# The Personal AI
# Book: Language
# Chapter URL: https://programmer.ie/books/language/29-chapter/
# Prose fingerprint (sha256, normalized): 98e2abde2d1a7f7d19b5f3d228ddddb1ea0eb52fa36baa55327bccb0be58a73a
# Snapshot generated at build time by Hugo. This is the complete rendered prose
# of the chapter; it is not a transcript of the source file.

At twenty past nine, Iris begins an architecture review. The channel already holds a standing policy that Iris declared, observations from earlier reviews, and a provisional inference about derivations. It can also create a temporary view and apply two narrowly authorised boundaries. All of these mechanisms now coexist, but coexistence is not a reason to consult all of them. This is a constructed session, not a report of a person using the complete system.
Iris says: I have ten minutes. In this review view, show material that bears on the decision. Keep the decisive equations and assumptions visible; fold the supporting steps, with the full calculation available.
Before retrieving the accumulated personal state, the system has a prior question: which part of it is needed to satisfy the request Iris has just made? If it retrieved everything first and decided afterwards what to ignore, it would already expose more personal information to the current computation than the task may need.
Chapter 28 made consequential beliefs inspectable and correctable. That does not establish that more of them should exist. The closing question of Part III is therefore:
How much personal state does this communication actually require, and what must an additional item justify?
more personal state ≠ better communication The title The Personal AI names the temptation as well as the possibility. Several narrowly scoped mechanisms may produce a coherent experience without a comprehensive portrait of Iris. The central design question is not whether the system can accumulate a personal model. It is whether each additional claim improves an authorised communication decision enough to justify collecting, retaining and using it.
Six available inputs, not six necessities For this constructed session, six mechanisms or records are available. They are deliberately not six interchangeable facts about Iris. Some are evidence, some are instructions, and some are temporary decisions or controls. Treating them as one collection of preferences would lose the distinctions of origin and authority that Part III has developed throughout.
DECLARED POLICY — standing, task-scoped Architecture reviews: preserve decision-relevant technical detail and expand derivations where the decision depends on them. OBSERVATION — historical, condition-labelled Iris expanded derivations in previous architecture reviews; those sessions' task conditions remain part of the evidence. INFERENCE — provisional, separately authorised for use Expanded derivations might help in similar review conditions; the original evidence does not justify a global preference. SESSION OVERLAY — derived, ends with this review Keep decisive material visible; fold only supporting steps. FRICTION CONTRACT — separately accepted intervention Remind before opening another paper; allow immediate override. PORTABLE BOUNDARY — separately accepted service grant Temporarily filter the specified recommendation feeds; preserve direct access and report each destination's status. The ten minutes are current context, not durable knowledge about Iris. The current request authorises the particular presentation choices it names. The overlay records how those choices were applied within this bounded period. It does not create new permission. A time limit constrains the result, but it cannot decide which assumption, equation or source passage may safely disappear.
Iris separately accepts the reminder for this review. She also separately authorises holding consensus-research recommendations out of selected service feeds, while preserving direct opening and an easy local override. Both grants end when the review or session ends, or at 09:30, whichever comes first. The fact that all three activities occur within ten minutes does not merge their permissions. A request for a brief view is not a grant to control navigation or other services.
The portable rule also carries a prohibition on silently turning behaviour shaped by the temporary filter into a lasting preference. This is a condition on the destination’s implementation. It is not a promise a mediator can make on the destination’s behalf. If a service cannot honour it, the adapter must report the limitation rather than describe the boundary as faithfully enforced.
Availability does not mean admission to the model’s context. A rendering decision may need a currently applicable policy and the current task. It does not need previous session transcripts or inferred traits. Each mechanism should receive only its authorised inputs. The account of what actually influenced the result must stay inspectable, without displaying all retained history by default.
What contributes to the view? Three items arrive in this illustrative review: a library release mentioned in the architecture document, a consensus-paper recommendation in a selected feed, and a colleague’s note containing a calculation that Iris needs to check. Arrival makes each item a candidate. It does not make any of them automatically trustworthy evidence, or a reason to interrupt. Source identity, permission and the current task still govern what the channel may do with each.
The colleague’s note bears directly on the current decision. A proposed view keeps the decisive equation, the failure assumptions and their source links in the main explanation. Only the supporting derivation steps are folded, and the full calculation remains recoverable. This follows from the current request together with obligations to preserve the source. The standing policy is compatible with this, but it does not establish that the note or calculation is correct. Iris must still be able to inspect the evidence.
The ten-minute condition does not identify what is dispensable. If a step described as supporting actually changes the result, it belongs among the decisive material. The channel must restore it, or say that the requested degree of compression cannot be delivered faithfully. The output must not claim both to have met the ten-minute brevity request and to have preserved an unchecked dependency simply because the renderer labelled it optional.
The library release is relevant to the architecture document, but the constructed task does not require it in order to check the calculation. Iris can reach it outside the main explanation, and the system does not need to consult observations about her library interests to do so. This is a task-based selection within the illustrated design, not a report of the scripted run described below. In that separate run, an optional inference does influence whether the library release is folded. The distinction matters when assessing the value of subtraction.
The consensus recommendation is governed by the separately accepted portable boundary, not by the review’s presentation preference. Suppose the video service supports the exact temporary feed filter, but the news service offers only a broader control that would also restrict direct research access. The first service may report a bounded, checked application. The second must remain unapplied and report its limitation. A recommendation may therefore stay visible in the news service. Neither the model’s explanation of this view nor the adapter’s receipt establishes protection across every destination.
If Iris chooses to open another paper, the reminder applies only if its contract is still active. She may override it immediately. The system must not strengthen the intervention because the review is unfinished, infer that she is distracted, or let an expired reminder persist in a cached view. One authorised choice cannot silently expand the others.
The proposed explanation of the calculation’s presentation is short:
Why this view? — proposed decision account Source: colleague's calculation and linked assumptions Preservation: decisive equation and qualifications kept visible Current request: fold supporting steps for this review only Standing rule: preserve decision-relevant technical detail Observation: not consulted for this presentation choice Inference: not used for this presentation choice Other controls: reminder and feed boundary handled separately Evidence status: correctness of the calculation not established The key distinction is between state the system possesses and state that actually influenced this decision. A trustworthy implementation would derive the displayed account from a recorded decision trail. A fluent explanation assembled after rendering is not evidence that the stated inputs were the ones used.
Try subtracting before adding Now exclude the earlier observation and inference from the constructed view. Its stated basis remains the source’s preservation obligations, Iris’s current request and an applicable declared policy. Those two optional records are not necessary to explain this specific presentation. The word constructed matters here. Without running controlled versions of the same renderer, the exercise cannot establish that removing the records leaves its actual output unchanged.
The operation needs a name:
not consulted ≠ disabled ≠ deleted An item not consulted for this decision may still be retained for a separate, authorised purpose. Disabling its use in this view is not the same as deleting it. Deleting a retained item is also not proof that a cached rendering, a dependent inference or a remote copy has disappeared. A defensible subtraction must identify which input is excluded, which derived outputs are invalidated, what is recomputed, and which earlier effects remain irreversible or unconfirmed. Binding accessibility, preservation and user-authored restrictions are never part of an optional-state ablation.
This is not a measured ablation. To test whether an optional record adds value, the source, task, model, retrieval corpus, output budget and binding policies must all be held fixed while that record’s admissible use changes. The relevant difference is not simply whether two paragraphs look alike. It is whether one version helps Iris reach a correct, adequately supported decision with acceptable costs in attention, correction and privacy.
The conclusion is local to this task and to this record’s proposed use. Chapter 24 separated retention from present applicability. Evidence can legitimately support later continuity without belonging in today’s model context. Equally, speculative future usefulness is not an adequate reason for indefinite retention. The system must review the justification for storage separately from whether the current renderer consults the item.
There are plausible cases for additional state. If Iris gives a vague request, a previously authorised, scoped rule might save a clarifying question. A retained accessibility requirement should not have to be restated at every interaction. A carefully supported inference might justify offering an optional alternative view. None of these possibilities establishes that broad behavioural surveillance or a fixed user type is necessary. Each is a different proposed benefit, with a different permission and a different baseline against which to test it.
For every proposed personal input, ask five questions: what decision needs it; what evidence and scope support it; what authority permits this use; what difference does it make against a simpler alternative; and when must its effect end? A satisfactory answer identifies the specific decision and its source-linked consequences. It does not rely on a broad description of Iris as the explanation.
Separation is necessary to the proposal, not a safety proof The information still needs distinct origins and roles:
declared ≠ observed ≠ inferred evidence strength ≠ permission to use durable history ≠ current context ≠ session overlay goal ≠ intervention authority policy ≠ interpretation ≠ verified enforcement supported scope ≠ applied scope valid component ≠ valid composition The description an expert who likes detail and dislikes distraction collapses the entire record into a persona. It cannot reveal that one observation arose under deep-study mode, that one inference is provisional, that an equation cannot be folded, or that a remote service has not confirmed expiry. Iris would have no precise claim, permission or dependent use to challenge. Separate records support local correction. They do not guarantee that the resulting correction succeeds.
Separate namespaces are necessary but not sufficient. A wrongly scoped inference may be correctly labelled as inferred. An adapter may faithfully preserve an instruction’s origin while misinterpreting its exceptions. Two independently valid instructions may conflict when applied together. Composition therefore needs an additional check on the resulting communication and its side effects, not merely validation of each input record.
Consider a second constructed review. Iris requests a view she may share outside her team, and she will read it on her phone. Every input below might be accurately recorded and legitimately available. The composer still has to show that the output preserves public, task-critical qualifications and does not disclose restricted material:
Declared: terse first views Observed: she normally expands derivations herself Inferred: familiarity with consensus basics in this review context Current: viewing on mobile Boundary: do not expose internal project history The composer combines brevity, recoverable detail, inferred familiarity and limited screen space into an aggressively compressed explanation. It correctly excludes internal project history, but it silently omits a non-confidential timing qualification that determines whether the recommendation applies. Every personal input stayed within its own scope, yet the rendered account still misstates the source. A successful privacy check does not substitute for a successful fidelity check, and neither establishes benefit to the reader.
local validity ≠ compositional validity Valid inputs do not establish a valid final view. A compositional gate would inspect the actual representation for missing decision-critical assumptions, unsupported claims, access violations, unauthorised uses of personal state and stale dependencies. Its own verdict must remain fallible. Passing the checks it runs is not proof that Iris’s intent or the source’s meaning survived.
The book has not earned a universal precedence hierarchy for personal state. Preservation, accessibility and authority restrictions constrain which outputs are permissible. Within those constraints, each declaration or temporary instruction needs an applicability check. Iris’s present request can guide the folding of supporting material, but it cannot redefine decisive evidence. If two authorised instructions still conflict, the system should expose the conflict or fall back to a safe view. It should not convert recency or model confidence into new permission.
Correction has the same limits. Iris can reopen the supporting steps for this view without creating a standing preference. A change to a standing rule needs a distinct declared decision. A corrected inference must reach its dependent uses, and a changed portable boundary must reach its enforcement points. Repairing a record does not prove that its effects have been repaired, and keeping records separate does not prove it either.
The claim that separation enables more local repair and fewer silent conflicts remains untested. It suggests a concrete experiment rather than a conclusion.
Testing whether the personal layer earns its cost Freeze a set of source-grounded architecture-review tasks. Include one with an easily missed qualification, one with a changed standing instruction, one with a stale inference, one with a temporary overlay, and one with an unsupported remote boundary. Independent reviewers should identify the evidence and the permitted effects for each task before examining system outputs. Test at least four conditions: a generic task-aware interface with manual controls; the same interface with applicable declared policy; declared policy plus optional authorised observation and inference; and the full composed layer, including temporary and cross-service controls. Use matched sources, task budgets and models, and repeat the test with held-out cases and changed conditions. This is a proposed design. No such comparison has been run for this book.
Measure decision correctness and missing qualifications independently of preference or clicks. Also record forbidden disclosures, unauthorised interruptions or restrictions, stale-state reuse, unverified revocation, time to correction, prompt and context footprint, retention burden and end-to-end latency. The full layer succeeds only if its added capabilities deliver a measurable benefit without breaching the hard constraints. If the task-only or declared-policy conditions perform as well, the additional personal state has not earned its operational cost. An automated check can test invariants, but reader benefit requires reader evaluation.
A compiler-assisted comparison could also keep the source and task fixed while varying which permitted context the mediator consults. Any change in source interpretation must be recorded separately from any change in delivery. Folding an explanation is a presentation choice, whereas reclassifying a temporary request as a durable preference is a new inference. Source, profile, model and context identities must accompany the diff, because changed outputs alone cannot identify the cause. Evaluate the resulting response on the same task, including correction effort and inappropriate inference. This extends the proposed subtraction test. It reports no personalisation gain and licenses no additional memory.
More state creates more obligations An additional belief is both a possible aid and a new obligation. It needs a legitimate origin, a supported scope, a retention justification, a temporal standing and authorised uses. If it shapes the view, the decision must be traceable to that belief, and its dependent outputs must be repairable. If it does not shape the view, continued storage requires its own justification. Being harmless in one rendering does not establish harmlessness in a larger collection.
Combination creates disclosures and inferences that may not be visible in any single record. Joining Iris’s derivation expansions with a quiet-feed rule to conclude that she dislikes distraction creates a new personal claim, with a new scope and purpose. That claim cannot inherit permission from the two underlying records. Nissenbaum’s account of contextual integrity helps explain why provenance and permitted information flows matter together. Data supplied appropriately in two contexts is not automatically available for unrestricted recombination in a third. That research supplies a privacy framing. It is not experimental validation of this composer.
Record count is therefore a poor measure of minimisation. One broadly reusable inference may reveal more and alter more decisions than several narrow, short-lived declarations. Dropping origin, scope and expiry metadata to save bytes can make a smaller record less accountable. The relevant accounting covers the fields collected, the contexts crossed, the decisions influenced, the time retained, the parties able to inspect it, and the effort needed to correct or revoke its effects.
The proposed burden is not that every personal item must prove its worth on a single screen. It is that a particular collection and use must have a defensible purpose and an observable contribution, or a binding necessity. Possible benefit does not authorise collection; authorised collection does not establish benefit; and successful rendering does not prove that every retained input was needed. In the human–AI interface literature, Amershi and colleagues argue for feedback, correction and control over time. Those are useful design references. They are not evidence that retaining more personal state improves this book’s proposed system.
available state ≠ necessary state authorised state ≠ beneficial state unused now ≠ useless forever None of these distinctions establishes that fewer records always communicate better. They prevent the opposite assumption from becoming the architecture’s default.
A path without optional personal conditioning Iris should have a usable way to request communication without optional behavioural predictions, inferred preferences or other learned state that the task does not require. The operation excludes those inputs from the current decision. It does not promise deletion of their authorised historical records, and it does not automatically revoke unrelated user-authored protections. This is a no optional personal conditioning mode. It is not a claim that the system can answer without context, policy or memory of applicable permissions.
source + explicit current request and conditions + preservation and accessibility obligations + binding user policy + privacy / data-use boundaries and authority restrictions ↓ communication without optional personal conditioning A non-personalised path is not an ungoverned path.
This is a first-class interface mode, not a degraded fallback. It can still follow Iris’s explicit ten-minute instruction, honour standing accessibility requirements, preserve the calculation’s decisive assumption and offer the source. It may also require showing more detail or asking a clarifying question rather than relying on a removed inference. That extra interaction cost belongs in the comparison, not outside it.
The exclusion must reach the computed effects of optional state, not only the next prompt. A cached summary, ranking, plan or inferred dependency may still reflect the removed input. The system must identify and invalidate known dependants, regenerate affected views under the reduced policy, and report unknown or unreachable effects as unconfirmed. A checkbox showing the predictor as off cannot, by itself, establish that the current display is unconditioned.
stop using ≠ invalidate derivatives ≠ recompute current state ≠ undo past effects Even repairing the current view cannot recreate the history in which the predictor was never active. Its earlier contributions to cached representations, subscription ordering or dependent beliefs need their own accounting, and any unresolved effects must still be reported as unresolved.
The switch must also state what remains binding. It cannot silently remove Iris’s separately authorised protective boundary, her accessible presentation requirements, applicable explicit policy, or the provenance needed to enforce restrictions on temporary learning. Such records may be personal in an ordinary sense, but they are not optional behavioural prediction. A more exact label than generic is therefore without optional learned personal conditioning.
If she also wants an existing rule revoked, that is a separate decision whose effects must be confirmed. Changing how a view is conditioned cannot be reported as deletion of history, cessation of all future inference, or removal of every distributed control.
Iris may compare a view produced with optional conditioning against one produced without it, holding the source, task, allowed actions and binding rules fixed. The comparison should show the actual differences in the decision record, not just two plausible paragraphs. With stochastic generation, an uncontrolled difference may be sampling noise. Repeated paired runs or deterministic decision traces are therefore needed before attributing an effect to personal state. This comparison supports investigation. It does not by itself establish a causal improvement for a reader.
Conceptual guide to the chapter; no new empirical result is asserted. Open full-size diagram. What should survive this session? At 09:30, or sooner if the review or session ends, the overlay, reminder and temporary feed grant lose their authority. That change should invalidate queued or cached actions derived from them. Where a remote control was applied, termination requires a removal or expiry signal and an appropriately scoped check. A destination that is offline remains unconfirmed, not safely cleared. The end of local authority and the end of every distributed effect are different events.
The session must not reappear as a durable trait through its behavioural residue. Iris’s expansion of a calculation under a ten-minute review overlay is an observation about that conditioned interaction. It is not an unconditional preference for brevity or detail. Any separately authorised retention must keep those conditions attached. A learning pipeline cannot strip that provenance and later present the pattern as an independent justification for changing policy.
The standing requirement remains in force unless Iris revises it. The earlier history keeps its own justified lifetime. Unused inferences do not become more authoritative merely because they have sat in storage through another successful-looking session.
This is continuity with a limit. The system should preserve what has a legitimate continuing role and let go of what belonged to this moment. It should not use today’s construction as an excuse to accumulate another layer tomorrow.
Running it Running it. Use npm run example:personal -- all in the reference implementation (ernanhughes/project-language). The ch28 selector is an end-of-review slice, not a standalone execution of the whole Part III composition. The full scripted run includes composition, an optional-inference comparison, correction, invalidation, an inspection-only generic comparison, revocation receipts and a later fresh interaction. It does not enact every detail of the hand-described review above.
What it shows: the runtime keeps declared policy, observation, inferred proposal, session overlay, friction contract and portable boundary as distinct supplied inputs, together with source, task, preservation and accessibility constraints. It refuses an overbroad consensus belief and keeps a decisive equation and its assumptions visible. In this scripted fixture, an optional inference folds the library release. Its decision record indicates that removing the inference would make the release visible. The correction invalidates known dependent plans, and a subsequent compilation makes the release visible. That result differs from the illustrative task-based exclusion earlier in the chapter. The two are separate constructions, not one observed end-to-end reader session.
Here, generic means that the run’s optional personal conditioning is excluded. It does not mean that all declared or governance state has been removed. The plan may still report personalised, because authorised standing policy, current instructions or a live overlay still apply. The example separately revokes a service control and ends the overlay and temporary contract. A later interaction reuses the same in-memory store and the applicable standing declarations. That demonstrates continuity within a process. It does not demonstrate durable storage or reliable remote expiry.
What it does not show: that the proposed amount of state is necessary or sufficient, that the composite improves a person’s decisions, or that corrections are cheaper. The generic comparison occurs after the corrective recompilation and reports no difference at that point. It cannot be cited as evidence that the optional inference never mattered. The alternative-view record is an inspection aid, not a controlled experiment. Iris, the services and the meeting are fixtures, and Pi uses a scripted model. The receipts show selected host mechanisms and checks. They do not show that a person noticed, understood or benefited, nor that every distributed effect ended.
How much is enough? Part III has built a possible personal layer from explicit policy, evidence of earlier interactions, scoped inference, temporary adaptation, user-authorised friction, portable boundaries and visible correction. Composition increases capability and creates combined failure modes. A fluent result cannot show that an input was justified, that a narrow claim stayed narrow, that the source survived compression, or that a revoked intervention stopped everywhere.
The constructed review establishes a design possibility. Source obligations, a current request and a few applicable rules can sometimes supply a sufficient reason for choosing a view without invoking a broad prediction about Iris. It establishes no measured sufficiency, reader benefit, optimal retention period or universal minimum record count. The experimental question is whether added personal state improves task outcomes enough to justify its collection, retention, correction and cross-context costs, when compared with strong simpler interfaces.
The next part examines the stronger temptation. If a system can anticipate Iris’s communication choices, perhaps it can predict more of her behaviour. That claim needs its own target and evidence. It does not follow from this session, and prediction would still not confer permission to speak or act for her.
personal state helps communication ≠ personal state predicts the person ≠ system reproduces the person Part IV must define its target before moving from communication support to claims about predicting or reproducing a person. Those are different questions, with different baselines and no implied permission to act. Part III ends with a deliberately narrower standard: How little of you does good communication actually require, and how would we know that the rest helped?
References Nissenbaum, H. (2004). Privacy as contextual integrity. Washington Law Review, 79(1), 119–158. (Used for the narrow point that appropriateness of information flows depends on context; not evidence that the proposed architecture works.) Amershi, S., Weld, D., Vorvoreanu, M., et al. (2019). Guidelines for human–AI interaction. Proceedings of CHI 2019, Paper 3. DOI: 10.1145/3290605.3300233. (Used as HCI design background on correction, control and adaptation; not evidence of this system’s benefit.)