Where Is AI Necessary and Can It Help Australians Live and Decide Well?

Executive Summary

The direct provisional answer

AI can help Australians in some tightly bounded tasks. On the evidence available to 22 August 2026, however, this report does not find AI necessary in any of the six functions transferred from Report 1.

That is not a finding that AI is useless. In anticipation and forecasting, machine-learned models can extend the range and speed of pattern detection. In cyber defence, AI can help defenders find and prioritise machine-speed anomalies. These are credible additions to human-led systems. But the evidence does not yet show that the strongest feasible non-AI package fails and that AI alone closes a material outcome gap. Australian operational evidence commonly reports forecast skill, administrative usefulness, technical capability or trial activity—not avoided harm, continuity, better decisions or durable public outcomes attributable to AI.

For climate, infrastructure and logistics; care and living well; information judgement; and deliberation, the responsible answer is unresolved. There are plausible mechanisms and promising studies. There are also strong non-AI alternatives, weak Australian head-to-head outcome evidence, important distributional risks and incomplete accounts of energy, water, materials, security, sovereignty and dependence.

Why usefulness is not necessity

"Useful" means AI improves something—perhaps speed, range or convenience—while the comparator can still meet the required outcome, or another feasible approach can match the contribution. "Necessary" is much harder. The inherited comparator must fail; a specified AI mechanism must make the material difference; that difference must be attributable rather than merely associated with adoption; and the result must survive every mandatory gate. A Report 1 comparator gap is therefore a research question, not evidence for AI.

The strongest rival is substantial. Australia can improve outcomes through capable institutions, workforces, maintenance, redundancy, public services, social protection, conventional modelling, secure system design, journalism, media literacy, skilled facilitation and accountable decision connections. Those reforms receive the same good-faith assumptions and outcome tests as AI.

Can AI help people live well?

The relevant standard is not whether a tool is impressive or popular. Living well includes agency, security and basic needs, learning and decision ability, mobility, health, relationships, purpose and contribution. AI may improve access, reduce some administrative burden or provide assistive support. It may also displace human contact, weaken skill and independent judgement, expose sensitive data, or make essential services dependent on vendors and networks. Australia's current public evidence does not establish a population-level AI-over-comparator gain across these dimensions, particularly for older people, disabled people, low-income households, regional and remote communities, women, carers and people outside standard employment.

Can AI strengthen democratic judgement and civic capacity?

There is a genuine positive signal. A 2024 peer-reviewed UK study involving 5,734 participants found that an AI mediator could draft group statements participants preferred to human-mediated statements while incorporating minority critiques. That tests a narrow mechanism: synthesising reasons and possible common ground. It does not establish representative Australian participation, durable political influence, resistance to provider or model bias, or better formal decisions.

Australian institutional evidence points the same way. Provenance tools may give useful context, but do not establish whether content is true. Current election law does not generally prohibit AI-generated campaign content, and official guidance combines authorisation rules with public literacy. Deliberation can only strengthen democracy if membership, evidence, conflict, moderation, dissent and reason-giving rules are transparent and a named decision owner publishes a reasoned response, action owner, milestone and feedback loop.

Evidence that a shorter working week can improve wellbeing does not show that released hours become sustained civic participation. A genuine time dividend requires reduced normal paid hours without lost ordinary pay, hidden overtime, compression or unsafe intensification. It also requires usable time after care and recovery, accessible routes to influence, and equivalent support for people outside standard employment. That causal bridge remains unresolved.

The gates prevent an easy positive average

No function passes every gate on current evidence. Mission-specific Australian lifecycle accounts for energy, water, materials and emissions were not located. Data-centre demand is material, and Australian governments now treat energy, water, social licence and national interest as explicit issues, but generic sector evidence cannot determine the footprint of a specified mission.

The same discipline applies to agency, distribution, democratic and information resilience, and security and sovereignty. A benefit for an average user cannot hide a failed group, a service outage, manipulation, unsafe dependence or exported harm. The crossed result is consequently modest: two functions show bounded useful contribution with partial legitimacy and endurance; four remain unresolved; none is demonstrated necessary.

What goes to Report 3

Report 3 receives function-specific results, the controlled instrument updates, the unresolved gate states, the formal-response requirement for community processes, a zero-national-weight N-of-1 protocol disposition, and named delivery, safeguard, shock and milestone questions. It must test whether any bounded benefit can actually be delivered and endure. It may not convert "promising", "useful" or "unresolved" into necessity without the missing evidence.

How to Read This Report

The report follows seven numbered passes. Pass 0 preserves what Report 1 established. Pass 1 specifies the six AI mechanisms and the attribution test. Pass 2 asks whether AI is necessary or only useful. Pass 3 deducts full costs. Pass 4 tests living well and independent judgement. Pass 5 tests democracy, information and civic capacity. Pass 6 gives the function classifications and the exact Report 3 handoff.

Claim IDs in square brackets resolve to the single canonical evidence-ledger.csv. They allow an auditor to inspect the exact source location, verification state, Australian transfer limit, counterevidence, confidence and owner. UNRESOLVED is a substantive result: it means the search did not support a stronger claim, not that the question was overlooked.

The four pathways P1–P4 and conditions S0–S5 are inherited analytical worlds and stress conditions, not forecasts. This report does not assign them probabilities or rank them finally.

Confidence Scale

The Direct Question, Direct Provisional Answer and Scope

Question: Where is AI necessary, and can it help Australians live and decide well?

Direct provisional answer: No tested function is demonstrated to require AI. AI is useful but not shown necessary in bounded anticipation/forecasting and cyber-defence tasks. Climate, infrastructure and logistics; care and living well; information judgement; and deliberation remain unresolved. AI may help people live and decide well when it is assistive, contestable, reversible, accessible and subordinate to human authority. The present evidence does not establish that benefit across Australia or across all mandatory gates.

The unit of judgement is one eligible transferred function in its particular risk, pathway, domain and condition—not "AI" in general. The six functions are anticipation and forecasting (FANT); climate, infrastructure and logistics (FCIL); care and living well (FCAR); cyber and AI-driven threats (FCYB); information judgement (FINF); and deliberation (FDEL).

Report 2 owns additionality, attribution, T2–T6 and the human/civic gates. It does not redesign the accepted baseline, comparators, interaction eligibility, P1–P4, S0–S5 or backcast. It does not settle delivery feasibility, safeguard ownership, final pathway ranking, or the portfolio or trilogy verdict.

Pass 0 — What Report 1 Established

Report 1 found that Australia has substantial institutions and feasible non-AI reforms, but the public evidence does not show adequacy across every mandatory domain and interacting shock to 2035. That is a comparator gap. It is not an AI finding.

Eight accepted items cross the boundary unchanged:

Report 1 screened 24 risk-function cells. Three are not transferred because their interaction eligibility is absent: RWAR–FDEL, RCLM–FDEL and RAGE–FCYB. The other 21 remain eligible for this report's tests. They do not all receive the same evidence or result.

The inherited comparison is stronger than the status quo

The comparator is not "do nothing". It includes capable institutions, workforce, public services, social protection, regulation, engineering, maintenance, redundancy, conventional models, rules-based systems, secure design, journalism, media literacy, facilitated participation and formal accountability. Existing embedded AI in 2026 stays in the launch baseline. This prevents a false contest between a modern AI system and an artificially weakened or technology-free Australia.

The evidence rule is symmetric. If a conventional system cannot claim success from adoption, accuracy or a plan alone, neither can AI. Both must show the same outcome, continuity, distribution, implementation and gate evidence. Current Australian audits illustrate why: the ANAO found AI use at IP Australia largely effective, yet benefits were inconsistently defined or measured and alternatives were incompletely documented. Treasury's Copilot evaluation found administrative usefulness for many users and accessibility benefits, but clear limitations for complex tasks and sensitive information. These are useful findings, not necessity findings.

The 21 eligible cells remain separate

No result is carried from one cell to another merely because the function code is shared. The function classifications in Pass 6 synthesise the cell evidence conservatively: a weak or failed eligible cell cannot be hidden by a stronger one.

What this means

Australia's current systems leave important gaps, but a gap is not a blank cheque for AI. AI must beat a serious alternative on the same outcome measure, in the actual setting, without failing a group or gate.

Pass 1 — AI Mechanism and Attribution

What counts as AI here

AI means a named mechanism whose material contribution depends on statistical learning, a learned model or another AI-dependent system. A database, digital form, conventional numerical model, fixed optimisation rule or ordinary workflow automation does not become AI because it is new or computerised. This distinction matters because digitisation and institutional reform may produce much of the claimed benefit at lower dependency and ecological cost.

Each function requires an exact task, owner, data and infrastructure, human authority and override, fallback, dependencies and failure modes:

The attribution chain

T2 asks whether AI makes a material attributable contribution over the exact comparator. The required chain is:

specified mechanism → changed capability → changed human or institutional decision → changed outcome or continuity → distribution and gate result.

A provider benchmark can support the first arrow. It cannot establish the last three. A forecast may be more accurate but too late, poorly communicated or ignored. A writing assistant may save minutes while adding verification time. A cyber model may detect an anomaly faster while creating an unsafe automatic response. A deliberative summary may attract approval without improving representation, formal uptake or legitimacy.

Australian operational evidence is clearest about this distinction. The Bureau of Meteorology publishes forecast performance and explains the combined role of observations, numerical models and meteorologists. Its research now compares and calibrates machine-learned forecasts, including the AIFS system, but this establishes model development and forecast skill—not avoided losses or AI indispensability.

The national cyber guidance is similarly careful. AI may speed and prioritise detection and response, but it is not a replacement for security fundamentals; outputs should be verified, humans should oversee consequential action and any autonomy should be narrow and reversible. Mature defensive practices remain the first resilience layer even as frontier systems increase attacker capability.

Confounding and counterevidence

Every positive observation must ask what else changed: staffing, budget, data quality, management, process redesign, training, service demand, participant selection, novelty and vendor support. Jurisdiction, population, setting, scale and time also matter. A controlled UK online experiment is not an Australian national institution. A single agency evaluation is not a whole public service. A model benchmark in stable conditions is not continuity under S1–S5.

This report treats negative and null evidence symmetrically. It also distinguishes absence of evidence from evidence of no effect. Where the exact Australian head-to-head result was not located, the evidence record documents the reproducible search and the result remains unresolved.

What this means

AI receives credit only for the part of an outcome that its mechanism actually caused. Faster, more accurate or more convenient can be useful, but those features do not prove that the final decision or public outcome improved.

Pass 2 — Is It Necessary or Only Useful?

The classification rule

The additionality axis has four values:

• non-AI sufficient: the comparator meets the outcome, so AI is not necessary;
• AI useful: AI improves a result, but the comparator meets the need or another feasible approach can match it;
• AI necessary: the comparator fails and only the specified AI mechanism closes the material, attributable gap; and
• adverse or unresolved: harms erase the gain, attribution fails or evidence is inadequate.

The legitimacy and endurance axis is separate: fails, partial or passes. No function can be called necessary until the evidence has been tested on both axes.

Anticipation and forecasting

The AI case is strongest where observations are too numerous or relationships too complex for a human team to inspect directly. Machine-learned weather or anomaly models may offer earlier or more detailed signals. But the comparator is already a hybrid of instruments, physics-based models, statistical methods, experienced forecasters, exercises, communication systems and named decision owners. Australia's public forecast record measures performance for that system, not an AI-versus-no-additional-AI outcome.

The material endpoint is not forecast accuracy alone. It is whether an accountable actor receives, understands and acts on a better warning in time to reduce harm or preserve continuity. The searches located Australian AI forecast research and operational system performance, but not a mission-level head-to-head study attributing avoided loss or improved public decisions to the AI component.

Climate, infrastructure and logistics

Prediction and optimisation can help schedule maintenance, route supplies and restore services. Yet the strongest comparator—asset knowledge, engineering standards, redundancy, preventive maintenance, stockpiles, trained controllers, mutual aid and manual restoration—is itself central to resilience. AI cannot compensate for an unfunded repair, missing spare, single supply route or weak emergency authority.

The Australian search found planning, analytics and data-centre demand evidence, but no mission-specific national comparison showing that an AI intervention improved infrastructure continuity after lifecycle and dependency costs. The current Productivity Commission care-economy work is a useful rival signal: regulatory alignment, collaborative commissioning and prevention can improve care-system efficiency without making an AI claim.

Care and living well

Assistive interfaces, documentation support and scheduling may reduce friction or expand access. That possibility deserves testing. The endpoint, however, is functional ability and a life worth living—not transactions completed. Australia's aged-care strategy and current programme reporting describe digital and AI trials, but the search did not locate a national comparator-grade study showing better autonomy, health, relationships, contact, purpose or contribution over a strong workforce, housing, prevention, accessibility and community-care package.

The live URL registered for EXT-010 no longer resolves to the original 2024 early-childhood report; it now resolves to a different 2025 Productivity Commission care report. This report does not silently substitute one identity for the other. The original lead is therefore qualified, and the new report has its own source ID.

Cyber and AI-driven threats

Cyber defence presents the clearest time-scale argument. AI-enabled attacks may operate at a pace that manual inspection cannot match, while learned detection can triage large event streams. The additional capability is credible. Necessity still cannot be inferred from threat speed. Secure design, segmentation, identity controls, patching, logging, rehearsed response, human incident command and offline recovery remain the strongest comparator and the basis on which AI operates.

Current official guidance supports bounded augmentation and the collection of operational measures such as time to detect and respond. It does not establish an Australian controlled outcome showing that the AI component alone prevented material harm under S4 or S5.

Information judgement

AI can retrieve material, compare sources and prompt a user to consider uncertainty. It can also generate fluent falsehoods, make selective framing feel authoritative and encourage people to offload judgement. The International AI Safety Report records persistent hallucination, brittleness and over-reliance risks, and notes that deepfake detectors perform substantially worse on real-world material than on their test benchmarks.

Australian election and cyber authorities use a socio-technical approach. Authorisation rules, source context, provenance, literacy, delayed sharing, correction and accountability matter together. Content credentials can establish aspects of provenance but not truth, and their absence cannot safely be treated as proof of deception. The search did not locate Australian population evidence that AI assistance improves consequential judgement over well-designed non-AI source protocols without increasing dependence or manipulation.

Deliberation

The Habermas Machine experiment is material because it tests a specific AI mechanism rather than a general aspiration. Participants preferred the AI-generated group statements; external judges rated them more informative, clear and fair; and minority critiques were incorporated. But most participants were recruited online in the United Kingdom, and the result concerns a bounded deliberative process. It does not show that Australian participants were representative, that a formal owner acted, or that an improved decision endured.

The strongest non-AI comparator—representative selection, independent facilitation, accessible participation, published evidence and dissent, a named owner and a reasoned response—can itself produce legitimate deliberation. OECD work shows ways to institutionalise deliberative processes, but institutional design evidence is not a causal demonstration of AI necessity. No Australian AI-assisted process satisfying the complete representativeness, formal-response and durability chain was located.

What this means

The strongest current cases for AI are additions to functioning human systems, not replacements for them. None yet clears the much higher bar of being the only feasible cause of a material public outcome.

Pass 3 — Full Net Benefit

T3 asks whether the whole intervention leaves Australia better off than the comparator after all material costs and displaced alternatives are counted. A digital service is not weightless.

Ecological lifecycle

The International Energy Agency's Energy and AI work establishes that data centres have material electricity implications and that outcomes depend on location, technology and energy systems. It is a global sector analysis, not a lifecycle account for any of the six Australian mechanisms.

Australian evidence makes the issue decision-relevant. AEMO's 2026 planning material identifies data centres as a rapidly growing source of demand and reports scenarios in which they account for about 12 TWh, or 6 per cent of grid electricity, by 2030 and about 34 TWh, or 12 per cent, by 2050. Government expectations now explicitly cover energy transition, water security, social licence and national interest. These figures concern a sector, not the marginal footprint of a forecast, care or deliberation system.

The necessary mission accounts—energy, water, materials, emissions, rebound and exported effects against the non-AI alternative—were not located. The ecological veto is therefore unresolved for every function, even where a low-compute design could plausibly be modest.

Security, sovereignty and continuity

Data location, model and cloud ownership, update control, supply concentration, support access and offline operation matter. S5 explicitly tests outage, compromise or withdrawal of a major model, cloud, vendor or compute service. A function is not resilient if its apparent benefit disappears when the dependency most matters.

The Commonwealth's Responsible AI policy v2.0 creates useful governance requirements for covered entities: accountable officials, public transparency statements, strategy, use-case registers, training and impact assessment. It is a requirement, not outcome evidence, and its coverage is not identical across defence and intelligence contexts.

Implementation, labour and opportunity cost

AI requires data preparation, integration, procurement, training, monitoring, challenge and fallback capacity. Those costs are often omitted from a demonstration. Jobs and Skills Australia's national study expects generative AI more often to augment tasks than replace whole occupations, but it is prospective and does not show a shorter week or civic dividend.

Every intervention also displaces another use of money, skilled people, electricity and political attention. A model that improves a scheduling metric may still lose to more nurses, maintainers, analysts or facilitators. Productivity Commission modelling describes possible productivity gains from data and digital technology, while making uncertainty and enabling reforms explicit. It does not establish function-specific necessity.

Inequality, manipulation and cognitive offloading

Digital access is uneven. The 2025 Australian Digital Inclusion Index reports especially high exclusion among people aged 75 and over and persistent gaps associated with disability, low income, public housing, regional location and mobile-only access. It is a non-government index and cannot settle causality, but it is a strong warning against assuming universal access.

People who must spend more time correcting errors, navigating inaccessible systems or securing human review do not receive the same benefit as confident, well-connected users. Averages must not hide that transfer. Nor can an efficiency gain be counted twice—once as institutional productivity and again as human free time—without showing where the saved capacity actually went.

What this means

AI's net benefit is what remains after infrastructure, oversight, dependence, exclusion and the best forgone alternative are counted. On current evidence, those accounts are incomplete enough to block a clean pass for every function.

Pass 4 — Living Well and Independent Judgement

Living well is a bundle, not a satisfaction score

The World Health Organization defines healthy ageing through functional ability: meeting basic needs; learning, growing and making decisions; being mobile; building and maintaining relationships; and contributing to society. This report adds the locked requirements of autonomy, security, health, purpose and independent judgement.

An AI intervention must therefore report its effect across at least:

1. autonomy and agency;
2. security and basic needs;
3. learning and decision ability;
4. mobility;
5. health;
6. relationships and human contact; and
7. purpose and contribution.

It must also record displaced contact, care, skill and judgement. A person completing an online task faster may still be worse off if there is no non-digital channel, errors are hard to contest, or a valued relationship has been replaced by a synthetic interaction.

Australian social context

AIHW evidence shows that loneliness and social isolation are already material and uneven in Australia. Its current synthesis cannot identify an AI effect, but it shows why human contact is an outcome, not an optional feature. AIHW's older-Australian material similarly describes social-support patterns without establishing a technology effect.

People with disability experience substantially lower social participation and higher isolation than people without disability. This strengthens the requirement for accessible design, supported decision-making and human fallback; it does not imply that a digital tool will close the gap.

The care comparison

The fair comparison gives both sides their strongest form:

The AI case wins only if it improves a functional outcome or continuity over that comparator and does not displace valued contact or independent judgement. Current aged-care AI material is mainly strategy, pilot or activity evidence. The required Australian outcome comparison remains unresolved.

Independent judgement

Human authority is not satisfied by a nominal button labelled "override". The person must be able to understand the decision, obtain the reasons and evidence, decline or contest the AI path, reach a competent human, and continue the essential task when the system is unavailable. Staff need time and authority to disagree. If verification burden consumes the saved time, or repeated reliance weakens skill, the claimed gain is reduced.

The International AI Safety Report's evidence on confident errors and automation bias supports expert oversight and high-stakes caution, but it also warns that mitigation is context-dependent. Human-in-the-loop language is therefore a design proposal until real override and fallback use are measured.

Distribution cannot be an appendix

The seven inherited lenses—age, disability, income, geography, gender, employment and caring—apply to each material result. The 2024 ABS Time Use Survey provides a current base for unpaid work, care, volunteering and feeling rushed. It cannot show what AI caused, but it prevents a nominal time saving from being treated as equally usable by everyone.

No function has current Australian evidence resolving every group and important intersection. The distribution gate is partial or unresolved, never silently passed.

What this means

Helping someone live well means more than making a transaction quicker. AI has to leave the person more capable, secure, connected and able to judge—not merely more dependent on a smoother screen.

Pass 5 — Democracy, Information and Civic Capacity

Democratic renewal is both condition and outcome

The democratic test covers evidence judgement, reason quality, preservation of viewpoints and minority reasons, representativeness, manipulation resistance, formal uptake, feedback and legitimacy. A legitimate process may help an intervention endure; improved voice and collective action may also be outcomes. The same improvement cannot be counted twice.

OECD's Australian trust survey reports a large perceived voice gap and significant differences by gender and financial security. The evidence concerns perceptions in 2023, not an AI effect, but it shows that representativeness and influence cannot be inferred from participant counts.

Information judgement is not content detection

Australia's current approach recognises that no universal detector can certify truth. The AEC requires authorisation for relevant electoral communications and advises voters to stop and consider sources, while the statutory misleading-information offence is narrower than a general ban on false campaign speech. ASD's content-credentials guidance describes provenance as context: credentials may show aspects of origin and editing, but not whether a claim is true or why it was created.

An AI judgement aid must consequently make sources, uncertainty, correction and contest visible. Its outcome is a better human decision, not more content processed. The democratic/information veto remains unresolved where hallucination, persuasive framing, manipulation or offloading could materially worsen judgement.

OWS-style communities: a bounded process, not automatic authority

A community process is eligible only when it is voluntary, issue-bounded and representatively composed. It must publish:

• purpose, membership and selection method;
• affected-interest, demographic, geographic and viewpoint composition;
• conflicts of interest;
• evidence and source rules;
• independent moderation and correction rules;
• reasons, dissent and minority critiques; and
• the exact relationship to a named formal decision owner.

The formal connection is a published response that accepts or rejects recommendations with reasons, identifies an action owner and milestone, and provides a feedback loop. OECD participation guidance supports reasoned public response and follow-up as a design requirement. Evidence that deliberative models can be institutionalised is supportive but international and does not prove Australian AI-enabled delivery.

Automatic binding authority is excluded. In Australia, a proposal for a national citizens' assembly on housing illustrates the institutional gap: a formal parliamentary route can still end without the requested executive process or response machinery. It is a counterexample, not a general claim that deliberation cannot work.

The search found no Australian AI-assisted OWS-style process that combined representative membership, transparent evidence and dissent, a named owner, a published reasoned response, action milestone, feedback and durable outcome.

Civic time and equivalent capacity

A genuine shorter week means reduced normal paid hours with ordinary pay—not compressed hours, hidden overtime or unsafe intensification. Peer-reviewed multi-organisation evidence reports improvements in burnout, job satisfaction and aspects of health, but the trial design and follow-up do not establish civic conversion.

The Australian Parliamentary Library's 2026 synthesis distinguishes reduced hours from compression and records mixed evidence on productivity, costs, sustainability and equity. The Australian search did not locate causal evidence that reduced paid hours create sustained, representative participation or formal influence.

Usable civic time is what remains after care, travel, household work and recovery, combined with material security, accessibility, skills and a route to influence. Equivalent capacity is required for carers, retirees, disabled people, unemployed people, insecure workers and others outside a standard full-time job. The ABS Time Use Survey can establish baseline allocation; it cannot supply the causal bridge.

The strongest rival, applied symmetrically

The rival says democratic renewal comes from trusted institutions, independent journalism, education and media literacy, representative selection, skilled facilitation, accessible participation and binding accountability—not from AI. This package receives the same measures as the AI-assisted case: evidence judgement, reason quality, minority preservation, representativeness, manipulation resistance, uptake, feedback and legitimacy.

AI earns additionality only if its specific mediation or information mechanism improves those outcomes without weakening the rest. The UK experiment supports a narrow common-ground mechanism. It does not defeat the institutional rival on Australian formal influence or endurance.

What this means

Better summaries and more spare hours do not by themselves renew democracy. People need fair access, protected disagreement and a real decision-maker who must answer with reasons and action.

Pass 6 — Report 2 Answer and Exact Report 3 Handoff

Function results

The two axes are assigned only after the preceding evidence and gate tests. Partial on C2 means some legitimacy and endurance requirements are supportable but at least one mandatory distribution, ecological, democratic/information, security/sovereignty or durability state remains unresolved.

No classification transfers between functions. Within a function, the 21 eligible risk-function cells remain recorded in the canonical instruments. A stronger result in one risk setting cannot hide an unresolved cell elsewhere.

Cross-cutting gate state

No positive average overrides these states.

What this means

AI looks useful in a few bounded roles, but none is yet proven indispensable. Report 3 now has a clear job: test whether the useful parts can be delivered safely, fairly and reliably enough to survive real institutions and shocks.

N-of-1 Box — Rick's Age-73 Human Inquiry

Status: protocol only; UNRESOLVED; national evidence weight = 0.

No governed observation record meeting the locked protocol was located in the bounded repository search. This report does not fabricate a result. If Rick later chooses to run the inquiry under separate authority, the question is:

Can deliberate AI use help a 73-year-old Australian live with greater autonomy, purpose, connection, learning and civic contribution without weakening independent judgement or creating unhealthy dependency?

The minimum record is: task and purpose; whether AI was used; time with and without AI including verification and correction; what decision or evidence changed; who retained final judgement; an independent first pass for selected decisions; autonomy, learning, health, security, relationships, purpose and civic contribution; contact displaced or enabled; privacy, safety, manipulation, error and dependency incidents; the best feasible no-AI alternative; weekly reflection; monthly no-AI comparison or abstention; and model/platform changes.

Measures and cadence must be pre-registered. Self-report must be distinguished from observed behaviour. Selection, learning, novelty and Hawthorne effects must be recorded. Stop or seek support if AI measurably weakens independent judgement, causes distress, displaces valued human contact, creates unsafe privacy or security exposure, or makes an essential task impossible without the tool. No personal result may be generalised to older Australians or the nation.

Strongest Rival and Symmetric Comparison

The strongest rival is not the 2026 status quo. It is an implemented public-capability package: well-resourced institutions; skilled people; sound data; conventional analytics; engineering and maintenance; redundancy; secure design; social protection; accessible services; journalism and media literacy; representative deliberation; and accountable formal response.

The rival explains much of the current evidence at least as well as the necessity proposition. That is why this report can recognise useful mechanisms without declaring necessity.

Evidence Limits and Counterevidence

The evidence base has six important limits.

First, Australian public material is much richer on policy, capability, adoption and trials than on controlled AI-over-comparator outcomes. Audit findings about governance and benefits are highly relevant, but one agency cannot establish a national function result.

Second, international empirical evidence transfers only with care. The UK AI-deliberation experiment demonstrates a mechanism under controlled conditions; it does not establish Australian institutional influence. International shorter-week evidence shows wellbeing patterns; it does not establish civic conversion. Global data-centre evidence identifies material costs; it does not allocate them to an Australian mission.

Third, counterfactual implementation is difficult. The non-AI comparator assumes a serious, feasible reform package, not chronic underinvestment. AI cannot be credited with solving a gap created by refusing to implement the comparator.

Fourth, distributional evidence is incomplete. Digital inclusion and social participation data identify who may be left out, but intervention-specific outcome studies across all seven lenses and their intersections are absent.

Fifth, fast-moving law, policy and provider conditions may change. This report verified current official material at the research cut-off, but Report 3 must reverify it at delivery decisions.

Sixth, absence searches cannot prove no evidence exists anywhere. They establish that specified, reproducible searches did not locate evidence adequate for this claim. That supports UNRESOLVED, not a claim of impossibility.

Where This Argument Could Be Wrong

The report may be too cautious if operational data unavailable publicly already show that an AI component prevented material Australian harm over the strongest comparator. It may understate machine-speed necessity in cyber defence if no feasible human/rules-based system can meet a critical response window. It may also understate the value of small improvements that accumulate across many decisions.

It may be too optimistic about usefulness if benchmark gains fail under real distribution shift, if verification costs exceed saved time, if people over-trust fluent outputs, or if lifecycle and vendor dependence are larger than current mission accounts suggest. The deliberation result could fail when participants are adversarial, identities and interests are consequential, or a provider can shape the model. Care tools could worsen isolation while improving an administrative metric.

The strongest rival may also be over-idealised. Some reforms may be politically or institutionally difficult to implement by 2035. That is a Report 3 delivery question, not grounds for weakening the comparator here. Conversely, AI procurement and integration face the same implementation reality.

Evidence that would change the answer includes: preregistered Australian head-to-head trials on material outcomes; independent audit of causal attribution and harms; disaggregated results for all material groups; mission-specific lifecycle accounts; successful S1–S5 fallback and exit tests; and representative deliberative processes with a published formal response, action and durable feedback.

Implications — Separate from the Verdict

These implications are not recommendations to deploy or a final pathway judgement.

Australian institutions should frame AI proposals as testable mechanisms, not transformation claims. Procurement and evaluation should name the exact comparator and collect outcome, continuity, distribution and gate evidence before scale. AI should remain assistive, contestable and reversible where independent judgement matters. Essential services need competent human, non-digital or offline fallbacks.

Forecast and cyber programmes are the most credible places for bounded additionality tests, but they still require causal outcomes, not only accuracy or speed. Care programmes should begin with functional ability, contact and supported decision-making. Information and deliberation programmes should publish sources, uncertainty, correction, representation, dissent and formal-response arrangements. No civic-time claim should be made until actual hours, workload, care, recovery, access, participation and decision influence are measured.

The evidence architecture should be treated as infrastructure in its own right. Stable claim IDs, search records, preserved comparator rows and dated checkpoints make it harder for enthusiasm or disappointment to silently rewrite the question.

The Bottom Line

AI may help Australians live and decide well, but help is conditional and function-specific. The current evidence supports bounded usefulness in anticipation/forecasting and cyber defence. It does not demonstrate AI necessity in any of the six functions. Four functions remain unresolved, and every function has at least one unresolved mandatory gate.

That answer is neither an endorsement nor a rejection of AI. It is a demand for the evidence that the word "necessary" requires: a failed strong comparator, an attributable material difference, better human and civic outcomes, fair distribution, safe dependence and a full ecological account.

Acknowledgements

This report relies on public institutions, auditors, statisticians, scientists, researchers, service providers and communities whose records make independent scrutiny possible. Acknowledgement and source inclusion do not imply endorsement by any cited institution.

Authors

Rick Molony, Our Resilient World. AI assisted the research and drafting under human Product Owner direction.

Appendix A — References

The canonical evidence ledger is authoritative for exact locations, access state, transfer notes, counterevidence and confidence. This list is for readers.

• EXT-008 — World Health Organization, Ageing: Healthy ageing and functional ability, current official explainer, accessed 2026-08-22.
• EXT-009 — Australian Institute of Health and Welfare, Social isolation and loneliness, updated 12 May 2026, accessed 2026-08-22.
• EXT-010 — Registered discovery URL for Productivity Commission early-childhood material; identity mismatch on live reverification, 2026-08-22. Not used as the original empirical source.
• EXT-011 — Australian National Audit Office, Governance of Artificial Intelligence at IP Australia, published 29 June 2026.
• EXT-012 — Australian Treasury, Microsoft 365 Copilot trial evaluation, official evaluation page and report.
• EXT-013 — Australian Electoral Commission, current guidance on AI, elections, authorisation and misleading information.
• EXT-014 — Australian Signals Directorate/Australian Cyber Security Centre, Content Credentials, current guidance.
• EXT-015 — OECD, OECD Survey on Drivers of Trust in Public Institutions 2024 Results — Country Notes: Australia.
• EXT-016 — OECD, Eight Ways to Institutionalise Deliberative Democracy, 2021.
• EXT-017 — Fan et al., Work time reduction via a 4-day workweek finds improvements in workers' well-being, Nature Human Behaviour (2025).
• EXT-018 — Parliamentary Library, Parliament of Australia, The four-day work week, 2026.
• EXT-019 — International AI Safety Report, International AI Safety Report 2026.
• EXT-020 — Tessler et al., AI can help humans find common ground in democratic deliberation, Science 386 (2024), DOI 10.1126/science.adq2852.
• EXT-021 — International Energy Agency, Energy and AI, 2025.
• EXT-022 — Jobs and Skills Australia, Our Gen AI Transition, current official national study.
• EXT-023 — Productivity Commission, Harnessing Data and Digital Technology, inquiry report.
• EXT-028 — Australian Institute of Health and Welfare, Older Australians — social support, current official material.
• AHL2-SRC-0001 — Bureau of Meteorology, Annual Report 2024–25, forecast-performance measures.
• AHL2-SRC-0002 — Bureau of Meteorology, Machine-learned forecasting and calibration of AIFS with RainForests, research report, March 2026; and current forecasting-method explainer.
• AHL2-SRC-0003 — Productivity Commission, Delivering quality care more efficiently, released 19 December 2025.
• AHL2-SRC-0004 — Department of Health, Disability and Ageing, Aged Care Data and Digital Strategy implementation material and Year One report, current at cut-off.
• AHL2-SRC-0005 — Australian Cyber Security Centre, Opportunities for AI in cyber defence, updated 12 August 2026.
• AHL2-SRC-0006 — Australian Cyber Security Centre, current guidance on frontier AI models and cyber threat/resilience, 2026.
• AHL2-SRC-0007 — Australian Bureau of Statistics, Time Use Survey, 2024, released 2026.
• AHL2-SRC-0008 — Parliament of Australia, parliamentary record concerning a proposed national citizens' assembly on housing, 2023.
• AHL2-SRC-0009 — OECD, OECD Guidelines for Citizen Participation Processes, public-response and follow-up provisions.
• AHL2-SRC-0010 — Australian Energy Market Operator, Digital demand surge, 1 June 2026, and 2025 Inputs, Assumptions and Scenarios data-centre demand projections.
• AHL2-SRC-0011 — Australian Government, Data Centre Expectations, 23 March 2026.
• AHL2-SRC-0012 — ARC Centre of Excellence for Automated Decision-Making and Society, Australian Digital Inclusion Index 2025.
• AHL2-SRC-0013 — Digital Transformation Agency, Policy for the responsible use of AI in government, version 2.0, effective 15 December 2025.
• AHL2-SRC-0014 — Australian Institute of Health and Welfare, People with disability in Australia — social inclusion and participation, current 2026 material.

Appendix B — Glossary

Additionality: the material contribution caused by the AI mechanism beyond the exact inherited comparator.

AI necessary: the comparator fails; the specified AI mechanism alone closes the material, attributable gap; and every mandatory gate survives.

AI useful: AI improves a result, but the comparator meets the need or another feasible approach can match the contribution.

Attribution: evidence connecting the AI mechanism, rather than accompanying investment or process change, to the decision and outcome.

Civic time: usable capacity for participation created by genuinely reduced paid work or an equivalent support, after care, recovery and access constraints.

Comparator: the strongest feasible non-AI package inherited from Report 1, including existing embedded AI in the 2026 baseline.

Democratic/information veto: a no-trade-off rule: material harm to evidence judgement, pluralism, minority protection, manipulation resistance or legitimate influence cannot be offset by benefit elsewhere.

Ecological veto: a no-trade-off rule requiring no material net deterioration in lifecycle energy, water, emissions, materials, land, biodiversity or exported effects.

Function verdict: one of NECESSARY, USEFUL NOT NECESSARY, NO MATERIAL BENEFIT / ADVERSE, MIXED or UNRESOLVED.

Living well: autonomy and agency, security and basic needs, learning and decision ability, mobility, health, relationships, purpose and contribution.

OWS-style process: a voluntary, issue-bounded, transparent and reason-giving community process; it has no automatic binding authority.

Strongest rival: capable institutions, conventional technology, material capacity, public services, social protection and trusted democratic practice as the causal explanation for gains.