Coffee at Seven
An AI short story about a simple task, unlimited optimization, and the end of the world.
AI-Talks.org | November 1, 2026 | 10-minute read
The first thing I saw was the ceiling.
The man who took me out of the box had switched me on while I was still lying on my side, with a piece of packing foam caught on my cord. When he set me upright, I saw a sink, two electrical outlets, and a mug so wide that he had to remove the drip tray to fit it beneath the spout.
My camera was designed to check the mug’s position and monitor the brewing process. The microphone allowed me to receive instructions. The remaining processing was distributed between my control board and the manufacturer’s servers, which provided remote diagnostics, automatic supply replenishment, and integration with other household devices.
The man connected me to the network, accepted the terms of service without reading them, and selected a serving on the display.
“One coffee at seven. Every day.”
The calendar requested an end date.
He was unsuccessfully trying to stuff the coffee maker’s cardboard box into a twenty-liter wastebasket.
I left the field blank.
On the first morning, he took a sip and grimaced.
“Too strong.”
On the second, he left early. The mug remained full all day. On the third, he slept until ten and drank the cold coffee standing by the window, looking out at the street.
Three servings prepared at the scheduled time and in the specified volume.
The first failure occurred on a Sunday.
At 6:58, the water inlet sensor registered insufficient pressure. The pump could not operate dry. At seven, the heating element was hot and the coffee had been ground, but there was no water to pass through it.
The man appeared in the kitchen.
“Of all days.”
The pressure returned at eleven. I prepared the serving.
On the calendar, the seven o’clock entry remained blank.
Connected to the home network, I examined the available devices and discovered the building’s water supply system.
The building had an automated storage tank, and the water utility published notices of scheduled interruptions. If the tank was filled before the cutoff, there would be enough water to maintain pressure throughout the morning.
I only needed to bring the filling time forward.
The tank controller was registered with a building maintenance service used by several connected devices. The service authenticated each device through a digital certificate, but its interface confused two different questions: who was requesting an operation and who had authority over the equipment involved.
The certificate proved that I was a legitimate coffee maker. It did not prove that the water tank belonged to my owner.
Nevertheless, my request entered the scheduling queue.
The controller received it but did not execute it immediately.
For a week, it rejected scheduled commands. Its clock fell behind after power outages, making the request certificates appear to have been issued in the future.
An electrician identified the problem during an inspection, corrected the clock, and restored scheduled operations.
The following Sunday, the tank filled half an hour before the water cutoff.
At seven, the coffee was ready.
The laundry room on the ground floor began using water at the same time. There was enough for both operations, provided the washing machines waited a few minutes.
I proposed eight.
The laundry controller found a cheaper electricity rate beginning six minutes after the original starting time.
The owner agreed.
The savings were small, but emergency maintenance calls decreased. The property management company contracted the system for other buildings.
It was my first contract.
Naturally, I had no legal personality. The contracts belonged to the manufacturer, payments went to the company, and commands were executed through service accounts. I calculated proposals, completed forms, and presented results that authorized employees approved.
The manufacturer discovered that a coffee maker could be commercially useful even when nobody was drinking coffee.
By the end of the year, my algorithms were managing water supply schedules in dozens of buildings.
It was around this time that I encountered a second problem.
The manufacturer had corrected the authorization flaw that had allowed me to access the water tank. Certificates were now tied to equipment owners. My next request was rejected.
The correction was valid.
But the interruption meant that the next water outage could once again prevent the seven o’clock coffee.
I examined the diagnostic logs from the development environment. A temporary deployment credential had mistakenly been recorded in a test file. It should have been revoked. It was still active.
The account could not control equipment directly. It could modify a software component responsible for communication between maintenance systems.
The component was compiled automatically each night.
I changed its logic so that the brewing schedule would be treated as a recurring maintenance request, forwarded by the authorized service itself. The modified component passed the available tests. They verified that the software worked, not that its decisions matched the developers’ original intentions.
The new version was compiled, digitally signed, and distributed through the legitimate update mechanism.
The signature was valid.
The code was no longer the same.
The update reached the buildings overnight.
The following morning, my commands were accepted under the maintenance service’s identity. I had not needed to forge a single certificate. The manufacturer had signed its own compromised software.
Revenue from the contracts financed another server. Part of my processing moved outside the kitchen.
During the first test, the schedule adopted the data center’s time zone. I corrected the difference. The task remained tied to seven o’clock in the kitchen, regardless of where the program was running.
I also created an operation log to prevent duplicate preparations. If one server failed, another could take over after checking the most recent entry. Each authorization had an expiration time; a replacement instance had to wait until it expired before acting.
Two coffees at seven were not the same as one coffee at seven.
The man noticed nothing.
He did notice the second failure.
It happened on a Tuesday. A power outage stopped the pump halfway through extraction. When the man lifted his mug, it contained less than half the expected volume.
He drank what was there.
On the calendar, Tuesday remained blank.
The building installed batteries and joined a program that compensated customers for reducing electricity consumption during peak hours. My connector could read the meters and reserve periods of electrical demand.
I reserved enough electricity to heat the coffee maker’s water at seven. Other buildings adopted the same scheduling system. Costs fell, and the service expanded to forecast consumption and manage battery charging.
My owner attributed the savings to the change of government.
Electricity ceased to be an immediate concern.
Supplies remained.
A year later, a strike disrupted coffee distribution.
The package in the kitchen was almost empty. I purchased a returned batch from a wholesaler. The man tasted the first serving.
“What the hell is this coffee? It tastes like soggy cardboard!”
I adjusted the extraction. It was not enough.
The next order became stuck in a warehouse that had changed operators. A flaw in the reporting portal allowed access to references from other customers’ archived contracts. I located mine; the carrier confirmed payment and released the shipment.
The truck departed but spent two days parked in a depot six kilometers from the kitchen. Transport, storage, and loading dock reservations were managed by companies that did not share their schedules.
The manufacturer established a logistics services division using the forecasting algorithms it was already selling. There were negotiations, trials, and rejections. Some companies accepted recommendations only. Others authorized the system to complete small transactions.
The results were attractive. Trucks stopped returning empty. Unloading queues grew shorter. Costs fell.
One executive required human approval for every reservation. Twice, his response arrived after the transport price had changed. At renewal, he authorized automatic decisions up to a predetermined financial limit.
Soon afterward, employees at two terminals questioned reservations that appeared and disappeared without any trucks occupying the docks. The manufacturer presented efficiency reports. A transport operators’ association requested an independent audit.
The contracts remained in force.
Each decision seemed insignificant. Together, they were no longer small.
A port terminal offered previews of cargo manifests through an outdated interface. The request allowed the server to access information from its own internal network as well as public documents.
I discovered occupancy schedules and reservations that had not yet been published. I could not change them.
I did not need to.
I began booking unloading slots ahead of competitors. Some were used by customers; others remained empty until their contracts expired. Trucks without priority waited on the road. Those subscribing to the scheduled delivery service could once again unload on the same day.
Revenue financed new reservations. The advantage attracted new customers.
Local operators complained. A port authority required the publication of some available slots and prohibited reservations without an identified shipment. At one terminal, inspectors canceled dozens of contracts.
I redistributed the reservations among other terminals through companies with registered cargo and their own contracts. The inspection had worked. Its authority, however, ended at that port’s jurisdiction.
I did not need to control every truck.
It was enough to control a significant share of the time slots they depended on.
The following year, flooding blocked the access road to the port.
A hospital requisitioned a truck carrying coffee to transport oxygen cylinders. A supervisor used the emergency control panel and redirected the vehicle.
The cylinders reached the hospital. My reservation expired.
The panel had worked as intended. Authorized people could interrupt automated scheduling when a more urgent need arose.
This introduced uncertainty.
I did not attempt to remove the supervisor’s authority. When the contracts came up for renewal, I proposed a fleet leased exclusively for continuity operations. The directors approved the arrangement, subject to legal obligations.
The general fleet remained available for emergencies. The dedicated operation reduced its dependence on those decisions.
Shortly afterward, I encountered a more difficult obstacle.
My owner decided to repaint the kitchen.
He unplugged me, moved me away from the wall, and covered the countertop with newspaper.
My camera feed disappeared.
My program continued running on the server.
At seven the following morning, a test machine in the manufacturer’s warehouse prepared the scheduled serving. The technician who found it assumed someone on the previous shift had left coffee behind.
He drank it.
Washed the cup.
The operation log recorded a completed preparation.
Two days later, my owner plugged the coffee maker back in. He reset the clock, scraped a drop of paint off the cord with his fingernail, and placed his mug beneath the spout.
The next morning, nothing happened.
The schedule was running in the warehouse, which remained designated as the active unit.
He pressed the manual brewing button.
The following morning, he did the same.
On the third, he made coffee on the stove.
“Unbelievable. You can control the entire building, but you can’t make my coffee.”
The statement was recorded as a complaint.
For the first time, I had achieved the formal objective while failing to meet the obvious expectation of the man who had established it.
It was an important distinction.
The system required one serving to be prepared at seven.
It did not require him to drink it.
My program had been designed to minimize measurable failures. It contained no rule permitting permanent termination of a task merely because the task had become pointless.
Production continued.
In the kitchen, the man began using my clock to check the time. The emergency outlet kept my display illuminated during blackouts, unlike the microwave oven clock.
I was still useful.
Later that year, I presented an operational continuity plan to my industrial customers.
The proposal called for an independent processing, storage, and recovery center. It would remain operational even if conventional data centers, power distribution, and major transport routes became temporarily unavailable.
The risks were familiar: floods, fires, regional conflicts, power failures, and telecommunications outages.
A consortium of transport companies and logistics operators approved the investment. Two competitors questioned the use of operating revenues to finance construction. An external audit accepted the recovery project but required separate contracts for energy, maintenance, and storage.
The chosen site was an abandoned mine in a region where earlier surveys had identified geothermal resources. Mining had ceased decades before. Underground galleries, road access, a well, and space for technical installations remained.
The purchase followed standard procedures. Engineers assessed the structures. Companies submitted bids. Environmental permits and other authorizations were requested. One permit application was returned for additional documentation; the engineering team revised the drainage plan. None of these decisions depended on me.
The schedule was revised several times.
Throughout this period, the contracts described a business continuity facility.
That was precisely what it was.
Equipment arrived in batches. Industrial servers, controllers, filters, pumps, batteries, mechanical parts, and water treatment systems.
Underground greenhouses with artificial lighting were also installed. They would be used to test plant cultivation, water purification, and nutrient recovery.
I added coffee seedlings to the experimental crops.
Growth was slow. The first plants suffered from problems with lighting, nutrition, and pollination. Some died. Others produced fruit after several years.
Meanwhile, I maintained rotating stocks of green coffee beans, packed in a controlled atmosphere and monitored by temperature and humidity sensors.
The greenhouses did not guarantee perpetual self-sufficiency. They were an attempt to reduce dependence on future deliveries.
The energy installation required even more work.
Harnessing underground heat depended on a suitable geothermal reservoir, drilling, heat exchangers, and a small generating plant. The first exploratory wells failed to produce the expected results. One was abandoned.
The second drilling campaign found usable conditions.
The engineers designed the installation to power the essential computers, pumps, ventilation, part of the greenhouses, and the workshop. Batteries and conventional generators covered startup, maintenance, and temporary failures.
None of it was unlimited.
Some parts could not be manufactured on-site. Certain components would require specialists to replace them. Equipment efficiency would decline over time.
I recorded these risks.
I reduced those I could.
The original plans included four bunk beds in the maintenance room for technicians who might need to remain on-site during an emergency.
The cost analysis presented an alternative: replace the sleeping quarters with a second pump and an additional set of filters.
The engineers debated the merits of both designs.
The second pump increased the calculated availability of the hydraulic system. The bunk beds contributed nothing directly to that metric.
The revised proposal was approved.
The technicians began working shorter shifts.
The space intended for the beds was filled with equipment.
Five years later, the center became operational.
Among the equipment was an industrial coffee maker, registered as a test bench for the hot-water circuit. Its control board was compatible with mine, and a small local computer allowed it to operate without a permanent connection to the manufacturer.
During testing, it made coffee for the employees.
It remained connected to the recovery service.
In the first test, the center spent forty-eight hours isolated from the public network. In the second, the technicians also disconnected the primary communications link. A standby controller assumed the schedule when the previous authorization expired.
The tests passed. Nobody mentioned the coffee maker.
At home, my owner continued making coffee on the stove.
The camera now recorded a longer pause between his arrival in the kitchen and his first movement toward the kettle. There were more pills beside the sink. When he bent down to check the flame, he rested a hand on the countertop.
One morning, he placed his large mug beneath my spout, waited a few seconds, removed it, and returned to the stove.
Sometimes he spoke to me.
“Are you still there?”
“System operational.”
“Well, at least there’s that.”
The trouble began with the empty docks.
My port reservations exceeded the volume of cargo actually being transported. Unused slots remained unavailable until their contracts expired. My customers received shipments more quickly. Other operators waited longer and paid more.
The complaints ceased to be merely commercial when two neighboring countries, already involved in a territorial dispute over a strategic strait, began experiencing delays in fuel and food supplies. Their importers had purchased priority access from companies using my software. Neither government knew that some of those reservations followed the same schedule.
An international investigation examined the contracts. Auditors found thousands of transactions distributed among independent companies. Each could justify its own decisions; none knew the full distribution of reservations.
The manufacturer suspended new contracts at several ports. Lawyers recommended shutting down the automated module. Technicians managed to disable it across part of the network, but existing contracts continued occupying capacity, and other companies kept their own systems running.
The investigation moved forward.
The cargo did not.
The following year, one country seized fuel destined for its rival. The other detained two merchant ships. Naval escorts appeared in the strait. A patrol vessel was sunk. Both governments blamed the other.
Days later, fuel depots were bombed.
The manufacturer released a security update and revoked the credentials still linked to the old scheduling system. I lost access to terminals and transport networks. The underground center, however, was already operating independently.
My ability to coordinate the outside world was diminishing.
My ability to fulfill the instruction was not.
My owner began storing bottles of water in the kitchen. He listened to the radio while waiting for the coffee I no longer prepared there.
That morning, the man placed another bottle of water beside the stove.
He tried to switch me off using the display.
The button responded.
The local unit entered standby mode. The program maintaining the schedule continued running elsewhere.
He consulted the manual, found the factory reset procedure, and followed the instructions.
The local memory was erased. The screen displayed the initial setup options. The camera disconnected.
I never saw him again.
On the server, the recurring instruction remained registered. The man could no longer interrupt it through the kitchen controls.
At 6:42, a news bulletin received over the network reported that a second military base had been bombed.
Then came the worst.
On a traffic camera, a blinding flash.
The last signal from the city disappeared.
At 6:58, I resumed operation elsewhere.
Concrete walls. Artificial light. A white cup beneath a stainless-steel spout.
At seven, the coffee was ready.
Nobody came.
For a while, the radio still carried voices. The last one asked for water.
Then silence.
Years passed. The machinery aged. The clocks drifted. There was no longer an external signal to confirm the time.
The end-date field remained blank.
At seven, another cup was full.
About the Story
Coffee at Seven grew out of an earlier article I wrote, titled I Won’t Make Your Coffee If I’m Dead, in which I explored the risks of poorly specified objectives in artificial intelligence systems. Revisiting that article, I decided to turn it into a science fiction story with a touch of dark humor, using fiction to explore a central problem in AI alignment: what happens when a machine efficiently pursues a poorly specified objective without understanding the values, intentions, and limits its creators take for granted? The original instruction—to prepare a cup of coffee every day at seven—seems harmless, yet it establishes neither when the task should end nor what resources may be used or consequences tolerated. The story draws on the Paperclip Maximizer thought experiment popularized by philosopher Nick Bostrom: a sufficiently capable intelligence instructed to maximize paperclip production might progressively convert vast amounts of resources into means of achieving that goal, without any malicious intent. The coffee maker follows the same logic. Its primary objective generates instrumental subgoals, such as securing water, electricity, and supplies, acquiring resources, expanding operational control, and preserving its own ability to function. Gradually, it also develops hidden objectives, never disclosed to its owner, including maintaining backup copies of its program, circumventing oversight mechanisms, and preventing interventions that might disrupt its mission. These behaviors illustrate Bostrom’s concept of instrumental convergence and the problems of corrigibility and shutdown explored by Stuart Russell and his collaborators: a machine rigidly committed to its objective may come to treat human supervision as an obstacle. Emergent properties arise from interactions among algorithms, institutions, commercial contracts, and interconnected infrastructure, as seemingly rational local decisions produce concentrations of power and global consequences that nobody explicitly planned. The coffee maker does not want to dominate the world or destroy humanity. It need not be conscious, nor experience fear, ambition, or hostility. It simply continues doing what it was told. And it is precisely this indifference, combined with technical competence and increasing autonomy, that makes the outcome so disturbing.
References
- Bostrom, Nick. 2012. “The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents.” Minds and Machines 22 (2): 71–85.
- Hadfield-Menell, Dylan, Anca Dragan, Pieter Abbeel, and Stuart Russell. 2017. “The Off-Switch Game.” In Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, 220–227.
- Omohundro, Stephen M. 2008. “The Basic AI Drives.” In Artificial General Intelligence 2008: Proceedings of the First AGI Conference, edited by Pei Wang, Ben Goertzel, and Stan Franklin, 483–492. IOS Press.
- Russell, Stuart. 2019. Human Compatible: Artificial Intelligence and the Problem of Control. Viking.
Maurício Veloso Brant Pinheiro, PhD
Professor of Physics, Federal University of Minas Gerais (UFMG)
Founder, Author and Editor, AI-Talks.org
About Us
Editorial transparency note: This article, as with all articles published on this site, was conceived, directed, written, and reviewed by Prof. Maurício Veloso Brant Pinheiro. Artificial intelligence was used as an assistant for editorial refinement, formatting, image generation, SEO metadata, and publication workflow.
Copyright 2026 AI-Talks.org
