Stewardship in the Age of AI

Artificial intelligence is the largest change to reach the production of the civic record in a generation, and the people who keep that record can feel the ground moving. This paper is a set of observations from inside the work, written for everyone who keeps a Hansard or a civic record in English. In short: AI helps a great deal with the mechanical part of the work and is unreliable in the interpretive part; the change is real but slower than it is sold to be; and the experienced people who feel most exposed are, in practice, the ones who make AI work at all. Beneath it sits a question rarely asked at the point of purchase — who, in the end, controls the record? I describe what I see and leave the conclusions to the reader.

01The change that is really arriving

Artificial intelligence is the largest change to interface with the production of the Hansard record since mechanical and then digital audio recording replaced shorthand as the sole means of capturing a record of proceedings. A technology now exists that turns hours of recordings into formatted draft text in minutes, and it is improving quickly. The people who have spent careers producing Hansards and civic records can feel the ground moving, and they are not wrong to. Any honest account has to begin by saying so.

Let us be plain: this is a real change. It will alter how the record is produced, it will change some of the roles of the people who produce it, and institutions that treat it as a passing fashion will be caught out. The more useful truth, though — the one a manager actually needs — is that the change is real but uneven, and slower than it is sold to be. We will come to the 'why'. Firstly, let's discuss who 'we' and 'you' are, because the striking thing about this moment is how many people are facing it at once.

A shared problem, across many jurisdictions

The Hansard or civic record is kept in a great many places that rarely compare notes: the parliaments of the United Kingdom; Canada; Australia's Commonwealth, state and territory parliaments and councils; New Zealand; the legislatures of the Pacific and of the Commonwealth Caribbean; the parliaments of South Africa, Kenya, Ghana, Nigeria and others; India and the chambers of Singapore, Malaysia and Sri Lanka; and the multilingual official reports of Europe's institutions, which meet the same change via a different door. Most keep the record under a name borrowed from a single London printer, Thomas Curson Hansard, who began printing the parliamentary debates two centuries ago. They inherited a shared idea of what an official record is and what it is for — and now, within a few short years, every one of them is meeting the same change and hearing the same pitch from the same kind of vendor. This paper is written for all of them.

The smaller and more remote the jurisdiction, the sharper the stakes. A large parliament can absorb a poor procurement decision; a small legislature that hands its record to a distant vendor and is then left stranded has far less room to recover. The questions are universal, but they press hardest on those with the least leverage.

What qualifies me to raise them is straightforward: I have worked as a court and parliamentary reporter in Australia since 1989, and the practice I lead now builds AI tooling for the same work. That is the whole of my standing — close enough to the technology to know what it can do, and close enough to the practice to know what it must not break. I describe what I see.

02The record is not the recording

It is tempting to treat the production of a record as a transcription problem: audio goes in, text comes out, and the better system is the one whose text more closely matches the audio. This is a misunderstanding, and almost every poor procurement decision in this field can be traced back to it.

A civic record is not a recording of what was said. It is the official, authoritative account of what an institution did — and it carries weight that an ordinary transcript does not. These records are consulted years later, cited as authority, and trusted precisely because the institution stands behind them.

This is why the record has always been an edited report, not a raw transcript — a point older than any technology now being sold. From its earliest days the Hansard was never a verbatim record: by long-standing convention it omits repetitions and false starts, corrects obvious slips, and renders what was meant while preserving the substance of what was said. The editorial judgement is not a flaw in the record, it is the record. Strip it out and you do not have a purer Hansard, you have a transcript, which is a different and lesser thing.

That distinction is the foundation of what follows. When a vendor offers to produce a civic record by automated transcription, the first question is not 'How accurate is it?' but 'Accurate to what?' — to the audio, or to the standard the institution has always held its record to? They are not the same target, and a system built for the first can quietly erode the second.

03What AI does well — and it is a great deal

None of what follows should be read as scepticism about the technology. The benefits of AI in civic-record production are real and in some cases overdue. An organisation that refuses them out of caution will be outcompeted by one that adopts them well, and serves its institution less well in the process. Where the technology genuinely helps turns out to be a great deal of the work.

The first draft

The single largest benefit is the automated first draft. Modern speech-to-text, properly configured, can convert hours of proceedings into a workable draft in minutes rather than the many hours a human would take to produce the same first pass. This is transformative for turnaround, and turnaround matters: a Hansard that appears the same day serves democracy better than one that appears a week later, and a court transcript that is available quickly can keep a matter moving. Used this way, AI does not replace the skilled editor; it removes the most mechanical and fatiguing part of their work and lets them spend their time where their judgement actually adds value.

Search, indexing and access

A second benefit is what becomes possible once the record is machine-readable at scale. Decades of proceedings can be made searchable; a citizen can find what their representative said on a subject; an institution can index and cross-reference its own history in ways that were previously impractical. This is a genuine expansion of public access to the democratic record, and it is the type of thing AI is very good at.

Accessibility

A third benefit, and to my mind a particularly important one, is accessibility. Real-time captioning powered by AI can open a proceeding to a deaf or hard-of-hearing participant who would otherwise be excluded. The technology is not yet good enough to be left unattended in high-stakes settings — more on that below — but as an aid that extends the reach of skilled captioners it is already opening doors that were closed.

Relief from drudgery

Finally, and least discussed, AI can relieve skilled professionals of work that wastes them. The experienced Hansard editor is a scarce and valuable resource. Every hour they spend on pure mechanical transcription and formatting is an hour not spent on the judgement calls only they can make. Used well, AI is not a threat to that workforce but a way of pointing it at the part of the job that actually requires a human.

04The line: mechanical work and interpretive work

If the benefits cluster so clearly, why is there any difficulty at all? The difficulty comes if, in the future, the same technology that excels at the mechanical parts of the work is asked — or quietly allowed — to make the interpretive judgements that constitute the record. And that is where the danger lives.

The difficulty has a single source, and it is the one idea most worth taking from this paper. The production of a civic record divides into two kinds of work. There is mechanical work: converting sound into a first draft of words. And there is interpretive work: deciding how an interjection is attributed, whether a witness's correction supersedes their earlier answer, when a stumble is preserved because it matters and when it is cleaned away because it does not, how a procedural moment is rendered, and a hundred other judgements that vary by chamber, by jurisdiction, and by context.

AI is well-suited to the mechanical work, and the records that hold up are the ones that embrace it there. AI is not suited to the interpretive work, and where the record endures, a human has remained the interpretive lens through which it is made. That is the line. It is not a line between using AI and not using AI — the systems that work best will use AI heavily. It is a line within the work, between the part where a machine produces a draft and the part where a person, accountable to the institution, decides what the record will say.

The reason the line matters so much is the one established above: the record is the edited data, not the raw data. The interpretive judgements are not overhead to be automated away, they are the act of producing the record. A system that automates them will not make the record faster; it will make a different, lesser record while appearing to make the same one. The degradation is invisible at the demonstration and obvious only later, in the cases where the judgement was wrong and no human ever saw it.

This is also why the field is genuinely hard, and why generic transcription accuracy figures mislead. The difficulty in civic record production was never mainly the audio. It is the hundreds of small conventions, exceptions and judgement calls that turn a transcript into an institution's record — and these differ between one jurisdiction's standing orders and another's, between a Senate estimates hearing and a local council meeting.

05Disruptive — but not yet

Return now to the manager's real question, the one beneath all the others: how big, and how fast? The honest answer has two halves, and most accounts give only one. The vendors give the first half — it is big, it is here, it is inevitable — and the anxious give the second — it is dangerous, hold it back. From inside the work, it is both halves at once: it is big, and it is slower than it looks, and the reason for the slowness is the most useful thing a planner can know.

The mechanical layer falls quickly. Automated drafting is already good enough to transform turnaround, and it will only improve, and it will arrive cheaply. An institution that has not begun investigating making AI transcription available for the first draft is already behind, and one that refuses to offer it at all is choosing to be outpaced. Making it available, though, is not the same as imposing it — a distinction that turns out to matter a great deal, and one to return to below.

The interpretive layer is a different curve entirely. The conventions, the judgement calls, the accountability, the institution-specific knowledge that turns a transcript into the record — these have proven stubborn, and there is good reason to think they will stay stubborn for a long time, because they are not really a technology problem. They are a problem of judgement and responsibility, and judgement does not improve on the same curve as raw transcription. The easy 50 per cent comes fast; the harder 50 per cent comes slowly, or not at all — and the harder 50 per cent is the part that is actually the record.

This is why 'disruptive but not yet' is the truest description on offer. Not 'not ever' — the change is real and parts of it are here. But the wholesale replacement of the skilled human that the boldest vendor pitches imply is not arriving next budget cycle, and a manager who plans as though it might will make expensive mistakes. The realistic horizon is gradual: the mechanical work largely handed to the machine quickly, the interpretive work changing shape slowly and staying human for the foreseeable future. That is a horizon an institution can plan against — worth more than either the hype or the alarm.

There is a second reason for gradualism, and it is the one managers most often miss because it is not about the technology at all. Even where AI could be pushed faster, it usually should not be — the limiting factor is not the machine but the people, and an institution that forces the pace past what its workforce can absorb does itself real damage. Drive adoption by mandate, retire the established methods overnight, and you do not get faster transformation; you get resistance, attrition, and the quiet departure of the people who hold the institution's memory. An experienced reporter or editor cannot simply be let go and rehired a year later when the new system turns out to need supervising; the knowledge leaves with them and does not return. A prudent institution treats the human pace as a design constraint, not an obstacle: it moves as fast as the technology allows on the mechanical layer, and only as fast as its people can absorb on everything else. That is not timidity; in my experience it is how the transitions that succeed are run — and it leads straight to a question about the people themselves.

06The people who keep the record

No account of this disruption is honest if it does not speak directly to the people in the room: the reporters, editors, captioners and clerks whose careers are bound up in the work, and who have the most reason to feel afraid. They deserve more than reassurance, and more than being treated as a line item in a transition plan. They deserve the truth, which is more interesting than either comfort or alarm.

The truth begins with a fact that the long-serving among them already know in their bones: change has always been the condition of this work. The parliamentary and civic record has been produced by a long succession of technologies, each of which felt total in its day and is now a museum piece. Each transition was a disruption. Each was feared. Each was absorbed. And through every one of them, the record continued — because the thing that persisted was never the tool but the human judgement about what the record should say.

This is what it means to be a steward of the record rather than an operator of a particular machine. The steward's craft was never the shorthand or the tape or the keystrokes; it was the judgement those tools served — knowing what the record should say, and taking responsibility for its truth. The tools are replaced every generation; the judgement is the constant, handed from one practitioner to the next. Seen that way, AI is not the end of anything, it is the next tool the judgement will outlast. The experienced practitioner is not made obsolete by a machine that drafts faster; they are made more central, because the work that remains is the part that was always hardest and most human, too often buried under the labour of getting words onto a page. Take the mechanical capture away and what is left is the judgement, which is what humans were always best at.

How the work changes — honestly

None of this means no job changes. It would be its own dishonesty to pretend otherwise, and the people in the room would see through it at once. So, plainly, what we see coming:

The work bifurcates rather than vanishes. The mechanical share is largely ceded to the machine; the interpretive share becomes the whole job. The reporter who once spent most of their time transcribing and a little of it judging now spends most of it judging — the part that was always the point.

New roles appear that did not exist. Someone must verify the machine against the conventions, decide where the line sits, and train and correct the system — and the only people who can do that are the experienced practitioners, because the knowledge lives in them. AI does not remove the expert; it makes the expert the one person who can supervise it.

Some roles do contract, and it should be said. A team that once needed many hands for the mechanical work might need fewer for capturing and more for governance, and the net may be smaller. The honest frame is not that nothing changes, but that the work which remains is more skilled, more central, and more defensible — and the people and institutions that lean into that, rather than away from it, are the ones who come through well.

The real danger is institutional impatience, not the technology. The thread is dropped — not by AI, but by an institution that sees the fast, cheap mechanical win and assumes the interpretive layer will follow on the same curve. It cuts its experienced people to bank the saving, and discovers two years later that it automated the easy part, degraded the hard part, and lost the only people who could have caught it. Threading the needle is as much a question of management nerve as of technology.

That is how humans thread the needle: not by resisting the disruption and being left behind, and not by surrendering to it and hollowing out the record, but by recognising that the steward's value moves up the chain — from producing the draft to governing the record — and by deliberately keeping, and retraining, the people who carry the thread. The institutions that do this will adopt AI faster and more safely than the ones that panic, precisely because their people are not afraid of being erased by it.

Horses for courses

There is one more thing to say to the people in the room, and it is the part most often left unsaid, because it is inconvenient to a vendor and awkward to a manager in a hurry: the established methods are still valid, and the people who work them well should be respected, and permitted, to keep doing so.

A reporter who captures a proceeding by the trained ear and the practised hand, a steno writer at their machine, an editor who works from the audio in a way refined over 30 years — these are not relics to be retired the moment a faster tool appears. They are producing a true record by a proven method. A new way of working does not make the old way wrong; it makes it one option among more than one, and for some practitioners in some settings the established method remains not merely valid but better. An institution that declares that craft obsolete overnight does real harm to people who, very often, cannot simply be moved on and should not be — the longest-serving and most knowledgeable in the workforce, the holders of the conventions and the institutional memory. To force them off a method that still works, on a timetable set by the procurement cycle, is to treat the carriers of the record as an obstacle to its production. That is backwards.

Horses for courses, then: make the new tool available to those who want it, let the established methods continue for those who work them well, and let the change come by adoption rather than decree. Those who take up the AI draft and AI formatting assistance because it genuinely saves them from drudgery will use it the better for having chosen it; those who keep their method will keep producing a true record and keep their knowledge in the building. Both are stewards, and neither is a problem to be solved. It is also, not by coincidence, the surest route to a transition that holds.

07Where it goes wrong

If that is how it goes right, it is worth being equally plain about how it goes wrong. The risks of AI in civic records are, with one exception, simply the consequences of letting automation cross the line from mechanical to interpretive work. They are worth naming individually because each tends to be discovered the hard way.

Accuracy under real conditions

Vendor accuracy figures are typically produced under favourable conditions: clear audio, a single speaker, a neutral accent, predictable vocabulary. Real proceedings are none of these. They have crosstalk, regional and non-native accents, technical and legal terminology, procedural language, and proper nouns the system has never seen. Accuracy that is impressive in the demonstration can fall away precisely where the record matters most. The danger is not the overall error rate. It is that the errors are not spread evenly — they cluster in the hardest passages, the crosstalk and the unfamiliar names and the technical terms, the very places a human editor would have known to slow down and check. And the vendor's headline accuracy figure hides exactly this.

The convention problem

Every civic record sits inside a dense web of conventions: how speakers are named and titled, how interjections and interruptions are handled, how procedural events are marked, how a quotation read into the record is rendered, how corrections are managed. These conventions are mostly unwritten, vary by institution, and are exactly what a long-serving editor carries in their head. An automated system does not know them unless taught, and teaching them is far harder than improving raw transcription. A system that produces fluent text in the wrong conventions produces something that looks like the institution's record and is not.

The interpretive black box

When a human editor makes a judgement call, another human can ask why, and the answer can be given, examined and overruled. When an automated system makes the same call, the reasoning is often opaque even to the vendor, and there may be no mechanism for the institution to see what was decided, let alone change it. A record whose editorial decisions cannot be inspected or appealed is no longer fully the institution's record. It has acquired an author the institution cannot question.

Deskilling

The slowest risk is the one the previous section described from the other side: if the mechanical work is automated and the interpretive work is quietly automated alongside it, the pipeline of skilled people thins, and an institution that has lost them cannot easily reverse a bad decision, because the stewards who would step back in are no longer there. The capability, once gone, is expensive and slow to rebuild — and in some small jurisdictions, may not be rebuildable at all.

08The question beneath the question: sovereignty

All of the above concerns how a record is made. There is a further set of questions concerning who controls it — and in practice these are rarely raised at the point of purchase, because procurement conversations naturally gravitate to features, accuracy and price. Yet the answers are largely determined then, by the architecture chosen, and they are hard to revisit afterwards. Not asking is itself a kind of answer.

It helps to recall that the relationship between an institution and the producer of its record is two centuries old. Across all of that time one thing held constant: the institution owned the words. The producer was an employee or contractor, and the record belonged to the parliament or the council. What is new in the present moment is that the words can begin to belong — functionally, through where they are processed, what they train, and what formats they are locked into — to the producer rather than the institution. Sovereignty, in this context, simply means keeping the words where they have always belonged.

It unpacks into five questions. None is exotic; each is simply unasked.

Residency — where does the record physically live? Where is the audio held while it is processed, and where does the finished record reside? On home soil or in a data centre whose location the institution has never been told?

Inference — where does the thinking happen? Even if the finished text rests on home soil, where does the AI actually run? If transcription routes through an offshore commercial service, the audio of a closed or sensitive proceeding transits a third party's systems.

Derived value — who owns what the record becomes? A parliament's accumulated, human-corrected record is among the most valuable training assets imaginable: decades of verified text with every convention and proper noun settled by editors who got it right. When a vendor processes that corpus, two questions usually go unasked — does the institution's record train a vendor's general product, and who owns the model that results? If the institution's own decades of editorial labour silently become the vendor's competitive moat, the institution has funded its own dependency.

Continuity — what survives when the vendor does not? Vendors fail, are acquired, lose contracts, or raise prices at renewal once switching has become impractical. For most software that is an inconvenience. For the system that produces the official record it is a failure of a democratic function: if the record cannot be produced, accountability breaks. The questions are whether the institution can extract its data in a documented, open format, and whether the system remains operable by someone else — or whether the record stops the day the vendor does.

Interpretive control — can the institution see, and overrule, the editorial decisions? This is the sovereignty face of the black-box risk from the previous section. A record carrying legal weight is made of editorial choices. If those choices are made in a system the institution cannot inspect or override, it has ceded interpretive sovereignty over its own record. The question is whether, when the system decides, the editor can see why, change it, and audit it afterwards.

Taken together, these five are what turn 'sovereignty' from a slogan into something an institution can actually see. And the reason they sit beneath the benefits-and-risks discussion rather than beside it is that they are decided earlier and last longer than any feature. An institution can change its mind about how much it automates. It cannot easily change its mind about where its decade of records was processed, what they trained, and who can still produce them.

09What we notice

This is not a procurement checklist, and it will not end as one. But across the work, certain things recur — patterns that separate the systems that serve an institution well from the ones that quietly do not. We offer them as observations. A reader who turns them into questions of their own will likely find them useful.

The human stays on the interpretive side of the line. Where the record holds up, the first draft is automated freely and a skilled, accountable person still makes and owns the editorial judgements that constitute it.

Accuracy is known under real conditions. The institutions least often surprised are the ones that saw the system tested on their own difficult audio — real accents, crosstalk, procedural language — rather than on a demonstration set.

The conventions are honoured. Where the record reads as the institution's own, the system has been taught that institution's specific conventions, and a human has confirmed they are met.

The five sovereignty questions have answers. Residency, inference, derived value, continuity and interpretive control — where these are settled in writing rather than left to the sales conversation, the institution tends to keep control of its record.

There is a real way out. The records that survive a vendor's departure are the ones held in open, documented, portable formats, with a genuine path to produce them by other means. Where there is no exit, the dependency is the product.

The people are brought, not driven. The transitions that hold are the ones where AI was made available rather than imposed, where established methods were allowed to continue alongside it, and where experienced people were kept and retrained rather than moved on. Adoption by consent is slower to start and far more durable than adoption by decree.

I should be candid that this list also describes, more or less, how I build my own tools — so weigh it accordingly. I offer it as things I have noticed, not as a standard I am entitled to set. The institution is the one that has to live with the record it ends up with, and the judgement about what matters is the institution's to make, not ours.

10The next few years

The institutions that keep the record are at a real inflection. The technology now arriving is good enough to be useful and good enough to do harm, and the difference between the two is being settled largely by procurement choices made in the next few years: where the human stays, what is tested, and who controls the record.

The case for optimism is the one most often seen when it goes well. Used on the mechanical side of the line, AI makes the record faster, more searchable, more accessible, and less wasteful of scarce skill, while the interpretive judgement — and the sovereignty — stays with the institution and its people. That is not a compromise between adoption and caution; it is adoption done well. The institutions that get there are the ones that asked the question beneath the question early, while they still had the freedom to choose the answer, and that kept faith with the people who carry the record across the change.

Two centuries ago the record of parliament was set in type by a printer named Hansard, and the words belonged to the parliament. The technology has changed beyond recognition since — quill to shorthand to tape to silicon — and the words still belong to the institution and the people who keep it true. That is the one thing every disruption has left standing, and AI will be no different.

Next in this series: the metadata problem

This is the first of a series. The next CAL Note takes up the problem most often named as the real limit on what AI can do in this work: speaker attribution — diarisation, the question of reliably knowing who said what. It is usually treated as an acoustic problem, to be solved by ever-better models listening to ever-messier audio. We think that is the wrong place to look. The information needed to get attribution right is already being produced in the room. The difficulty has never been that the answer does not exist; it is that it keeps being captured in forms a machine cannot read easily. The next CAL Note will be written for the same readers, in the same spirit: observations from inside the work, offered to everyone who keeps the record, wherever they keep it.