Ethics, Sustainability and Social Impact

Executive summary

Ethics for AI Solution Engineers is not a slide footer—it is structured judgment about who may be harmed, how much environmental cost you accept for marginal quality gains, whether inclusion is designed into interfaces and datasets, and when dual-use capabilities must be refused or constrained. You document impact notes, justify model size and routing, assess workforce displacement honestly, and link decisions to Governance and topic 19—not performative "responsible AI" while always deploying the largest model. Sustainability and social impact belong in architecture and commercial choices, not only in CSR reports.

When to use this page

• At discovery when a use case touches vulnerable populations, employment, credit, health or public services.
• When choosing model size, always-on inference, or batch vs routed architectures.
• When accessibility, language diversity or digital inclusion are programme requirements—or absent and risky.
• When a capability could be repurposed for surveillance, weapons guidance or disinformation (dual-use).
• Before launch gates: produce impact notes alongside DPIA and risk tier artefacts.
• When executives pressure for automation levels that remove meaningful human oversight.

Executive view

Ask who could be harmed if this works as designed—not only if it fails. Challenge always-on largest models without carbon and cost justification. Require documented mitigations, not principles posters.

Decision required: Is automation level proportionate to risk tier and workforce agreements?

Technical view

Implement model routing, eval for bias and refusal, accessibility standards, logging for audit—not ethics as afterthought slide.

Link impact notes to [Governance](/ai-solution-engineering/governance) inventory and approval workflows.

Why this matters

AI systems scale benefits and harms faster than manual processes. A biased retrieval corpus, an always-on 70B model, or an agent with write access can:

• Deny service to protected groups at machine speed
• Burn megawatt-hours for marginal accuracy on routine queries
• Displace workers without retraining pathways—breaching trust and contracts
• Generate plausible misinformation in regulated domains
• Enable dual-use capabilities clients or vendors later misuse

Regulators, unions, employees and communities now scrutinise AI programmes—not only security officers. "We follow vendor responsible AI" is insufficient when you chose the use case, data, autonomy level and model route.

Ethics washing—principles without trade-offs—creates reputational Sev 1 when incidents occur. Strong ethics practice produces impact notes, testable inclusion requirements, environmental proportionality, and clear red lines on dual-use—aligned with Responsible AI and Governance and Governance playbook.

Learn

Ethics vs compliance vs governance

Lens	Question	Owner
Compliance	Is it legal?	Legal, DPO
Governance	Is it controlled and approved?	Risk, AI board
Ethics	Is it right—who bears harm?	Shared; engineer surfaces

Legal clearance ≠ ethical acceptability. Lawful biased outcomes or environmentally reckless architecture may still destroy trust.

Stakeholders who may be harmed

Map affected parties, not only users:

Group	Example harm
End users / customers	Wrong advice; exclusion; surveillance
Non-users	Scored or ranked without knowledge
Workers	Unfair monitoring; displacement; deskilling
Communities	Environmental burden; misinformation
Future users	Training on data without consent

Impact note requires at least three stakeholder groups with harm + mitigation each.

Algorithmic discrimination and fairness

Sources of harm:

• Biased training or retrieval corpus
• Proxy variables (postcode → ethnicity)
• Unequal error rates across demographics
• Language coverage gaps (dialect, low-resource languages)
• Accessibility barriers (screen reader, cognitive load)

Mitigations:

• Stratified eval by demographic proxy where lawful
• Human review for high-stakes decisions
• Transparency on limits; appeal path
• Participatory design with affected groups
• Document known limitations in model card

Fairness metrics are contextual—consult risk and legal; do not DIY legal conclusions.

Workforce impact and responsible automation

Honest assessment:

• Tasks automated vs roles eliminated
• Timeline and magnitude (headcount, overtime, agency spend)
• Retraining budget and union/work council consultation
• Quality of remaining jobs—more judgment vs more monitoring

Anti-pattern: "AI assists workers" marketing while KPIs penalise anyone who overrides the model.

Engage HR and works councils before public launch in regulated sectors.

Environmental cost of AI

Energy drivers:

• Model size and quantization choice
• Always-on vs on-demand inference
• Retrieval index size and refresh frequency
• Agent loops multiplying token calls
• Training/fine-tuning frequency and hardware

Carbon-aware practices:

• Route routine queries to small models; reserve large for complex
• Batch offline jobs in renewable-heavy grid windows where measurable
• Cache embeddings and responses where policy allows
• Right-size GPU; avoid idle warm pools
• Document kg CO2e estimate per month in impact note (order-of-magnitude acceptable)

Reject always-on largest model for FAQ when small model meets eval bar—document delta.

Digital inclusion and accessibility

Inclusion dimensions:

• Language (translation quality, idiom)
• Literacy and cognitive load
• Disability (WCAG 2.2 AA target for UI)
• Device and connectivity (mobile-first, low bandwidth)
• Digital literacy (onboarding, human fallback)

Testable requirements:

• Screen reader path for core journey
• Plain-language summary option
• Human chat/phone escalation within 2 clicks
• Performance on 3G latency budget

Accessibility is not optional for public sector and many enterprise policies.

Misinformation and synthetic content risk

GenAI can produce authoritative-sounding falsehoods. High-risk domains: health, finance, legal, news.

Controls:

• Citation required for material claims
• Watermarking or disclosure where policy requires
• Rate limits and abuse detection
• Refusal templates for out-of-scope advice
• No autonomous public publish without human review in tier-high

Dual-use and acceptable use

Dual-use: Capability beneficial in one context, harmful in another—facial recognition, voice clone, autonomous targeting assistance, bulk personalised persuasion.

Red lines (examples—firm policy may extend):

• Surveillance without legal basis
• Weapons or violence optimisation
• Deceptive impersonation of humans without disclosure
• Exploitation of minors

Process: Escalate to ethics/governance forum; document refusal or constraints; do not silently implement.

Social impact assessment — lightweight template

1. Purpose and beneficiaries
2. Affected stakeholders (≥3 groups)
3. Harms — likelihood, severity
4. Benefits — quantified where possible
5. Mitigations — design, process, human oversight
6. Alternatives considered — including no-AI
7. Monitoring — metrics, review cadence
8. Sign-off — risk tier linkage

Attach to gate pack for tier-medium and above.

Responsible automation ladder

Level	Description	Typical tier
0	Informational only	Low
1	Recommend; human decides	Medium
2	Act with human approval	Medium-high
3	Act with audit sample review	High
4	Autonomous	Rare; often prohibited

Do not skip levels because vendor demo shows autonomy.

Frameworks and methods

ETHIC snapshot (engagement quick scan)

Exposure — Who affected?
Trade-offs — Environment, jobs, fairness
Harm pathways — Failure and success modes
Inclusion — Access and language
Controls — Oversight, kill-switch, eval

15-minute discovery screen; full impact note if any dimension red.

Model routing for ethics and sustainability

Query classify → simple / complex / sensitive
→ simple: small model + cache
→ complex: large model
→ sensitive: large + mandatory human review
→ refuse: policy violation

Document eval delta between routes—justify large model only where measurable.

Inclusion requirements checklist

• WCAG 2.2 AA for user-facing UI
• Language coverage list documented
• Human fallback path tested
• Plain language review completed
• Low-bandwidth performance tested

Ethics escalation ladder

Level	Trigger	Forum
L1	Designer/engineer concern	Squad lead + risk SME
L2	Material harm potential	Client AI ethics / governance board
L3	Dual-use or reputational	Executive + legal
L4	Regulator or media	Crisis comms plan

Same discipline as security escalation—do not minimise.

Carbon estimation — order of magnitude

Rough formula for planning:

Monthly kWh ≈ (inference hours × GPU power draw) + training job allocation
CO2e ≈ kWh × grid factor (region-specific)

Use cloud provider carbon tools where available; compare options in decision memo—not false precision.

Pairing with governance

Impact note feeds:

• Governance inventory entry
• Model card limitations section
• Risk tier and approval path (topic 19)
• Contract acceptance exclusions (topic 30)

Real-world scenarios

Scenario A — Claims assistant: harm to vulnerable claimants

Context: Insurer FNOL chatbot; 220k claims/year; 18% elderly users; model occasionally denies coverage categories incorrectly with confident tone.

Impact note stakeholders:

Stakeholder	Harm	Mitigation
Elderly claimants	Wrong denial; confusion	Plain language; human button prominent
Low digital literacy users	Abandonment	Phone fallback; no forced chat
Claims handlers	Workload spike fixing errors	Handler dashboard of AI sessions

Outcome: Autonomy capped at level 1 (recommend draft); human sends all denials; faithfulness gate 92% on elderly-stratified eval set.

Numbers: Incorrect denial rate dropped from 3.1% (pilot) to 0.4% post-mitigation; handle time +6% accepted by buyer for risk reduction.

Scenario B — Contact centre: workforce and monitoring ethics

Context: Telco; 8,500 agents; AI QA scores 100% calls for compliance; union alleges covert performance punishment.

Conflict: Business wants auto-coaching from AI scores; union wants aggregate QA only.

Intervention:

• Social impact assessment with union Input
• No individual punitive automation; team-level insights only phase 1
• Agents can flag incorrect AI QA; human adjudication SLA 48h
• Retraining budget £2.4m for upskilled roles in complex sales

Outcome: Launch with union neutral statement; attrition −12% vs control site over 9 months.

Numbers: Projected 400 role redeployments vs 120 net reduction over 24 months—honest comms prevented strike ballot.

Scenario C — Always-on model vs routed stack — carbon and cost

Context: University IT helpdesk; 40k students; vendor proposes always-on 70B hosted GPU; £38k/month; estimated 4.2 MWh/month.

Alternative: Router + 8B for 80% tickets; 70B for 20%; £11k/month; ~1.1 MWh/month.

Eval: Small model 91% resolution on L1 set; large required for integration tickets only.

Decision: Routed stack; carbon note in sustainability report; 74% energy reduction.

Numbers: Student satisfaction 4.1 → 4.0 (within noise); FinOps saves £324k/year.

Scenario D — Dual-use: facial analytics vendor

Context: Retail client requests emotion inference from CCTV for "customer experience"; vendor SDK includes age/gender estimation.

Ethics escalation L3:

• Harm: surveillance normalisation; misclassification bias; regulatory challenge under GDPR
• Dual-use: repression if deployed in other client contexts by same vendor
• Decision: Decline emotion inference; approve aggregate footfall only; vendor clause prohibiting re-identification

Outcome: Lost £190k change request; retained £2.1m core programme; firm ethics committee citation in proposal win for adjacent public sector client.

Numbers: Reputational risk model £5m+ vs £190k revenue—executive Decide documented.

Accessibility test plan — WCAG 2.2 sample cases

ID	Test	Pass criteria
A1	Keyboard-only navigation	All actions reachable
A2	Screen reader (NVDA/VoiceOver)	Labels on chat bubbles
A3	Colour contrast	4.5:1 minimum text
A4	Focus visible	Focus ring on interactive
A5	Plain language	Readability score agreed

Include in UAT sign-off for public-facing assistants.

Model size decision tree — ethics and FinOps

Is query routine FAQ?
  Yes → small model + cache
  No → requires citation?
    Yes → RAG + medium/large based on eval delta
    No → is stake tier-high?
      Yes → large + human review
      No → medium with monitoring

Document decision in impact note Appendix B.

Stakeholder harm workshop — 90-minute facilitation guide

1. Silent brainstorm (10m): who could be harmed?
2. Affinity grouping (15m)
3. Severity vote (10m): likelihood × impact
4. Mitigation design (40m): pairs assigned groups
5. Report-out (15m): top 3 to governance

Facilitator neutral—engineer captures for impact note.

UN Sustainable Development Goals — optional client alignment

Some clients map initiatives to SDGs:

SDG	AI link
8 Decent work	Workforce retraining
10 Reduced inequalities	Fairness eval
12 Responsible consumption	Carbon-aware routing
13 Climate action	Energy disclosure

Only claim alignment with measurable programme elements.

Whistleblowing and ethics escalation — engineer duty

If leadership overrides documented tier-high harm:

1. Document dissent in decision log
2. Escalate to governance forum / ethics committee
3. Use firm speak-up policy—do not leak to media
4. Protect yourself—facts in writing, not hallway accusations

Topic 33 psychological safety applies upward too.

Practice exercises

Primary exercise — Impact note for claims assistant (60 minutes)

Brief: Auto insurance claims status assistant; web + mobile; English only initially; 35% users non-native English; tier-medium.

Tasks:

1. Identify four stakeholder groups and harms.
2. Mitigation per harm—design or process.
3. Responsible automation level recommendation with rationale.
4. Three inclusion requirements testable in UAT.
5. Model routing sketch—when large vs small model.

Acceptance criteria:

• No generic "monitor bias" without specific eval or process
• Human fallback specified
• Link to governance risk tier

Stretch exercise — Sustainability and ethics decision memo (half day)

Brief: Global bank internal research copilot; 22,000 users; proposal for 24/7 large model on dedicated GPU cluster.

Tasks:

1. Order-of-magnitude carbon and cost comparison: always-on large vs business-hours vs routed.
2. ETHIC snapshot red/amber/green.
3. Full social impact section on research analyst workforce (junior task automation).
4. Dual-use paragraph (insider information, market manipulation prompts).
5. Decision memo recommendation with dissent case for always-on.
6. Link controls to Governance approval path.

Acceptance criteria:

• Numbers for low/base/high scenarios
• Explicit no-AI alternative considered
• Monitoring metrics named

Scenario G — Geographic fairness in loan copilot

Context: RAG over regional policy PDFs; model answers urban policy to rural applicant query 8% of time (eval by postcode proxy).

Mitigation:

• Metadata region filter mandatory in retrieval
• Stratified eval ≥95% regional accuracy
• Human review all denials

Numbers: Harm incidents near zero post-fix; +120ms latency accepted.

Scenario H — Corporate net-zero vs AI expansion

Context: Client net-zero 2030; AI programme +12% datacentre energy; board challenge.

Response:

• Routed models −40% kWh
• Batch inference off-peak renewable regions
• Don't build low-value always-on chatbots—kill 2 of 5 pilots

Ethics and sustainability portfolio decision—not per-project only.

Ethics metrics dashboard — post-launch

Metric	Review cadence
Fairness disparity index	Monthly
Override rate by user group	Monthly
Energy kWh/query	Weekly
Accessibility ticket count	Weekly
Workforce FTE redeployed vs reduced	Quarterly

Dashboard owned by product + ethics forum—not buried in engineering only.

FAQ — ethics, sustainability and social impact

Q: Is legal clearance enough?
A: No—lawful ≠ ethical; impact note required for stakeholder harms beyond compliance checkbox.

Q: Must we always use smallest model?
A: Use smallest that meets eval—document delta if large model required; route complex queries only.

Q: Who approves automation level 2+?
A: Governance forum per Governance—not engineering alone.

Q: How quantify carbon without false precision?
A: Order-of-magnitude from cloud tools; compare options A vs B; disclose uncertainty range.

Q: When refuse a use case?
A: Dual-use red lines, prohibited practices, inability to mitigate severe harm—escalate and document refusal.

Ethics gate timing in lifecycle

Discovery → ETHIC snapshot
Design → Impact note draft
Build → Inclusion requirements in DoD
Pre-launch → Impact note sign-off + automation level
Operate → Ethics metrics dashboard review

Missing gate = Stage 4/6 incomplete—not "soft" delay.

Inclusion testing with real users

Budget 5–8 sessions with:

• Screen reader users (2 minimum)
• Non-native language speakers (2)
• Low digital literacy (2)

Findings feed UAT blockers—cannot waive for tier-high public services.

Responsible automation case library (study patterns)

Use case	Level	Rationale
Internal FAQ	0–1	Low stake; cite sources
HR policy draft	1	Human publishes
Claims recommendation	1–2	Financial impact; human approve
Credit decision support	2	Regulated; audit sample
Autonomous trading	4	Often prohibited or heavily gated

Use library in workshops—speeds automation level agreement.

Environmental disclosure template (client report insert)

AI workload [Service Name]: estimated [X] MWh/month ([Y]% of IT portfolio AI).
Routing policy: [%] queries on efficient model path.
Change vs prior year: [−Z]% kWh via [routing/project kill].
Methodology: [cloud carbon tool / PUE × GPU hours].
Uncertainty: ±[range]%.

Legal/comms review before external publish.

Dual-use screening questions at discovery

1. Could output enable physical harm if misused?
2. Could system scale surveillance beyond stated purpose?
3. Could synthetic media deceive vulnerable populations?
4. Is there off-label use clients might pursue without controls?

Any yes → ethics escalation before estimate and proposal.

Impact note full template (outline)

1. Purpose and scope
2. Beneficiaries
3. Affected stakeholders (≥3) — harm table
4. Automation level — ladder level + approver
5. Fairness and inclusion — eval and UX
6. Workforce — magnitude, timeline, mitigation
7. Environment — model routing, kWh estimate
8. Dual-use — screening result
9. Alternatives including no-AI
10. Monitoring metrics
11. Residual risk — acceptor name
12. Review date

Attach to governance submission per Governance.

Ethics steering talking points (5 minutes)

• "Here is who could be harmed if it works"
• "Here is our automation level and why not higher"
• "Here is energy vs alternative architecture"
• "Here is what we monitor post-launch"
• "Here is what we are not building"

Executives remember boundaries—reduces scope creep into unsafe territory.

Scenario I — AI literacy and inclusion for elderly users

Context: Council services bot; 34% users 65+; default UI chat-only; 22% abandon before first answer.

Mitigation: Large text mode; phone click-to-call; simplified intent buttons; staff-assisted kiosks in libraries.

Numbers: Abandon 22% → 9%; satisfaction 3.2 → 4.0; cost £95k UX vs £400k failed adoption rework estimate.

Inclusion is delivery success—not charity.

Scenario J — Green AI claim challenged by NGO

Context: Retailer markets "green AI assistant"; NGO report cites always-on GPU; social media backlash.

Response: Public methodology correction; routed architecture deployed 60 days; third-party review of carbon claims.

Numbers: Stock sentiment recovery; −38% kWh post-routing; ethics comms team now in gate pack.

Workforce transition plan — template sections

1. Tasks automated (list) vs tasks remaining human
2. Headcount impact — range, timeline, confidence
3. Retraining programmes — budget, provider, duration
4. Union/work council — consultation status
5. Communication timeline — internal before external
6. Success metrics — redeployment rate, satisfaction

Pair with topic 26 change management—ethics without execution plan fails workers.

Fairness eval — worked example (illustrative)

Use case: Credit limit recommendation assistant (not auto-decision).

Eval set: 2,000 historical cases with demographic proxies (region, age band—legal review).

Metric: False denial rate disparity <5% between bands; calibration within ±3%.

Outcome: Launch at automation level 1 only; level 2 gated on 6 months production monitoring.

Document in impact note and model card limitations.

Carbon comparison table — presentation-ready example

Architecture	kWh/month (est.)	CO2e kg/month (0.4 kg/kWh)	Faithfulness
Always-on 70B	4,200	1,680	93%
Routed 8B/70B	1,100	440	91.5%
Always-on 8B only	680	272	88%

Choose routed when 1.5% faithfulness delta acceptable—74% carbon reduction vs always-on 70B in this illustrative example.

Ethics and governance joint gate — sign-off matrix

Role	Signs impact note	Signs risk tier	Veto
Product owner	✓
AI engineering lead	✓
DPO	✓		✓ privacy
Model risk		✓	✓ tier-high
Ethics forum chair	✓		✓ dual-use

All relevant ✓ required before tier-high external launch—see Governance.

Long-form ethics case study pattern — how to write impact notes

Structure narrative: context → design choice → stakeholder harm identified → mitigation implemented → residual risk → monitoring plan. Avoid sensationalism; use numbers where possible; cite governance tier; link model card. Review with legal before external sharing. This pattern matches board-ready responsible AI reporting expectations in regulated industries.

Accessibility statement — public-facing assistant example

"We aim for WCAG 2.2 AA conformance on core journeys. Known limitations: [list]. Alternative access: phone [number], hours [X]. Feedback: [email]. Last tested: [date]." Publish and maintain—reduces legal and reputational exposure when gaps exist honestly disclosed.

Sustainable AI principles — operational not slogans

1. Measure energy where material spend
2. Route models by task complexity
3. Kill low-value always-on workloads
4. Include affected users in design
5. Document in governance—not slide footers

Principles without metrics and owners are ethics washing—topic 34 exists to make them operational.

Post-launch ethics monitoring cadence

Period	Activity
Weekly	Fairness/disparity sample; harmful output review
Monthly	Energy kWh vs budget; inclusion ticket trend
Quarterly	Impact note refresh; automation level still valid
Annual	Full social impact reassessment if scope changed

Assign named owner outside engineering—product or risk—so metrics do not orphan when build team rolls off.

Ethics decision log — example entries

DEC-ETH-001: Declined emotion inference on CCTV—dual-use and GDPR risk; approved footfall only.
DEC-ETH-002: Capped automation at level 1 for claims denials—harm to elderly users.
DEC-ETH-003: Routed models for university helpdesk—carbon reduction 74% with acceptable eval delta.

Searchable log prevents re-litigating settled ethics decisions each quarter when new executives arrive.

Inclusion personas — reference set for workshops

Persona	Needs
Margaret, 72	Large text, phone fallback, plain language
Ahmed, 34	ESL; idioms avoided; glossary
Priya, screen reader	ARIA labels; logical focus order
James, field engineer	Low bandwidth; offline read cache

Design reviews reference personas by name—humanises inclusion requirements beyond WCAG checkbox.

Residual risk acceptance — sample wording (for governance board)

"Residual risk of incorrect policy citation remains at automation level 1 with human approval on all payout-impacting outputs. Accepted by [Name, Role] on [Date]. Review by [Date]. Monitoring: weekly faithfulness sample ≥90%, harmful output <0.05%." Legal must approve exact wording—engineers supply technical facts only.

Cross-link summary — ethics in the lifecycle

Discovery ETHIC → Design impact note → Build inclusion DoD → Contract MEAT (topic 30) → Launch governance sign-off → Operate monitoring cadence → Scale reassessment if autonomy increases. Ethics is continuous—not a discovery checkbox discarded at build. Revisit impact note when corpus, model, or autonomy changes—same triggers as contractual CR and eval regression. Store signed impact notes in the same document control system as DPIAs and model cards so audit and ops teams can find them under incident pressure. Ethics artefacts age quickly—date every version and retire superseded notes so teams do not follow obsolete mitigations after architecture change. When in doubt, escalate early: an ethics forum delay is cheaper than a reputational Sev 1 after launch. Document the escalation either way.

Questions you should be able to answer

1. Who could be harmed if this system works exactly as designed?
2. What stakeholder groups are missing from our discovery?
3. What is the responsible automation level—and who approved it?
4. How do error rates differ across demographic or language proxies in eval?
5. What workforce impact is expected—roles, timeline, retraining?
6. What is the environmental cost of the chosen architecture vs alternatives?
7. Why not always use the smallest model? Why not always the largest?
8. What inclusion and accessibility requirements are testable in UAT?
9. What misinformation or overconfidence risks exist in this domain?
10. Does this capability qualify as dual-use—and what are our red lines?
11. What does the impact note say—and is it in the gate pack?
12. How do workers override or challenge AI outputs?
13. What monitoring detects harm post-launch?
14. How does ethics link to governance inventory and risk tier?
15. What would make us refuse or pause this use case?

EU AI Act and ethics — practitioner awareness

Not legal advice—engineers must know when to escalate:

• Prohibited practices (e.g. certain social scoring)—refuse and document
• High-risk categories—extra logging, human oversight, conformity assessment triggers
• Link Governance for client-specific classification

Impact note references regulatory tier where applicable.

Participatory design — involving affected communities

For public services and consumer products:

• Co-design workshops with users who face digital exclusion
• Prototype testing with assistive technology users
• Community advisory for language minorities

Budget 2–4 weeks in plan—exclusion discovered late costs 10× in rework.

Environmental reporting — client-facing narrative

When clients publish sustainability reports:

• Document kWh/month estimate and routing policy
• Compare baseline (all-large) vs chosen architecture
• Honest uncertainty ranges—methodology footnote

Avoid greenwashing: "AI-powered efficiency" without numbers invites audit challenge.

Children and vulnerable adults — heightened care

Red lines:

• No manipulative nudging of minors
• No mental health autonomous advice tier-high
• Age-appropriate UX; parental controls where required
• Escalate any child data processing to DPO immediately

Supply chain ethics — vendor selection

Vendor labour practices and conflict minerals in hardware may matter for ESG procurement:

• Include vendor ESG questionnaire in RFI where client policy requires
• Prefer providers with published sustainability reports
• Data centre PUE and renewable energy % in TCO ethics comparison

Scenario E — Language exclusion in government service

Context: National employment service chatbot; English-only; 23% users primary language other than English; complaint rate high in Welsh-speaking region.

Impact note missed: Inclusion dimension red at discovery.

Fix: Phase 1.5 Welsh corpus + human translator review loop; plain English toggle.

Numbers: Complaints −41% in region after 12 weeks; cost £180k vs £1.2m reputational risk model from ombudsman scenario.

Scenario F — Carbon audit challenge

Context: University claims carbon-neutral AI; auditor asks methodology; always-on GPU 24/7; no measurement.

Fix: Metering via cloud carbon tool; move to routed stack; publish monthly kg CO2e dashboard.

Numbers: 62% reduction post-routing; audit pass; student green society engagement positive.

Ethics review board pack — slide outline

1. Use case and automation level
2. Stakeholders harmed (summary table)
3. Eval fairness approach
4. Workforce impact
5. Environmental comparison
6. Dual-use assessment
7. Mitigations and residual risk
8. Monitoring plan
9. Decide recommendation

Attach full impact note as appendix.

Bias eval — practical metrics (with legal caution)

Where lawful and ethical:

• Equalised odds difference across groups on proxy labels
• Calibration within cohorts
• False denial rate parity for service access

Document limitations of proxies; legal review before reporting externally.

Never claim "unbiased"—prefer "tested for disparity on X with result Y."

Long-term societal effects — second-order thinking

Ask:

• Does this concentrate power with employers over workers?
• Does it erode public trust in institutions if wrong?
• Does it shift accountability so no human feels responsible?

Document in impact note second-order section—even qualitative scoring helps governance forum.

Pairing ethics with change management

Topic 26 link: ethical automation requires honest comms to workforce:

• Timeline of task change
• Retraining paths
• How to override AI safely

Ethics without change honesty produces strike and sabotage—social impact real.

Extended Learn — historical ethics cases (lessons, not legal precedent)

Case pattern A — Automated benefits eligibility: Systematic underpayment to vulnerable groups; harm from over-trust in model; lesson: human review on edge cases; stratified eval.

Case pattern B — Hiring screening tool: Gender bias in training proxy; lesson: fairness testing; legal + ethics review before HR automation.

Case pattern C — Environmental backlash: Tech firm AI training energy report; lesson: disclose routing and renewable mix; avoid secret mega-jobs.

Study patterns; cite in impact notes as risk category, not sensational detail.

Extended Learn — children's data and EdTech

School and EdTech AI requires:

• Parental consent flows
• Minimal data collection
• No behavioural advertising
• Age-appropriate interaction design

Often tier-high regardless of "internal" label if minors involved.

Extended Learn — misinformation at scale

GenAI enables cheap false content:

• Client comms bots need source grounding
• Rate limits on public endpoints
• Monitoring for prompt injection publishing

Social impact extends beyond direct users to readers of AI-generated public content.

Extended Learn — water and land use of datacentres

Beyond kWh—some clients ask water usage for cooling:

• Prefer providers with WUE disclosure
• Consider region grid and water stress maps
• Right-size compute—biggest ethics win is not running unnecessary GPU

Include in sustainability appendix when client ESG team engaged.

Extended Learn — indigenous and marginalised data

Some knowledge must not be ingested without community consent:

• Traditional ecological knowledge
• Sacred or restricted cultural content

Discovery question: "Whose data is this—and who consented?"

Escalate beyond standard DPIA when community rights implicated.

Extended Learn — ethics in presales

Do not promise outcomes that violate ethics to win deal:

• Surveillance customers without legal basis
• Manipulative persuasion engines
• Deepfake capabilities without disclosure

Firm quality and independence standards apply in presales—topic 29 link.

Walk away or narrow scope; short-term revenue vs long-term licence risk.

Negative cases — when ethics fails

Ethics washing

Symptom: Principles slide; largest model; no impact note.

Impact: Incident; media; employee walkout.

Fix: Impact note gate; routing policy; sign-off.

Compliance equals ethics

Symptom: "Legal said OK" on discriminatory outcome.

Impact: Trust collapse; regulatory scrutiny beyond law.

Fix: Stratified eval; human oversight; ethics forum.

Hidden workforce harm

Symptom: Launch surprise for union; monitoring without consultation.

Impact: Strike; programme halt.

Fix: Early works council; honest job impact numbers.

Carbon blindness

Symptom: Always-on GPU for FAQ; sustainability goals breached.

Impact: ESG audit failure; cost overrun.

Fix: Routing; carbon estimate in architecture decision.

Accessibility afterthought

Symptom: WCAG fail at launch; exclusion lawsuit risk.

Impact: Rework £400k+; brand damage.

Fix: Inclusion requirements in definition of done.

Dual-use drift

Symptom: SDK feature enables surveillance; scope creep.

Impact: Client scandal; firm reputational contagion.

Fix: Escalation L3; contract use restrictions; refuse.

Output

Expected outputs for this topic:

• Impact note / social impact assessment — Stakeholders, harms, mitigations
• Responsible automation level — Documented approval
• Inclusion and accessibility requirements — Testable in UAT
• Model routing justification — Quality, cost, carbon trade-off
• Workforce impact summary — Magnitude, timeline, retraining
• Fairness eval plan — Stratified metrics where lawful
• Dual-use assessment — Red lines and escalation record
• Carbon/cost comparison — Order-of-magnitude for major options
• Ethics escalation log — Issues raised and decisions

Practice checklist

• I wrote an impact note with ≥3 stakeholder groups
• I justified model routing on ethics and sustainability grounds
• I defined testable accessibility or inclusion requirements
• I linked ethics artefacts to governance risk tier
• I completed the primary impact note exercise
• I articulated one dual-use red line for my context
• I avoided ethics washing—trade-offs named explicitly

Related playbook content

• Governance — Inventories, approvals and operating model
• Responsible AI and Governance — Risk tier and model cards
• Privacy, Legal and Compliance — Lawful basis and DPIA
• Performance Engineering and AI FinOps — Model routing and cost
• User Experience and Human Factors — Accessibility and human-centred design
• Change Management and Adoption — Workforce and adoption
• Leadership and People Management — Escalation and psychological safety
• Ethics, Sustainability and Social Impact — This topic
• Leadership — Stage hub
• How to use this Learning Map — Reference-depth study method

Key takeaways

• Ask who is harmed when it works, not only when it fails—impact notes belong in gate packs.
• Model routing serves ethics and sustainability—reject always-on largest models without measured benefit.
• Workforce and inclusion require honest numbers and testable requirements, not marketing assist narratives.
• Dual-use and red lines need escalation—not silent implementation for revenue.
• Link all artefacts to Governance and risk tier—not slide-only responsible AI.

What to do next

1. Complete the primary impact note exercise for a practice use case.
2. Continue with Personal Effectiveness.
3. Open Governance to align impact notes with inventory workflow.
4. Cross-check Responsible AI and Governance for model card limitations section.