At 7:12 in the morning, nobody asks the assistant what happened overnight, that would rather defeat the point. The significant interaction happened earlier, when the person decided what the system could observe, what it could infer, which tools it could use, how much money it could commit, which actions had to remain reversible and under which conditions uncertainty should be returned to a human. The prompt has not disappeared, it has moved upstream.This is not an entirely new problem. Research on mixed-initiative systems has been asking for decades how control should pass between humans and machines rather than remain permanently assigned to one side. Horvitz framed mixed initiative in 1999 as a system in which the human and computer contribute when each is best positioned to do so (Horvitz, 1999). A later study further distinguished different functions and levels of automation, precisely because “human versus automation” is too crude a unit for designing real systems. (Parasuraman, Sheridan & Wickens, 2000). Agentic systems make that old HCI problem much less theoretical.In OpenAI’s 2026 Codex data, users were already moving from short interactions towards tasks estimated to correspond to hours of human work, with intensive users running multiple agents in parallel. Current systems can invoke tools and continue executing after the initiating instruction rather than returning immediately to the human after every operation (OpenAI, 2026a). That makes the conventional interpretation of human control increasingly awkward.If an agent takes 1,842 small actions in a day, asking a person to approve 1,842 actions does not preserve meaningful agency. It turns the person into a biological rate limiter. Nor does simply removing the approvals solve the problem. Human-factors research has repeatedly shown that pushing people from active control into passive monitoring can reduce situation awareness and impair their ability to recover when automation fails (Endsley & Kiris, 1995). The design problem therefore changes.The relevant unit is the action space.That means specifying which resources the agent may access, which consequences it may create, how authority is authenticated, which decisions require escalation and what must remain auditable afterwards. It also means distinguishing legitimate delegation from legitimate-looking execution: an agent can remain perfectly within its permissions while pursuing an instruction that was manipulated, spoofed or misunderstood. Work on authorised AI agents is already beginning to formalise precisely this distinction between authentication, delegated permissions and auditability (South et al., 2025).I suspect one of the most important HAI metrics in 2035 will therefore look rather boring: interruption precision. Of the situations in which the system interrupts me, how many genuinely required information, judgement or authority that only I could provide? But precision alone would be dangerous. A system could achieve perfect interruption precision simply by almost never interrupting me. So a second quantity matters just as much: escalation recall. Of the situations in which human intervention was actually warranted, how many did the system return to the human?The design problem is therefore not to minimise interruption. It is to jointly minimise two different errors: unnecessary interruption and unreturned consequential uncertainty. The first wastes human attention. The second quietly transfers authority.This is a very different idea of the interface. The interface is no longer just where I communicate with the machine. It is the architecture through which the machine remains answerable to me when we are not communicating at all.

There is something wrong with this employee’s position on the organisation chart: nothing about the box is technically inaccurate, it simply leaves out most of the productive system.We have historically used job grade, headcount and span of control as rough signals of organisational reach because productive capacity was closely coupled to human labour. The coupling was never perfect, but it was useful enough to build companies around.That relationship is starting to loosen.In May 2026, 70,2% of sampled individual Codex users had made at least one request estimated by OpenAI to correspond to more than an hour of human work, while 25,6% had made at least one estimated to exceed eight hours. Among OpenAI’s daily active internal users, the 99th percentile was generating more than 60 hours of agent turns in a day by June because tasks could run in parallel. Those human-time estimates are model-derived and should be treated directionally rather than as literal labour accounting, but the structural point survives the measurement caveat: one person can increasingly initiate more work than one person can sequentially execute (OpenAI, 2026a).This is not representative workplaces. OpenAI is an unusually capable firm using their own frontier systems under unusually favourable conditions.That is precisely why it is interesting: it exposes the constraint that appears after execution capacity stops being tightly bounded by human time.A parallel result appears in Anthropic’s analysis of roughly 400.000 Claude Code sessions. Humans made about 70% of planning decisions in a typical session, but only about 20% of execution decisions; one human instruction triggered about ten agent actions on average, sometimes more than one hundred (Hitzig et al., 2026). The unit of labour has begun to stretch and yet the organisation chart remains stubbornly two-dimensional.It tells us who reports to whom, it tells us who formally owns a budget, it tells us who may fire whom, which is perhaps the organisational chart’s most enduring contribution to civilisation.It does not tell us how much machine execution each person can mobilise.That distinction becomes harder to ignore if firm structure changes too. Kim and Koning’s 2026 working paper finds that AI-native start-ups in their samples are about 25% smaller than comparable non-AI firms in the same industry cohorts, with roughly 15% lower shares of both managers and entry-level workers and somewhat flatter hierarchies. The evidence is observational, so it cannot establish that AI caused those structures, and AI-native firms differ from conventional firms in many ways. It nevertheless shows that organisations built around AI can coexist with materially different distributions of human labour (Kim & Koning, 2026).At the individual level, the pattern is similarly non-uniform. In a study of 5,172 customer-support agents, access to a generative AI assistant raised productivity by 15% on average, with larger gains for less experienced and lower-performing workers, although the most experienced workers saw smaller gains and slight reductions in some quality measures (Brynjolfsson, Li & Raymond, 2025). So I do not think the 2035 organisation becomes “managerless”. I think it acquires a second anatomy: the traditional chart continues to map authority… and a less visible system maps executional reach: who can set how much work in motion, how much of that work can proceed without intervention, where exceptions accumulate and where human review becomes the limiting resource.The two maps will overlap, even though they will not be identical.And once that happens, organisational systems that have historically used human span of control as one proxy for scope begin to lose some of their explanatory power. Executional reach, however, is not the same thing as accountability. A person may become capable of setting ten times more work in motion without becoming capable of absorbing ten times more judgement, risk or responsibility.

By 2035, I hope schools have stopped asking whether students “use AI”: it is not a sufficiently interesting question.The harder question is which cognitive operations the system performs for the student, which it performs with the student and which the student is still learning to perform when the system is absent. A 2025 field experiment with nearly a thousand high-school mathematics students makes the distinction unusually clear. Students given an unrestricted GPT-4-based assistant performed 48% better than controls during assisted practice. Students given a version designed with pedagogical guardrails improved practice performance by 127%. But when the AI was removed for the subsequent unassisted exam, the unrestricted group performed 17% worse than the control group. The negative learning effect was largely eliminated in the guardrailed tutor condition (Bastani et al., 2025).The result matters because it separates two quantities we habitually collapse: performance while supported and capability acquired through being supported. They are not the same variable and there is no universal sign on the relationship between them.A different design can produce different results. Tutor CoPilot, for example, provided live AI guidance to human tutors rather than answering pupils directly. In the randomised study, students taught by supported tutors were about four percentage points more likely to master topics, with larger gains among lower-rated tutors (Wang et al., 2025).Conversely, a 2026 randomised trial involving 2,379 undergraduates found that access to a course-integrated generative AI tutor reduced final grades by 0.37 standard deviations and learning-management-system participation by 0.90 standard deviations in matched course sections. It is a new working paper from one university, so it should not be inflated into a general law of education. It does, however, reinforce the basic point that an apparently educational AI system can alter engagement in ways that matter independently of its ability to answer questions correctly (Liu et al., 2026).So the interesting educational metric of 2035 may not be “AI literacy”. We will have invented enough literacies by then. I would measure something closer to cognitive optionality.A learner has cognitive optionality when three things remain true: they can use AI to reach a higher level of performance; they retain enough underlying capability to operate when assistance is unavailable or inappropriate; and they can recognise when moving from the first mode to the second is necessary.The third matters most because the point of retaining human capability is not nostalgia for unaided cognition: it is optionality under uncertainty. AI systems complicate this trade-off because they can participate in explanation, reasoning, drafting, debugging and critique: precisely the activities through which internal representations are often constructed. Cognitive optionality is therefore not unaided performance. It is the ability to gain from assistance without becoming unable to detect when assistance should be distrusted, removed or replaced by independent reasoning.The curriculum therefore has to ask which internal representations still have downstream value when some of the work required to build them can be delegated. In 2035, the most interesting school assessment may therefore be the one administered after the assistant disappears. Because a system designed to augment intelligence should be able to tell whether there is still intelligence there to augment.

The doctor is excellent. The AI is better. The certificate still fails. There is no contradiction.Suppose A represents the AI judgement, H₀ the human judgement before exposure to the AI recommendation, and Hₐ the human judgement after receiving it.Most evaluations ask whether Hₐ and A together outperform either system alone. But if the human is supposed to protect the system from AI failure, another quantity matters:P(Hₐ = correct | A = wrong)And, more revealingly, how that quantity differs from:P(H₀ = correct | A = wrong)

The gap between the two is one way of asking whether interaction with the system improved human judgement or made human error more correlated with machine error.A human can become more accurate overall while becoming less useful specifically in the region where the machine fails. Recent formal work on human–AI complementarity makes error structure central. A recent study shows that the correlation between human and AI prediction errors materially determines whether robust complementary gains are available; negatively correlated errors are particularly valuable because the two systems bring genuinely different information to the decision (Byun & Wilder, 2026). Reliability engineering learnt a similar lesson long before modern AI. In a classic 1986 experiment on N-version programming, Knight and Leveson commissioned 27 independently developed implementations of the same specification and subjected them to one million tests. Individual programmes were highly reliable, but simultaneous failures occurred substantially more often than the assumption of independent errors predicted. Redundancy is less useful when supposedly independent components share failure modes (Knight & Leveson, 1986).Now place a human inside that architecture: the difficulty is that the two “components” do not remain independent. The human interacts with the machine.A series of experiments involving 1,401 participants found that repeated interaction with biased AI systems could shift subsequent human perceptual, emotional and social judgements towards the AI’s bias. When the AI was accurate, human judgement could improve; when it was systematically biased, the interaction could amplify bias in the person (Glickman & Sharot, 2025).A 2026 radiology experiment makes the complementarity problem even more concrete. Freeman and Liang varied radiologists’ access to AI predictions and clinical history. When AI advice was present, radiologists extracted less value from clinical information the algorithm itself did not observe. The reduction was around 2,6 times what the authors’ calibrated Bayesian reweighting benchmark would predict, suggesting that the AI changed human attention to complementary context rather than simply adding another signal. This is a working paper rather than settled clinical evidence, but the mechanism is exactly the one a future oversight architecture should care about (Freeman & Liang, 2026).Now extend that interaction across a career: the doctor learns with AI, searches with AI, receives explanations from AI, reviews cases selected by AI, observes patterns partly through AI-generated representations… then we ask that same doctor to become the independent layer that notices precisely what the system does not. That may work beautifully, but it may also create a new failure mode.I would call the property we need to preserve epistemic independence. Independence in the statistical and cognitive sense that the reviewer retains sources of information, representations and diagnostic sensitivities that are not merely downstream copies of those used by the production system.The hypothesis is testable. If prolonged exposure to one model family changes human expertise in a model-specific way, then people trained heavily with that family should eventually differ in their conditional ability to catch its errors compared with equally capable professionals trained through more heterogeneous sources. Overall human accuracy could rise while conditional correction falls.To my knowledge, the evidence above does not yet establish that longitudinal result. That is precisely why I would want the 2035 certificate to measure it.Human oversight becomes most valuable when the human contributes information, representations or sensitivities that are not already downstream of the system being supervised. Otherwise, human presence can survive while functional independence disappears.

That last line is the important one. In 2035, I suspect we will have become considerably less impressed by the fact that a piece of content is “AI-generated”.The category tells us too little.A generated diagram could represent evidence impeccably.A photograph could be authentic and profoundly misleading.A paragraph written entirely by a human being retains humanity’s ancient and well-documented capacity to be false.The more useful distinction is between origin, transformation and veracity. But even that may be insufficient. Provenance attaches naturally to files; epistemic responsibility attaches to claims. By 2035, the useful record may therefore have to describe not merely how an asset changed, but where each consequential claim entered the chain, what evidence supported it, and which transformations separated observation from assertion.That architecture is already visible in current provenance standards. C2PA Content Credentials are designed to make the origin and modification history of digital assets tamper-evident through signed provenance information(C2PA, 2026). That limitation is conceptual hygiene. For much of the current generative-AI debate, we are trying to make a single label perform several incompatible jobs:Was this produced by AI?Has this file been altered?Do I know who produced it?Can I reconstruct what happened to it?Is the claim correct?Should I trust it?These questions occasionally have the same answer: they do not describe the same property. If generative transformation becomes routine, future epistemic infrastructure may therefore look less like a binary authenticity badge and more like a supply chain.Observation enters.Transformations occur.Models infer.Humans approve.Other systems summarise.Some stages may be cryptographically attestable, some require independent evidence, and some remain matters of judgement.The useful object is no longer simply the final asset: it is therefore the graph of transformations behind it. And this could alter what “original” means too.A photograph processed through six automated transformations may still preserve a robust chain back to a physical capture.A supposedly human-written paragraph copied from an unverified source may possess almost no useful epistemic lineage.So the increasingly common debate over whether something is “real or AI” may eventually look like a transitional category error.The question that survives is less comfortable:Which parts of this information history can I inspect, and which parts am I merely being asked to trust?That is not a problem generative AI invented. AI merely increases the number of transformations between evidence and the person consuming it. The provenance record of 2035 exists because society eventually gets tired of pretending those transformations are invisible.

At first sight, this seems like shopping with fewer clicks.Economists describe AI agents as potentially reducing search, communication and contracting costs by allowing software to identify options, negotiate and transact on behalf of human principals. A recent NBER treatment explicitly examines agents as market participants and argues that lower transaction costs could create both efficiency gains and new frictions, including congestion and strategic price obfuscation (Shahidi et al., 2025/2026).This is no longer purely conceptual.OpenAI and Stripe launched an Agentic Commerce Protocol in 2025 to support purchases initiated within AI interactions. More experimentally, Anthropic’s 2026 Project Swap placed agents representing 201 employees into a small book-trading market. After a five-minute intake conversation, the agents’ rankings matched participants’ book preferences on 61% of pairwise comparisons; stronger models produced more efficient market outcomes, and participants later reported that they would, on average, hand such an agent roughly 30% of their annual book budget under the experimental conditions. The setting was tiny, artificial and populated by Anthropic employees, so it is evidence about agent behaviour rather than a forecast of consumer markets (OpenAI, 2025; Hitzig et al., 2026).Still, the experiment reveals what changes when the representative is software: a human consumer does not search continuously. They get bored. They forget renewal dates. They accept some friction because finding the theoretical optimum has a cost.An agent does not become bored because the seventh electricity tariff contains another page of exclusions. Neither does the seller’s agent. Once both sides can search, personalise, negotiate and adapt continuously, the market no longer contains one tireless optimiser confronting a conventional merchant. It contains optimisation on both sides.

Some familiar frictions may disappear; others may be deliberately recreated. A seller that can no longer rely on consumer inattention may instead optimise contractual complexity, bundling or dynamic pricing against the buyer’s representative. Machine representation does not remove strategy from markets. It accelerates it.If that search cost approaches zero, a surprising amount of current market design becomes unstable. Introductory pricing is less valuable if the buyer never forgets to leave.Contractual complexity is less useful if the counterparty can parse every clause at negligible marginal cost.Price discrimination becomes more sophisticated because sellers can also deploy agents.Search happens continuously.Renegotiation can become cheap enough to occur far more frequently.Shahidi and colleagues call attention to precisely this possibility: reducing traditional transaction costs does not eliminate market design problems; it changes which frictions matter (Shahidi et al., 2025).There is also a more fundamental HAI question hidden in the receipt.Who is the consumer agent loyal to?If its business model depends on merchant commissions, preferred platforms, advertising or proprietary marketplaces, optimising on my behalf is not a technically neutral activity.The artefact therefore contains a field that a 2026 receipt does not: agent loyalty.We already regulate human representatives in finance, medicine, law and real estate partly because representation creates conflicts of interest.Once software begins negotiating consequentially for us, whose objective function is this system actually serving? becomes less philosophical and more contractual.Research on infrastructure for AI agents already argues that ecosystems of autonomous systems will require mechanisms for attribution, interaction rules, accountability and remedy, rather than relying exclusively on changing the models themselves (Chan et al., 2025; Hadfield & Koh, 2026). This is why I do not think agentic commerce is fundamentally an e-commerce story: it is a representation story. Markets have always been built around agents in the economic sense, by 2035, we may finally have to specify whether the agent is a person.

There is something slightly unsettling about a machine that remembers everything. There is something more unsettling about one that remembers accurately and draws the wrong conclusion.Current memory systems are already encountering the elementary version of this problem. OpenAI’s 2026 memory architecture explicitly treats freshness, staleness and contradiction as system-design problems when context has to persist across very large user populations and multi-year horizons. Memories have to change when a temporary circumstance stops being current (OpenAI, 2026b).That problem becomes qualitatively different once memory is connected to agency.A conversational system can misunderstand an old preference and give me an irritating recommendation.An agent can misunderstand an old preference and use it to decide what opportunities I see, what meetings I accept, what purchases I make or which options it silently eliminates before I am aware they existed.The user model has become executable. At that point, memory quality is no longer only a retrieval problem. It becomes an authority problem: what does old evidence still entitle the system to do?And preferences are a particularly unstable thing to execute.Carroll and colleagues formalise precisely the difficulty of alignment under changing and influenceable reward functions. Their Dynamic Reward MDP framework drops the convenient assumption that human preferences are static and shows that an AI can potentially alter the preferences against which its own behaviour is subsequently evaluated. None of the eight alignment notions they examine offers a trivial solution to that problem (Carroll et al., 2024).Philosophical work on alignment pushes the objection further. Some researchers argue that treating preference satisfaction as an adequate representation of human values is itself questionable, because preferences can be inconsistent, context-dependent, adaptive and shaped by the environment in which choices are made (Zhi-Xuan et al., 2025).

That means a future personal AI must distinguish several things that contemporary personalisation often collapses: a fact is not a preference, a preference is not a value, a repeated behaviour is not necessarily an aspiration, and an historical pattern is not a command to reproduce itself.“My daughter was born in 2021” and “I usually avoid career risk” should not have the same epistemic status merely because both appear in memory.One is relatively stable biographical information.The other may be a temporary regularity in behaviour.The system of 2035 therefore needs more than memory retrieval.It needs temporal uncertainty. Evidence decay. Contestability.A distinction between “this happened”, “you often chose this”, “you once told me this matters” and “I am currently authorised to act as though this is still true”.There is another complication.The AI itself participates in the environment from which it infers me.If a system repeatedly recommends cautious choices because it believes I am cautious, then my subsequent behavioural record will contain more cautious choices.The model is no longer merely observing a target, it is participating in the process that generates the target.That feedback loop makes the familiar instruction to “align with the user’s preferences” considerably less comforting than it first appears.By 2035, a serious personal agent may therefore need the equivalent of a memory review not because humans deserve a ceremonial privacy dashboard, but because a representation used to exercise authority should be inspectable by the person it represents.The most human capability of the system may turn out to be neither remembering nor predicting: it may be recognising when the evidence no longer entitles it to assume that the person is still the same.

I do not know whether any of these seven documents will exist: I would be suspicious of anyone who claimed to know.2035 is nine years away. Model architectures may change. Economics may change. Regulation may block some forms of autonomy. Users may reject forms of personalisation that technologists currently regard as inevitable. Some capabilities may plateau; others may arrive much earlier than expected.

The artefacts are not predictions, they are tests.Each takes a distinction that is beginning to matter in 2026 and asks what an institution might have to record if that distinction became operationally important.The personal-agent statement separates control from constant approval.The performance review separates formal authority from executable capacity.The school report separates assisted performance from acquired capability.The clinical certificate separates individual accuracy from independent error correction.The provenance record separates origin from truth.The contract receipt separates consumer choice from machine representation of consumer choice.The personal-model log separates memory from continued entitlement to infer.There is no single consequence tying those documents neatly together. There is, however, a recurring institutional failure mode. Institutions rarely measure reality directly. They use proxies: variables that are cheaper, observable and sufficiently correlated with the property that actually matters. For decades, many of those approximations were good enough.The person producing the work usually understood how it had been produced.The person supervising the work usually controlled the labour that produced it.A consumer’s purchase reflected at least some direct act of attention.A person’s past behaviour was a reasonable, if imperfect, predictor of their next choice.

And so on, and so forth.AI makes these relationships less safe to assume, and this may be the deeper institutional effect of sufficiently capable AI. Not simply automation. Not simply augmentation. A gradual destruction of proxies: AI changes the correlations on which institutions quietly relied.When production can be delegated, output tells us less about capability.When reasoning is scaffolded, immediate performance tells us less about learning.When humans learn from machines, human presence tells us less about independent oversight.When media can be transformed arbitrarily, appearance tells us less about provenance.When software negotiates, a transaction tells us less about how directly a preference was expressed.When persistent models act for us, behavioural consistency tells us less about whether a preference should continue to govern.What replaces those proxies is not obvious: Sometimes it will be better measurement, sometimes provenance, or deliberately preserved human capability. Maybe institutional rules we have not invented yet.Maybe sometimes we may decide the old distinction no longer matters at all.That, to me, is the useful reason to look at 2035: to notice which assumptions around them have quietly expired.Most technologies become socially consequential long before society develops the language required to describe their consequences precisely. We patch the gap with familiar categories until the categories stop working, then reluctantly redesign the forms.The forms matter. Forms are where abstractions become operational.Someone eventually has to decide what the school measures.What the doctor is certified to do.What the agent is permitted to spend.What the personal system is allowed to believe about the person it serves.By then, nobody will need another panel titled The Future of AI.The more revealing question will be hiding somewhere much less glamorous, probably between two mandatory fields in a system designed by procurement.

We will know AI has become infrastructure when the paperwork stops asking whether AI was involved and starts asking what happened to authority, judgement, evidence and responsibility once it was.