HUMANITARIAN AI RISK ASSESSMENT

2026 Annual Review

By Peter Collins · Evidence cut-off: 8 September 2026

Read the full annual review below. This is the author’s published assessment; its methodology, evidence limitations and publication notice are included.

HUMANITARIAN AI

RISK ASSESSMENT

2026 ANNUAL REVIEW

A benchmark against the 2025 baseline, with verified developments to September 2026

PETER COLLINS

with research assistance from Claude (Anthropic)

Copyright © 2026 Peter Collins. All rights reserved.

No part of this publication may be reproduced, distributed, or transmitted in any form without prior written permission of the author, except for brief quotations in critical reviews.

First edition: 2026 Annual Review. Historical baseline: Humanitarian AI Risk Assessment, 23 September 2025.

IMPORTANT NOTICE

This book is a general-public risk-education resource. It is not legal, medical, financial, engineering or regulatory advice, and must not be relied on as a substitute for professional advice specific to your circumstances. It has not undergone independent formal peer review; it was produced by the author with AI research assistance, drawing on publicly available sources current to the evidence cut-off stated in the Methodology chapter. Every effort has been made to cite verifiable sources and to distinguish observed evidence, expert judgement, and forecast. Errors may nonetheless exist; corrections are welcomed via the contact details in the About the Author section.

This publication does not claim to be the most advanced risk assessment ever produced, a formal standard, or independently peer reviewed, and none of those claims should be inferred from its tone or scope.

Foreword by Peter Collins

I'm Peter Collins. In September 2025 I commissioned and led a Humanitarian AI Risk Assessment to do one thing well: identify the AI hazards that matter most to people's lives, rights, livelihoods and the planet we share, and set out what should be done about them.

This book is the first annual review of that work. A risk assessment that is never revisited is just an opinion with a date stamped on it. So a year on, I've gone back through every hazard, checked what has actually happened against what we projected, and updated the picture honestly — including the places where I don't yet have good enough evidence to update a score with confidence, and say so plainly rather than guess.

Some things moved faster than expected: verified battlefield use of autonomous weapons, measured (not projected) data-centre resource consumption, and audited figures on AI-enabled fraud. Some things moved more slowly, or in a direction the 2025 baseline didn't fully anticipate: UK data-protection law loosened some automated-decision restrictions even as a regulator's statutory AI code of practice remains a year away, and feared worst-case election disruption from deepfakes has, so far, been less severe in mature democracies than the volume of synthetic content might suggest.

This is not a think-piece, and it is not a campaign document. It is a working record for the public, employers, and policymakers who want an honest account of where AI risk actually stood at this point in 2026 — grounded in named, dated, checkable sources, with the limits of that evidence stated as clearly as the findings themselves.

Executive Summary

This is the 2026 annual review of the Humanitarian AI Risk Assessment first published on 23 September 2025. It re-examines sixteen distinct hazards (consolidated from twenty numbered entries in the 2025 document, four of which duplicated others) against verified developments to the September 2026 evidence cut-off.

What has genuinely changed

• Lethal autonomy in warfare moved from projection to documented battlefield practice (Ukraine/Russia conflict), alongside major Western defence contracts with frontier AI companies.
• AI data-centre resource consumption is now measured, not projected: ~448 TWh of electricity in 2025, on track to double by 2030, per UN University research.
• AI-enabled fraud is now independently audited for the first time: the FBI's IC3 recorded $893.3 million in 2025 losses under a dedicated new reporting category.
• UK automated-decision law loosened (DUAA 2025, in force Feb 2026) ahead of the statutory AI oversight code meant to accompany it (not due until 2027).
• The EU AI Act became generally applicable from 2 August 2026, but the highest-risk obligations most relevant to critical infrastructure were pushed back to December 2027.

What the evidence does not support

• A sudden mass-unemployment shock — the clearest 2026 labour data shows a slower, generational effect (reduced hiring of young workers in exposed roles) rather than large-scale layoffs of incumbents.
• Confirmed large-scale election disruption from deepfakes in mature democracies, despite a sharp rise in deepfake volume.
• A change in the near-term likelihood of catastrophic or existential AI risk — the evidence shows capability growing and safety frameworks multiplying, without a clear net movement in either direction.

Honest gaps in this edition

Three hazards — AI in Health & Biotech, Loss of Human Autonomy, and AI in Nuclear Command & Control — did not receive dedicated primary-source research for this review and are carried forward from the 2025 baseline with that limitation stated. They are flagged for full research before the next edition rather than scored on inference.

Methodology

This review keeps Peter Collins's Master Matrix from the 2025 baseline for comparability: each hazard is scored on Likelihood (1–5) and Severity (1–5), multiplied to a Risk Score (1–25), banded as Low (1–4), Moderate (5–9), High (10–16) or Critical (17–25). A single L×S number can hide a low-likelihood, catastrophic-severity tail risk behind a moderate mid-range score, so hazards such as Existential Risk, Biological Risk and Nuclear Command & Control are flagged as tail risks regardless of their numeric band.

Evidence cut-off for this review: 8 September 2026. Sources were prioritised in this order — primary legal and regulatory text and official statistics; peer-reviewed and multi-expert reports (such as the International AI Safety Report 2026); respected institutional reports with disclosed methodology; journalism, used only to evidence a specific reported incident and cross-checked where possible. No claim in this book rests on an unopened source, an unverifiable statistic, or a vendor marketing figure presented as independent fact — where sources disagreed or lacked disclosed methodology (as with several competing deepfake statistics), that disagreement is stated rather than resolved by picking the larger number.

Baseline integrity note: the 2025 source document numbered twenty hazards, but four pairs substantially duplicated the same source–pathway–harm chain (critical infrastructure appeared twice, cybersecurity appeared twice, warfare and autonomous weapons overlapped heavily, and misinformation and democratic disinformation overlapped heavily). This review merges each pair into one hazard, bringing the working total to sixteen, and records that decision here rather than silently changing the 2025 count.

Where 2026 evidence for a hazard was not obtained to this review's standard, the hazard is marked accordingly and its score is carried forward from 2025 rather than adjusted on inference — this affects three hazards, listed in the Executive Summary.

The 2026 Hazard Register

Sixteen hazards, each showing the 2025 recorded baseline alongside the 2026 current assessment. "Why this one now?" sections report verified developments since the 2025 cut-off, with sources listed at the end of each entry.

Hazard 1 — AI-Enhanced Warfare & Autonomous Weapons

Merges 2025 Hazard 1 (AI-Enhanced Warfare) and Hazard 11 (Autonomous Weapons Systems / LAWS & Swarming Drones) — both traced the same driver-pathway-harm chain and are treated as one hazard for 2026 to avoid double-counting.

Why this one now?

The clearest verified development since the 2025 baseline is that lethal autonomy has moved from doctrine and prototype into contested battlefield use. Independent reporting through 2026 describes Ukrainian and Russian forces fielding drones that can identify and strike targets with reduced or, in some reported cases, no continuous human control once cued onto a target — a step change from earlier remotely piloted systems.

State military procurement has scaled in parallel. In 2026 the US Department of Defense contracted with a group of major technology and AI companies to accelerate autonomous and AI-enabled military systems, and separate commercial AI providers confirmed large contracts to make frontier models available on classified defence systems.

Diplomatically, the gap between capability and governance remains wide. Lethal autonomous weapons systems (LAWS) are still not covered by any binding international treaty; discussions continue under the UN Convention on Certain Conventional Weapons (CCW), with a further Review Conference scheduled for November 2026. Major military powers continue to resist binding restrictions, favouring flexibility over new law.

Who may be harmed

Combatants and civilians in active conflict zones (documented civilian drone-strike casualties in Ukraine), and — as an ambient risk — anyone in a future conflict where responsibility for a lethal decision cannot be clearly attributed to a human.

Controls — what changed

No material change in binding control since 2025: the core immediate-term recommendation (retain meaningful human control over the use of lethal force, and support a negotiated instrument at the CCW Review Conference) is unmet and now more urgent given confirmed battlefield autonomy.

Humanitarian reflection by Peter Collins

What was a projection in 2025 is now a documented practice. The world did not decide to cross this line through treaty or public debate; it crossed it through wartime necessity, and the law is still trying to catch up.

Evidence

• The Intercept, "It's Time to Rein in Lethal AI Drones" (26 Aug 2026)
• Small Wars Journal, "Fully Autonomous Drones Reportedly Kill in Ukraine" (Aug 2026)
• Autonomy Global, "Ukraine War Becomes Live Test Bed for AI-Enabled Autonomous Weapons" (Jan 2026)
• UN Convention on Certain Conventional Weapons, Seventh Review Conference schedule (Nov 2026)

Hazard 2 — Loss of Privacy & Data Exploitation

Corresponds to 2025 Hazard 2 (Loss of Privacy).

Why this one now?

UK data protection law changed materially in the review period. The Data (Use and Access) Act 2025 came into force on 5 February 2026 and replaced UK GDPR Article 22 with new Articles 22A–22D, broadening the circumstances in which a decision can be made about a person by an automated system alone, subject to new safeguards (notification, a right to make representations, human review on request, and contestability). The stricter pre-2026 regime is retained only where special category data — health, ethnicity, biometrics — is involved.

The ICO is now under a statutory duty (SI 2026/425, in force 12 May 2026) to produce the UK's first statutory Code of Practice on AI and automated decision-making; a consultation on draft guidance closed in May 2026, with the final code expected in 2027. In parallel the ICO is engaging directly with eleven major foundation-model developers over how personal data is safeguarded during training.

This is a genuine loosening of a protection that existed at the 2025 baseline, not merely a clarification, and it applies while the promised statutory code that is meant to constrain the new discretion is still a year or more away.

Who may be harmed

Anyone whose personal data trains or is processed by an AI system, with disproportionate exposure for people who cannot easily contest an automated decision (benefits, credit, employment, insurance).

Controls — what changed

Immediate term: organisations should not treat the DUAA 2025 change as a green light — human review rights under Article 22C still need to be operationally real, not nominal. Medium term: the ICO's statutory Code of Practice, once published, is the control to track.

Humanitarian reflection by Peter Collins

A right that existed at the 2025 baseline has been narrowed, with the safeguard meant to replace it not yet written. That sequencing — permit first, codify protection later — is itself a risk worth naming.

Evidence

• GOV.UK / legislation.gov.uk, Data (Use and Access) Act 2025, commencement 5 Feb 2026
• Information Commissioner's Office, guidance on AI and automated decision-making (2026, under review)
• The Data Protection Act 2018 (Code of Practice on AI and ADM) Regulations 2026, SI 2026/425

Hazard 3 — Existential Risk (Misaligned AGI / ASI)

Corresponds to 2025 Hazard 3 (Existential Risk), which the 2025 baseline treated as a ceiling risk with maximal severity but low near-term likelihood.

Why this one now?

The most authoritative update available is the second International AI Safety Report, published 3 February 2026, chaired by Yoshua Bengio and produced with over 100 experts nominated by more than 30 countries plus the UN, OECD and EU. It reports that general-purpose systems now converse fluently across languages, generate working code, produce realistic images and short video, and solve graduate-level mathematics and science problems, with 'reasoning' models increasingly common.

The Report is explicit that this capability growth has not been matched by proportionate control: models are still less reliable on long, multi-step tasks, still hallucinate, and — on the specific tail risks that matter for this hazard — the number of companies publishing frontier AI safety frameworks has more than doubled since the 2025 edition, but funded societal-resilience measures still have large, unaddressed evidence gaps.

Nothing in the 2026 evidence base supports moving this hazard's near-term likelihood up sharply; nothing supports moving it down either. It remains, correctly, a tail risk that is easy for a mid-range score to hide, which is why it stays in the separate tail-risk register regardless of its L×S product.

Who may be harmed

In the near term, no one — this is a governance-adequacy risk, not a realised harm. Its severity is scored on the basis that if triggered, harm would be irreversible and global.

Controls — what changed

Immediate/near term: compute governance, staged deployment and independent evaluation before release, and continued growth in the number and rigour of published frontier safety frameworks (now over 100 companies, up from under 50 in 2025) — this is the one area with documented movement.

Humanitarian reflection by Peter Collins

This is the one hazard where I would rather be accused of over-caution than under-caution. The 2026 evidence doesn't show an imminent event; it shows a capability curve still rising faster than the safety literature can independently verify.

Evidence

• International AI Safety Report 2026, Executive Summary and Key Update on technical risk-management approaches (3 Feb 2026)

Hazard 4 — AI-Controlled Critical Infrastructure

Merges 2025 Hazard 4 (AI-Controlled Critical Infrastructure) and Hazard 18 (AI in Critical Infrastructure — Energy, Transport) — the 2025 document ran near-identical treatments of the same hazard under two numbers.

Why this one now?

The clearest documented change is the growth in the infrastructure itself, which is now large enough to be a systemic dependency in its own right. A United Nations University assessment (June 2026) found global data-centre electricity consumption reached roughly 448 terawatt-hours in 2025 — more than all but ten countries in the world — projected to reach around 945 TWh by 2030, and noted that in Ireland, data centres already accounted for 21% of national metered electricity in 2023.

The concentration this represents — a small number of hyperscale operators and cloud providers underpinning an increasing share of national grids, water systems and, per the EU's July 2026 action plan, cybersecurity resilience planning — is itself the infrastructure-dependency hazard identified in 2025, now measurably larger.

On governance: the EU AI Act's general application from 2 August 2026 does now give the AI Office direct enforcement powers over general-purpose AI providers, but the high-risk system obligations most relevant to critical-infrastructure deployments (Annex III) were pushed back under the Digital Omnibus package to December 2027 — a genuine delay against the 2025 assumption that these controls would be live sooner.

Who may be harmed

Populations dependent on power, water and transport networks with growing AI/data-centre coupling; communities local to data-centre siting facing direct resource competition.

Controls — what changed

Medium term: the delayed EU high-risk obligations (now Dec 2027) should be treated as the binding trigger date for infrastructure operators, not the general 2 Aug 2026 application date many are currently working to.

Humanitarian reflection by Peter Collins

The infrastructure got bigger before the guardrails caught up. That is the single most measurable change since 2025.

Evidence

• United Nations University (INWEH), "Rising Emissions, Depleting Water and Vanishing Land" (3 June 2026)
• European Commission, Digital Strategy — AI Act application timeline and July 2026 action plan on Cybersecurity and AI
• Al Jazeera, "What came into force with the EU's AI Act this week — and what didn't" (6 Aug 2026)

Hazard 5 — AI Hacking & Cybersecurity Threats

Merges 2025 Hazard 5 and Hazard 20, which were the same hazard (AI-related cybersecurity/model/data threats) recorded twice.

Why this one now?

The International AI Safety Report 2026 states plainly that AI's largest current role in cyber-offence is scaling the preparatory stages of an attack — reconnaissance, phishing content, code generation — rather than executing full attacks autonomously. That is a more precise, and slightly less alarming, picture than a 2025 reader might have feared, and it is a genuine evidence-based refinement rather than a change in the underlying hazard.

Fraud built on AI-generated synthetic media is the area with the strongest documented 2026 growth. The FBI's Internet Crime Complaint Center broke out AI-enabled fraud as its own reporting category for the first time in roughly 25 years of the report's history, recording $893.3 million in adjusted losses across 22,364 complaints in the 2025 IC3 report — the only independently audited figure in a space otherwise dominated by vendor marketing statistics that lack disclosed methodology.

Regulation has moved to meet this: the EU AI Act's Article 50 content-labelling obligations became binding on 2 August 2026, the US TAKE IT DOWN Act's 48-hour takedown duty has applied since 19 May 2026, and 47 of 50 US states now have some form of deepfake-specific law.

Who may be harmed

Organisations and individuals targeted by AI-scaled phishing and impersonation fraud; disproportionately, older adults (39% of FBI-recorded AI-fraud losses fell on people 60+) and employees in finance/approval workflows.

Controls — what changed

Immediate term: verify high-value requests (payment, credential, HR) out-of-band from the channel the request arrived on — this remains the single most effective procedural control identified across 2026 sources, regardless of detection-technology quality.

Humanitarian reflection by Peter Collins

The honest 2026 picture is two-track: state-level offensive AI use is still mostly about speed and scale of existing attack types, not new autonomous capability — but AI-enabled fraud against ordinary people is now large enough to be independently audited by the FBI in its own right.

Evidence

• International AI Safety Report 2026, cyber-offence chapter (3 Feb 2026)
• FBI Internet Crime Complaint Center, 2025 Internet Crime Report — AI-enabled fraud category
• Digital Applied, "Deepfake Statistics 2026: Fraud and Detection Data" (July 2026), citing FBI IC3 and EU AI Act Art. 50 and US TAKE IT DOWN Act commencement dates

Hazard 6 — Deepfake Manipulation & Synthetic Media Fraud

Corresponds to 2025 Hazard 6 (Deepfake Manipulation & Synthetic Media).

Why this one now?

Independent research through 2026 continues to show humans cannot reliably detect deepfakes unaided: a University of Florida study (Feb 2026) found human classification accuracy for deepfake face images statistically indistinguishable from a coin flip, consistent with a 2025 iProov finding that only 0.1% of a 2,000-person sample correctly identified every deepfake and genuine sample shown to them.

Verified incident tracking (Resemble AI's H1 2026 Deepfake Threat Report) documented 821 verified attacks drawn from 1,760 news reports in the first half of 2026 alone, versus 1,567 verified incidents across the whole of 2025 — a meaningful acceleration, though the same source cautions that most incidents disclose no financial figure, so public loss totals should be read as a floor, not a full estimate.

As noted under Hazard 5, the regulatory response (EU AI Act Article 50 labelling, US TAKE IT DOWN Act, 47-state US legislation) has caught up faster on non-consensual and fraud content than on election-related synthetic media, where the Knight Columbia analysis found deepfakes affected 38 countries' elections in the review period but with less real-world impact in mature democracies than had been feared in 2025.

Who may be harmed

Fraud victims (corporate and individual); subjects of non-consensual synthetic imagery, which independent trackers still find is the largest single category of deepfake harm by volume.

Controls — what changed

Immediate term: treat voice and video as unverified by default in any approval workflow; near term: monitor uptake of provenance/labelling standards now legally required in the EU from August 2026.

Humanitarian reflection by Peter Collins

This is the hazard where the 2025 warning has most clearly become the 2026 reality — not because the technology changed unpredictably, but because adoption simply continued on the trajectory already visible a year ago.

Evidence

• University of Florida deepfake-detection study (Feb 2026), as reported by Adaptive Security
• Resemble AI, H1 2026 Deepfake Threat Report
• Knight Columbia, election-deepfake impact analysis, cited in StationX "Deepfake Statistics [2026]"

Hazard 7 — Biased & Discriminatory AI

Corresponds to 2025 Hazard 7 (Biased & Discriminatory AI).

Why this one now?

The clearest documented 2026 development is in recruitment. The ICO published "Recruitment Rewired" on 31 March 2026, drawing on evidence from over 30 employers who engaged voluntarily between March 2025 and January 2026 plus a review of privacy notices and data protection impact assessments; it wrote to 16 organisations it judged likely to be using automated decision-making on candidates, with recommendations, which those organisations have committed to acting on.

This is a regulator actively finding and correcting practice, not merely publishing principles — a genuine, if modest, strengthening of oversight since the 2025 baseline, and the clearest sector-specific evidence available for this review.

Against that, the broader UK legal environment for automated decisions loosened in the same period (see Hazard 2): DUAA 2025's replacement of Article 22 permits more solely automated significant decisions than before, with the stricter regime retained only for special-category data. Bias risk and automated-decision permissiveness are moving in different directions at once, which is exactly the kind of tension a single-number risk score can hide.

Who may be harmed

Job applicants, credit and insurance applicants, and anyone subject to automated screening — with the ICO's own recruitment work now providing a live UK evidence base rather than only historical or US-sourced examples.

Controls — what changed

Near term: the ICO's Recruitment Rewired recommendations to the 16 identified organisations are a concrete, trackable control with a named regulator and named cohort — worth monitoring for compliance follow-through in the next review.

Humanitarian reflection by Peter Collins

Regulators are starting to do the unglamorous work of checking specific employers' specific systems, rather than only issuing guidance. That is real progress, even while the general legal permission for automated decisions has widened.

Evidence

• Information Commissioner's Office, "Recruitment Rewired" (31 March 2026)
• Data (Use and Access) Act 2025, s.80 (in force 5 Feb 2026 via SI 2026/82)

Hazard 8 — AI in Health & Biotech (Clinical Safety)

Corresponds to 2025 Hazard 8 (AI in Health & Biotech).

Why this one now?

This hazard did not receive dedicated primary-source research for this review beyond the cross-cutting International AI Safety Report 2026, which notes reasoning-model performance gains on graduate-level science tasks and improved protein-design applications, alongside persistent unreliability on long, multi-step tasks — a mixed signal for clinical deployment specifically.

Under this review's evidence standard, a hazard sheet without dedicated sector research cannot be assigned an Evidence Strength above C for its 2026-specific claims. This entry is flagged for full primary research (MHRA guidance, NHS AI deployment data, clinical-safety incident reporting) before the next edition, rather than scored on inference.

Who may be harmed

Patients subject to AI-assisted diagnosis, triage or treatment-planning tools; provisional pending dedicated research.

Controls — what changed

No control update issued this review; carried forward from 2025 pending dedicated evidence gathering.

Humanitarian reflection by Peter Collins

I would rather tell you honestly that this one wasn't researched to the same depth as the others than pad it with confident-sounding filler. It's flagged for the next edition.

Evidence

• International AI Safety Report 2026 (general capability findings only — not sector-specific)

Hazard 9 — Economic Displacement & Job Automation

Corresponds to 2025 Hazard 9 (Economic Displacement & Job Automation).

Why this one now?

US outplacement tracker Challenger, Gray & Christmas recorded AI as the leading stated reason for job cuts in March 2026 (15,341 announced cuts, 25% of that month's total) and AI-attributed cuts remained a top-three cause through April 2026 (around 21,400 cuts, 26% of that month's total) — the clearest verified US evidence that AI-attributed displacement, as reported by employers themselves, has become a persistent line item rather than a one-off event.

Independent labour-market research complicates a simple 'mass unemployment' narrative. Goldman Sachs (April 2026) estimated a net 16,000 US jobs eliminated by AI per month (~192,000 a year) — measurable, but far below the scale some 2025 commentary anticipated. Research cited via NBER found workers aged 22–25 in the most AI-exposed occupations saw a 13% employment decline between November 2022 and December 2025, driven mainly by reduced hiring of new entrants rather than layoffs of existing staff — a 'quiet contraction' pattern rather than visible mass redundancy.

The same research found nominal wages in high-AI-exposure industries rose 16.7% since late 2022 versus 7.5% economy-wide, consistent with a bifurcating labour market: rising rewards for adaptable incumbents alongside a shrinking door for new entrants.

Who may be harmed

Younger and entry-level workers in AI-exposed occupations, disproportionately (evidenced 13% employment decline for 22–25 year-olds in exposed roles); administrative, content-production and entry-level financial-services roles.

Controls — what changed

Medium term: the 2025 baseline's recommendation for augmentation-first design and funded retraining remains unmet at population scale; the entry-level hiring effect specifically suggests transition support needs to target new labour-market entrants, not only displaced incumbents.

Humanitarian reflection by Peter Collins

The 2025 fear of sudden mass unemployment hasn't materialised as such — but a quieter, generational effect has: young people finding it harder to get a first foothold. That may be the more durable story.

Evidence

• Challenger, Gray & Christmas job-cut reports (March–April 2026), cited in AI Job Displacement Statistics 2026 roundups
• Goldman Sachs labour-market note (April 2026)
• NBER-cited research on 22–25 age-group employment decline, cited in "GAGI" working paper (2026)

Hazard 10 — AI-Enabled Misinformation, Disinformation & Democratic Integrity

Merges 2025 Hazard 10 (Misinformation & Mass Manipulation) and Hazard 17 (Disinformation in Democratic Processes) — the 2025 document itself flagged these as overlapping.

Why this one now?

The World Economic Forum's Global Risks Report 2026 ranks misinformation and disinformation as the single most severe short-term global risk for the second consecutive year, explicitly citing AI-generated deepfakes and synthetic audio sophisticated enough to deceive informed audiences.

Against that, the Knight Columbia analysis referenced under Hazard 6 found that in the 38 countries tracked, election-related deepfakes had less measurable real-world electoral impact through the review period than had been feared in 2025 — evidence of resilience building (fact-checking, platform labelling, public awareness) partially offsetting rising content volume, though the WEF ranking suggests experts still judge the underlying trend as worsening.

Provenance and labelling regulation is now partially live: the EU AI Act's Article 50 transparency duties for AI-generated or manipulated content took effect on schedule on 2 August 2026, giving regulators a concrete tool that did not exist at the 2025 baseline.

Who may be harmed

Electorates and information ecosystems broadly; specific harm remains hardest to attribute to individuals compared with the fraud-focused hazards above.

Controls — what changed

Near term: EU Article 50 labelling enforcement data (not yet available at this review's cut-off) will be the key indicator to track for the next edition.

Humanitarian reflection by Peter Collins

Two things are true at once: experts rank this the world's top short-term risk, and the worst-case electoral disruption feared in 2025 has not clearly happened yet in mature democracies. Both deserve to be reported, not just the more alarming one.

Evidence

• World Economic Forum, Global Risks Report 2026
• Knight Columbia election-deepfake analysis (2026)
• European Commission, EU AI Act Article 50 application from 2 August 2026

Hazard 11 — AI's Environmental & Resource Burden

Corresponds to 2025 Hazard 12 (AI's Environmental & Resource Burden).

Why this one now?

This hazard now has the strongest single new evidence source of the whole review: a United Nations University assessment (3 June 2026), co-authored by the 2026 Stockholm Water Prize laureate, found global data-centre electricity use reached approximately 448 terawatt-hours in 2025 (more than all but ten countries), producing roughly 189 million tonnes of CO2 and consuming around 4.5 trillion litres of water, with electricity use projected to reach 945 TWh by 2030 — roughly a doubling in four years.

The report's central methodological finding matters as much as the headline numbers: carbon, water and land footprints do not always move together, so a 'low-carbon' AI deployment is not automatically low-water or low-land, and evaluating sustainability on a single metric can hide where the burden actually lands — often on regions already under water or land stress. The report separately projects AI-driven electronic waste could reach 2.5 million tonnes annually by 2030.

Some efficiency gains are real and documented: Microsoft and AWS reported 25% and 37% improvements in water-use efficiency respectively between 2022 and 2025. But the UN report explicitly names a rebound effect — cheaper, more efficient AI drives higher total usage, increasing the aggregate footprint despite per-query efficiency gains.

Who may be harmed

Communities local to data-centre siting facing water and electricity competition (Ireland's data centres already exceeded total metered household electricity use in 2023); lower-income countries with limited capacity to safely process AI-hardware e-waste.

Controls — what changed

Immediate term: the 2025 baseline's call for environmental standards on AI build-out (closed-loop cooling, circular hardware) remains the correct control; this review upgrades the hazard's likelihood band given documented rather than projected 2025 consumption figures.

Humanitarian reflection by Peter Collins

This is the hazard where the raw numbers moved most since 2025 — not because anything unexpected happened, but because what was a projection in 2025 is now measured, audited consumption.

Evidence

• United Nations University (INWEH), environmental footprint report (3 June 2026)
• PBS News / Associated Press coverage of the UN University report (4 June 2026)
• IndexBox, "AI Data Center Water Consumption" (2026), citing Microsoft and AWS sustainability disclosures

Hazard 12 — Loss of Human Autonomy & Overdependence on AI

Corresponds to 2025 Hazard 13 (Loss of Human Autonomy & Overdependence on AI).

Why this one now?

This hazard did not receive dedicated primary-source research for this review. The nearest relevant cross-cutting evidence is Microsoft's Q1 2026 Work Trend Index finding that 17.8% of the global working-age population was using AI at work (up from 16.3% the previous quarter) and that only 19% of workplace AI users were classified as 'Frontier' (highly capable, deeply integrated) users — suggesting most workplace AI use remains shallow rather than deeply autonomy-displacing, but this is adoption data, not autonomy or dependency data, and should not be over-read.

Flagged for dedicated research (cognitive-offloading studies, over-reliance incident data) before the next edition.

Who may be harmed

Provisional pending dedicated research.

Controls — what changed

Carried forward from 2025 pending dedicated evidence gathering.

Humanitarian reflection by Peter Collins

Adoption is not the same as dependency, and I don't want this book to blur that distinction just to fill a page.

Evidence

• Microsoft Work Trend Index, Q1 2026 (adoption data only — not autonomy-specific)

Hazard 13 — Concentration of AI Power (Corporate & Geopolitical)

Corresponds to 2025 Hazard 14 (Concentration of AI Power).

Why this one now?

The clearest 2026 evidence for this hazard sits inside Hazard 1: a small number of companies (SpaceX, OpenAI, Google, Nvidia, Microsoft, Oracle, Amazon Web Services, and others) now hold direct contracts to supply frontier AI capability into US and allied military systems, and the same small set of hyperscalers identified under Hazard 4 and Hazard 11 control an increasing share of the compute and data-centre infrastructure the rest of the economy depends on.

This is consistent with, not a departure from, the 2025 baseline's concern — but the 2026 evidence (defence contracts, data-centre concentration figures) makes the dependency more concrete and traceable to named companies than the 2025 document was able to show.

Who may be harmed

States and public institutions dependent on a small number of private compute/model providers for both civilian infrastructure and now, directly, military capability.

Controls — what changed

Medium term: the 2025 baseline's call for interoperability, multi-cloud and model/data portability remains unmet and is now more urgent given the direct defence-sector concentration documented this year.

Humanitarian reflection by Peter Collins

This hazard used to feel abstract — 'market concentration' as an economics problem. Seeing the same handful of company names appear under the warfare, infrastructure and cybersecurity hazards in this edition makes it concrete.

Evidence

• The Intercept, "It's Time to Rein in Lethal AI Drones" (26 Aug 2026) — named defence contracts

Hazard 14 — AI-Driven Surveillance & Erosion of Civil Liberties

Corresponds to 2025 Hazard 15 (AI-Driven Surveillance & Erosion of Civil Liberties).

Why this one now?

The most concrete UK development is regulatory rather than a new deployment: the ICO's UK-wide programme now includes a live facial-recognition audit of five police forces, run alongside its wider AI-and-biometrics strategy, per the UK AI regulation tracking sources reviewed for this edition.

This is genuine, if partial, oversight progress since 2025 — an active audit is a stronger control than a policy statement — but this review did not obtain the audit's findings (not yet published at the evidence cut-off), so the hazard's likelihood/severity judgement is carried forward unchanged pending that publication.

Who may be harmed

Individuals subject to live facial recognition and predictive-policing deployments; protest participants, per the 2025 baseline's original concern about chilling effects.

Controls — what changed

Near term: track publication of the ICO's five-force facial-recognition audit findings for the next edition.

Humanitarian reflection by Peter Collins

An audit that hasn't reported yet is a control in progress, not a control delivered — I'm noting it honestly as unfinished business.

Evidence

• UK AI Regulation Tracker (Regula), citing ICO AI and biometrics strategy and live facial-recognition audit of five police forces (2026)

Hazard 15 — AI-Enabled Biological Risks (Synthetic Biology)

Corresponds to 2025 Hazard 16 (AI-Enabled Biological Risks).

Why this one now?

The International AI Safety Report 2026 addresses this directly: it finds general-purpose systems can help produce laboratory instructions, troubleshoot experimental procedures, and answer technical questions relevant to biological and chemical threats — a dual-use capability the report treats as a genuine, if still developing, concern rather than a resolved one.

The Report frames this as a difficult policy trade-off rather than a simple escalation: the same capabilities that could lower barriers to misuse also power legitimate defensive and scientific work (the Report separately notes AI-assisted protein-design gains), and it does not claim the misuse risk has been operationally demonstrated at the 2026 cut-off — only that the enabling capability exists and is improving.

Who may be harmed

Population-scale, in a low-likelihood/high-severity tail-risk scenario; no documented 2026 incident to point to, which is itself worth stating plainly.

Controls — what changed

No material change from the 2025 baseline's recommended controls (dual-use research licensing, screening for synthesised DNA orders); this review found no evidence either that these controls have been implemented at scale or that the risk has crystallised.

Humanitarian reflection by Peter Collins

It would be easy to sensationalise this one. The honest 2026 position is: the capability is real and growing, and there is no confirmed incident — both of those sentences need to stay in the book together.

Evidence

• International AI Safety Report 2026, biological and chemical risks section (3 Feb 2026)

Hazard 16 — AI in Nuclear Command, Control & Communications

Corresponds to 2025 Hazard 19 (AI in Nuclear Control).

Why this one now?

This hazard did not receive dedicated primary-source research for this review. It is closely related to Hazard 1's confirmed findings on military AI adoption and warrants its own dedicated evidence pass (nuclear-state doctrine statements, NC3 modernisation programmes, arms-control-forum positions) before the next edition rather than an inferred score.

Who may be harmed

Provisional pending dedicated research; scored on catastrophic-severity precautionary grounds consistent with the 2025 baseline.

Controls — what changed

Carried forward unchanged from 2025 pending dedicated evidence gathering.

Humanitarian reflection by Peter Collins

Of everything in this book, this is the hazard I most want properly researched before the next edition rather than guessed at now.

Evidence

• No dedicated 2026 source obtained for this review — carried forward from 2025 baseline

2026 Priority Ranking

Ranked by 2026 current risk score. Ties are shown openly rather than given a false rank order; several hazards remain tied at the top of the Critical band.

Conclusion & Call to Action

A year on, the honest headline is not that AI risk exploded in some unpredictable new direction. It is that the trajectories named in 2025 continued — some faster than expected (autonomous weapons, resource consumption, audited fraud losses), some more slowly or with more nuance than feared (mass unemployment, election disruption). Governance moved too, but unevenly: some genuine strengthening (the ICO's active recruitment and facial-recognition audits, the EU AI Act's transparency duties going live), alongside at least one clear loosening of an existing protection (UK automated-decision rules) ahead of the oversight meant to accompany it.

Three actions for the next twelve months:

• Track the specific, named triggers this review identifies for the next edition: the EU's delayed high-risk AI obligations (Dec 2027), the ICO's statutory AI Code of Practice (due 2027) and its facial-recognition audit findings, and the CCW's Seventh Review Conference on autonomous weapons (Nov 2026).
• Commission dedicated research on the three hazards this edition could not properly assess — AI in health and clinical safety, loss of human autonomy, and AI in nuclear command and control — before the 2027 review.
• Resist the temptation, in either direction, to round a mixed picture up to a single story. This year's evidence supports neither 'AI risk is spiralling out of control' nor 'the 2025 warnings were overblown' as a fair one-line summary — both are true in places, and a humanitarian risk assessment owes its readers the more complicated, more honest version.

This remains a living document. It will be reviewed again in 2027.

Glossary

AGI / ASI — Artificial General Intelligence / Artificial Superintelligence — hypothesised AI systems matching or exceeding human capability across most or all cognitive tasks.

CCW — UN Convention on Certain Conventional Weapons — the main international forum discussing regulation of autonomous weapons.

DUAA 2025 — The UK's Data (Use and Access) Act 2025, which amended UK GDPR's automated-decision-making rules.

GPAI — General-purpose AI — an AI model or system capable of performing a wide range of tasks, the focus of the EU AI Act's GPAI obligations and the International AI Safety Report.

ICO — The UK's Information Commissioner's Office, the data-protection regulator.

LAWS — Lethal Autonomous Weapons Systems — weapons that can select and engage targets without direct human input.

L×S — Likelihood multiplied by Severity — the Master Matrix risk score used throughout this book (1–25).

Select Bibliography

European Commission, Digital Strategy — "AI Act" (regulatory status, application timeline, July 2026 Cybersecurity and AI action plan). digital-strategy.ec.europa.eu

Al Jazeera, "What came into force with the EU's AI Act this week — and what didn't", 6 August 2026.

SundAI, "EU AI Act 2026: What Applies Now & What Comes Next", last verified 4 September 2026.

GOV.UK / legislation.gov.uk, Data (Use and Access) Act 2025, and The Data Protection Act 2018 (Code of Practice on AI and ADM) Regulations 2026, SI 2026/425.

Information Commissioner's Office, "Recruitment rewired", 31 March 2026; ICO AI and biometrics strategy; live facial-recognition audit of five police forces (2026).

International AI Safety Report 2026, Executive Summary and Key Updates, published 3 February 2026, chaired by Yoshua Bengio.

United Nations University (INWEH), environmental footprint of AI report, 3 June 2026; UN News coverage, 5 June 2026; PBS/AP coverage, 4 June 2026.

The Intercept, "It's Time to Rein in Lethal AI Drones", 26 August 2026.

Small Wars Journal, "Fully Autonomous Drones Reportedly Kill in Ukraine", August 2026.

Autonomy Global, "Ukraine War Becomes Live Test Bed for AI-Enabled Autonomous Weapons", January 2026.

UN Convention on Certain Conventional Weapons, Seventh Review Conference, scheduled November 2026.

FBI Internet Crime Complaint Center, 2025 Internet Crime Report (AI-enabled fraud category).

Resemble AI, H1 2026 Deepfake Threat Report; full-year 2025 Deepfake Threat Report.

University of Florida deepfake-detection study, February 2026 (as reported by Adaptive Security).

World Economic Forum, Global Risks Report 2026.

Knight Columbia, analysis of election-related deepfakes across 38 countries (2026).

Challenger, Gray & Christmas, job-cut reports, March–April 2026.

Goldman Sachs, labour-market research note, April 2026.

Microsoft, Work Trend Index, Q1 2026.

UK AI Regulation Tracker (Regula), last reviewed 26 August 2026.

About the Author

Peter Collins commissioned and leads the Humanitarian AI Risk Assessment series, an independent, evidence-led effort to track AI-related risk to people, communities and the planet, and to name proportionate controls decision-makers can act on. This is the 2026 annual review of the assessment first published on 23 September 2025.

Back to GOA resources