Cross-note synthesis

Open questions across the notes

A running register of questions the workshops have raised but not yet answered, and standing expectations for future notes to confirm or break. Public, general synthesis — safe to draw on without checking. It generalises the lessons; the sources below link back to the particular workshops (some members-gated).

56 insight notes so far. Not every workshop becomes one — recurring sessions and smaller gatherings usually don't. For some of the larger workshops a transcript is turned into an insight note so its lessons can be shared — with the room's consent, through time, and with people who weren't there.

What the workshops have raised but not yet answered. Each question names the note that raised it; a question closes when a later workshop answers it, and the answer will be linked here. Below the register sit standing expectations — patterns the current corpus suggests, stated so future notes can confirm, qualify, or break them.

Open questions

  • Who is accountable when a practitioner and an automated recommendation disagree? Raised by Growing with AI (2026-05), which argues the liability question deserves explicit attention rather than being left implicit. Beyond the Principles (2026-05) gives part of the answer in structural terms — name the operator, the implementer and the senior decision-maker, and keep judgements about people human — but the live case where a human and the tool actively disagree is still open. Getting Hands-On With High-Risk AI (2026-06) advances it for the measurement case by designing the disagreement out — the tool never decides, only gathers evidence; deterministic code scores; a qualified human attests each piece — so the accountable human is never in symmetric dispute with the tool. Where a system does issue a recommendation a person must weigh, though, the question stays live.
  • When must AI use be disclosed, and where is the threshold? Raised by Beyond the Principles (2026-05): groups kept hitting the line between material contribution (disclose) and cosmetic help (no disclosure) across recruitment, health information and everyday documents, with no settled rule. Open.
  • How is consent made meaningful under a power asymmetry? Raised by Beyond the Principles (2026-05): when the affected person cannot freely decline (a candidate before a panel, say), what counts as a genuine, non-punitive fallback rather than a nominal opt-out is unresolved. Getting Hands-On With High-Risk AI (2026-06) offers a partial design answer — consent as a granular, revocable envelope chosen layer by layer, with accommodations so people can show up in the way that represents them best — but whether someone facing a hiring decision can truly decline remains the unresolved core. Open.
  • Who shares in the value of the data practitioners generate — and what would make sharing rational? Raised by Growing with AI (2026-05): without benefit-sharing models and common data standards, resistance to sharing is rational, not backward. Open.
  • How should organisations govern how AI agents interact — with each other and with people? Raised by AI for the Systems You Inherited (2026-05). Constraining a single agent is increasingly well understood; governing the interactions is not. Authoritative by Source (2026-06) answers part of it for a single agent's transactions — gate consequential actions in deterministic code, and make each call authorised, provenance-traced and auditable — and sharpens the rest, warning that chaining agents compounds risk because one agent's error becomes another's trusted input. From Pilot to Production (2026-06) adds the production-side answer for the single-agent case — route every agent through one governed control plane, acting under verified user identity within the requesting person's own permissions rather than a shared service account, so each action stays authorised and auditable — while keeping the chain question open by composing narrow, deterministic single-purpose skills rather than trusting agents to coordinate freely. How to govern that chain end-to-end stays open.
  • Where does support come from for practitioners who don't know what they don't know? Raised by Growing with AI (2026-05), which names government and research-and-development bodies as candidates. Open.
  • Can AI reduce bias in human-led selection without importing the bias in its own training data? Raised by Unlocking career potential (2025-07), which flags both directions at once: AI as a check on biased processes, and bias in the data behind the tools. Open — no later note has tested it.
  • How do teams keep AI-written artefacts — documentation, instructions, specifications — current, rather than drifting the way legacy code did? Raised by AI for the Systems You Inherited (2026-05). Open.
  • What makes a contestability or feedback loop actually reach someone empowered to act on it? Raised by When the Algorithm Goes Rogue (2026-03): the recurring failure mode is that signals from affected people never reach a decision-maker — sometimes because someone has an incentive to suppress them — and what structural design reliably prevents that is unresolved. Open.
  • What kind of explanation actually lands with a non-technical decision-maker? Raised by Beyond the Black Box (2026-03): mathematically sound explainability techniques exist but can be more complex than the prediction itself, and procured black-box tools resist them. The note offers a partial answer — a chain-of-trust model in which each layer verifies the one below rather than the decision-maker absorbing the maths — but what reliably works for a high-stakes decision is unresolved, and it connects to the contestability gap above.
  • What reliably converts access into sustained adoption? Raised by People and Culture Leadership (2026-06): licence rollouts spike then fade without permission, purpose and a "safe to fail" environment — and Tourism data (2026-03) shows the same gap on the data side, where abundance does not become insight for operators without capacity to use it. What specifically closes the gap, beyond naming the conditions, is open.
  • How do organisations capture AI productivity at junior levels without breaking the experience pipeline? Raised by People and Culture Leadership (2026-06): AI lets junior staff produce sophisticated work faster, tempting cuts to entry-level roles — but experienced judgement is built through those roles, and some sectors that cut them have reversed course. AI and the changing nature of information work (2026-03) offers a partial reframe — move everyone "up a layer" so the entry-level skill becomes reviewing and directing a draft, and rebuild rather than delete the formative pathway — but how to build that reviewing capability where the discarded first-draft work was itself the only training remains open.
  • Who owns AI workforce transformation when it sits across IT, HR and leadership? Raised by People and Culture Leadership (2026-06): treating AI as IT's responsibility, or a single AI officer's, leaves the hardest workforce and culture questions unowned. Which function actually carries it, and with what mandate, is open.
  • Whose values should an AI express, and how should value drift be monitored? Raised by How AI Speaks Our Values (2025-11): models are not neutral, their stance varies by language and model, and safety guardrails shift faster than the underlying weights — so what baseline of values is the right one, and what infrastructure tracks how a model's stance changes over time, is open.
  • How does government stay discoverable and authoritative when AI assistants are the default interface to it? Raised by Government information in the age of AI (2025-12): users increasingly get official answers via AI rather than official sites, agencies have lost visibility of how their content is reused, and misinformation outpaces correction. Authoritative by Source (2026-06) answers part of it — expose authoritative data as machine-readable, provenance-traced, auditable tools — but how to rebuild public trust and visibility at the interface stays open.
  • Can a middle power convert hosting compute into genuine strategic influence — and at what energy and sovereignty cost? Raised by AI and the geopolitics of compute (2025-12): with chips and frontier models out of reach, hosting data centres is floated as the available lever, but whether it yields real influence, given grid limits and dependency on foreign labs, is unresolved. Open.
  • How should multilingual output variability be contextualised so it improves advice rather than degrading it? Raised by How do LLMs behave in languages other than English? (2025-03): the same model returns materially different answers depending on the language of the prompt, which is a useful signal to a user who knows to look for it and a hidden failure to one who does not — what practice reliably turns that variability into better rather than worse advice is unresolved. Open.
  • How is the preference for sovereign or local AI to be reconciled with the pull toward convenient consumer cloud tools? Raised across OpenSI-CoSMIC and SCOTi Flow (2025-04), which treat data control as non-negotiable, against Practical AI for risk professionals (2025-06), which leaned on a consumer cloud assistant for government risk work: when convenience justifiably overrides sovereignty, and who decides, is unresolved. Open.
  • What makes public-sector innovation survive a change of leadership or machinery of government? Raised by Scaling and sustaining innovation (2025-06): initiatives are pulled back into a "gravity well" of bureaucratic rationality and left fragile when a sponsor moves on — what structural design outlasts the individuals who start it is unresolved. Open.
  • Are AI chat transcripts discoverable business records, and how should FOI, records and classified-information rules treat them? Raised by AI and the changing nature of information work (2026-03): senior public servants hold their hardest problems back from AI tools while this is unclear, so the most consequential work stays away from the tools until the guidance is settled. Open.

Standing expectations

  • The trust-journey adoption pattern will recur beyond agriculture — peer demonstration beating feature lists, small proofs preceding strategy (Growing with AI). Partly confirmed: a third, unrelated domain reframes the same pattern — adoption as adaptive change that needs trust and a "safe to fail" environment, where access alone produces only a fading spike (People and Culture Leadership) — and a concrete proof of concept again anchors co-design (Tourism data). The new sharper claim to test is that access is not adoption — provision rarely becomes sustained use without deliberate design. Confirmed again by AI and the changing nature of information work, where licensed tools go unused and a cut-down deployment convinces sceptics the technology is useless.
  • Organisational constraints will outlast technical ones in every domain the program touches — each new note should find tool restrictions, governance and process friction binding before model capability does (Uncovering barriers to practical collaboration). Confirmed again by the data-estate constraints of Tourism data, the workforce, cost and dependency constraints of People and Culture Leadership, the literacy, fear and model-cost constraints of Beyond the Black Box, and the procurement, accreditation and data-readiness constraints of Authoritative by Source — all binding ahead of model capability. From Pilot to Production is the most direct confirmation yet: its whole thesis is that the hard part is no longer building but finding, adapting and governing, with a security checklist, procurement posture and identity model — not model capability — deciding what reaches production. Confirmed from the outset, too: the program's earliest public-sector rooms already located the obstacle in structures, incentives and culture — the "gravity well" of bureaucratic rationality, fragility to a change of leadership, and access as a whole-of-organisation IT and data-sharing hurdle — rather than in the technology (AI and the future of the public service, Scaling and sustaining innovation, The NSW Trend Atlas), so this is a long-standing finding the later notes extend, not a recent discovery.
  • The augmentation boundary will keep moving — tasks the current notes treat as judgement work (architecture trade-offs, interpretation, quality assurance) will come under the same automation pressure translation did; future notes should show the line being renegotiated, not fixed (Key issues across the notes).
  • "Compare against the status quo, not perfection" will travel — future notes from other domains should keep finding that the fair benchmark is the flawed existing process, not an ideal one, and that AI's case often rests on beating a baseline that was never as good as assumed (Key issues across the notes). A note that holds a tool to an absolute standard instead would qualify it.
  • In high-stakes public deployments, the decisive failures will keep being set upstream — in the business case, procurement and incentive design, before any model is built (When the Algorithm Goes Rogue). Future public-sector notes should keep locating the leverage points before the build, not in the model; a note where the model itself was the decisive failure would qualify it.
  • Right-sized models will keep beating frontier defaults for constrained tasks — Beyond the Black Box argues smaller, cheaper, more controllable models often win on cost-benefit, especially as below-cost hyperscaler pricing corrects; future delivery notes should keep deciding model choice on fit and cost rather than capability alone. A note that genuinely needed a frontier model for a constrained task would qualify it.
  • The durable asset will keep being structured, authoritative data — not the model or the protocol (Authoritative by Source), echoing Tourism data's "structure the foundation before the AI". Future data-infrastructure notes should keep finding the hard, lasting work in making data machine-callable and trustworthy rather than in the AI layer on top; a note where a model or interface choice, not the data, was the decisive asset would qualify it.
  • AI will keep showing up as non-neutral, not just inaccurate — How AI Speaks Our Values found expressed values varying by language and model and drifting as guardrails change; future notes that probe model behaviour on value-laden tasks should keep finding standpoint and drift rather than a neutral default, and the practical advice should keep moving from "check the answer" toward "notice and monitor the stance". A note that found a genuinely value-neutral model would qualify it. Confirmed earlier than first thought: a March-2025 session had already found a model's answer shifting with the language of the prompt (How do LLMs behave in languages other than English?), so the standpoint-not-just-accuracy pattern predates the later values-focused notes rather than emerging with them.
  • The macro constraints will keep being read structurally and sceptically — AI and the geopolitics of compute and Exploring the economics of transformative AI both located the limiting factors in physical and economic substrate — energy, supply chains, capital cycles, labour bottlenecks — and interrogated headline forecasts rather than accepting them. Future macro-scale notes should keep landing on substrate-and-incentives over raw capability; a note that treated model capability as the binding macro constraint would qualify it.
  • A convergence check on ourselves: if notes from new domains stop landing on "augmentation, not autopilot" — or stop keeping consequential judgements about people human, as Beyond the Principles does — that is signal, not noise: it would suggest the earlier unanimity was partly an artefact of the shared note-writing rubric (index). Still holding: Practical AI for Policy People lands on the same boundary from the individual-skills side — "augment, not offload", an "80% product" the human completes and stays accountable for, and never sending on what you have not read — a convergence worth watching precisely because it is so rubric-friendly. Vibe Coding for Curious Humans lands on the same boundary from the no-code builder side — a generated prototype is explicitly a starting point, not a product, and shipping it to real users without a human review of the output is treated as a line not to cross. Getting Hands-On With High-Risk AI is the strongest confirmation yet from the highest-stakes end — a system built expressly so that judgements about people stay with an accountable human, the AI confined to gathering evidence — and, being so rubric-aligned, exactly the kind of convergence to keep watching rather than take as vindication. From Pilot to Production lands on the same boundary from the production/government side — a "night shift" in which the machine only prepares and a human always decides, under a standing rule that nothing the AI surfaces goes out unread — again rubric-friendly, so more a convergence to watch than a vindication.

Sources

Last compiled 2026-07-03 from 56 published note(s).

Questions and expectations drawn from:

Shared with the AI CoLab Alliance — see join.aicolab.org.
This note was synthesised from the workshop recording with AI assistance and human review. It captures general lessons only, is not attributed to individual participants, and was reviewed for sensitive content and shared with the presenters before publication.