What we fear
when we fear AI.

A manager watches a tool finish a task in seconds. A worker watching the same demonstration wonders whether the task will still be theirs. What happens to the job depends on what the employer does next.

That gap matters. A demonstration can establish that a system produces a useful result under particular conditions. An employment claim needs evidence about adoption, costs, demand and the decisions employers make. Moving from one to the other without those steps makes an uncertain future sound settled.

Fear can be well founded before the outcome is certain. A worker does not need an economy-wide unemployment forecast to worry about a client, a role or a route into a profession. But understanding that concern means being precise about what is changing.

Work: from a task to a livelihood

Consider a hypothetical employer using a large language model (LLM) to prepare a first draft. The time saved may allow more work, less overtime or fewer paid hours. Someone still has to define the assignment, check the result and handle mistakes. Which effect dominates depends on the work and the choices around it. The example identifies questions; it supplies no employment estimate.

A 2024 study illustrates why capability and adoption differ. Svanberg and coauthors model the cost of automating computer-vision tasks. Under the study’s cost assumptions, tasks that firms could economically automate accounted for 23% of compensation for vision tasks. The estimate covers computer vision and uses task compensation as its denominator. Measuring LLM-related job loss would require employment data. [1]

For an employment claim, ask what was measured: tasks a model might perform, firms buying a tool, vacancies, payroll employment or earnings. Then ask who is covered and over what period. These measures can tell different parts of the story. They cannot be substituted for one another.

Trust: what would make an answer worth relying on?

Imagine a summary that cites a real report but attributes a finding to the wrong population. The link works. The claim is still unsupported. Verification requires reading the relevant passage and checking the date, denominator and conditions.

Trust therefore depends on the use. A mistaken suggestion for a story title and a mistaken statement in a benefits decision have different costs. A useful assessment asks how often a system fails on the actual task, whether failures can be detected, and who can correct them.

Reliability is one part of that assessment. Safety concerns possible harm; trust also involves the people and organizations deploying the system. A tool can be useful while the terms of its use deserve challenge. Our concern is whether reliance is justified, not whether a product inspires confidence.

Who gets to decide?

A person choosing an assistant has options that an employee required to use one may lack. Ask who chooses the system, defines acceptable performance, reviews an error and bears its cost. Can an affected person question a result or refuse a use?

These are questions about human control. They cut across work and trust: a more accurate system does not, by itself, settle who should have authority to deploy it. An objection to the terms of adoption should not automatically be read as ignorance of the technology.

Infrastructure: where the costs land

The physical buildout requires its own accounting. Total data-center electricity is not the same quantity as electricity attributable to AI. A proposed grid connection is not annual consumption. A national share does not describe the burden on one community.

Our data-center research keeps those boundaries visible. It separates forecasts from historical estimates and asks which company, place and period a figure describes. That is the starting point for evaluating claims about electricity, water and local impacts.

About Fear AI

Fear AI publishes research and writing about work, trust and infrastructure. We examine specific claims, follow them to their sources and make the difference between a finding and an interpretation visible. The name leaves room for concern, curiosity and criticism.

Here, AI names the broad subject. When a claim concerns an LLM, computer vision or another particular technology, we name it. Evidence about one should not silently become a verdict on them all.