This is the fourth installment of Where We Stand: Position, Proximity, and Power in AI Safety.
When I wrote Becoming a Cyborg Evaluator, I was excited about what happened when human judgment met machine capability. AI could hold more information than I could, find patterns across documents, translate technical language, and help me build tools that would once have required far more time and resources. It did not replace my judgment, but it extended my reach.
I still believe that. I use AI constantly, and Anthralytic would not exist without it. The Conditions Web is only one example of something I can build because these tools have expanded what is possible for a very small business.
What I did not ask clearly enough in that first piece was what happens to the work around the tool. Increased capability does not necessarily produce less work. Sometimes it simply expands the amount of work one person can imagine taking on.
That question becomes more important when the person using AI is not also the person deciding how the job is designed. A worker may choose how to write a prompt, but an employer determines the caseload, the performance target, the deadline, and what happens to any time the tool saves.
The technology changes the work. Power determines who benefits from the change.
There is more than one worker in the system
Because this series is about where people stand in relation to AI, it helps to distinguish three worker positions.
The first is the worker alongside AI, the person using it to write, analyze, code, summarize, translate, or complete administrative work. This is the position I know most directly, and it is where most conversations about productivity begin.
The second is the worker managed through AI. They may never enter a prompt, but software helps assign their schedule, track their activity, evaluate their performance, or recommend whether they should be promoted or disciplined. The OECD calls this algorithmic management, and its research shows that these systems are already common across the countries it studied. In this position, AI is not primarily a tool the worker controls. It is part of the institution looking back at them.
The third is the worker inside AI. Data labelers, content moderators, reviewers, and other workers prepare training data and correct the outputs that make an automated product appear automated. The International Labour Organization has called attention to these often overlooked data workers. Their labor is part of the system, even when the interface is designed to make the labor disappear.
These positions can overlap, but they do not carry the same power. A consultant choosing when to use a chatbot has more agency than an employee whose work is being scored by software, and both are positioned differently from a data labeler working through a platform they do not control.
This piece focuses mainly on the worker alongside AI because that is the position many of us are entering now. Even there, the central issue is not simply whether the tool works. It is who gets to decide what happens once it does.
Who owns the saved hour?
I asked a version of this question in Who Is Better Off When AI Speeds Up the Workflow?. A case manager who finishes documentation faster has produced a real efficiency gain, especially if documentation is the part of the job keeping them from being present with clients. The question is where the saved time goes.
It could become a longer conversation with a client, an actual lunch break, or a workday that ends on time. It could also become another intake, a larger caseload, a responsibility that used to belong to another position, or an hour spent checking whether the AI summary quietly changed something important.
AI does not choose among those outcomes. The workplace does.
The evidence so far does not tell a single story. A Zoom-commissioned survey found that 76 percent of AI users reported saving at least thirty minutes a day, and nearly three quarters said they would use the time for a dedicated lunch break. That is the hopeful version, and there is no reason to dismiss it. AI can give people time back.
An eight-month UC Berkeley Haas study found a different pattern. Workers at a technology company moved faster, attempted a wider range of tasks, kept multiple threads active, and allowed work to spread into times that had previously served as natural pauses. Nobody necessarily had to order them to do more. Greater capability widened their own sense of what they could take on, and that expansion gradually began to look like the new normal.
I recognize that pattern. AI allows me to attempt work that would previously have stayed on the someday list. That freedom is valuable, but it does not always make the list shorter. Often, it makes the list more ambitious.
The nonprofit sector adds another complication. People enter social impact work because they care, and that commitment can make additional capacity difficult to protect. If AI frees an hour, using it to serve one more person may feel like the obviously ethical choice. Do that often enough and the exceptional effort becomes the expected caseload. The worker remains just as tired, but the dashboard improves.
The work that does not disappear
The clean productivity story assumes that a task automated is a task removed. Sometimes that is true. More often, the work moves or changes shape.
Someone still has to prepare the information, decide what the system should do, check the output, correct errors, protect confidential material, explain the result, and accept responsibility when something goes wrong. A Workday survey of 3,200 employees found that nearly 40 percent of reported AI time savings was lost to correcting, rewriting, and verifying low-quality output. Workday sells workplace technology, so its research should be read with that interest in view, but the finding names a form of work that most productivity calculations leave out.
The reviewer is now expected to recognize an error they did not create, inside language polished enough to make the error difficult to see. In consequential work, that can require as much attention than producing the original result would have. The machine performs the first pass, while the human inherits the uncertainty and the liability.
This is why “a human remains in the loop” is not much of a safeguard by itself. The human needs enough time to review the work, enough knowledge to recognize a problem, and enough authority to reject the system’s answer. Without those conditions, human review becomes ceremonial. The person is present mainly so someone can be held responsible afterward.
There is also a slower question about skill. We do not need to preserve every repetitive task simply because doing it manually once taught us something. Calculators did not destroy mathematics, and spellcheck did not end writing. Still, workers need to retain the underlying skills required to judge whether the system is wrong. If someone can approve an AI-generated analysis but can no longer conduct or explain the analysis themselves, the organization has not eliminated risk. It has hidden dependence inside a faster workflow.
Start by finding out where the time went
The practical difficulty is that most organizations are not measuring any of this. They can see that more reports were produced or more cases were closed, but they cannot see whether the worker spent the saved time on higher-value work, AI cleanup, a larger workload, or recovery.
A worker can begin with a simple two-week time audit. This does not need to become another elaborate tracking system. At the end of each day, record which tasks involved AI, how much time the tool appeared to save, how much additional time went into prompting and verification, and what filled the remaining space. Also note whether the change improved the actual work, especially for the person the work was meant to serve.
The important part is not the precision of the minutes. It is the pattern. Did AI remove work, relocate it, or expand the job? Did it create time for judgment and relationships, or did it make the day denser? Did the worker finish earlier, or simply produce more before leaving at the same time?
If you are doing this for yourself, keep the record as a tool for understanding your own work and deciding what boundaries you need. If you manage people, be careful about asking employees to disclose time savings. To someone who reports to you, saying “AI saved me three hours” can sound like volunteering for three more hours of work. Any shared audit should be voluntary, anonymous where possible, and explicitly separated from performance evaluation.
Most importantly, do something with the answer. If time is being lost to correction, the organization may need a different tool, better training, or a workflow in which AI is used for fewer tasks. If the role has expanded permanently, the job description, staffing, and compensation should be reconsidered. If the tool has created genuine capacity, some of that gain should remain with the worker rather than being automatically converted into a higher target.
Renegotiate the work, not only the workflow
Organizations often introduce AI by redesigning the workflow while leaving the employment relationship untouched. The task changes, but responsibility, evaluation, compensation, and workload are expected to sort themselves out.
They rarely do.
Before adoption, a team should be able to write down what is changing, who will verify the output, what will happen to any time saved, which skills must remain human, what worker data the system will collect, how someone can challenge an AI-assisted assessment, and when the arrangement will be reviewed. This does not require a forty-page policy. It requires an honest agreement about the job.
Workers should also be involved before the tool is selected, not only trained after the decision has been made. The person doing the work often knows where automation would genuinely help, where it would add risk, and which apparently inefficient steps are carrying judgment or trust that the workflow diagram cannot see.
In Measuring the Wrong Thing Faster, I argued that the nonprofit sector often treats acceleration as impact because speed is what its dashboards can measure. A worker-centered assessment would look beyond volume. It would ask whether the job became more sustainable, whether workers retained meaningful judgment, whether the people served felt more understood, and whether the efficiency gain strengthened the mission rather than merely increasing throughput.
That does not make this an argument against AI. It is an argument for using AI well enough that the benefits reach the people doing the work. Saved time could become rest, flexibility, professional development, more thoughtful analysis, or the relational parts of a job that administrative work has crowded out. Those are productivity gains too, even when they do not appear as another completed unit.
I still want the cyborg version of work I imagined in that earlier piece, where machine capability extends human judgment rather than displacing it. I understand now that this outcome does not emerge automatically from having a good tool. It depends on how the work is governed and whether the worker has enough power to shape what happens next.
If AI saves an hour and the worker receives only a higher target, the technology did not give them time. It gave the institution more capacity.
When you are the worker, safety means having a voice in how the job changes, retaining the judgment necessary to challenge the system, and sharing in the value your AI-assisted labor creates.
Anthralytic helps mission-driven organizations use strategy, evaluation, data, and AI to understand their impact and make better decisions.
