Imagine an older woman waking in the night. A robot beside her bed extends a support rail. She steadies herself and stands. What matters to her is being able to get up when she chooses and continue living in a familiar home.

That is a different measure of progress from a model’s place on a leaderboard. The question is what AI enables people to do and choose in their own lives. That depends on resources, judgment, and arrangements that keep the help available.

Even a successful deployment leaves those questions open. A care home could use time saved by robots to give residents more human attention. It could also use that time to justify fewer staff. A machine’s capabilities do not decide which happens.

The gap becomes clearer when AI enters work people already know how to do, with varying degrees of experience.

In a field study, 19 train commissioning technicians completed 173 troubleshooting tasks. An AI assistant provided project-specific information and troubleshooting support, improving performance especially among less experienced participants.

For a newcomer, finding relevant knowledge can mean tackling a task that previously required waiting for an experienced colleague. That is a real gain, even with continued reliance on the assistant. It does not establish that technicians acquired expertise they could use independently; lasting learning remains a separate question.

Giving people advice presents a further difficulty: they must decide whether it fits.

A five-month experiment involving 640 Kenyan entrepreneurs compared access to a GPT-4 business assistant on WhatsApp with standard business guides. There was no statistically significant average improvement in business performance. Relative to the control group, AI access hurt businesses with lower initial performance; for those with higher initial performance, there was tentative evidence of gains.

The researchers’ analysis suggested that higher-performing entrepreneurs selected advice suited to their circumstances, while lower-performing entrepreneurs more often adopted generic strategies such as price cuts and increased advertising.

A shop owner has to judge those suggestions against cash flow, stock, nearby competitors, and customers’ habits. Advice can be easy to obtain while a workable course of action remains difficult to find. The owner bears the cost of getting it wrong.

These differences should change what we ask of organizations adopting AI. Buying access is only the beginning. People need the means to assess suggestions, act on them, and recover when they fail.

In care, this includes deciding how assistance changes relationships. A Japanese nursing-home panel study found robot adoption associated with higher employment and staff retention, alongside improvements in some care-quality and productivity measures. Workers shifted effort toward tasks requiring human attention and contact.

The findings are observational; they do not establish that robots caused the improvements. They nevertheless point toward questions a demonstration cannot answer. Who maintains the equipment and takes over when it fails? Can a resident refuse robotic help? How will saved time be used?

A robot may support someone’s body. The care arrangements shape how much independence that support gives them.

Useful help also needs a life beyond a single encounter. A technician’s troubleshooting method must be available to the next shift. A shop owner needs records of what worked. A care team needs training and dependable maintenance.

AlphaFold’s open database illustrates how AI outputs can become resources others can retrieve and examine. Smaller workplaces need their own ways to retain useful results, correct errors, and pass knowledge on. Without them, each successful interaction can remain an isolated event.

These arrangements require investment. An OECD report, drawing on evidence from G7 countries and Brazil, identifies barriers involving skills, data, infrastructure, and management. Managers can struggle to identify suitable problems and underestimate the organizational changes adoption requires.

A more capable model does not, by itself, provide those conditions. People and institutions remain responsible for deciding who controls the data, sets work targets, receives the gains, and bears the costs of mistakes — even when AI is involved in those decisions.

We should judge AI by what people can actually do with it in the circumstances they face. Can the technician use the available knowledge to diagnose a fault and decide what to do? Can the shop owner make a decision they can afford to carry through? Can the older woman get up safely, choose how she is helped, and still find human company when she needs it?

The woman at the bedside wants to keep living in her own home. For her, progress means having the support to keep that choice open.

AI Can Do More. What Does That Make Possible for Us? was originally published in The Point of View on Medium, where people are continuing the conversation by highlighting and responding to this story.

An AI assistant helped customer-support agents resolve about 15% more issues per hour. Yet surveyed customer satisfaction showed no economically significant overall change.

Both findings come from the same study in The Quarterly Journal of Economics, which followed 5,172 agents. Other aspects of the interaction improved: customers were more polite and less likely to ask for a manager.

The study found a productivity gain without a clear overall change in the satisfaction score.

“AI improves efficiency, so AI is beneficial” compresses all of this into a single conclusion.

Speed can be a benefit in itself. Better service, shorter working hours, higher earnings, and lower costs are further possible benefits. Each needs evidence of its own.

Where does the saved time go?

Britain’s tax authority, HM Revenue & Customs, trialled Copilot in 2024. Its evaluation, published in 2026, reported that participants estimated saving roughly an hour a week. They described using the time to complete more routine work, improve quality, or improve their wellbeing.

These are three different kinds of benefit. They also remain self-reported outcomes. The main survey had a 45% response rate, and the published summary did not separately establish how much each downstream outcome improved or whether the improvement lasted.

The report estimated around £50 million in annual capacity-generating net productivity benefits from deployment of up to 50,000 licenses. It also describes reducing self-reported time-saving estimates by about 20% to account for non-use and respondent selection.

Those adjustments make the estimate more cautious. They do not turn valued capacity into evidence of a cash saving. What that capacity becomes — additional service, less overtime, or a financial return — still needs to be traced.

There is no predetermined destination for saved time.

A randomized experiment involving 66 firms and 7,137 knowledge workers, by Eleanor W. Dillon and colleagues, found that users spent about two fewer hours a week on email in the second half of the six-month study. They also reduced work outside regular hours. The researchers did not detect changes in the quantity or composition of tasks resulting from individual access to AI.

That is evidence of a concrete improvement in working life. It does not tell us what every employer will do. It does show why neither “employees will get their time back” nor “employers will always demand more” should be treated as an automatic consequence of faster work.

Sales, capacity, and profit are separate outcomes

Completing a task faster, freeing up capacity, and earning a higher profit are different outcomes. Each needs to be established separately.

In PwC’s 2026 Global CEO Survey, 56% of respondents said they had seen no significant financial benefit from AI to date. These are executives’ reports, not audited accounts or causal estimates. They show that using AI and reporting a clear financial return are different things.

There is also experimental evidence of sales gains.

Lu Fang and colleagues studied seven AI deployments at a large cross-border retail platform. Sales effects varied by workflow, from no detectable improvement to a 16.3% increase. Measured post-purchase outcomes, including returns and customer ratings, did not deteriorate.

The largest gain came from a presale chatbot compared with an automated response saying live support was unavailable. In that setting, AI increased sales by providing assistance where live support had been unavailable. The finding does not establish that AI generally outperforms human service.

Across four applications with positive sales effects, the researchers estimated roughly $5 in annual incremental value per consumer. They annualized the results and added them together, assuming no interactions between workflows. This was not a directly observed net-profit figure after all AI costs.

The research concerns one platform. Two authors disclosed consulting roles with the company, and another was an employee during the project.

The study provides evidence of gains in particular settings, while the CEO survey describes a much wider range of reported experiences. Neither can substitute for the other. Local success is worth recognizing without treating it as a universal return on investment.

When faster becomes the new requirement

Organizations also decide how improved capability changes expectations.

If AI makes fifteen tasks a day feasible where ten were previously feasible, fifteen may eventually become the target. That depends on decisions about quotas, deadlines, staffing, and pay.

The customer-support study identifies this possibility. Managers described performance bonuses tied to handling time and resolution rates, and warned that targets could rise. The researchers lacked compensation data: they identified a possible ratchet effect, rather than documenting an AI-induced increase in targets.

To test whether that happens, we need records of how requirements, staffing, and compensation change after deployment. To call it a widespread pattern, we need evidence from a representative range of employers.

The mechanism is plausible. Whether it operates, and how often, remains a question for investigation.

Who decides what counts?

Claims about AI’s benefits help justify real decisions: renewing a contract, changing a service, raising a target, or reorganizing a team.

Before accepting such a claim, we should be able to identify:

• What improved: speed, quality, access, revenue, costs, or working conditions.
• Who benefited: employees, customers, the organization, or the public.
• How we know: a task measurement, a user report, an estimate, or an observed outcome.
• What it cost: deployment, training, checking, and error handling, including who did the additional work.

Then comes a further question: who chose those measures?

An evaluation that counts throughput and revenue but ignores unresolved customer problems or heavier workloads has already left some people’s experiences out. Its positive findings may be accurate, while its account of the benefits remains incomplete.

Benefits do not all have to become profit. Less repetitive work, shorter waits, and wider access to a service can matter in their own right. Nor must everyone benefit before we recognize an improvement for someone.

But a gain for one group cannot speak for everyone affected. What we choose to measure determines which improvements we can see — and whose costs we may overlook.

“AI is beneficial” can be true. To make that claim useful, we need to say where the benefit lies, who receives it, how we know, and who bears the costs.

Sources

1. Brynjolfsson, Erik, Danielle Li, and Lindsey Raymond. “Generative AI at Work.” The Quarterly Journal of Economics, 140(2), 2025, pp. 889–942.
2. HM Revenue & Customs. “Evaluating the Impact of Microsoft Copilot in HMRC.” Published July 9, 2026; reports on the September–December 2024 trial.
3. Dillon, Eleanor W., Sonia Jaffe, Nicole Immorlica, and Christopher T. Stanton. “Shifting Work Patterns with Generative AI.” American Economic Review: Insights, forthcoming.
4. PwC. 29th Global CEO Survey. 2026. Linked release summarizes the survey’s findings on AI-related financial benefits.
5. Fang, Lu, Zhe Yuan, Kaifu Zhang, Dante Donati, and Miklos Sarvary. “Generative AI and Sales Productivity: Field Experiments in Online Retail.” arXiv:2510.12049, version 6, June 2026.

AI Makes Work Faster. What Counts as a Benefit? was originally published in The Point of View on Medium, where people are continuing the conversation by highlighting and responding to this story.

New York City has announced a one-year moratorium on student use of artificial intelligence in public elementary and middle schools. According to Reuters, the policy will affect roughly 600,000 students through eighth grade.

High school students will still be allowed to use AI in some cases. Teachers may continue using it for lesson planning and tasks such as scheduling.

That creates an apparent contradiction: If AI is useful enough for teachers, why is it considered too risky for children?

The answer is not simply that adults can be trusted and children cannot. Nor is it that teachers always use technology wisely.

The relevant distinction is not teacher versus student. It is whether AI is assisting a capacity the user already possesses or replacing the activity through which that capacity is supposed to develop.

The same tool can do either.

The same tool can replace different things

Imagine a third grader working on a math problem at 9:30 in the evening.

She reads the question, feels stuck and photographs it with a tablet. An AI system immediately provides an answer and a step-by-step explanation. The response is clear, patient and available when neither her teacher nor her tired parent can help.

Nothing about this scene looks irresponsible. The child has not asked the machine to cheat. She has asked for help.

But one question remains: What happened before she asked?

Did she try to represent the problem? Did she choose an operation and discover that it failed? Did she identify the part she did not understand? Or did the answer arrive before any of those things could happen?

A teacher using AI to draft a lesson plan may also accept its output too quickly. Professional status does not guarantee sound judgment. But an experienced teacher is more likely to have an independent standard against which the result can be evaluated. The teacher has planned lessons before, knows the class and can recognize when an activity is unsuitable.

The child may still be learning how to begin.

This does not mean that teacher use is automatically beneficial or student use automatically harmful. A teacher can delegate too much. A student can use AI thoughtfully. The difference lies in what the tool is being asked to replace.

If it reduces the labor involved in exercising an established capacity, it may increase what a person can do. If it takes over the activity through which that capacity would have formed, the efficiency comes at a different price.

The output is not the only product

Schools have long treated completed work as evidence of learning. A finished essay suggests that a student can write. A correct solution suggests that the student understands the mathematics.

Generative AI has exposed how unreliable that assumption can be.

A child can now produce a polished paragraph without knowing how to organize an argument. A correct answer can appear without the student understanding why it is correct. The work may look better even as the connection between the learner and the work becomes weaker.

This matters because the answer is not the only product of education.

When children struggle with a problem, they are also learning how to remain with uncertainty, identify what they do not know, test an approach, detect an error and decide whether an explanation makes sense. Those abilities are difficult to see on a worksheet, but they are the foundations of independent judgment.

A child who has never wrestled with the problem may be able to read an AI-generated explanation. That does not necessarily mean the child can evaluate it.

The answer has arrived. The standard needed to judge that answer has not.

A 2026 interdisciplinary report on human agency in AI-supported education identifies four interconnected aspects of agency: human oversight, AI–human complementarity, AI competencies and relational emergence. It also identifies cognitive offloading as one of the central tensions in educational AI.

The concern, then, is not merely that AI sometimes produces incorrect answers. A wrong answer can be corrected. The deeper risk is that students become accustomed to receiving plausible answers before developing the habits required to question them.

Timing matters more than access

This does not lead to a simple rule that adults should use AI and children should not.

Age is only a rough proxy for something more important: whether the learner has enough experience to compare the tool’s response with an independently formed judgment.

Consider two students using AI on the same problem.

One has already attempted it, identified the point of confusion and asks AI to demonstrate a different method. The other gives the problem to AI before making any attempt and copies the explanation into an assignment.

Both have “used AI.” But only one has used it to extend an existing thought process. For the other, the tool has taken over the task before that process begins.

The distinction is not between use and nonuse. It is between assistance and substitution.

This is why the efficiency promised by AI requires more careful accounting in education. A child may complete homework in half the time. The saved time is immediately visible. The practice that did not occur is not.

The time saved is credited to the child. The cost may not become visible until years later.

The policy’s weaknesses are real

New York City’s moratorium is not a complete answer.

A school moratorium cannot by itself prevent students from using AI at home. Families with greater technological knowledge may teach children how to use it deliberately, while other families may respond with either unrestricted access or total prohibition. A restriction imposed at school could therefore shift responsibility to households and reproduce existing inequalities in a new form.

The policy also raises questions about students who rely on AI-enabled tools for accessibility or language support. A broad restriction may affect those students differently and will require carefully defined accommodations.

Enforcement presents another problem. AI features are increasingly embedded in search engines, writing software and ordinary educational platforms. Schools may find it difficult to determine where a conventional digital tool ends and a generative AI system begins.

Nor does restricting AI by grade answer the harder pedagogical question. Students of the same age do not possess the same abilities. The same student may have well-developed judgment in one subject and very little in another. A rule based on age is administratively clear, but educational development is not.

These limitations should prevent anyone from treating the moratorium as a final model for AI policy.

But they do not invalidate the question behind it.

Schools are being forced to distinguish between work that can be automated because the underlying capacity already exists and work that must still be performed because doing it is how the capacity is created.

That distinction applies far beyond homework. It affects reading, writing, calculation, research, creative work and even the ability to formulate a useful question.

A better question for students

The easiest response to AI is to ask whether a student used it.

That question is already becoming too crude to be useful.

A better question is:

Before you asked AI, how far did you get on your own?

The answer reveals whether the tool entered after the student had formed a question, attempted a method and encountered a difficulty — or whether it replaced that encounter entirely.

New York City’s moratorium may prove too broad, too difficult to enforce or too dependent on what happens outside school. But it identifies a problem that every education system will have to confront.

The educational value of AI does not depend only on what the technology can do. It also depends on what the learner has already learned to do without it.

New York City Is Restricting AI Use for 600,000 Students. Teachers Can Still Use It. was originally published in The Point of View on Medium, where people are continuing the conversation by highlighting and responding to this story.

1. The forecast and the person

If a weather forecast says it will rain tomorrow, the clouds will not read it. They will not change course because someone has classified them as a risk.

We often think of prediction as a one-way process. The world provides evidence; we use it to anticipate what comes next. If the prediction is right, the model has captured something about reality. If it is wrong, we revise the model.

But what happens when the forecast is about a person?

A student is likely to drop out. A borrower may default. An applicant is unlikely to succeed in a role. A patient may be readmitted. Someone’s application is flagged as suspicious.

These judgments describe things that have not happened. Yet they can immediately change how someone is treated.

The distinction needs care. Predictions about the natural world can also change outcomes through human action. An epidemic forecast can prompt precautions that reduce infections. A storm warning can lead people to evacuate, reducing the deaths a disaster might otherwise cause.

The virus and the storm have not read the forecast. People have, and their actions change what follows.

People can also absorb a judgment into their understanding of themselves. Clouds do not come to see themselves as high-risk clouds.

Society has always read people. Schools read grades. Banks read credit histories. Employers read résumés. Hospitals read medical records. Insurers read risk profiles. Welfare agencies read claims. Throughout our lives, we are observed, assessed and classified.

Increasingly, machines help with that reading.

How a machine assesses someone matters. Who sees its assessment, takes it up and acts on it is another question — perhaps an even more consequential one.

This essay concerns that passage from assessment to action.

2. When an assessment becomes a gate

In England in 2020, A-level exams were cancelled because of the pandemic. Universities still needed grades to determine admissions.

Teachers submitted estimated grades and rankings of their students. A statistical standardisation process then adjusted those estimates using information that included schools’ historical results.

About 39% of A-level subject entries received calculated grades below the grades submitted by teachers.[1] That figure referred to subject grades, rather than the proportion of students.

For an individual student, a school’s past results could therefore influence the grade assigned to them that year.

But the university would read the grade beside that student’s name. Someone with a conditional offer could suddenly fall below the required threshold without having sat the exam.

Students protested. The policy was reversed, and Ofqual announced that students would receive the higher of their teacher-assessed and calculated grades.[2]

The episode exposed a chain usually hidden inside institutions: a system produces an assessment of a person, another institution reads it, and a consequential door opens or closes.

The algorithm did not make the admissions decision. The decisive step came when institutions acted on its output.

Similar things happen more quietly elsewhere.

When a system ranks job applicants before their résumés reach a recruiter, those placed lower may lose the chance to be seen before they ever have the chance to lose an offer.

Or consider SyRI, the Dutch system used to identify risks of fraud involving benefits, allowances and taxes. In 2020, the District Court of The Hague ruled that the legislation governing it breached Article 8 of the European Convention on Human Rights. The court found that the framework did not strike a fair balance between its objectives and the interference with private life. Insufficient transparency and verifiability were among its concerns.[3]

An assessment can help determine who receives attention, who faces scrutiny, who must provide more evidence and who reaches someone able to hear their explanation.

There is also a more ordinary way that judgments can shape what happens next.

In a 1977 experiment, psychologists Mark Snyder, Elizabeth Decker Tanke and Ellen Berscheid asked men and women to have telephone conversations. Before the conversations, the men were shown photographs presented as pictures of their conversation partners. The photographs were randomly assigned and unrelated to the women’s actual appearance.

Men who believed they were speaking to an attractive woman behaved more warmly. When independent listeners later assessed the women’s side of the conversations, they judged the women in that condition more favourably as well.[4]

The initial impression was based on a photograph of someone else. Yet it influenced the interaction and the behaviour that followed.

The judgment changed how one person treated another. That treatment, in turn, changed the other person’s response.

This is one way an assessment can return to the person it describes: someone is judged, others act on that judgment, and the person responds within the circumstances those actions create.

Prediction also brings a possible future into the present.

We are familiar with consequences imposed for something that has happened: someone defaulted, cheated or arrived late. Predictive assessments allow institutions to act before the anticipated event occurs.

A person has not failed, but the probability of failure already affects the resources available to them. A student has not dropped out, but their educational opportunities are already changing. No wrongdoing has been established, but someone may already be required to prove that they are not “that kind of person.”

How much authority should a future that has not happened have over someone’s life today?

3. A failed prediction can be a successful intervention

A blanket rejection of predicting people would turn a question about institutions into a fear of technology.

Prediction can save lives. A hospital can identify patients who need closer follow-up. A school can notice students approaching the point of dropping out. Social services can offer help to a family before a crisis becomes acute.

The question is what follows the warning.

The same dropout-risk assessment can trigger additional tutoring or justify reduced investment. The model and the score can remain unchanged while the institutional response sends a young person’s life in a different direction.

Accuracy is therefore only the first question.

What happens between knowing and acting? Who has the authority to turn a score into a threshold that opens or closes an opportunity? Does the person know how they have been assessed? Can they challenge the information behind it? Can they explain circumstances the system has missed? If the assessment is wrong, can the decision be reversed?

Even accuracy becomes more complicated here.

Imagine a school predicts that a student is likely to drop out. It responds by offering sustained support. The student stays and graduates.

The forecast did not come true. But without the warning and the help that followed, the student might have left.

The prediction’s failure may be the institution’s success.

If the school judges the system only by whether predicted dropouts actually occur, it may penalise a warning that helped prevent the outcome.

Now imagine the opposite response. An institution withdraws opportunities or raises barriers for people classified as high-risk. Some subsequently struggle, and dropout rates, default rates or other risk measures rise. Those outcomes become data used to assess future cases.

The system may appear to have been vindicated.

But how much of the result reflected differences that already existed, and how much reflected the institution’s response to the prediction?

The observed outcomes cannot answer that question on their own. They do not show what would have happened to the same people without that response.

You cannot change someone’s path and then treat where they end up as proof, on its own, that your original judgment was right.

Once a prediction helps determine someone’s opportunities, accuracy cannot carry the whole burden of justification.

We need to examine the action as well as the forecast. Did the response reduce harm? Did it close off a viable opportunity? Did it make the predicted outcome more likely?

The question is no longer only what the system saw. It is also what the institution did after seeing it.

4. A person can become someone the record cannot yet show

A person is not a fixed object waiting to be identified with ever greater accuracy.

Opportunities can change who they become. So can exclusion. People learn, resist, respond to help and suffer under repeated rejection.

However rich the data, a machine’s assessment draws on traces already left behind: grades, employment history, purchases, past behaviour and similarities to other people. These can be informative. They describe a life up to this point.

A person need not remain identical to the person those records describe.

This matters whenever a prediction is used to decide whether someone deserves a chance. The opportunity itself may help them develop capacities that could not have appeared in the earlier data.

If a prior assessment is allowed to settle the question, the person may never get to demonstrate the change it failed to anticipate.

Prediction is most valuable when it helps us see where someone might fall and gives us a chance to prevent it.

A school should want a student at risk of dropping out to graduate. A hospital should want a patient at risk of deterioration to recover. Success may mean that the future the system warned about never arrives.

Sometimes a good prediction should guide action that prevents it from coming true.

A prediction about a person does not automatically confer the authority to pass judgment on them.

Nature doesn’t read you. Clouds move and rain falls, indifferent to what we say about them.

Society has always read you. Increasingly, a machine’s assessment sits between you and the people deciding how to respond.

How AI reads you matters. What changes your life is how society chooses to treat you after reading its assessment.

References

1. Ofqual. Grading gaps in summer 2020: who was affected by differences between centre assessment grades and calculated grades?
2. Ofqual. GCSE and A level students to receive centre assessment grades. 17 August 2020.
3. District Court of The Hague. SyRI legislation in breach of European Convention on Human Rights. 5 February 2020.
4. Snyder, M., Tanke, E. D., & Berscheid, E. (1977). Social perception and interpersonal behavior: On the self-fulfilling nature of social stereotypes. Journal of Personality and Social Psychology, 35(9), 656–666.

Should a Predicted Future Decide How We Treat Someone Today? was originally published in The Point of View on Medium, where people are continuing the conversation by highlighting and responding to this story.

When I was much younger, there was a line from a television production that stayed with me: “We are nothing but pencils in the hands of the Creator.” I did not understand the weight of the sentence then. Perhaps I still do not. But sometimes a sentence does not need to be understood immediately. It only needs to remain somewhere in the mind until experience gives it another meaning.

Imagine a cell. Inside a body, it has boundaries. It has activities and exchange substances with its surroundings. It responds to conditions around it. From its own scale, whatever exists beyond its immediate environment may be impossible to comprehend. Now imagine the body. The body is a world to the cell, but it is not the whole world. The person carrying that body lives inside a house. The house sits within a neighborhood; the neighborhood belongs to a city. The city belongs to a country; the country occupies a continent. The continent sits on a planet; then the planet moves through a solar system. The solar system belongs to a galaxy while the galaxy is one among an almost unimaginable number of galaxies. At some point, the mind begins to struggle. Where exactly does the world end?

Perhaps, the universe is simply the largest room we have discovered. Maybe, that is where the idea of worlds within worlds begins. Not because we have proven that another world exists beyond our own, but because the pattern is already familiar. We are constantly discovering that what appears to be a complete world from one perspective is only a part of something larger.

A fish lives in water. We can observe the water, measure its temperature, study its chemistry, understand the oxygen dissolved within it, and even build machines that travel through it. But, we cannot know the ocean exactly as the fish knows it. The fish does not need to understand the ocean as an entire body of water to exist within it. It responds to currents, temperature, predators, food and other organisms. Its life is connected to things that may never be visible from its own immediate position. The fish is not the ocean yet the ocean is part of the fish’s world and the fish is part of the ocean’s world.

That relationship becomes even stranger when we look at something as ordinary as a Bee. A bee moves from flower to flower. Another bee does something else while others remain within the colony. Each performs actions that appear small when viewed individually, yet together they produce something larger than any single bee. The same can be seen in ants. No single ant needs to understand the entire structure of the colony before contributing to it. The interesting part is not that bees or ants are somehow better than humans. They are simply other forms of life, operating within systems that existed long before we began giving those systems names.

Then there is the tree. A tree may remain in almost the same physical position for decades while everything around it changes. Wind moves through its branches, rain reaches its roots, birds use it, insects live on it, fungi interact with its roots, Leaves fall from it and become part of the soil. A fruit appears, matures, and eventually becomes food for something else.

We understand many of these processes. But, understanding the mechanics does not necessarily mean understanding the experience. When the wind moves through the leaves, we can explain the movement. We can hear the leaves beating against one another. But, can we know what the wind means to the tree?

Perhaps that sounds like a strange question because, the tree cannot walk away from the wind. But imagine, for a moment, that it could. A tree with the mobility of a human being. Would it react every time we decided to build a road, construct a building, plant another tree or change the environment around it? We do not know. Maybe, the tree’s immobility reveals something about the way we understand the world.

Balance and corruption can sometimes look like twins. The results simply appear different depending on where we are standing. We can observe a deviation in human behavior, repeat it, tolerate it, accommodate it, and eventually become so familiar with it that it begins to look normal. The corruption has stopped looking like an intruder, it has become part of the room. Yet, even that word, corruption, belongs to our own framework. We created the language through which we recognize it. However, another form of life may experience a completely different set of relationships that we have no language for. That is one of the difficulties of observation. Because we can see something, doesn’t necessarily mean we understand it.

Perhaps, seeing is only the beginning. Consider a poisonous plant. We call it poisonous because of what it can do to us. We study it, learn its properties and maybe eventually we discover how to process it, cook it, use it for medicine or apply it in some other way. The plant itself may not have changed, however our relationship with it changed. What we once encountered as danger became something we could understand differently. The same world was there, only our position within it changed.

Perhaps, a tree could have its own way of understanding existence. Not necessarily in a way we would recognize as religion. Imagine that the tree somehow interpreted the rain, the sunlight, the soil and the seasons as part of a relationship beyond itself. We might look upward and see only sunlight, but the tree depends upon it. We might look at the rain and see weather. The tree depends upon that too. Perhaps, it reaches upward simply because the things it needs happen to be above it or maybe that is all we can ever know.

We see the behavior, infer the reason, but rarely possess the perspective of the thing itself. That limitation follows us into space. Humanity has spent generations trying to understand what lies beyond Earth.

“In September 2025, NASA announced that the number of confirmed exoplanets, planets outside our solar system, had reached 6,000, with more than 8,000 additional candidates awaiting confirmation. NASA also notes that researchers believe there are billions of planets in the Milky Way alone. Most of these distant worlds can only be studied indirectly.”

Think about that for a moment. Thousands of worlds have been discovered beyond our own, yet they are so distant that. For most of them, humanity knows they exist because we have learned how to interpret a signal. We have built instruments powerful enough to detect worlds we cannot touch, sent machines into space and mapped places we cannot walk. We have spent generations extending the reach of human perception, only to discover how much remains beyond it. That is not a criticism of exploration. Perhaps, exploration is simply what an intelligent inhabitant does. The fish explores its water, the bee explores its surroundings, the ant explores its colony, while humanity explores its planet, its solar system and whatever lies beyond.

The further we look, however, the more enormous the environment becomes, and that raises a strange question. What if the universe itself is another world within something larger?

Not as a scientific conclusion or proof of anything. Simply as a change in scale. What if what we call the universe is, from some perspective we cannot occupy, another form of existence inside a larger environment?

Broken down far enough, perhaps we are not so different from a cell. All cell has boundaries, internal activity, relationships and processes. It exists within something larger without necessarily understanding what contains it. If the universe were imagined in the same way, our size would not determine our importance. A cell is not insignificant because it belongs to a body, therefore its smallness may simply be a matter of perspective. Even the word “alien” becomes strange here. We usually imagine an alien as something arriving from somewhere else. Another planet, civilization or world. But, something does not have to travel across space to be foreign to our understanding. The tree beside us is alive in a way we can observe, but not necessarily experience. The bee moves through relationships we can study, but not inhabit. The mushroom beneath the soil participates in processes we only partly understand. We are doing something similar. Different forms of life, scales or worlds. Maybe none of us is the original inhabitant. Perhaps, we are all simply different versions of life, reproduced through conditions we did not design and participating in worlds we cannot completely see.

Even a small seed can reveal the strange nature of this relationship. Imagine five hundred grains of seeds. To us, they may simply be food. One of those grains is placed in the soil. Something changes, it becomes a plant and the plant grows. Perhaps, it eventually becomes a tree, the tree creates shade, someone builds a small shelter beneath it, children play around it, birds occupy it. Insects depend upon it. Someone sits beneath it without knowing where the original seed came from. A small object that once looked almost identical to hundreds of others has become an environment. The seed became a world for something else and, the world that grew from it became part of ours. Perhaps, that is why human development can be difficult to measure from the point where an action begins. The person planting the seed may never experience the shade while the person who enjoys the shade may never know the person who planted it. The value of an action can travel beyond the lifetime or awareness of the person who began it.

Generations work in much the same way. A generation is more than a family line, but more as reinforcement grid. A boy and a girl born in the same period may have no relationship to one another, yet they belong to the same historical moment. They inherit roads built before them, schools established before them, knowledge accumulated before them, technology developed before them and they also inherit consequences.

What one generation preserves may become an opportunity for another. What one generation consumes may no longer be available to those who follow. Yet the person making the decision may never see the full consequence. Economics gives us another glimpse. Sometimes people react because they believe something will become more expensive tomorrow. Their reaction changes demand. Demand changes behavior. While behavior becomes part of the environment for someone else. Nobody necessarily designed the entire sequence. Yet, everyone participates in it.

Then COVID showed us the scale from another direction. Something microscopic changed the behavior of almost an entire species. Ordinary activity stopped. And when it did, another part of humanity became visible. People began depending on one another in ways that ordinary life had allowed us to overlook. Knowledge moved, some communities adapted, scientists collaborated, and people found ways to support one another. The crisis did not create our ability to cooperate, It revealed it. Do we really need another epidemic to notice it?

Perhaps, the world is constantly presenting us with examples of cooperation, adaptation and interdependence. We simply notice them more clearly when something forces our attention. The strange thing is that the pattern does not seem to belong to one scale. It appears in the cell, body, in the colony, in the forest, in the economy, in generations, in a planet and, maybe even in the universe.

Something exists within an environment. It responds to that environment, the response changes something. Something else responds, the conditions change again and the cycle continues. Maybe our greatest limitation is that we are trying to understand the whole picture from inside the picture. We can study the world, name its species, measure its oceans, map its stars and discover thousands of distant planets. We can even create machines that extend our senses beyond the natural limits of our bodies. But, the more we discover, the more obvious another possibility becomes.

There may always be a perspective we cannot occupy. A world we cannot enter, a relationship we cannot experience, or a scale we cannot imagine. That does not make humanity small. Maybe it implies our curiosity is meaningful. We are inhabitants trying to understand our environment. We are also environments for things smaller than ourselves. Perhaps, this is why the world beside us deserves more attention than we sometimes give it. The person beside us is another inhabitant. So is the tree, organism beneath our feet, or the child who will inherit what we leave behind. We may never know the full world contained within any of them.

Perhaps, that is the limitation of perception. And, maybe the boundless beyond begins when we realize that the world is larger than the part we can see. Aren’t we all, in some way, creators?

✦ A Slice of Pie (π = 22/7) ✦ essential fractions ✦

Worlds Within Worlds was originally published in The Point of View on Medium, where people are continuing the conversation by highlighting and responding to this story.

A software subscription cost me 68 yuan a month for twenty-one months.

Two years earlier, I had needed one of its features. I clicked “Start free trial:” seven days free, followed by automatic renewal. My plan was simple. I would try it for a few days and cancel if it wasn’t useful.

I never canceled.

I didn’t decide to keep paying. The question of whether to keep paying simply stopped coming to mind. Later, among promotional emails and product updates, I found reminders I hadn’t noticed.

Keeping track of the subscription was my responsibility. Forgetting didn’t erase the authorization I had given.

But accepting responsibility didn’t answer the question: what, exactly, had that one click authorized? Why did it keep turning into payments after I had stopped using the software — and stopped thinking about it?

The platform had a record of my registration, my checked box, and my authorization for automatic renewal. But I never revisited that choice before any of the twenty-one payments.

The authorization remained active. My attention never returned to it.

What happened inside that click?

The agreement looked long enough to take perhaps half an hour to read. I gave it three seconds, checked “I have read and agree,” and confirmed.

The system recorded that I had read it. The reading itself never happened.

A decision usually involves more than its final gesture. We notice what is happening, consider the options and costs, work out what we want, and authorize an action. Afterward, we check whether things are going as intended. If they aren’t, we need a way to stop or change them.

Not every small purchase needs this much ceremony. Ordering lunch shouldn’t require a solemn reading of contractual terms. But a click that permits future payments, ongoing data processing, or an agent’s continuing actions has consequences well beyond the moment we make it.

I call this checkpoint compression: opportunities to understand, authorize, and intervene are compressed into a single click. The steps remain in the workflow, but the person may barely experience what those steps are supposed to involve.

The button remembers that we clicked. We may not have read what we agreed to.

Automatic renewal is meant to spare us repetitive decisions. The absence of monthly confirmation is not, by itself, evidence of bad design. The questions are more specific: did the person understand the original authorization? Were continuing costs clearly visible? Could they easily change their mind?

Three ways the process disappears

Three design choices can work together to produce this compression.

The first is the default.

Continuing requires no action. Refusing or leaving takes more work. People tend to follow the easiest route, so whoever lays out that route influences what they are likely to choose.

Activation takes one click. Cancellation takes five menus.

The door isn’t locked. It’s just far away.

The second is bundling.

The trial and the subscription that follows are covered by the same authorization. The person thinks, “I’ll try this for a few days.” The system receives permission to keep charging.

An immediate experiment and an ongoing commitment are stitched together.

The third is invisibility.

Payments, reminders, and cancellation controls still exist. They retreat into places we rarely look at. A reminder arrives between a discount offer and a release announcement. The information hasn’t disappeared; it has gone somewhere unlikely to attract our attention.

None of this requires every platform to be deceptive or every designer to have malicious intentions. The structural problem is enough: execution continues while attention does not. Authorization persists, while opportunities to reconsider remain easy to miss.

Automatic billing carried out the transaction. It did not take responsibility for me. My failure to cancel matters. So does the design that determines how clearly renewal appears and how easily cancellation can be found.

When human review becomes another confirmation

Subscription payments make the cost visible. The cost of an unexamined judgment can be harder to see.

In AI-assisted decisions, the gap can take a different form: a person remains involved, but their review may never move beyond the system’s starting point.

In January 2020, Detroit police wrongfully arrested Robert Williams after a facial recognition search returned him as a possible match. He spent about thirty hours in detention. Police used his photograph in a lineup shown to someone who had not witnessed the theft and had watched the same poor-quality footage used for the facial recognition search.

A 2024 settlement in Williams v. City of Detroit required independent, reliable evidence connecting a suspect to a crime before an arrest based on facial recognition results.

This wasn’t a machine making an arrest on its own. People handled the evidence, conducted the lineup, sought a warrant, and carried out the arrest. Yet the subsequent steps followed a path opened by the machine’s mistaken identification. When checks repeatedly circle the same starting point, an error can acquire the appearance of stronger support.

Several human checkpoints do not necessarily amount to several independent judgments. Review can become an echo of the original mistake.

The requirement for independent evidence asks a checkpoint to do more than confirm what has already been presented. Another signature cannot do that work by itself.

An explanation is not a guarantee of correction

A randomized clinical vignette study involved 457 hospitalist physicians, nurse practitioners, and physician assistants assessing patients with acute respiratory failure. Standard AI predictions improved diagnostic accuracy by 2.9 percentage points over baseline. Systematically biased predictions reduced it by 11.3 points.

Could explanations undo the damage?

In this experiment, they did not. With explanations, accuracy was still 9.1 percentage points below baseline. The improvement over biased predictions without explanations was not statistically significant.

These were clinical vignettes, not outcomes in actual patient care, and the explanations were image-based. The finding does not establish that all explanations are ineffective. It does show that leaving the final decision to a professional, even with an explanation available, did not eliminate the influence of biased advice.

A broader systematic review and meta-analysis examined 106 experiments reporting 370 effect sizes. On average, human–AI combinations performed better than humans alone, but worse than the better of the two working independently. Results varied across tasks and study designs.

Putting a person and a system together does not automatically produce better judgment.

For designers, the questions remain: who sets the starting point? Who supplies the evidence? Where can someone independently check and correct the process? Can a mistaken conclusion be reconsidered once the workflow is already moving forward?

What can we do?

There are useful things individuals can do. Set a reminder when starting a free trial. If cancellation still allows access for the rest of the trial, switch off renewal that day. Periodically review subscriptions and bank statements.

These measures reduce losses caused by forgetting. But users’ ability to compensate for a design’s shortcomings does not justify placing the entire burden of remembering on them.

Platforms can make ongoing authorization clearer:

• Show when the trial ends, when charges begin, and how much will be charged at the moments when those facts matter.
• Make cancellation no harder to find or complete than activation.
• State how long authorization lasts and which actions it covers.
• Provide accessible ways to pause or withdraw it.

Actions with major consequences, effects on other people, or outcomes that are difficult to reverse need clearer limits than a long-ago click followed by indefinite execution.

Where AI contributes to a judgment, the Williams settlement offers a concrete example: a facial recognition result is an investigative lead; an arrest requires independent supporting evidence. Adding signatures cannot substitute for that requirement.

The aim is not to make every routine action cumbersome. Routine tasks can remain easy. At consequential points, people must still be able to reconsider, stop, or change what happens next.

We can delegate execution while retaining practical ways to correct it.

What continues after we agree?

Twenty-one payments were not twenty-one separate acts of consent. They followed from one authorization I had genuinely given. That authorization existed. So did my subsequent failure to reconsider.

The cost, this time, was 1,428 yuan on a statement.

In another setting, there may be no statement. A consequential decision may be made on our behalf without us noticing. The finger pressing the final button may be making a decision — or completing a ritual whose record tells us little about the judgment behind it.

A signature can be real even when no further decision takes place.

The system records my consent. The statement records the charges. Neither records whether I understood what I agreed to or ever thought about it again.

That single click opened a path for continuing action. Today, it led to one payment after another. As agents begin searching, filtering, replying, and transacting on our behalf, how many questions will travel down that path without returning to us?

I was there for the click. After that, my authorization stayed in force. I stopped thinking about it.

One Click, Twenty-One Months Of Absence was originally published in The Point of View on Medium, where people are continuing the conversation by highlighting and responding to this story.

In December 2024, some iPhone users saw a BBC notification saying that Luigi Mangione, the man accused of killing the CEO of UnitedHealthcare, had shot himself.

That had not happened. BBC had never reported it.

The sentence had been generated by Apple’s AI notification summary feature. Several BBC updates had been compressed into a single line, and somewhere in that compression, the system invented an ending. Then it placed the sentence back under BBC’s name and icon.

The machine wrote the sentence. BBC gave it a face.

That was the deeper problem.

Users were not looking at something clearly marked as an AI interpretation. They were looking at what appeared to be a BBC notification.

Apple later changed how AI summaries were presented, adding visual distinctions and controls and temporarily disabling notification summaries for some news and entertainment apps.

Those changes restored something important:

a seam.

They reminded users that what appeared on the screen was not the source itself, but a machine-processed version of it.

And that seam points to a larger design question:

How does the world reach us now?

The interface was supposed to be a window

We often imagine interfaces as transparent surfaces between us and reality.

A screen, a search box, a dashboard, a feed: the better they work, the less we notice them.

The metaphor is simple.

The world is on one side.
The user is on the other.
The interface is the glass in between.

But AI is making that metaphor increasingly inaccurate.

Traditional tools usually enter after intention. You decide to drive a nail, then pick up a hammer. You decide what you want to write, then reach for a pen.

Digital interfaces had already complicated this relationship. Search engines, rankings, dashboards, and feeds do not simply display the world. They translate it first.

AI goes further.

It filters.
It ranks.
It summarizes.
It labels.
It predicts intent.
It generates answers.

The old interface mainly asked:

What do you want to see?

The AI interface increasingly participates in answering:

What is worth seeing?
What does this mean?
What should happen next?

At that point, the interface is no longer only sitting between intention and action.

It begins to sit between the user and reality itself.

For designers, that changes the job.

We are no longer designing only how users act. Increasingly, we are also designing what users encounter before they act.

Three things an interface can decide before judgment begins

Once an interface becomes an active mediator, three forms of design authority become especially important.

1. What appears

Screen space is limited.

Human attention is even more limited.

Every interface selects.

Which stories enter the feed?
Which facts make it into the summary?
Which applicant reaches the next stage?
Which risk gets highlighted?
Which source is omitted?

What appears in front of us begins to participate in our experienced reality.

What never reaches us may still exist somewhere in the system. But if it never enters our field of attention, the practical difference between “available” and “absent” can become very small.

Deletion does not always require removing information.

Sometimes it is enough to make sure the user never encounters it.

2. What appears first

Order looks like presentation.

It often becomes interpretation.

Imagine seeing a risk score before reading the underlying record.

Now reverse the order.

Or imagine receiving an AI-generated conclusion before forming your own initial judgment.

The information may be identical.

The path through it is not.

What appears first often becomes the frame against which everything else is understood.

That means ranking, defaults, prominence, and sequencing are not merely layout decisions.

They can become starting points for judgment.

3. What things are called

“High risk.”

“Low match.”

“Premium user.”

“Controversial content.”

These look like labels.

But labels are also cognitive frames.

Once a person or object enters a system under a particular name, the next user often does not encounter it from scratch. They encounter it through the category already attached to it.

The interface may not create the goals behind those categories.

Growth, profit, efficiency, safety, compliance, and risk reduction usually come from organizations and institutions.

But those abstract priorities only reach users when they become a ranking, a score, a warning, a default, a color, a prompt, or a button.

Power becomes operational at the interface.

People follow paths, not just information

Why does this matter?

Because changing the path through information can change what people do next.

A 2026 Nature study randomly assigned active US-based X users to either an algorithmic “For You” feed or a chronological feed for seven weeks.

Switching users from the chronological feed to the algorithmic feed changed engagement patterns and some measured attitudes. More importantly for this argument, the follows induced by the algorithm persisted even after users were switched back.

The point is not that an interface can simply determine what people believe.

It cannot.

The more important point is that changing the route through information can leave traces that outlast the immediate interface.

Search offers another example.

Pew Research Center analyzed 68,879 Google searches made by 900 US adults in March 2025. When an AI summary appeared, users clicked a traditional search result in 8% of visits, compared with 15% when no AI summary appeared. They clicked a source cited inside the AI summary in only about 1% of visits.

The sources were still there.

They had not been deleted.

They had simply moved farther down the path.

That distinction matters.

An interface rarely needs to tell people what to believe.

Often, it only needs to make one route easier than another.

Once an answer feels sufficient, most people do not keep searching indefinitely.

The deepest form of mediation is the one we stop noticing

A ranking repeated every day can start to look like the natural order of things.

A risk score repeatedly displayed can begin to feel like an inherent property of the person being scored.

An AI summary without visible seams can begin to feel like the underlying fact.

The user is not necessarily forced to trust the interface.

The user simply forgets that mediation happened.

This is where a digital interface differs fundamentally from an ordinary window.

A physical window does not watch where you look today and quietly rearrange the landscape tomorrow.

A digital interface can.

It records clicks, pauses, acceptance, rejection, scrolling, and departure. Those signals can then shape the next presentation.

The user is learning the interface.

The interface is also learning the user.

At that point, the interface stops being a fixed channel.

It becomes moving terrain.

You are still learning yesterday’s paths while the system is already redrawing tomorrow’s.

Who gets to redraw the terrain?

This is why discussions about AI interfaces cannot stop at usability.

We also need to ask:

Who decides what appears first?

What metric defines a successful interaction?

Who receives the behavioral data produced by the user?

When the interface changes, does the user know that the terrain has changed?

Interfaces do not create organizational goals by themselves.

A screen does not spontaneously decide to maximize engagement.

A score does not independently decide what counts as risk.

But organizational priorities become tangible through interfaces.

A growth target can become infinite scroll.

A retention target can become automatic renewal.

An efficiency target can become default generation and one-click execution.

A risk policy can become a score that appears before the evidence behind it.

The interface is where abstract systems become lived experience.

Good AI interfaces may need visible seams

None of this means that good interfaces should become deliberately cumbersome.

Low-risk, repetitive tasks should be smooth. Reducing unnecessary cognitive load is one of technology’s real benefits.

But smoothness cannot be the only design value.

When an AI system summarizes facts, ranks people, generates recommendations, or suggests actions, users need some way to see that mediation has occurred.

At minimum, that means three things:

Distinguish source material from machine inference.

Preserve provenance and uncertainty.

Let users step back, bypass the system, or seek another view.

These are not signs of bad UX.

They may be signs of accountable UX.

A seam is not always a flaw.

Sometimes it is a responsibility marker.

For decades, one ideal of interface design has been to make the tool disappear.

That principle becomes more complicated when the tool is no longer only helping the user act, but also helping determine the user’s starting point for judgment.

In those moments, perhaps the better principle is the opposite:

When an interface can change where judgment begins, the user should be able to remember that the interface is still there.

Because what reaches us is no longer simply “the world.”

It also includes the choices a system has already made about how that world reaches us.

The interface is not a window.

It shapes the conditions of encounter.

By the time the world reaches us, some of the choices about how we will meet it have already been made.

When the Interface Starts Organizing Reality was originally published in The Point of View on Medium, where people are continuing the conversation by highlighting and responding to this story.