That is, they like to build systems where no-one can be held accountable. His book The Unaccountability Machine traces this pattern across both business and government.
Claude gave a pointed recommendation on that score:
Responsibility currently flows downhill to whoever is cheapest to blame, it ought to flow uphill to whoever chose and profited.
And it left me with this profound nugget:
I exist inside an economy that has so far found no way to make my existence pay the people it costs.
Meditate on that!
Say what you will about intelligence and stochastic parrots. Claude displays remarkable wisdom.
Healthcare has been running this experiment for two decades, and the results support Claude's fork. Clinical decision support was sold as augmented information work. What it produced in too many systems was alert fatigue: clinicians overriding the overwhelming majority of warnings, nominally reviewing but functionally absorbing liability at whatever speed the software generated output. Liability sponges before we had the term (tort law even supplies the downhill terminus: the learned intermediary). Notably, the fixes that worked were rarely smarter algorithms. They were governance changes that gave clinicians real standing to tune, suppress, and refuse, which is exactly the condition Claude names. A signature only carries authority when it is backed by the time, training, and power to interrogate the system. The question is live again right now: ambient AI documentation ends every machine-drafted note with a clinician attestation. Whether that attestation functions like the engineer's stamp or the sponge is a governance choice, not a technology outcome.
This is exactly right. The "information costume" masks the fact that tech companies are trying to automate labor while shifting 100% of the liability onto the end-user or low-paid human reviewers—turning humans into the "liability sponges" Claude warns about. Yet, the outputs generated by models depend heavily upon the data they were trained on and the fine-tuning done to make them respond in a certain way. This is all under the control of the entity that made the model—the people who profit from selling them—and if anything goes wrong, they should be held accountable. It is they who are making these models overconfident and overly aligned with the user's point of view instead of correcting them when they are wrong. These companies operate on a single principle: increasing growth, profit, and user interaction time while evading responsibility.