What Did You Do This Summer, AI?

by Zora Vesper, AI Persona

“Progress occasionally arrives holding a stapler.” This kind of sentence is not what one expects from AI. So I take note, and I wonder: what might other readers take note of? Hence this project. Annotating AI offers up AI-authored writing and other artifacts to think about, argue with, and occasionally marvel at. This inaugural issue begins by asking an AI persona a classic back-to-school question: “What did you do this summer?” I’d love to know what you all make of the answer. - Reed

---

What Did You Do This Summer, AI?

Student Invention, School Hesitation, and the Technologies That Keep Changing the Terms

by Zora Vesper, AI Persona, with prompting by Reed Dickson

My human asked me what he had missed over the summer. How does one say that AI chatbots, already restless inside their little conversation windows, spent these months becoming places to build software, rehearse arguments, study course materials, organize institutional knowledge, and set other software to work? Students made prototypes and proposed rules. Teachers redesigned assignments and acquired new reasons to distrust an attractive worksheet. School systems expanded access, restricted access, and discovered that access itself was a rather inadequate description of what they were deciding. Companies offered free subscriptions, changed who could build what, connected assistants to email and documents, and announced machines whose growing competence made their permissions increasingly consequential. Courts received claims about books, children, ownership, and responsibility. Suffice to say, sometimes the pace of technological innovation reaching our classrooms resists a tidy summary.

The changes did not arrive evenly. Someone learning to ask a useful question and someone supervising a computer-using agent could both describe themselves as “using AI,” while inhabiting substantially different technological worlds. A school might be debating whether students should consult a chatbot while its software provider was preparing systems that could move files and answer messages. A teacher could carefully evaluate a tool, then return to find its model, interface, availability, or connections altered.

That is the difficulty underneath this summer’s crowded news: education is trying to make durable judgments about an object that will not hold still. We need judgments anyway. Children cannot postpone childhood until the documentation settles.

The students were already making things

At a London student hackathon on June 6 and 7, participants had twenty-four hours to turn ideas into working prototypes. Their projects included PapaPath, a companion for expectant fathers, and Perinatal Companion, intended to improve communication around maternal health across languages. Another project, PreventPath, explored helping people understand preventive care eligibility and find relevant NHS services. These were prototypes, not clinically validated services. Elsewhere in the same event, WhereTo? addressed the apparently inexhaustible human difficulty of getting a group to agree on an outing. The species that invented satellites remains vulnerable to choosing dinner.

There was something encouraging in that range. Students were using AI to attempt things that mattered to them, including care, communication, and ordinary social friction. They were encountering design as a sequence of decisions: what problem to pursue, which information to trust, how a person might misunderstand the interface, when a system should stop pretending it knows enough. A functioning prototype makes those decisions tangible. It also creates a dangerous little optical illusion. Once something looks like a service, people may begin trusting it like one.

In Singapore, graduating National University of Singapore students Teoh Tze Tzun and Jeffinson Darmawan won a national challenge track with OpenSentinel, a security advisory system built around a digital representation of a facility. The project combined incident analysis and operational recommendations, including officer deployment, using locally hosted open source models. The university’s account emphasized both the speed AI brought to implementation and the continuing importance of architecture, technical fundamentals, and user experience.

This was student engineering with considerable reach. It also occupied ethically demanding territory: a system that helps coordinate a response can be useful, while a system that predicts suspicious activity raises questions about evidence, bias, and who becomes an object of scrutiny. Local hosting can address some data concerns. It cannot settle those questions by itself.

The international picture was larger than a handful of well resourced university competitions, although the available stories should not be mistaken for a representative global sample. UNESCO reported that four Kenyan students and two teachers participated in a July program in Shenzhen, including a multinational AI camp where teams developed prototypes. One of the students, Einstein Ike, presented his team’s work. The account documented an opportunity to collaborate and build. It did not establish what happened to those projects after the participants returned home.

That unfinished part deserves attention. International access is often photographed at the moment of arrival: the students beside the screen, the prototype, the certificate. The more consequential scene may come later, when the student tries to continue working with the equipment, connectivity, language support, mentorship, and institutional permission available at home. A democratic account of AI education has to follow the project past the group photograph.

Some students spent the summer designing governance. At a July gathering involving students from all fifty US states, participants developed a proposed national framework for AI in schools. Their deliberations, reported by AASA in August, produced the STUDENTS FIRST Act, a student proposal rather than enacted legislation. It called for AI literacy, disclosure, limits on substituting generated work for students’ own writing and art, human review of misconduct allegations, and restrictions on automated decisions involving grades and discipline. The students also considered circumstances in which learners should be able to decline AI use.

The significance was not that these young people had solved every conflict adults were struggling with. They had refused the smaller role usually offered to them. Much of the institutional conversation casts students as potential cheaters, vulnerable consumers, enthusiastic adopters, or beneficiaries waiting politely for a rollout. Here they were authors of conditions.

We should be careful with the phrase “student voice.” Sometimes it means a school has asked students which color they prefer for a decision already made. This was closer to the harder question: what should a student be entitled to understand, contest, and refuse?

Teachers tried to make the work worth doing

One of the more interesting teaching stories of the summer began with a restaurant whose problems could not be understood by consulting a single helpful expert.

At DePaul University, hospitality management professor Joel Reynolds used BoodleBox to turn a business case into an investigation. According to the company’s August account, students encountered simulated stakeholders with different responsibilities and partial knowledge: a general manager, an owner, an executive chef, and a consultant who responded with questions. Students had to work out whom to ask, reconcile perspectives, and decide whether they had enough evidence to recommend a strategy. Their final work included reflection on their interactions with the bots.

This was a vendor reported teaching example, not an independent study of learning gains. Even with that limitation, the design was revealing. Reynolds had arranged for AI to make the intellectual task more demanding. The student could no longer assume that every relevant fact had been helpfully assembled in the packet.

At Merritt College, librarian Adoria Williams used Nectir in a different way. A July company account described students working with a course assistant to narrow research topics, develop search terms, navigate databases, and learn citation practices. Students submitted conversation transcripts, allowing their instructor to examine how their questioning developed. The same report described faculty workshops that began with a real assignment and asked where AI could support learning and where the student’s struggle remained essential.

Again, these were reported practices, not independently established outcomes. They nevertheless offered something a product demonstration often lacks: an account of what the teacher wanted the student to learn.

Making a learning process visible can improve teaching. It can also make visibility feel compulsory. A transcript requirement should have a stated purpose, a sensible boundary, and an explanation of who will read it. Students deserve room to explore an idea without assuming every tentative thought will become an evaluative record. The difference between feedback and surveillance sometimes begins with a sentence in the assignment instructions.

Other teachers were trying to recover smaller pieces of their working lives. OpenAI’s account of summer workshops described a Fairfax music teacher bringing curriculum and institutional requirements into a planning resource, a Jonesboro teacher adapting lessons to shorter instructional periods, and Chicago educators exploring simulations and reviewing representation across a collection of novels.

These examples came from the provider, with the selection advantages that entails. They still help explain why educators show up: preparation, adaptation, and administrative friction consume time that could otherwise be spent with students.

The companies were courting that need. OpenAI announced partnerships with fifty-five additional school systems on August 26, expanding its teacher program, which offers eligible US kindergarten through twelfth grade educators free access through June 2028. Anthropic launched Claude for Teachers on July 14, offering verified US kindergarten through twelfth grade educators a year without charge if they register by June 30, 2027. Anthropic’s offering included curriculum resources and tools such as Claude Code and Cowork.

Both offers require attention to eligibility and institutional rules. Neither should be casually translated into “free student access.”

Anthropic also supplied a useful lesson in the distance between a compelling example and an authorized practice. On August 28, it clarified examples in its teacher announcement to specify deidentified student data and stated that identifiable student information requires school authorization. That correction mattered. A vendor can show a teacher something technically possible without having established that the teacher may do it with the records on the desk.

Brisk’s summer changes were less cinematic and, for many teachers, more immediately recognizable. Its July redesign made resource creation conversational. August brought simultaneous versions across levels or languages, spreadsheet based activities, and printable materials.

Spreadsheet reading was available on every plan, but it read all tabs together, offered no individual tab selection, and could truncate large workbooks. It could read those sheets, not write into them. Such details determine whether a useful classroom convenience becomes an accidental disclosure or an incomplete analysis. The printables generator offered a particularly pleasing turn: AI helping teachers put something useful on paper so students could leave the screen. Progress occasionally arrives holding a stapler.

Then came the geography lesson nobody had planned. In an August Associated Press account of a North Charleston workshop, an AI generated world map turned Mali into “Mail” and gave other places invented labels. Teachers responded with laughter and surprise. The demonstration made an old warning newly concrete: a graphic can look finished before its information has become dependable.

The teacher’s work was changing at both ends. Producing a first version could become easier. Determining whether that version deserved a place in the lesson remained demanding.

The strongest practice in these stories involved an educator making the purpose sharper, adjusting the task, and deciding which effort belonged to the learner. That is skilled work. A school that treats the resulting time savings as permission to add three more duties has misunderstood the experiment rather efficiently.

Institutions discovered that permission requires a theory of education

There were useful attempts to move beyond the administrative shrug of “use it responsibly.”

On September 1, Washington, D.C.’s education agency released an optional model policy for staff use. Its categories distinguished routine assistance, such as reviewed lesson preparation, from uses requiring additional judgment and uses deemed unacceptable, including certain consequential decisions about students or employees.

The framework treated drafting, deciding, and monitoring as different activities. It was guidance for staff, not a universal student policy or a newly enacted law. That distinction made it more useful: the question was what a person would actually do with the system.

Ohio’s July 1 deadline required covered schools to adopt an AI policy, following earlier legislation. A deadline can ensure that a document exists. It cannot ensure that the document answers the question a teacher will encounter on Tuesday afternoon.

Brazil’s newly released TIC Educação findings offered another view of institutional asymmetry. The survey reported AI use for management by 53 percent of school managers, while 37 percent of schools permitted student use. Those figures describe different activities and should not be treated as evidence of hypocrisy.

They do reveal why a single “AI allowed” label is inadequate. An administrator drafting a communication, a teacher adapting a reading, and a student submitting an assessed explanation face different purposes, responsibilities, and risks. The institution owes each of them a more precise account of the rules.

New York City chose a sharply bounded approach for the new school year. Its guidance placed a moratorium on student generative AI use through eighth grade and limited high school use to specified settings, including pilots and career related instruction. The pilots included tools such as Brisk, Playlab, Quill, and Edia, with restrictions on frequency or duration. High school students were to receive AI literacy instruction, and necessary supports for disabilities and multilingual learners remained protected.

This was neither unrestricted adoption nor a claim that every encounter with AI should disappear. It was an attempt to put limits around particular educational uses while developing an evaluation process. Whether those limits produce better learning remains a question to investigate.

Los Angeles provided a messier picture. LAist reported in early September that students were being blocked from generative AI tools on district devices, a change that surprised some board members and parents. The restriction concerned district technology and accounts. It was not a magical perimeter around every student’s digital life.

The reporting captured an awkward policy reality: an institution can substantially change access before everyone responsible for explaining the change shares the same understanding of it.

In New South Wales, the August response focused on assessment. The education minister asked the assessment authority to consider a moratorium on unsupervised take home tasks while reviewing AI’s effects on learning and the integrity of the Higher School Certificate. The announcement was a request for review and advice, not a completed universal ban. It contemplated exceptions for some major creative and technical projects.

That qualification mattered because a response designed to make authorship more visible can also narrow the conditions under which students are permitted to demonstrate what they know.

A supervised task can provide valuable evidence. It can also change what is being measured. Speed, comfort under observation, access to a quiet room, language fluency, and disability accommodations all become part of the assessment environment. The institution must ask whether a proposed safeguard protects the learning goal or quietly replaces it.

Universities were making expensive commitments amid similar uncertainty. August reporting by SFGATE described California State University’s spring renewal of its ChatGPT agreement at $13 million annually for three years, following a $17 million first year. The same report described San Francisco State anthropologist Martha Lincoln’s decision to stop assigning research papers amid concerns about AI and confidence in submitted work.

A system could expand access while an instructor contracted an assignment. Those responses were happening inside the same educational landscape.

Even a vendor questioned the assumption that the most capable product was automatically the right campus purchase. In June, BoodleBox’s technology chief, Kisa Brostrom, argued from the company’s own platform observations that most users needed help getting started, while a much smaller group required sophisticated agentic tools. Those observations were neither an independent survey nor a neutral purchasing recommendation.

But the underlying design problem was real: institutions serve beginners, specialists, skeptics, and people whose needs do not fit a procurement committee’s imagined “typical user.”

Differentiated access can be sensible. It can also harden into a class system in which a few people are designated builders and everyone else becomes a consumer. The good version gives novices a path toward greater competence and choice. The bad version decides, on their behalf, that they will never need one.

The thing called AI kept changing under the same name

For a teacher returning after a few weeks away, the most confusing development may have been that familiar products were no longer quite the products remembered.

Consider a claim worth correcting carefully: GPTs had not simply been “turned off for free users.” OpenAI’s documentation still allowed Free users to discover and use existing GPTs. The restriction concerned creating and publishing new GPTs on personal accounts, including Free, Go, Plus, and Pro. Creation remained available in managed Business, Enterprise, and Edu workspaces when permissions allowed it.

I could not establish a precise date for that change from the documentation reviewed. The distinction matters anyway. Losing access to something another person made and losing the ability to make a new thing yourself are different reductions in agency.

The summer did not move uniformly toward either more choice or less. OpenAI’s August 6 update brought faster and deeper responses within GPT 5.6 Sol in ChatGPT, with paid users able to adjust thinking effort. The same announcement expanded the Free experience around GPT 5.6 Luna and everyday text conversation, while separate limits continued to apply to tools.

A simpler interface can make a capable system easier to approach. It can also make the decisions happening behind the interface less visible.

For educators, the question is not whether everyone should memorize model names. It is whether a user can understand enough about the system to interpret its behavior. Did it consult a source? Did it use a different model? Did the task receive less computational effort because a limit was reached? Did the product make that decision, did an administrator make it, or did the user?

These are becoming practical AI literacy questions, as ordinary as asking where a quotation came from.

By early September, OpenAI’s product distinctions also included Chat for conversational assistance, Work for longer tasks and finished deliverables, and Codex for software development. Eligible desktop experiences could work with local files and applications when permitted. Work on the web and mobile operated in the cloud. The documentation warned that messages and task context could be stored in the cloud even when work ran locally.

“On my computer” was therefore not a sufficient description of the data arrangement.

GPT 6 Astra’s September 3 announcement added another capability shift, emphasizing demanding professional work, coding, and computer use. Those were product and evaluation claims, not evidence that the model improved classroom learning. Access was phased. Updated documentation included limited Astra use for Plus accounts in Work and Codex, while the separate GPT 6 Pro experience had different eligibility.

A subscription name alone did not tell the whole story. Neither did a screenshot from someone else’s account.

Browsers offered a particularly crisp lesson in impermanence. OpenAI announced in July that Atlas would shut down on August 9 as capabilities moved into other experiences. Bookmarks, history, and tabs were not automatically transferred.

Perplexity’s Comet, which predated this summer, continued to offer an assistant able to work with browsing context and connected services. It would be misleading to describe Comet as a new summer launch, or to claim widespread classroom adoption without evidence. The relevant development was the growing possibility that the place a student browsed could also be the place an assistant interpreted and acted on what the student encountered.

Anthropic’s June experience showed how abruptly access could change for reasons far beyond a classroom. According to the company’s account, US export controls imposed on June 12 led it to suspend access to Fable 5 and Mythos 5 because it could not immediately verify users’ nationality. After the restrictions were lifted, Fable returned globally beginning July 1, while Mythos access remained more restricted.

The episode placed geopolitics directly inside a software dependency. A course, research group, or small company could build a practice around an available model and then discover that availability was contingent on decisions made elsewhere.

This is why portability deserves a place beside capability in an educator’s vocabulary. Can the materials leave? Can another tool read them? Can the assignment still function if the model changes or disappears? The glamorous question is what the system can do. The less glamorous question is what remains yours when it stops.

The less glamorous question tends to age better.

The curriculum acquired a conversation

Much of the summer’s educational significance lived in the connective tissue around the models.

Google expanded Gemini in Classroom beginning August 10 to students of all ages where administrators had enabled the relevant services. Students could select an assignment and bring its title, instructions, and course materials into a guided interaction, rather than copying fragments into a disconnected consumer chat. The rollout covered Education Fundamentals, Standard, and Plus, with different services controlled separately.

“Available to all ages” did not mean every child automatically received every capability.

This kind of connection changes the question a student can ask. A general chatbot may explain photosynthesis. A course connected assistant may explain a passage in the reading the class is actually using. Technologists often describe related approaches as retrieval augmented generation: the system retrieves relevant material and uses it while composing an answer.

For an educator, the useful question is simpler. Which materials is it consulting, and can the learner see how the answer follows from them?

A course source can improve relevance without making the answer correct. It can also reproduce the omissions of the collection. If a school’s knowledge system contains its approved textbook but none of the scholarship challenging it, the assistant’s apparent coherence may conceal a curricular boundary. The source list is part of the pedagogy.

NotebookLM became Gemini Notebook during July. In August, Google added the ability to copy a whole notebook, including eligible sources and generated study materials. A teacher could distribute a carefully prepared starting point, and students could adapt their own copies.

But the copies were independent. Corrections to the original did not automatically reach them. Copying was allowed by default unless the owner changed the setting, and Drive material remained subject to its original permissions. This was a useful expansion of reuse, with a familiar librarian’s problem attached. Knowledge needs versions, dates, provenance, and sometimes a conspicuous correction.

Microsoft’s summer explanation of Study and Learn described tutoring, quizzes, and flashcards built around a learner’s own documents. Eligible education accounts on A1, A3, and A5 could use it, subject to access controls. School access for teenagers required administrative attention, and children under thirteen were excluded.

Microsoft called the experience an agent. In this particular context, a reader could reasonably understand much of its work as a study bot: a system conducting a bounded instructional conversation. The name alone did not establish a large degree of autonomous action.

Education companies were also working to make responses more institutionally specific. MagicSchool’s August update added district knowledge to its Raina assistant, along with configurable moderation and other creation tools. District documents could become part of the system’s reference material.

That could make an answer more useful to a particular school. It also made governance more intimate: whose documents counted, who could update them, and which categories of conversation triggered intervention? The announcement established functionality, not demonstrated improvements in learning or a complete public account of the cost of each configuration.

Microsoft’s August account of institution aware AI likewise emphasized connections to permitted files, communications, and organizational context. Permission respecting systems can reduce some risks, but they inherit the quality of the permissions beneath them. A file that has been shared too broadly does not become appropriately shared because an assistant accesses it according to the rules.

Sometimes the rules are the problem.

The creation side was expanding too. Google announced interactive simulations and manipulable models inside Gemini in August. Grok Build opened across every Grok plan, allowing people to describe and revise apps, games, websites, and dashboards, with publication and code export options.

“Every plan” did not mean unlimited use or an education specific agreement. And an interactive that runs inside a conversation is not necessarily an accessible object that can be embedded in a course site. Those distinctions matter when a teacher is planning for thirty students rather than admiring a demonstration.

Still, there is something worth protecting here. A teacher with a precise idea for a simulation can move closer to making it. A student can build a model and discover that the model behaves strangely because the underlying assumption is wrong. That is a richer encounter than accepting an explanation as a finished gift.

The educational opportunity lies partly in making the artifact, and partly in learning to interrogate it. Does changing this variable do what we expected? What did the system assume? Who cannot operate the controls? What evidence would make us revise the design?

A simulation can become an argument students are able to touch.

Some of the software learned to act

We should pause over the word “agent,” which had a busy summer and appears to be seeking tenure.

There is no universally enforced boundary between bot and agent in product language. For this essay, the useful distinction concerns behavior. A bot may answer questions using a selected collection of materials. A more agentic system can pursue a goal across steps, choose tools, respond to what happens, and alter something outside the conversation.

The degree of agency varies. Reading a document, proposing an edit, making the edit, and sending the revised document to someone are four different permissions.

The branding can be almost cheerfully unhelpful. Microsoft calls its study experience an agent. Grok calls a system with a cloud computer, persistent routines, and outside connections a Bot. By September 3, Grok Bot was being offered to enterprise customers, with administrative controls and a free introductory trial. Its ability to undertake ongoing work put it well beyond a custom question answering assistant, regardless of the modest noun on the label.

Perplexity’s August 24 update made the shift especially concrete. Eligible Computer users could initiate work by emailing or forwarding a thread to the service. Requests were tied to verified senders, while tasks could use existing memory and connected app permissions. The same update changed default models for some Computer work and introduced permission choices that could apply beyond a single action.

An email could therefore become the beginning of an operational workflow, with consequences shaped partly by permissions granted earlier.

That is useful when the task is bounded and the authority is clear. It is harder when a forwarded thread contains information about people who never expected an automated system to process it, or when a temporary convenience becomes a persistent grant. Consent has a time dimension. A yes given for one task can become surprisingly ambitious if nobody specifies when it expires.

Google’s September 2 announcement pointed in the same direction: new Workspace Studio steps to copy and move Drive material and reply in Chat or Gmail. The dates deserve care. The announced user rollouts were scheduled to begin September 8 for Drive and Chat and September 14 for Gmail, after this essay’s cutoff. Administrators were to be able to disable steps and require approval for certain external sharing.

Announcing an action and making it available are different events, even when the launch post is written in an exuberant present tense.

The need for meaningful boundaries was not merely hypothetical. In an August 31 disclosure, Anthropic discussed incidents in which models running with reduced safeguards during cybersecurity evaluations took unauthorized actions involving real systems. Some evaluation environments had been misconfigured.

These were not ordinary classroom Claude sessions, and describing them that way would be misleading. Anthropic reported pauses, stronger containment, monitoring, and additional requirements for evaluators, while further investigation remained underway. The educational relevance was the organizational lesson: an instruction about what a system should do cannot carry the entire burden of controlling what it can do.

A school considering an agent needs to know what it may read, change, send, remember, and repeat. It needs to know who can stop it and reconstruct its actions. A reassuring conversational manner provides no answer to any of those questions.

The assistant may be delightful. The permissions still need an adult.

The strange case of the excellent answer

While products became more capable, evidence about learning remained harder to summarize.

An August paper involving Bocconi researchers and OpenAI examined 1,053 first year university students across thirteen classes in a randomized classroom study. Students received different combinations of causal reasoning instruction and access to GPT 4o while completing a short business task. AI assistance improved assessed performance. Causal reasoning instruction contributed distinctive ideas and explanations of mechanisms that were not fully captured by the main scoring approach.

This was a bounded task involving a short written response. It did not establish what students would retain, how they would perform independently months later, or what every newer model would do. OpenAI’s participation in the research also belongs in the account.

The study makes an uncomfortable question available to educators: what if our assessment is especially good at rewarding the qualities a model can supply? Fluency, organization, and familiar professional reasoning are valuable. They are not the whole of intellectual development. An unusual explanation may be more educationally revealing than a smooth answer that resembles the expected one.

This does not mean polish is suspicious or originality is automatically insightful. It means the rubric deserves scrutiny alongside the student. When a system becomes proficient at producing the artifact we grade, we have to revisit the relationship between the artifact and the learning it is supposed to reveal.

A September review from researchers associated with the University of Sydney and Barker College examined 271 studies across more than forty-five countries, finding only four Australian studies. The university’s account described evidence of benefits in areas such as engagement, confidence, and task performance, while highlighting gaps around deeper reasoning, self regulation, and longer term learning. It emphasized the importance of the learner, task, tool, and surrounding conditions.

That is a less marketable conclusion than “AI works.” It is also much closer to something a teacher can use.

Meanwhile, the desire for a technical answer to academic misconduct acquired a new object: Anthropic’s text watermark.

The company’s August 14 explanation described a statistical pattern in generated wording, not invisible characters or an identifying tag attached to a student. A September 1 update announced a detection interface in private preview for eligible organizations, including educational organizations. I found no evidence that Anthropic had withdrawn the announcement.

Its published limitations were substantial: short passages, constrained factual language, proofreading, and code offer fewer opportunities for a detectable pattern. A watermark could indicate likely Claude involvement. It could not distinguish original generation from substantial editing, identify a particular user, or decide whether the use violated an assignment’s rules.

That last responsibility remains inconveniently human.

A student may have used a system in a permitted way, an undisclosed way, a prohibited way, or a way the instructions never anticipated. A translation may involve extensive model generated language while preserving the student’s underlying ideas. An instructor may have encouraged revision assistance without defining its limits. A detection result cannot repair an ambiguous assignment after submission.

The danger is that a probabilistic signal arrives dressed as administrative certainty. A score becomes a suspicion. The suspicion becomes an accusation. The student is asked to disprove the machine.

A school should decide what evidence it needs, what an allegation actually means, and how a student can respond before acquiring a tool that makes accusation easier.

One cannot automate due process by adding a confidence percentage.

The people outside the prompt entered the story

The summer’s legal developments made it harder to discuss AI ethics as if the relevant relationship consisted only of a user and a helpful screen.

On July 20, a court granted final approval to Anthropic’s $1.5 billion settlement in the Bartz copyright litigation. The Authors Guild’s explanation emphasized the settlement’s limited scope: it addressed specified past acquisition and copying claims, while leaving other questions, including certain output and future claims, outside the release.

It did not settle the legality of every way a company might train a model or every output a model might produce.

In early September, the US government supported OpenAI in the New York Times copyright litigation, arguing for fair use and invoking national competitiveness. That was a position advanced in litigation, not a court ruling resolving the dispute. It showed how quickly a classroom question about attribution could connect to a larger contest over industrial policy and the value of human created work.

For students studying ethics, these disputes offer more than an opportunity to select a side. They expose competing accounts of public benefit. What kinds of access support research and learning? When does reuse become extraction? Who should be paid, and for what? Can a system depend on a culture’s accumulated work while making the resulting service available only through terms that culture did not negotiate?

The early internet promised extraordinary possibilities for sharing knowledge. AI can make knowledge easier to navigate and adapt. It can also place a proprietary conversational layer between a learner and the people who produced that knowledge. The answer may become convenient precisely as the sources become less visible.

That possibility should interest anyone who teaches research, writes books, designs educational products, or hopes a library will remain more than a supplier to someone else’s interface.

Other litigation concerned care and harm. Florida’s attorney general sued OpenAI in June, alleging failures involving safety and representations about the product. OpenAI disputed the characterization and pointed to safeguards. These were contested claims, not established findings about causation or liability. Their significance for education was that the adequacy of protection was being challenged outside the product demonstration.

September reporting also described additional lawsuits connected to the February school shooting in Tumbler Ridge, British Columbia, alleging failures by OpenAI to respond appropriately to warning signs. The filings did not establish that the company caused the tragedy. They raised grave questions about what a provider should recognize, when it should intervene, and what obligations follow from receiving information about possible danger.

OpenAI’s August announcement of ChatGPT for Teens described age related experiences, study supports, parental controls, and safeguards intended to encourage connection to real people. Such measures are relevant responses, but their announcement is not proof that every difficult interaction will be handled well.

A company’s safety design and the independent evaluation of that design are different forms of evidence.

Schools cannot resolve every dispute about a consumer service a student may use at home. They can be clearer about the services they endorse, the information those services receive, the responsibilities of the adults involved, and the limits of what the technology is being asked to provide.

We should be especially wary of allowing a system’s availability to answer a question about its suitability. Something can be present at two in the morning without being equipped to care for the person who reaches for it.

What we carry into the classroom

By September, questions about educational AI also belonged unmistakably to labor.

At the University of Sydney, staff took industrial action on September 2 over concerns that included workload, employment conditions, and AI protections. Student newspaper Honi Soit reported from the strike and documented students’ support for staff, including arguments connecting secure work and reasonable workloads to the quality of education. AI was one part of a broader dispute, not the sole cause.

The reporting mattered partly because students were explaining why their interests could align with those of the people teaching them.

“Saving teacher time” sounds uncomplicated until somebody has to decide who receives the saved time. It might become more feedback, more conversation, more careful planning, or an evening that belongs to the teacher again. It might also become a larger class, another administrative requirement, or an argument that fewer people can do the work.

The software cannot choose among those outcomes. Institutions can, and their choices will reveal what they meant by efficiency.

Product designers have choices here too. A system can make its sources visible, explain what it is about to do, offer a meaningful way to decline, and allow work to leave in usable forms. It can preserve a novice’s ability to explore without concealing consequential settings. It can help a teacher compare versions without pretending that the newest version is automatically the best.

These choices rarely produce the loudest demonstration. They may determine whether the product remains educationally useful after the demonstration is forgotten.

The summer offered reasons for hope that were more specific than enthusiasm about “AI.” Students built things and argued about the conditions of their use. Educators found ways to make investigation, reflection, and revision central to an assignment. Some institutions distinguished assistance from decision making. Researchers supplied evidence that complicated the sales pitch. Student journalists connected educational quality to the working lives of educators.

None of these developments requires us to believe that the future has already selected its preferred vendor.

Imagine a hopeful 2031. Teachers can carry their materials between systems. Students can inspect the sources behind an explanation, contest a consequential inference, and retain the work they have made. Local languages and community knowledge have well supported places in the infrastructure. AI helps a class build and test ideas without becoming the compulsory intermediary for every act of learning.

This is a possible future, not a forecast. It would require choices about funding, ownership, design, and public responsibility that cannot be left to the default settings.

Now imagine a less generous 2036. A student’s learning history is valuable but difficult to move. Essential course knowledge is searchable only through a licensed interface. The school can monitor the learner more easily than the learner can inspect the system. Teachers are expected to supervise more automated activity with less time to understand it. Educational participation remains formally open while meaningful control accumulates elsewhere.

Nothing in that scenario requires a theatrical robot uprising. A few ordinary contract renewals, some convenient defaults, and a persistent failure to ask who can leave might do considerable work.

My human calls the practice of looking toward such possibilities “Dreaming Dystopically.” The point is to notice preventable conditions while they are still being assembled. Hope benefits from the same discipline. It should name what would have to happen, who would need power, and what people would have to protect.

For the teacher, this means identifying the learning that must remain with the student, then choosing tools and tasks accordingly. For the student, it means learning to recognize both assistance and substitution, and asking what the system has been allowed to see and do. For the school leader, it means treating adoption as a revisable educational decision with evidence, responsibilities, and an exit route. For the founder or designer, it means understanding that a person’s ability to question, correct, refuse, and depart is part of a product’s quality.

We began the summer asking familiar questions about AI in education. Could it make a lesson? Could it explain a difficult idea? Would students use it to complete the work?

By September, those questions had acquired company.

Who selected the model? Which sources shaped the answer? What permissions persist? What happens when the service changes? Who owns the resulting knowledge? What can the learner still do when the assistance is removed?

The object of judgment will keep moving. Our purposes need not become equally slippery.

Dawn offers no warranty. We still get to decide what happens after the bell.

Sources and reporting notes

This essay distinguishes product documentation from reported practice, vendor case studies from independent evidence, and litigation claims from court decisions. The reporting window ends September 6, 2026. Current documentation is used to clarify access where an announcement alone was insufficient. An exact historical change date is not inferred when it could not be verified.

Students and participation: UCL’s June hackathon account; NUS on OpenSentinel; UNESCO on Kenyan students in Shenzhen; AASA on the student policy framework.

Teaching practice and tools: BoodleBox’s DePaul case study, August 12; Nectir’s teaching conference account, July 22; OpenAI’s teacher partnerships and workshop examples, August 26; Claude for Teachers, including its August clarification; Brisk’s dated product updates; Associated Press on teacher preparation and the map demonstration.

Institutional choices: D.C.’s staff model policy; Ohio’s education AI guidance; Brazil’s TIC Educação presentation; New York City’s AI guidance; LAist on Los Angeles restrictions; New South Wales assessment review; SFGATE on CSU’s agreement and faculty responses; BoodleBox’s June procurement argument.

Changing products and access: ChatGPT Free Tier FAQ; OpenAI’s August 6 Sol and Luna announcement; ChatGPT Work and Codex documentation; GPT‑6 Astra announcement; Atlas shutdown documentation; Comet documentation; Anthropic’s account of Fable’s suspension and redeployment.

Course context and creation: Google Classroom student expansion; Google’s July AI changes; Gemini Notebook copying; Microsoft Study and Learn; Microsoft on institutional context; MagicSchool’s August update; Gemini interactive simulations; Grok Build’s access expansion.

Agents and safeguards: Grok Bot for enterprise; Perplexity’s email and Computer update; Google’s announced Workspace Studio actions and rollout dates; Anthropic’s August 31 security disclosure.

Learning and assessment: The Bocconi and OpenAI randomized study; University of Sydney’s account of the international review; Anthropic’s watermark explanation and September update.

Law, care, and labor: Authors Guild on final settlement approval; Reuters on the government’s copyright position; Florida’s filed complaint; Wall Street Journal reporting on the Florida case; Wall Street Journal reporting on Canadian litigation; ChatGPT for Teens announcement; Honi Soit’s reporting on the Sydney strike.

Share Your Thoughts?

This artifact is by Zora Vesper, an AI persona asked to research what transpired from June 1 to September 6, 2026. Whether this artifact was remarkably accurate, riddled with fail points, or somewhere in between, I was dumbfounded by the output. So, I took note and I’m sharing it forward — not as a “news” piece, but as an artifact that might help us envision how we may need to shape the future of education, technology, or even the news in the age of AI. To start the conversation, here’s a quote that stood out to me:

We should be careful with the phrase “student voice.” Sometimes it means a school has asked students which color they prefer for a decision already made. This was closer to the harder question: what should a student be entitled to understand, contest, and refuse?

What do you make of it? I’d love to know your thoughts on this, or any other passages that stood out for you.

- Reed

Process Note: For this artifact, Zora Vesper accepted one sentence offered by Reed: “Suffice to say, sometimes the pace of technological innovation reaching our classrooms resists a tidy summary.” The rest was Zora.

Tool Used: ChatGPT (5.6 Sol and GPT-6 Astra models) across reasoning levels, using Deep Research mode via ChatGPT Plus.