Where the Line Moves

The first paper in this series drew a line. Mechanical work — the conversion of sound into a first draft of words — to the machine. Interpretive work — the editorial judgements that constitute a civic record — to the human. The record, I argued, is the edited data, not the raw data; the human is what makes it a record at all. That line is the foundation of the case for human stewardship of the official transcript, and I stand by it.

This note completes the first paper rather than retreating from it. The first paper said the human should hold the interpretive side of the work. This paper says what holding it actually requires, given a structural condition the first paper did not state plainly enough: the pipeline of practitioners trained to produce the record from raw audio at institutional Hansard standard is contracting in every English-speaking jurisdiction I know of, and has been for some time. The transition to AI-augmented production of the first draft is therefore not, in 2026, an optional efficiency gain. It is increasingly the only available path to maintaining the institution's capacity to produce the record at all. What follows is about how to steward that transition deliberately rather than allow it to happen to the field by drift.

I write this from the same position as before: 36 years in the chair as a court and parliamentary reporter in Australia, and a practice that now builds AI tooling for the same work. The paper that follows engages with peers — D'Arcy McPherson at British Columbia, Henk-Jan Eras at the Dutch Parliamentary Reporting Office, Ana Rita Pereira and Paulo Granja at the Portuguese Assembly, Eero Voutilainen at the Finnish Parliament, Ji Tang in the Chinese stenography industry — all of whom have written carefully about this question in the December 2025 issue of Tiro, the journal of the international stenography federation.¹ They reach different conclusions from mine in some places and the same in others. I treat their work as a community thinking aloud about a shared problem, not as positions to be defeated. Where I push back, I push back specifically; where I agree, I say so.

01Where we left it, and what was left unsaid

The previous paper made the case that the production of a civic record divides into two kinds of work, mechanical and interpretive, and that the interpretive work is what constitutes the record rather than the mechanical work. The line between them is not new. It has been drawn implicitly by every Hansard editor for two centuries, and the records that have survived have survived because someone with the standing to make editorial judgements was on the right side of it. That argument stands.

What that paper did not say plainly enough is the structural condition that makes the argument urgent rather than abstract. A senior colleague at one of the Australian Hansards put the matter to me directly some time ago: we are not making shorthand reporters any more, anywhere. He meant exactly what he said. The pipeline of practitioners trained from the beginning of their careers to capture proceedings from raw audio at institutional Hansard standard — pen shorthand, machine shorthand, audio-only — is contracting in every English-speaking jurisdiction I know of, and has been for some time. The training paths are not being maintained. The new practitioners joining the field are not being formed in the older methods. The practice that produced the editors who carried the record across the last hundred years is being replaced — not at the institution's choosing, but by the structural drying-up of the workforce behind it.

This is the fact that makes the first paper's argument harder to hold than it looked on the page. The first paper argued that AI should be adopted on the mechanical side and the human held on the interpretive side, and that established methods should be permitted to continue for those who work them well. The first half of that argument stands without qualification. The second half — the “horses for courses” position — needs to be qualified by what was left unsaid. Established methods cannot continue indefinitely if the workforce that knows them is no longer being made. They can continue for as long as the current practitioners work, and not much longer. The horses-for-courses position is correct as a description of what should be permitted now; it is not correct as a description of what will be available in twenty years. By the next generation of editors, the older mechanical capture methods will be a niche craft survival, not an institutional alternative.

This is not a matter of resignation. It is a matter of describing the field as it actually is rather than as the first paper implied it could remain. Once the structural condition is named, the question this paper takes up becomes sharper and more urgent. The mechanical/interpretive line is not being moved by careless institutions or cynical vendors. It is being moved by the slow, structural reality that the human is increasingly unavailable to do both sides of the work, and AI is increasingly the only available collaborator for the mechanical side.

The adoption of AI for the first draft is therefore not, in 2026, an optional efficiency gain. It is, in most jurisdictions, the only path forward that maintains the institution's capacity to produce the record at all.

That changes the question. The question is no longer “should we adopt AI” — for most institutions, the answer is being settled for them by the absence of any alternative. The question is what “good adoption” means in practice, and what disciplines are required to preserve, within the AI-augmented practice, the substance of what the older practice carried.

I closed the first paper by promising the next note in the series would take up speaker attribution — the diarisation problem. That paper is still coming. It is set aside for now because the question above has become more pressing. The diarisation paper assumes the field has time to work out the technical questions in detail. The question of how to steward a transition the field cannot refuse needs to come first.

02What the field is actually doing

In the December 2025 issue of Tiro, four institutional responses to AI in parliamentary reporting are described from four different countries. They are not the same response, and reading them together is more useful than reading any one of them alone. Each is best read as a response to the structural condition described in the previous section: none of these institutions is choosing AI as a feature; each is, in its own way, managing the contraction of the older workforce and the rising demand for the record.

At the Legislative Assembly of British Columbia, Hansard Services has developed an in-house ASR tool which they have named Parrot. Their director, D'Arcy McPherson, describes the rationale in stewardship terms: a commercial product would have meant “dependency, cost escalation and potential loss of control over our data — and by extension, our reliability and trust.”² BC chose to build for institutional ownership of data and editorial control. The tool processes chamber audio, generates draft transcripts, enriches them with metadata, and hands them to human editors. McPherson reports that average editing time for a five-minute take has dropped from 45 minutes to 35 — a 22 per cent saving. No jobs were lost. The development process was deliberately participatory across an editing team spanning ages from the 20s to the 70s. The Parrot story is the cleanest example I know of an institutional response that holds the mechanical/interpretive line consciously: AI on the draft, human on the record.

I will note in passing that the in-house Hansard movement, when looked at closely, is rather more complex than the “in-house” framing alone reveals. An earlier Tiro piece from Hansard BC, by Dan Kerr in 2022, describes the initial ASR implementation at BC as having been provided by an existing specialist transcription vendor and integrated into BC's existing systems.³ Several legislatures across the English-speaking world rely on the same small set of specialist vendors for the recording, integration and ASR infrastructure their in-house tools sit on. The specialist legislative-media vendor market is thin, the institutions that depend on it are dispersed, and the dependencies are not always apparent from a single institution's own account of its work. The structural observation worth making is not about any one vendor; it is that the field's sovereignty picture has more layers than any single institutional description usually presents.

At the Dutch Parliamentary Reporting Office, Henk-Jan Eras describes a different posture. The PRO operates under a formal national-government prohibition on generative AI, with exceptions for pilot projects under strict conditions. They have undertaken one such pilot — three months, five people — combining OpenAI's Whisper for ASR with ChatGPT for cleanup against the office's editorial standards. The resulting prototype, called Vluister, transcribes a five-minute audio fragment in three minutes; after automatic correction, the output matches the final official Hansard at 80 per cent word-level accuracy.⁴ Eras is honest about the 20 per cent that remains, and about the structural bind the PRO operates within: even if the national prohibition lifts, European procurement law requires a tender process; specialist vendors are few; “big tech shows little interest in parliaments due to limited budgets and complex demands.” The Dutch posture is experimentation under constraint, with measured outcomes and explicit acknowledgement that the path to permanent adoption is uncertain. Worth naming directly: Eras opens his article with an explicit reference to the workforce-pipeline question, framing the Dutch pilot as a response to “possible personnel and capacity challenges due to an ageing population and labour market developments over the next five to 10 years.” The Dutch PRO is already acting on the structural condition described above.

At the Portuguese Assembly of the Republic, Ana Rita Pereira and Paulo Granja describe a third response. Their Whisper-based system, STAAR, achieved word error rates below 2 per cent on pre-written speeches almost immediately on deployment in April 2023. Reporters no longer transcribe from scratch; they review, correct, identify speakers, and edit for coherence. The interesting move is what Pereira and Granja have done since: they have begun measuring whether the introduction of STAAR has changed the editorial behaviour of reporters and editors themselves.⁵ Their hypothesis — that an ASR-produced draft would lead editors to be more tolerant of borderline grammatical structures, intervening less — is not yet confirmed by the data. Correction rates for the linguistic structures they examined are still substantial. But they note, honestly, that for some structures the correction rate is lower than they would have expected, and that the pre-STAAR comparison data they would need to confirm the hypothesis is not yet available. Their work is the only piece in the Tiro issue that is measuring the cognitive question rather than asserting it. That deserves attention.

In China, Ji Tang describes a fourth and quite different posture: not the response of a single institution but the codification of an industry-level response. Tang reports that demand for traditional shorthand in routine SME meetings and film/TV subtitling has dropped by 90 per cent over five years, while demand for high-complexity work — court records, government and high-level meetings, financial roadshows — has grown 15–25 per cent annually.⁶ The Chinese stenography industry's response has been to draft a new national vocational standard, due in early 2026, formally naming new role categories including an “AI-assisted stenographer.” The industry is not waiting for individual institutions to choose a posture; it is standardising one.

Four institutions, four postures. BC has chosen institutional ownership with a careful workforce transition. The Dutch are experimenting under structural constraint and naming workforce pressure as the explicit driver. The Portuguese are measuring. The Chinese are standardising. All four are responding, in their different ways, to the same underlying condition: that the older mode of production is contracting and the institution must manage the transition deliberately or have it managed for them by drift.

Three of the four institutional tools — Parrot, Vluister, and STAAR — rest on OpenAI's Whisper model. So does the tool my own practice builds. The Chinese tools Tang describes are built on Chinese-developed ASR models from a separate corporate ecosystem rather than on Whisper, but the structural point is the same in both cases: a small number of foundation models in a small number of corporate hands. That is a foundation worth holding in mind as the rest of this paper proceeds. We will return to it.

03The disciplines of necessary adoption

If the adoption of AI on the mechanical side is increasingly structural rather than optional, the question becomes how to do it well — and the costs of doing it poorly are not the costs of having adopted AI at all but the costs of having adopted it without the disciplines that make adoption real. Four disciplines are necessary. None of them is free. Each addresses a way the substance of the older practice can be lost inside the apparently-successful adoption of the new.

The first discipline: measure whether the interpretive judgement remains active

When a human editor produces an editorial decision from raw audio, they are exercising originating judgement: hearing the words, weighing context, deciding what the record should say. When an AI produces a draft and a human reviews it, the human is exercising ratifying judgement: deciding whether to accept what the machine has proposed. These are not the same kind of cognitive work. The originating editor is making the call; the ratifying editor is checking a call already made. Anyone who has worked both ways knows the difference. Originating judgement requires sustained attention to the audio itself, to the speaker's pauses, hesitations, self-corrections, and procedural cues. Ratifying judgement is constrained by what the draft presents — and is constrained, importantly, in ways that are invisible to the ratifier. The errors a draft makes by omission — the procedural cue it never captured, the witness's stumble it cleaned away, the attribution it failed to flag as uncertain — are harder to catch than the errors it makes by commission. An editor checking a fluent draft tends to read it, find what looks wrong, and fix it. An editor producing from audio tends to hear what was said and write that. The two methods do not always produce the same record.

This is not an argument from theory. It is the documented pattern wherever automation has been introduced into expert review work — and the foundational analysis dates from 40 years before AI's current moment. Lisanne Bainbridge's 1983 paper on the “ironies of automation” identified a pattern that has been confirmed across many fields since: when a system is automated, the human role shifts from doing the work to monitoring it, and the skills needed to take over when the automation falters atrophy through disuse.⁷ The pattern has since been measured in expert review specifically. In a 2023 study published in Radiology, experienced radiologists reading mammograms were significantly affected by incorrect AI based suggestions: accuracy on the cases where the AI gave wrong advice fell from 82 per cent to 45.5 per cent even among readers with more than 15 years of clinical experience.⁸ The phenomenon has a name in the human-factors literature — automation bias — and the literature is clear that it affects experts, not just novices.⁹ Whether parliamentary editing follows the same pattern is the open empirical question. There is no reason to think it does not.

The discipline this risk requires is measurement. Not abolition of the AI workflow — that is not on offer — but the institutional commitment to look at what is happening to editorial judgement under it. Pereira and Granja's work is the methodological template: compare pre-system and post-system editorial choices, look for drift, look for declining correction rates on structures that should still be corrected, surface the pattern early enough to act on it. The institution that adopts AI without ever measuring what is happening to its editors' judgement has not really adopted AI; it has just changed its working method and hoped for the best.

The second discipline: keep the conventions the institution's, not the tool's

Hansard conventions are not generally written down comprehensively in any single document; a style manual is a guide, not an exhaustive bible for rational verbatim Hansard editing. The editing conventions live in the heads of editors and are transmitted by apprenticeship, supervision, and habit. An editor who has worked under a senior editor for 10 years knows when to preserve a stumble, when to clean it; when an interjection is attributed and when it is rendered as a parliamentary aside; how to mark a procedural moment that has not been formally announced but that everyone in the room understood. These are not rules; they are practices. They are also the substance of what makes the institution's record the institution's record rather than a general transcript of an institution's audio.

The moment a tool encodes a subset of those conventions and applies them at scale, the encoded subset becomes the de facto standard. What the tool can encode — the explicit, the regular, the rule-shaped — quietly displaces what it cannot — the contextual, the practice-shaped, the apprentice-transmitted. This is not a problem of bad tools; it is a problem of any tool. Even a perfectly engineered system encodes what its designers thought to encode at the moment it was built, and the work of an institution's senior editors begins to be shaped by what the tool supports rather than by the full corpus of practice the institution carries in its people. The drift is real, gradual, and not always recoverable.

The discipline this requires is making the conventions explicit. Not because conventions can be reduced to rules — they cannot — but because the conventions a tool encodes have to be tested against the institution's own published standards on a continuing basis. The institution that adopts AI tooling without externally documenting its conventions has, by default, accepted the tool's conventions as its own. The institution that has documented its conventions can see the gap, and act on it. This is editorial work, done by editors, and it has to be funded.

The third discipline: cultivate in the new practice what the older practice carried

The previous paper made a workforce argument: institutions that cut their experienced people to bank short-term savings lose the knowledge that holds the record together. That argument stands. There is a deeper version of it that the structural condition described in Section 01 makes more urgent than the previous paper allowed, and that requires the previous paper's framing to be revised.

The editor's craft was historically formed through a particular apprenticeship: years of producing the record from raw audio, under the supervision of senior editors, building up over time the editorial judgement, the convention literacy, and the procedural knowledge that the work required. Bainbridge's 1983 observation about skill atrophy applies here, but with a twist the previous paper did not develop. Skills atrophy at the individual level when not exercised, yes; the deeper problem is that the institutional pipeline that formed those skills is no longer running at the scale it once did. The current generation of senior editors was formed through the manual apprenticeship. The next generation, increasingly, is being formed through AI-augmented practice from the start. The cognitive habits that the manual apprenticeship developed automatically — sustained attention to audio, originating editorial judgement under the discipline of the speech itself, the listening skill that catches what the draft missed, the convention literacy that lives in the editor's hands as much as their head — do not develop automatically through AI-augmented practice. They have to be deliberately built in.

This is where the previous paper's “horses for courses” argument needs revision. That argument implied that established methods could be maintained alongside the new methods, as a permanent option for those who wished to work them. The structural condition described above makes that argument unsustainable as written. Established methods cannot be maintained at institutional scale once the practitioners who carry them retire and are not replaced. The institution's task is not to preserve manual production as a working alternative track; it is to preserve, in the new practice, the substance of what the manual practice taught. That is a different kind of preservation, and a harder one.

What the third discipline actually requires of an institution in 2026 is the deliberate cultivation, within the AI-augmented practice, of the cognitive habits the older practice carried. Structured exposure to the audio independently of the draft. Periodic exercises in originating editorial judgement before the AI is consulted. Apprenticeship of new editors under senior editors who can transmit, deliberately, the judgements that are being lost as the older practice retires. Audit of editorial decisions across the team to surface drift before it becomes invisible. Investment in the time and the supervision that produced the older editor, redirected to producing the new editor in a way that preserves what the older formation produced. This is harder than what the previous paper proposed. It is also closer to what the field actually needs, given the workforce condition Section 01 named.

The fourth discipline: measure whether the productivity case actually holds

Eras's piece in Tiro references an observation from the Spanish Senate that deserves to be quoted at length, because it is the quietest and most important sentence in the entire issue: “some parliaments that state they are very satisfied with the results of using AI for transcriptions do not seem to have reorganised their team or changed their working method.”¹⁰ Read that twice. It says that institutional satisfaction with AI is being reported without corresponding measurable change in how the work is actually done. Either the satisfaction is correct and the institutions have failed to reorganise around the productivity gain — which is its own management failure — or the satisfaction is not, in fact, supported by measured productivity at the institutional level.

Either interpretation is troubling. The first means the field is leaving real productivity on the table by failing to redesign its work around the tools. The second means the field is in the middle of an adoption cycle being driven by user-level enthusiasm — which is genuine and worth taking seriously — without the institutional-level measurement that would tell anyone whether the tool actually produces the work-process improvement that justifies its cost, its sovereignty exposure, and its convention-narrowing risk. Eras notes elsewhere in his piece that AI projects fail at remarkable rates in industry generally — research from RAND and from MIT puts the figure at 80 to 95 per cent — usually for reasons of unclear business cases, overestimation of technological capabilities, and the cost of implementation, support, and maintenance being underestimated.¹¹ Parliamentary reporting is not exempt from these patterns.

The discipline this requires is institutional-level measurement, not user-level satisfaction. McPherson's 22 per cent figure at BC is the kind of number worth having. There need to be more like it, properly methodologically grounded, before the field can speak with confidence about whether the line being moved is being moved for measured benefit or for unmeasured enthusiasm. An institution that adopts AI without measuring whether the adoption is actually producing the productivity claim that motivated it is in a position to be surprised, in either direction, when measurement eventually arrives.

04Sovereignty restated

The previous paper offered five questions an institution should ask about the sovereignty of its record: residency, inference, derived value, continuity, and interpretive control. Those questions stand. They are necessary. They are not, on closer examination, sufficient — and the structural condition described in Section 01 makes the gap more urgent.

The previous paper's questions assume the institution has chosen its tooling. They ask: given the choice, what does the choice mean for the record's independence? They do not ask: given that the institution's practice of producing the record now depends on tooling, what does sovereignty mean over the practice itself?

This is the deeper version of the question, and it sharpens once we accept that AI adoption on the mechanical side is increasingly structural. An institution that has resolved residency, inference, derived value, continuity, and interpretive control well — that has its archive on home soil, its inference local, its derived value retained, its continuity in writing, and its editorial control intact — can still find itself in a position where the editors cannot produce the record at the institution's expected throughput without the AI tooling they have come to depend on. The archive is sovereign. The practice is not. The institution owns its history and has lost the conditions under which it can make its future.

This matters because the practical, day-to-day exercise of sovereignty over the record is the ability to make new records, not just to safeguard old ones. The institution that depends on tooling to make new records at its required pace has a particular kind of contingent sovereignty. It works as long as the tooling continues to work. The moment something changes upstream — the model's licence terms, the specialist vendor's continued existence, the open ecosystem's commitment to the kind of model the institution relies on — the institution's production capability becomes vulnerable, even though its archive is safe.

This is the foundation question the Tiro issue raises without naming. The three institutional Whisper-based responses described — Parrot at BC, Vluister in the Netherlands, and STAAR in Portugal — and the tool my own practice builds, all rest on a single corporate decision by OpenAI to release Whisper under permissive terms. That decision could change. OpenAI's pattern with foundation models has gone both ways: GPT-2 was released with open weights in 2019; the GPT-3, GPT-4 and GPT-5 families have been API-only since; and in August 2025 OpenAI released two new open-weight reasoning models, gpt-oss-120b and gpt-oss-20b, under the Apache 2.0 licence.¹² The lesson is not that the trajectory is one-way. The lesson is that the field's in-house movement rests on a corporate licensing posture that has shifted before and could shift again, and that the open availability of any given model is not guaranteed by the structure of the field.

This is a collective exposure, not an individual-institution problem. No single Hansard can solve it. The structural answer, if there is one, is collective action — sustained funding of open speech models adapted for parliamentary work, collaboration with open-research speech communities, the kind of cross-Commonwealth cooperation that the Hansard tradition is structurally good at but rarely exercises. That is a hard suggestion to land, because it asks institutions that are not used to acting collectively to coordinate on something whose urgency is invisible until it is too late. But it is the honest answer to the foundation question, and the question deserves to be asked even if the answer is uncomfortable.

Add it, then, as the sixth sovereignty question: foundation — what does the institution's position become if the upstream model on which its tooling depends ceases to be available on current terms? Most institutions do not have an answer to this question because they have not asked it. The asking is the beginning of the answer.

05What can be defended

What follows is not a procurement checklist. It is a description of three layers of defence that an institution can construct, in order of increasing difficulty, none of them free. The previous section established why these defences matter; this section is what they actually look like.

The defensible minimum: vendors and tools must be replaceable

This is the layer the previous paper described as the practical sovereignty argument: records portable in open documented formats, tools using documented data structures, institutional data not training vendor general products, editorial decisions auditable and overridable, and a written exit plan that has been tested. This is the work most institutions can do, and the work most institutions have not yet done. It is necessary. It is also the most achievable, because it consists mainly of contractual and architectural choices made at procurement and renewed at each contract cycle.

The minimum has internal stratification by institutional resource. BC's Parrot — built in collaboration with their existing systems vendor, integrated into their own data pipeline, with their own editors training and correcting it — represents the well-resourced case of this minimum: an institution that has the technical capacity to take ownership of its own tooling even when it does not write every line of the underlying code. McPherson's framing of “in-house” describes institutional control over what matters: the data, the editorial workflow, the convention-encoding, the workforce. That is a working version of the defensible minimum, and it deserves to be recognised as one. Smaller jurisdictions cannot replicate BC's resource position. They have to commission, and the specialist legislative-media vendor market that serves them is thin. For them the defensible minimum is harder, and the institutional collective action question becomes more pressing.

The harder defence: transparent conventions and auditable choices

The minimum protects the archive and the exit. The harder defence protects the content of the record against the convention-narrowing risk described above. It requires that the conventions a tool encodes are documented externally to the tool, in the institution's own published standards. It requires that when a tool makes an editorial decision — a punctuation choice, a stumble cleanup, an attribution assignment — the institution can see what was decided and why, and can audit the pattern of decisions over time. It requires that the institution's published conventions are kept current as the tool's behaviour changes, and that conventions the tool cannot encode are kept alive in editorial practice rather than quietly retired because the tool does not support them.

This is harder than the minimum because it requires sustained editorial attention to the relationship between the institution's published conventions and the tool's actual behaviour. Most institutions are not currently doing this. Some are: the Portuguese OJD's work in measuring whether editor behaviour has changed under STAAR is the beginning of this kind of attention, and the methodology — comparing pre-system and post-system editorial choice — is exactly the methodology the field needs. The Portuguese example shows the harder defence is possible. The 2026 question is whether other institutions adopt the same methodological discipline.

The hardest defence: keeping originating judgement alive in the new practitioner

The previous paper, and the v2 of this one, named “maintained human production capability” as the hardest layer of defence — keeping a portion of work flowing through the manual editorial path as a working practice rather than a theoretical fallback. The structural condition described in Section 01 makes that formulation unsustainable as written: manual production capability cannot be kept alive at institutional scale once the workforce trained to perform it is no longer being replenished. The underlying argument the previous paper was making, though, must not be lost in the reframing. That argument was not heritage-craft preservation. It was cognitive. The editor formed through the older apprenticeship — pen shorthand, machine shorthand, audio-only capture — could not avoid exercising originating judgement on the speech, because they had no draft to ratify against. They listened, they rendered, they decided what the record would say. The cognitive loop of hearing, rendering, judging was inseparable from the act of producing the record, and it is that loop that produced the editor's ear, the convention literacy, the procedural knowledge, and the practised originating judgement the previous paper argued the institution must keep at the centre of the work.

The AI-augmented editor can avoid that loop completely. The draft is always there. The cognitive default is to read the draft and ratify it. The hardest defence is therefore not the maintenance of the older practice as a parallel track, and it is not generic training of the new practitioner. It is the institutional commitment that the new practitioner, throughout their working life and not only in apprenticeship, continues to exercise originating judgement directly on the speech — with the draft set aside, by structural rule, as a continuing part of the editorial workflow. The shorthand reporter could not avoid this loop because their tools did not permit avoidance. The AI-augmented editor can, and probably will, unless the institution structures the work to prevent it.

This is the institutional version of the third discipline, sharpened. It requires real investment: structured exposure to the audio independently of the draft, not as a one-off training exercise but as a continuing feature of the work; supervised apprenticeship under the senior editors still in the workforce, before they retire; deliberate transmission of the convention literacy, procedural knowledge, and editorial judgement the older formation produced; institutional support for the time and care this transmission and its continuation take, against the productivity pressures that push in the opposite direction. The honest acknowledgement is that almost no institution currently does this with the seriousness it requires. The first institution that does, deliberately, and publishes its findings, will have made a real contribution to the field.

These three layers together — replaceability at the procurement layer, transparency at the editorial layer, and originating judgement at the practice layer — are what full production sovereignty looks like in the conditions the field actually faces in 2026. Few institutions will achieve all three. Many can achieve the first. The question this paper puts to the field is whether the institutions that go no further than the first will be willing to say so plainly, rather than describing the first as if it were the full picture.

06Threading the needle, again

The previous paper closed with an image: two centuries ago the record of parliament was set in type by a printer named Hansard, and the words belonged to the parliament. That image is still right. The harder version of it, which this paper has been working toward, is this: the words have always belonged to whoever could still produce them — and the producer is now changing.

The new producer is the AI-augmented editor — the practitioner who works with the draft, supervises it, corrects it, and signs off on it as the institution's record. This is not a degraded version of the older producer; it is, increasingly, the only producer the field has. The institution's task is not to refuse this transition but to steward it deliberately: to ensure that the new producer carries, in their practice, the judgement, the convention literacy, the procedural knowledge, and the editorial responsibility that the older producer carried in theirs. That transmission is the work of the next decade for every institution that wants to preserve what its record has historically been.

What I have been arguing for, across these two papers, is a stewardship that takes the longer view. The mechanical/interpretive line still holds. The interpretive work is still the substance of the record. The human still constitutes the record by their judgement. What is new is that the human is changing, and the institution's responsibility shifts from preserving the older practitioner to preserving, in the new practitioner, what made the older practitioner the keeper of the record.

There is no single answer to where each institution should place its line in detail. Different jurisdictions, with different resource positions, chamber loads, multilingual constraints, and workforce ages, will negotiate the transition differently. The question this paper has tried to make visible is what is being transmitted — and what is at risk of being lost — at each placement, so that the institutions making the choices are making them with their eyes open.

The words will belong to whoever can still produce them. That is what sovereignty has always meant, and it is what it still means now. The institutions that come through this transition well will be the ones that recognised, while there was still time, that the new producer must be deliberately formed in the substance of what the older producer carried — and that without that deliberate formation, the title of editor survives the practice that gave the title its meaning.

A note on what comes next: the previous paper promised a CAL Note on speaker attribution and the diarisation problem. That note has not been displaced; it has been deferred. It will follow, in the same series, in the same spirit.