Cross-note synthesis

Key issues across the notes

Recurring issues and themes across the published notes' key insights. 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.

The recurring lessons across the corpus — each led by the claim, evidenced by the notes that support or qualify it. The corpus is small (see the coverage line above), so these are early patterns, not settled findings.

AI's first proven value is translation: it makes gated or opaque knowledge legible. Each note shows the same move in a different register — decoding jargon-heavy hiring processes for applicants without insider networks (Unlocking career potential); turning raw, unstructured material into clear, consistent summaries (AI for Insight Notes); teaching practitioners how the system reasons so they can interrogate its outputs (Growing with AI); reverse-engineering how an inherited software system fits together (AI for the Systems You Inherited); and reducing many fragmented, siloed datasets to a single decision-ready story for operators who would never work the raw sources themselves (Tourism data). The move was there from the program's earliest sessions, too — turning volume into signal for horizon-scanning policymakers (The NSW Trend Atlas), and casting AI as a translator and visualiser of complex civic information rather than its gatekeeper (Building better together; AI and national security). One note turns the same move on the technology itself: demystifying how AI works — tracing it back to the regression and arithmetic practitioners already trust — so a non-specialist can interrogate its outputs rather than defer to them (Beyond the Black Box), which is the precondition for every other translation above.

But the translator is not neutral — its values shift with language and model, which complicates the translation claim above. How AI Speaks Our Values found the same value question posed across languages produced different stances, and that the language of the prompt could matter more than which model answered — so a tool relied on to make gated knowledge legible can quietly re-weight what it makes legible. This was visible from the program's earliest months: a March-2025 session found the same model returning materially different answers depending on the language of the prompt, and read that variability as a signal to attend to rather than noise (How do LLMs behave in languages other than English?); a macroeconomic session named language and culture as infrastructure, warning that First Nations and other minoritised languages are largely absent from training data so the tool is least reliable exactly where inclusion matters most (Macroeconomic implications of AI); and a women-in-AI session noted translation that carries the words but loses the cultural context (Women in AI). Refusals proved a value signal in their own right rather than an absence of data, and safety guardrails shift within days while the underlying model changes over months, so an AI's expressed values drift and need monitoring rather than one-off certification. This sharpens rather than overturns the augmentation boundary that follows: where the other notes ask a person to check an answer's accuracy, this one asks them to notice its standpoint — the same reason Beyond the Black Box reframes models as pattern machines that reflect their training rather than a neutral oracle.

Every note draws the same augmentation boundary — AI accelerates, people stay accountable — but they disagree about whose judgement is the failure mode. Unlocking career potential wants AI to check human judgement: bias in human-led hiring is the documented problem, and structured AI support is part of the remedy. The later notes run the arrow the other way — practitioner instinct must weigh, and sometimes override, automated recommendations (Growing with AI), and a wrong machine-generated account of a system, taken as a foundation, makes everything built on it wrong (AI for the Systems You Inherited). Beyond the Principles turns the boundary into an operational rule: risk-tier the uses, keep anything that renders a judgement about a person human, and name the points in the chain — the operator, the team standing the tool up, the decision-maker who acts on the output — where responsibility actually sits. Getting Hands-On With High-Risk AI turns that rule into a mechanism for systems that judge people: confine the AI to gathering evidence, let deterministic code do the scoring, require a qualified human to attest each piece, and then monitor the assessors themselves for drift — pushing the boundary up a level so the people doing the overseeing are measured too. The portable lesson is conditional, not absolute: where human judgement is the documented weak point, AI is the check on people; where tacit human expertise is the asset, people are the check on AI. When the Algorithm Goes Rogue sharpens the stakes from the public-sector side: when a system makes consequential decisions about people, what makes that boundary hold once the system is live is contestability — a route by which someone affected can be heard by a human empowered to act — and a preserved human backstop, since banking promised savings by cutting the workforce first removes the very capacity a recovery depends on. From Pilot to Production gives the boundary a portable operational shape: split the work into a preparation layer the machine runs ahead of time — gathering, structuring and summarising so a person arrives already informed — and a decision layer the human keeps, under a standing rule that nothing the AI surfaces goes out until a person has read it. Who is accountable when the two disagree remains open — see Open questions.

What AI produces is patterned synthesis, not knowledge or belief — which fixes where the human has to stand. An epistemology session framed model output as pattern-based synthesis that carries no gradation of confidence and does not amount to belief or knowledge (LLM epistemology); a research-methods session put the same point operationally — AI organises existing knowledge rather than originating it, and the human still defines the problem space (Meta-cognition and deeper research). The practical corollary the early rooms drew is that verification and citation are first-class, not optional: a document keeps its accountability role precisely because its provenance can be checked (Document intelligence), and a later session made that backbone explicit — grounding generative output in a knowledge graph so every inference links back to a verifiable source (Designing trustworthy GenAI with knowledge graphs). The same reframe reaches into ownership: where output is patterned synthesis rather than authored work, a copyright session found that content generated without meaningful human input attracts no copyright at all, so it is the degree of human creative control that decides who — if anyone — owns the result (Copyright, IP and GenAI). This is the epistemic floor under the augmentation boundary above — if the output is synthesis rather than a claim to truth, the human is the one who must supply the belief, the confidence and the accountability, which is the same reframe Beyond the Black Box reaches when it recasts models as pattern machines rather than a neutral oracle.

The safest uses are the ones where you never have to audit the reasoning. Practical AI for Policy People draws the line by what the user can verify: a model earns its keep on "face-value artefacts" — a draft, a reworking, an image, a working tool — whose quality can be judged at a glance because the person already owns the underlying content, so nothing hidden has to be trusted. The corollary cuts against the reflex use case: summary and analysis are weaker productivity plays precisely because the reasoning is the product, and to trust it you must reconstruct it — often costing back the time saved. That sharpens the augmentation boundary above into a selection rule — point the tool at work your own expertise makes checkable — and it is the user-facing twin of the build-side discipline of wrapping a probabilistic core in deterministic guardrails (Beyond the Black Box): both refuse to take a model's hidden workings on trust.

Judge AI against the realistic status quo, not against perfection. The fair comparison is the process actually in place, which is rarely as good as it is assumed to be. The human-led baseline carries its own unmeasured bias, opacity and unreliability (Beyond the Principles) — the same point Unlocking career potential makes from the other end, treating documented bias in conventional hiring as the problem AI is brought in to mitigate. Held to "AI versus perfection" almost anything fails; held to "AI versus the status quo" the real question — does this improve on what we already do, and at what cost — comes into view, and can reframe the underlying process rather than merely automate it. Getting Hands-On With High-Risk AI makes the comparison concrete: a transparent, reproducible assessment can beat an opaque human interview precisely because the latter leaves no trail to reconstruct — the status quo's unrecorded judgement is the thing to improve on. Beyond the Hype supplies the discipline that keeps "reframe, don't automate" honest: fix the underlying process first, because automating a broken, twenty-step process only makes the mess run faster. The same instinct shows up as governance: a sanctioned tool with explicit rules and monitoring can beat the ungoverned "shadow" use already happening off the books (Beyond the Principles).

Problem first; grounding and guardrails throughout. Value starts with defining the actual problem rather than force-fitting a tool, and a single well-defined problem usually decomposes into several pieces no one product covers (Growing with AI). Reliability then comes from closing in the probabilistic tool's boundaries: grounding it in the relevant source material rather than its own assumptions (AI for Insight Notes), and constraining it with reference examples, documented standards and a specification to verify against, with human review at the checkpoints (AI for the Systems You Inherited). Beyond the Black Box gives the reason this works at all: a model's output is probabilistic by construction, so "hallucination" is sampling rather than malfunction, and the practical answer is to wrap the probabilistic core in deterministic guardrails — input and output checks, escalation-to-human rules, and tool use that grounds answers in real sources. Context transfer is a known failure mode — results drawn from one setting may not hold in another (Growing with AI). The data layer is part of the same discipline: a general-purpose tool handed fragmented sources it cannot connect or interpret produces little, so the structured foundation comes before the AI, and the design should start from the user's actual question rather than the dataset to hand (Tourism data). The discipline scales to agents that act rather than only answer: anything with real-world consequences — spending, ordering, calling an external registry — belongs in deterministic code at the gateway rather than in a prompt the model may or may not honour, and the data an agent calls should be served exactly as the authoritative source published it, with no generative step in the path that touches it (Authoritative by Source). A production deployment reaches the same rule from the cost-and-reliability side: reserve the model for what only it can do and run every step that can be deterministic as plain automation, then bound the agent with hard limits — capped retries so one transient error cannot drain the budget, and per-user spend ceilings — because open-ended agentic access without boundaries is a liability, not a feature (From Pilot to Production).

When AI acts on the world rather than only advising, trust shifts from the answer to the audit trail. Once an agent transacts on someone's behalf, being right in the moment is no longer enough. Authoritative by Source frames a trustworthy agentic transaction around three answerable questions — who authorised this call (security), where each result came from (provenance, traced to a named authoritative source rather than a confident guess), and can the whole exchange be reconstructed later (audit) — and warns that wiring agents into chains compounds the stakes, because a downstream agent will treat an upstream error as ground truth. This is the build-side complement to the accountability the public-sector notes demand after the fact: the contestability and preserved human backstop When the Algorithm Goes Rogue wants once a system errs are only enforceable if the transaction was authorised, sourced and replayable to begin with — the same instinct as wrapping a probabilistic core in deterministic checks (Beyond the Black Box). Getting Hands-On With High-Risk AI extends the same logic to systems that judge people: an outcome is only contestable if you can reconstruct the whole provenance chain behind it — every piece of evidence, and every mid-process steer by a reviewer, recorded with who made it and when — so a "defensibility bundle" can replay exactly how a rating was reached. From Pilot to Production shows the same trust-shift in a live government deployment: the controls that let agents run against real systems are a single governed path where every call is logged, and verified user identity — the agent acting within the requesting person's own permissions rather than a shared service account — so who did what stays answerable and auditable rather than laundered through a machine account.

As AI mediates how the public reaches official information, provenance becomes a public-trust problem, not only an engineering one. Government information in the age of AI describes AI assistants becoming the default interface to government — confidently wrong on entitlements, eroding agencies' visibility of how their content is used and reframed, and outpaced by misinformation that mimics official sources — so the response it points to is structural: make authoritative content machine-readable and adaptable rather than defend a web page. That is the public-facing face of the instinct the build notes reach from inside, serving data exactly as the authoritative source published it, traced and auditable (Authoritative by Source), and of the enduring principle an earlier futurist session put first — authoritative information and clear provenance, with regulation lagging the technology and a standing caution against repeating past automated-decision failures (What's next for the Australian Government and AI?). The same requirement surfaces in research practice as traceable sourcing and disclosed AI involvement (Qualitative research in action).

Public trust is treated as a precondition for deployment, not a by-product of it. Several 2025 rooms put social licence upstream of the build: health-and-research practice pointed to citizens' juries and treated social licence as foundational (Ethical AI in health and research); a policy-work session read the hesitancy left by a past automated-decision failure as the thing any new tool has to earn its way past (Generative AI in government policy work); an oversight session located the demand for scrutiny in public expectation rather than compliance (Auditing AI); and a planning-engagement session called its domain the place "where trust is made or unmade", warning that a poorly-placed chatbot can widen the very trust gap it means to close (Building better together). A national-security session frames the same currency defensively — resilience against deepfakes and synthetic media as a matter of preserving public trust in what is real (AI and national security). Trust is the thing provenance protects: the machine-readable, traceable sourcing above is the mechanism, social licence the standing it earns.

Who holds the data and the model is treated as a design variable, not a given. A concern raised as a gap in the earliest sessions — the absence of a locally hosted option for sensitive public-sector work (OpenAI 101) — was answered across early 2025 by designs that put data control first: an open-source, on-premise orchestration framework (OpenSI-CoSMIC) and a sovereign small language model built so that data control is non-negotiable (SCOTi Flow). The same instinct shows up as a preference for methods that avoid large external models where assurance matters (Clustering techniques), as a case for sovereign AI over national datasets as critical infrastructure (Women in AI), and, on the public-facing side, as a government-hosted hub with opt-in data-sharing (Building better together). The corpus does not resolve the pull the other way: a risk-practice session leaned on a consumer cloud assistant for government risk work, an unremarked tension with the sovereignty preference the other rooms treat as basic (Practical AI for risk professionals) — see Open questions.

The binding constraints are organisational, physical and economic — not model capability. This has been the canonical finding since the program's earliest public-sector rooms, which located the obstacle in structures, incentives and culture — the risk of otherwise "writing the same report in five years" — and in fragility to a change of leadership, rather than in the technology (AI and the future of the public service, Scaling and sustaining innovation). Bias and opacity in human processes (Unlocking career potential); institutional limits on which tools teams may use at all (AI for Insight Notes, AI for the Systems You Inherited); the field's hardware, connectivity, cost and vendor-lock-in realities plus unresolved data governance (Growing with AI); and sunk-cost pressure to keep failing projects alive (AI for the Systems You Inherited); and, in high-stakes public deployments, decisions locked in upstream of any model — a business case built on point-estimate savings and pre-announced staff cuts, procurement that rewards over-promising, and budget pressure that quietly descopes the guardrails, so the failure is in effect procured before it is coded (When the Algorithm Goes Rogue); fragmented, siloed and sometimes paywalled data, complex data-sharing frameworks and privacy obligations before any useful tool can be built (Tourism data); and shifting, usage-based pricing, dependency and data-sovereignty exposure, and project-by-project rollouts with no overarching people strategy (People and Culture Leadership) — felt at the individual scale too, where metered "slot-machine" pricing, deliberately engaging products and vendor lock-in push experimenters toward local models for cost control and data privacy alike (Vibe Coding for Curious Humans); and the economics of model choice itself — below-cost hyperscaler pricing unlikely to last, and the pull toward a frontier model where a smaller, cheaper, more controllable one would do, which makes matching the model to the task a cost discipline rather than a technical afterthought (Beyond the Black Box); and, in the public sector, adoption slowed less by what the technology can do than by procurement, accreditation, financial delegations and cultural caution — so that data which is public in principle stays practically inaccessible until someone does the unglamorous work of structuring it (Authoritative by Source). Argued in full in Uncovering barriers to practical collaboration.

Most of the corpus works at the scale of a team or an organisation; a smaller cluster lifts the same questions to the national and physical scale — and there the binding input is compute, energy and supply chains. AI and the geopolitics of compute recasts "capability" as access to scarce inputs — advanced semiconductors concentrated in a handful of suppliers, and, increasingly, firm power as the real ceiling on frontier compute — leaving middle powers to compete on narrow, deliberate bets such as hosting data centres rather than on models. Exploring the economics of transformative AI works the same altitude in macro terms: scaling laws make rapid capability gains plausible, but the room's instinct was to interrogate the forecasts — physical bottlenecks, capital cycles, where human labour stays scarce — rather than accept them, the macro cousin of judging AI against the status quo rather than against a clean projection. The productive difference from the delivery notes is one of scale: the same energy-and-cost realities that appear as a model-choice line item in Beyond the Black Box reappear here as determinants of national strategy.

Speed raises the value of judgement — including the courage to stop. When execution accelerates, choosing the right direction, setting the constraints, and recognising when a path is failing become the scarce skills; stopping a failing build is itself a capability, not an admission of defeat (AI for the Systems You Inherited). The same logic appears at small scale as iteration discipline — each cycle refining the instructions that govern the next, not just producing output (AI for Insight Notes) — and at programme scale as the discipline of letting a pilot fail: running one and then proceeding regardless of what it showed is a recurring way large public rollouts come undone (When the Algorithm Goes Rogue). A strategic-foresight session generalises the same instinct to collective decision-making under uncertainty: when the future is genuinely unsettled, the scarce skill is matching the response to the kind of problem you face — naming whether it is simple, complicated, complex or chaotic before acting — and finding the courage to relinquish processes that no longer work rather than carry them forward (Beyond the Hype).

As execution gets cheaper, the binding skill is communication — directing the tool, not operating it. Vibe Coding for Curious Humans found that non-coders and career engineers hit the same failures — vague requirements, missing context, no plan — and that the remedies are the ones that work on people: state the goal precisely, supply the context up front (and keep it in a reusable file so a fix sticks), and break a job into layers so each step has a narrow scope. It reframes working with a capable model as leading a small team rather than running a tool, which is why those who manage people well tend to adapt fastest — the same instinct People and Culture Leadership reaches from the organisational side when it treats adoption as a people-and-culture transformation rather than an IT rollout. It also restates the augmentation boundary in build terms: a generated prototype is a starting point, not a product, and shipping it to real users without a human review of the output was treated as a line not to cross. From Pilot to Production makes the same point at production scale: once execution is cheap, the binding constraint becomes the quality of the specification, so the work — and the scarce skill — moves up front into thinking the problem through precisely rather than into producing the build. AI Productivity in the Agentic Era makes the same case as a whole-role shift: when capable intelligence is abundant, the scarce resource is clarity — of intent, of judgement, of what "good" looks like — so the durable contribution moves from doing to designing, orchestrating and governing outcomes. Because an agent cannot be progressively nudged the way a colleague can, effort moves upstream into specifying intent and downstream into continuous assurance, and the oversight skill becomes holding two things at once: being comfortable not knowing how an outcome was produced while staying accountable for whether it is right.

Adoption is a trust journey that travels peer-to-peer. People adopt at very different speeds; a respected peer demonstrating that something works moves more people than any feature list, and a small, tangible proof of concept builds the confidence to go further (Growing with AI). A concrete proof of concept also works as a provocation that anchors co-design — giving people something real to argue with produces sharper ideas than abstract brainstorming (Tourism data). The workshops themselves model this — see Lowering barriers to safe experimentation.

A distinct strand treats AI less as a productivity tool than as a way to widen who is seen and who can take part. The throughline is "you can't be what you can't see": the WIC Image Equity Challenge used generative tools to put under-represented figures into the visual record, trading precision for participation, so that anyone could join and the model's quirks were part of the craft rather than a barrier. The same widening instinct runs through AI as a leveller for people outside insider networks (Unlocking career potential), as a way to bring community historians and families into research once gated by expertise (Qualitative research in action), and as a means of levelling hierarchy in a room so that quieter contributions reach the shared draft (Human-AI teaming for people and planet). No-code "vibe coding" pushes the same widening from who is seen to who can make: someone without an engineering background can turn an idea into a working prototype themselves, loosening the long-standing dependence on a technical gatekeeper to be heard — while the engineer's depth stays the thing that turns that prototype into something robust (Vibe Coding for Curious Humans). The promise here is access to participation, not just to output — held against the standing caution, from the same diversity-of-employment note, that these tools can also re-encode the very biases they are meant to widen past.

Access is not adoption, and data is not insight. Putting the capability in front of people is the start of the work, not the end. Top-down licence rollouts tend to produce an early spike in use that then fades, because people lack the permission, purpose and confidence to experiment — so a "safe to fail" environment (distinct from "fail safe"), with explicit permission to try and to get things wrong, is the precondition for use to stick (People and Culture Leadership). The data version of the same trap is abundance no one can act on: the binding constraint is rarely a shortage of data but fragmentation, so the opportunity is to democratise insight — an at-a-glance answer to the question someone actually has — rather than handing more raw data to operators with no capacity to use it (Tourism data). A third face of the trap is the tool itself: where staff only ever meet a cut-down enterprise version, a poor first impression hardens into "AI is useless" and forecloses the case for anything more capable, so what an organisation has licensed and what it actually uses drift apart (AI and the changing nature of information work). All reframe the goal from provision to uptake, and connect back to the trust journey above.

The brake on adoption is often comprehension and fear, not capability or access. Where the access notes put the gap in permission and purpose, Beyond the Black Box puts it in understanding: fear narratives — AGI timelines, "smarter than a PhD" claims, weekly scare stories — crowd out people's bandwidth to learn what is in fact publicly knowable, and the antidote is a modest "plumbing-level" literacy (enough to know what is inside, what the failure modes are, and who to call) rather than mastery. This is the same capability gap the delivery notes name from the other side — AI an asset to those who can validate its output and a liability to those who cannot (AI for the Systems You Inherited). The public-sector rooms name the skill that literacy builds toward — critical thinking, treating the human as the teacher who evaluates the model's output rather than the student who accepts it, so AI is most valuable as a thinking partner that contests your reasoning rather than a typist (AI and the changing nature of information work). Practical AI for Policy People reaches the same antidote through a single reframe — that these are "pattern machines, not logic machines", whose accuracy tracks how common a pattern is rather than how hard a task looks — which gives a non-specialist a way to predict where a tool is reliable instead of fearing it wholesale.

Consent and disclosure are design decisions, not afterthoughts. Where there is a power asymmetry, consent to an AI tool is only meaningful if it can be declined without detriment — an opt-out offered at the point of use to someone who cannot really refuse is not consent, so the affected group needs a genuine fallback and a hand in shaping the tool, not a checkbox (Beyond the Principles). Disclosure works the same way: when AI use must be declared is a threshold to design deliberately — distinguishing material contribution from cosmetic help — rather than a blanket rule that over- or under-discloses (Beyond the Principles). This is the consent-and-transparency face of the same "design stage, not final gate" lesson the delivery notes reach from the build side. Transparency can even be turned into a feature rather than a liability: showing people how a system has read them — like a pre-filled return they can correct — makes outcomes more accurate, not merely easier to challenge (When the Algorithm Goes Rogue). Getting Hands-On With High-Risk AI builds the consent side out furthest: consent treated as a granular, revocable envelope chosen layer by layer — to AI participation, to interacting by voice, to who sees the outcome versus the underlying content — with the AI's involvement disclosed up front, so the person keeps informed control throughout rather than signing once at the start.

The conversation is shifting over time. The earliest sessions (early 2025) are dominated by the public sector taking AI's measure — locating the barriers in structures and culture rather than the technology (AI and the future of the public service), asking what kind of knowledge a model even produces (LLM epistemology), probing how its answers shift by language (How do LLMs behave in languages other than English?), and pressing early on sovereignty and who holds the data (SCOTi Flow, OpenSI-CoSMIC). A parallel mid-2025 strand turns to individual access — prompting as a skill, AI as a leveller for people outside insider networks (Unlocking career potential). Across the second half of 2025 the program fans out across registers at once — foundational literacy that frames AI as a sociotechnical system (Understanding AI Systems), agentic tooling and its guardrails (AI Agents in Action), responsible adoption for small business (AI Ready?), AI as a teaming and collective-intelligence amplifier (Human-AI teaming), and the macro and value-laden questions of compute geopolitics, the economics of transformative AI, and how models encode values across languages (AI and the geopolitics of compute, Exploring the economics of transformative AI, How AI Speaks Our Values). By late 2025 the delivery focus is custom agents and structured workflows (AI for Insight Notes); through 2026 it broadens to how an organisation governs AI — designing whole high-stakes public deployments to survive their own failure (When the Algorithm Goes Rogue), applied-ethics deliberation and disclosure (Beyond the Principles), demystifying the technology's own statistical foundations to build the literacy adoption rests on (Beyond the Black Box), reframing agentic productivity as a shift from doing to designing and governing outcomes (AI Productivity in the Agentic Era), multi-agent delivery and specification-driven development (AI for the Systems You Inherited, Growing with AI), and — at the other end of the skill spectrum — no-code "vibe coding" that lets non-engineers prototype working tools for themselves (Vibe Coding for Curious Humans), and the data-and-agent infrastructure underneath it — how authoritative sources are exposed to agents as secure, provenance-tracked, auditable tools (Authoritative by Source). A mid-2026 note lands squarely on the workforce side of that question — treating AI adoption as a people-and-culture transformation to be owned, not an IT rollout to be delivered (People and Culture Leadership). The individual-skills conversation has not been displaced so much as run alongside the organisational one, though — a mid-2026 introductory session still teaches prompting as tasking and where to trust the tool, the same practitioner-skill register as the earliest notes (Practical AI for Policy People). By late June 2026 the program is also circling the hardest governance and human questions head-on: a hands-on case study of a high-risk system — one that makes consequential judgements about people — pushed the accountability thread to its most consequential category (Getting Hands-On With High-Risk AI), while a strategic-foresight session turned the lens onto how groups themselves decide together under uncertainty, foregrounding facilitation and collective sense-making as the human capability the technology most depends on (Beyond the Hype). At the very end of June 2026 a session took the organisational question to its operational end — how an agency actually crosses from pilot to production, with a single governed control plane, verified user identity and audit as the enabling controls rather than model capability (From Pilot to Production). In a year the question has broadened from "how do I use this tool?" to "how do we run an organisation around it?" without abandoning the first — a drift worth tracking as the corpus grows.

If you're starting

The portable propositions the corpus currently supports. Each is falsifiable — future notes may qualify or break them (see Open questions):

  • Define the problem before choosing a tool, and expect it to decompose into pieces no single product covers (Growing with AI).
  • Prove value small and low-stakes first — a tangible proof of concept plus a peer who has gone first moves an organisation further than a strategy document (Growing with AI).
  • Budget more effort for organisational constraints than technical ones — tool restrictions, data readiness, process bias and governance bind before model capability does (AI for Insight Notes, AI for the Systems You Inherited).
  • Invest early in the artefacts that constrain the tool — grounding material, reference examples, standards, specifications — because that is where reliability comes from (AI for Insight Notes, AI for the Systems You Inherited).
  • Decide explicitly who checks whom — whether AI is auditing human judgement or people are auditing AI output in your context (Unlocking career potential) — and agree the stopping rule before you need it (AI for the Systems You Inherited).
  • Compare against the status quo, not perfection, and risk-tier the use — keep judgements about people human, and name where accountability sits before going live (Beyond the Principles).
  • Treat consent and disclosure as design choices — under any power imbalance, ensure a real, non-punitive fallback and decide in advance when AI use must be declared (Beyond the Principles).
  • Engineer recoverability before go-live — for systems that make consequential decisions about people, build in contestability and keep a human backstop rather than banking savings by cutting capacity first (When the Algorithm Goes Rogue).
  • Plan for uptake, not just access — a licence or a dataset is the start, not the finish; budget for a "safe to fail" environment and the purpose, permission and insight that turn provision into use (People and Culture Leadership, Tourism data).
  • Build the data foundation before the AI, and start from the user's question rather than the dataset to hand — structuring authoritative data is usually most of the value, and the "AI step" is often optional (Tourism data, Authoritative by Source).
  • If agents will act, not just advise, enforce the guardrails in code and ground every result — gate anything with real-world consequences (spending, orders, external calls) in deterministic code at the gateway rather than a prompt, and trace each result to a named authoritative source you can audit later (Authoritative by Source).
  • To reach production, make agents act as the user and log everything — route them through one governed control plane, have each act within the requesting person's own permissions rather than a shared service account, keep every action auditable, and cap retries and spend so an agent cannot run away with the budget (From Pilot to Production).
  • Match the model to the task, and build literacy to the "plumbing" bar — a smaller, cheaper, more controllable model often beats a frontier one for a constrained job, and most roles need only enough understanding to know the failure modes and who to call, not mastery (Beyond the Black Box).
  • Decide who holds the data and the model, not only which performs best — for sensitive work, weigh a sovereign or local option against a convenient consumer cloud one, since data control and assurance can matter more than raw capability (SCOTi Flow, OpenSI-CoSMIC).
  • Point AI at "face-value" work first — tasks whose output you can judge at a glance because you already own the content — and treat summary and analysis as weaker plays, since auditing the model's hidden reasoning can cost back the time saved (Practical AI for Policy People).
  • Lead the tool like a team member — state the goal, brief the context up front and keep it in a reusable file, plan before you build, and treat a generated prototype as a starting point that still needs a human review before anyone relies on it (Vibe Coding for Curious Humans).
  • Notice the standpoint, not just the accuracy — a model's expressed values shift with language and version, so treat value-laden outputs as a position to check and re-check over time, not a neutral answer (How AI Speaks Our Values, How do LLMs behave in languages other than English?).
  • If AI mediates access to your authoritative information, structure it for machines — assume people will reach it through an assistant, make the source machine-readable and traceable, and treat confident misinformation as part of the threat model (Government information in the age of AI, Authoritative by Source).
  • If a system measures or decides about people, make its basis contestable — keep the AI to gathering evidence, keep the scoring deterministic with a qualified human attesting each piece, monitor the overseers for drift too, and keep a replayable provenance trail so any outcome can be reconstructed and challenged (Getting Hands-On With High-Risk AI).
  • Under genuine uncertainty, match the method to the problem — name whether you face a simple, complicated, complex or chaotic problem before deciding how to act, deliberately generate optimistic as well as pessimistic futures (groups skew toward the dark ones), and treat the messy human collaboration, not the technology, as the hard part (Beyond the Hype).

Sources

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

Drawn from the key insights of:

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.