In August 2026, US security agencies warned that attackers were using AI-generated scripts to scout and probe internet-exposed controllers in water, energy and other sectors. They described reconnaissance and tool-building, not a service outage. Separately, attacks in July 2026 disrupted operations at water systems in several states, but officials did not link those to AI. Our end-2028 forecast is consequential AI-assisted service interruption across multiple systems. It extends beyond a bounded demonstration; the full chain is not established by the current cases. Recovery is Unresolved: comparable incidents can inform restoration and fallback, but the records reviewed here do not establish adequate drinking-water and sanitation continuity for the stated AI-assisted scenario. A May 2026 Dragos investigation also documented AI-assisted IT intrusion at a Mexican municipal water utility and an unsuccessful attempt to access an OT-adjacent interface. It did not establish an operational-system breach or household service interruption.

AI’s contribution

AI cuts the skill and time needed to target industrial controllers.

Other pathways we track

data-center water demand straining local supply; errors in AI-supported treatment decisions.

Selected connections

Depends on Electricity, heating and cooling · relied on by Healthcare and medicine

Sources

1. NSA and partner agencies, controller threat report (Aug 2026)
2. FBI and EPA alert on water-sector disruptions (July 2026)
3. CDC, drinking water advisories
4. Dragos, AI-assisted water-utility intrusion, May 2026
5. Google Threat Intelligence, agent-enabled attack workflows (September 8, 2026)

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• Critical systems that are hard to supervise

Now E1/H2/Unresolved: AI assistance is documented at early attack stages, but the reviewed cases do not connect it to a household service failure. Temporary loss of service can be costly or disruptive. Recovery duration remains unresolved.

End-2028 forecast and assumptions

End-2028 E3/H2/Unresolved forecasts consequential AI-assisted disruption across multiple systems, rather than a single bounded demonstration. It assumes attackers transfer increasingly capable tools to exposed controls despite segmentation and manual procedures. The July 2026 attacks are not attributed to AI; documentation lag is not treated as proof of hidden AI use. Substitute drinking water remains in the household baseline and does not settle sanitation continuity.

Forecast confidence: low. These are projections under the stated household circumstances, not findings that the future harm has already occurred.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

Household with no access to safe substitute water or workable sanitation: H3 (basic needs threatened). E and F unchanged.

• Forecast review triggers: Confirm the forecast with investigated AI-assisted actions linked to household-relevant service interruption across systems. Lower it if isolation, access controls and exercised manual procedures reliably prevent that chain. A single end-to-end incident or test would support a bounded demonstration, not by itself the full operational-scale forecast.
• Limits: the assessment uses the agencies' accessible summary; the full advisory and original utility notices have not been inspected.

Change history

October 19, 2026: initial household assessment.

October 19, 2026: updated with July 2026 water-system attacks and August 2026 federal advisory; scores unchanged.

October 19, 2026: end-2028 forecast revised from 2/2/U to 3/2/U; forecast bar 4 to 5. Current scores and household baseline unchanged.

A consumer-group study released in December 2025 recorded different offered grocery prices on Instacart. It observed offers, not completed purchases or established yearly household losses. Instacart denied using personal data in those tests and announced that it ended them on December 22, 2025.

Walmart reported in March 2026 that roughly 2,300 US stores used digital shelf labels. These displays update shelf prices; they do not by themselves establish individual pricing or collect shopper data. In October, Walmart said it does not and will not use individual customer data or the labels for personalized pricing. Amazon also denied using personal information to charge higher prices in a statement supplied to NPR. These are the companies’ stated positions, not independent audits.

Alpha Modus patent US 12,591,901, granted March 31, 2026 with priority dating to 2013, describes tracking shoppers, analyzing information and sending personalized product offers. Its disclosed devices include cameras and sensors; its claimed purchase option includes a limited-time reduced price. This establishes a patented capability, not deployment, widespread overcharging or proof of higher net grocery bills. A changing price shared by everyone is dynamic pricing; a price or offer tailored to a shopper is personalized pricing.

The FTC’s January 2025 preliminary findings described intermediaries offering pricing and promotion tools using consumer data. Their clients included grocery retailers, but that does not establish that every client used individualized prices. The FTC proposed a personalized-pricing enforcement policy in August 2026; that announcement is not a nationwide ban.

Enacted limits have different scopes and start dates. Maryland Chapter 154, approved April 28, 2026 and effective October 1, restricts covered food retailers and delivery services from using personal data to set higher prices for covered tax-exempt food; statutory exceptions apply. New Jersey signed its Fair Price Protection Act on July 23, 2026, restricting personal-data-based grocery pricing while preserving qualifying loyalty programs and discounts. Passage does not mean all provisions are already operating. Connecticut’s operative pricing provisions are in Public Act 26-130, section 11, effective July 1, 2027, rather than the earlier October 2026 date; the legislature’s revised summaries record that change. Seattle Ordinance 127515 was signed October 5, 2026; its new pricing chapter takes effect September 1, 2027. It covers specified grocery retailers, mixed-use stores and delivery providers, with defined exclusions. These are targeted protections, not a universal ban on all algorithmic pricing.

AI’s contribution

AI-driven systems set the varied prices.

Other pathways we track

wrong AI food-safety advice; AI-related supply disruption (scored under Logistics).

Selected connections

Depends on Work and income, Banking, Electricity, heating and cooling, Logistics · relied on by: no selected links

Sources

1. Consumer Reports and partners, data and methods
2. Company announcement ending item-price tests (Dec 2025)
3. FTC proposed policy statement on personalized pricing (Aug 2026)
4. Alpha Modus, US patent 12,591,901 (March 31, 2026; capability, not deployment)
5. Walmart, digital shelf labels in roughly 2,300 stores (March 2, 2026)
6. Walmart, statement rejecting personalized pricing (October 5, 2026)
7. FTC, preliminary surveillance-pricing study findings (January 2025)
8. Maryland, Chapter 154, enacted food-pricing law (effective October 1, 2026)
9. New Jersey governor, Fair Price Protection Act signed (July 23, 2026)
10. Connecticut, Public Act 26-130, section 11 (operative pricing provisions; July 1, 2027)
11. Connecticut legislature, revised summary noting pricing-law delay to July 1, 2027
12. Seattle, Ordinance 127515, enacted text and signature history (pricing chapter effective September 1, 2027)
13. NPR, Amazon statement added October 13, 2026 (original reporting carried by KTEP)

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• Personalized pricing
• Economic allocation shaped by systems people cannot reliably override

E2 is a historical demonstration; the named tests have ended. H1: the study recorded offered prices, not purchases or yearly losses. F1: the baseline shopper can buy elsewhere before paying.

End of 2028: E3 forecasts consequential AI-assisted pricing operating at scale; confidence is low. H2 reflects a persistent, meaningful added grocery cost. F3 reflects uncertainty about an adequate affordable alternative if competing stores adopt similar pricing. The current recovery answer describes today’s F1 workaround; the forecast does not assume that workaround will remain adequate.

Forecast assumptions

The forecast assumes competing retailers copy profitable methods where lawful and that state limits cover some, not all, shoppers. Digital labels, patents and profiling show capability, not current personalized overcharging. Retailer commitments, competition, enforcement and affordable alternatives could limit adoption. The forecast change follows a revised adoption assumption; it does not require a forecast tie-break.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

A household with a tight food budget, no affordable alternative shopping route and limited ability to travel or substitute could struggle to obtain adequate food if competing stores adopt similar pricing by 2028. For that household, the end-2028 conditional assessment is E3/H3/F3: adequate food access is threatened. Pathway development and recovery are inherited from the main forecast. This does not change the main household baseline or either bar.

Forecast review triggers

Confirm: a regulator, court or audit documents a major retailer setting individual prices from personal data, or a surcharge for buying without being identified. Lower: broad bans or federal rules are in force, and no use is documented, by the 2027 reviews.

Change history

October 19, 2026: initial household assessment.

October 19, 2026: end-2028 forecast raised to E3/H2/F3; bar 3 to 5, based on the assumption that competing stores may adopt similar pricing tools. Current scores unchanged.

October 19, 2026: clarified the patent description as shopper tracking and individual offers, without characterizing the patented system as AI-assisted; scores unchanged.

Berkeley Lab’s June 2026 report estimates total US data-center electricity consumption at 192 TWh in 2024 and projects 464 TWh in its 2028 reference case. The same report’s 2030 reference case is 649 TWh, or 11.8% of projected US electricity use. The report contains an explicit 2028 estimate; the 2030 number is not substituted for it. These figures cover all data-center electricity use, not AI alone, and do not directly predict household bills.

PJM’s July 14, 2026 results for the June 2028 to May 2029 delivery year put the capacity clearing price at the FERC-approved cap of $325 per MW-day. That was 2.5% below the previous auction’s $333.44 cap, so it is described here as sustained near-record pricing, not another annual price increase. PJM valued the cleared capacity at $16.4 billion but cautioned that this is not the total cost to electricity customers: self-supply and bilateral contracts hedge some exposure. Capacity is only one component of a retail bill.

The independent market monitor attributes a substantial part of recent capacity-cost pressure to data-center load, and Maryland’s consumer advocate traces capacity costs into household bills. Forecast load can affect capacity prices before the facilities are operating. Successive delivery-year auctions sustain expenses over time; their clearing prices are not added together as simultaneous household surcharges. Credits, hedging and cost-allocation reforms can reduce household exposure.

PJM reported committed capacity 6,831 MW below its reliability requirement for 2028–29, while retaining a 14.7% reserve margin. It said demand was growing faster than electricity supply. Falling short of the reliability standard means slimmer reserves and greater risk, not proof that outages will occur. New supply, backstop procurement and flexible large loads could improve the outcome before the delivery year.

AI’s contribution

AI-server growth is a major contributor to expanding data-center electricity demand; the AI-only share of household costs has not been measured.

Regional note

The assessed household remains in an affected PJM retail territory. Texas has a separate grid, market and household exposure. ERCOT reported more than 438 GW of large-load connection requests in June 2026, nearly 89% from data centers. Those requests are not operating demand or a forecast that every project will connect. Its Batch Zero framework limits connections to what the grid can reliably support and includes options for onsite generation and curtailment of large customers. These findings support tracking regional supply constraints; they do not establish that homes must be cut off before data centers.

Other pathways we track

AI-assisted attacks on grid control systems (plausible but undemonstrated; a household relying on powered medical equipment without adequate backup or a clinical bridge would face life-threatening risk if it happened); faulty AI recommendations in grid operations.

Selected connections

Depends on: no selected links · relied on by Food and essential goods, Water and sanitation, Healthcare and medicine, Transportation, Communication and digital security, Education and childcare, Emergency response and safety, Logistics

Sources

1. Berkeley Lab data-center energy report
2. Maryland Office of People's Counsel (Sept 2025)
3. Monitoring Analytics, market monitor presentation (June 2026)
4. Berkeley Lab, June 2026 update (institutional abstract)
5. Berkeley Lab, June 2026 report (full text; explicit 2028 reference case)
6. PJM, July 14, 2026 auction results for 2028–29: price cap and reliability shortfall
7. ERCOT, Batch Zero framework and large-load connection requests (June 18, 2026)

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• Data-centre electricity demand
• Critical systems that are hard to supervise

Current assessment remains E3/H2/F3. End of 2028: E3/H3/F3, with low confidence in the H3 household forecast. The pathway is already established, so greater demand alone does not require E4. H3 forecasts a threat to essential heating or cooling for the same household, rather than merely a larger discretionary expense. F3 remains because an adequate affordable substitute for essential grid power is difficult or uncertain.

Forecast assumptions

Forecast lag rule applied: H2 and H3 remain similarly plausible after considering persistent costs, supply constraints, household budget flexibility and countermeasures. The higher defensible forecast assumes repeated elevated electricity expenses through 2028 materially erode that flexibility, and that credits, cost-allocation reforms or other protections do not substantially offset the household burden. This is a forecast judgment, not a household outcome proven by the auction result. The baseline household and the cost pathway are unchanged.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

Household with little budget margin and no way to cut essential heating or cooling: H3 now and in the end-2028 conditional assessment. Basic needs can be threatened sooner than for the main household. E3 and F3 are inherited; no change to either main bar.

• Limits: no AI-only dollar figure; the monitor does not assign all capacity-cost growth to data centers.
• Forecast review triggers: Confirm: sustained retail pass-through despite protections, with evidence that the assessed households are struggling to maintain essential electricity. Rising arrears or disconnections are indicators to investigate; they do not alone identify data centers as the cause. Lower: cost-allocation rules, bill relief, supply additions or demand controls materially reduce household exposure before mid-2028. Adoption of a rule alone is not enough; check its effective date, scope and effect.
• Full sources: Maryland Office of People's Counsel (July 2025), credits and protections: https://content.govdelivery.com/accounts/MDOPC/bulletins/3ead2f8

Change history

October 19, 2026: initial household assessment.

October 19, 2026: end-2028 household harm forecast raised from H2 to H3; forecast bar changed from 5 to 6. The revised outlook assesses cumulative household expenses and persistent demand-supply pressure. Current scores unchanged.

A Senate investigation found that one large insurer's denial rate for post-hospital care rose from 10.9% in 2020 to 22.7% in 2022 as it automated reviews, and that another deployed AI to cut skilled-nursing spending. A bipartisan Senate letter in July 2026 said staff may have been pressured to follow machine-generated recommendations. Federal inspectors found plans overturned nearly all appealed nursing-facility denials, which points to bad initial decisions, though not to AI as the cause in each case. Insurers say humans make final decisions; human sign-off does not rule out heavy influence from a predictive tool. Medicare's own AI-assisted review pilot runs in six states, and we expect the reach and burden of AI-assisted review to increase through 2028. Wider deployment does not by itself prove inappropriate denials or grave injury.

Before and after care

Before care, the WISeR pilot began in six states in January 2026 and covers selected services in Original Medicare, not Medicare Advantage. KUOW reported in May 2026 that Washington hospitals described longer approval waits, including waits of 15 to 20 days at UW Medicine hospitals compared with one to three days previously. These reports concern the review process and do not establish that AI itself caused the delays. CMS says licensed clinicians determine non-payment recommendations. After care starts or is delivered, proposed class actions allege that UnitedHealth and Humana used nH Predict to terminate post-acute coverage, and that Cigna used PxDx to deny payment claims in bulk. The cited orders allow parts of those cases to proceed; they do not establish the allegations as fact. Humana’s August 2026 answer denies using nH Predict to make adverse coverage decisions. Cigna’s alleged algorithm is not treated as proof of AI use.

AI’s contribution

Predictive recommendations shape coverage decisions, and staff may be pressured to follow them.

Other pathways we track

clinical prediction errors; inaccurate AI-generated medical information.

Selected connections

Depends on Work and income, Banking, Electricity, heating and cooling, Water and sanitation, Logistics · relied on by Relationships and mental health

Sources

1. Senate Permanent Subcommittee on Investigations, report release (Oct 2024)
2. Subcommittee letter to UnitedHealthcare (July 2026)
3. HHS OIG, appealed nursing-facility denials (June 2026)
4. CMS, WISeR model (deployment)
5. Healthcare Finance News, UnitedHealth lawsuit advances (February 2025; reporting)
6. Barrows v. Humana, defendant’s amended answer (August 26, 2026)
7. Kisting-Leung v. Cigna, order on motion to dismiss (March 31, 2025)
8. KUOW, Washington hospital reports of WISeR approval delays (May 29, 2026)
9. Lokken v. UnitedHealth, order on motion to dismiss (February 13, 2025)
10. Barrows v. Humana, order on motion to dismiss (August 15, 2025)

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• Decisions that can no longer be reversed
• Sepsis prediction: a separate healthcare pathway

Now E3/H3/F3: official findings link automated and AI-assisted review to consequential denial patterns. Needed care is threatened, and successful appeals do not establish timely adequate care for everyone affected.

End-2028 forecast and assumptions

The end-2028 base assessment remains E3/H3/F3 for the same post-hospital patient. Greater reach and repeated review burdens worsen the outlook without establishing the next underlying maturity or harm anchor. The public forecast position is 6.5, comprising the base total of 6 plus a documented 0.5 display adjustment. This represents qualitative increased pressure between High and Critical, not a measured half-unit of harm. It does not assert E4, H4 or a change in recovery score.

Forecast confidence: low. These are projections under the stated household circumstances, not findings that the future harm has already occurred.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

Patient whose condition is unstable and depends on time-critical care with no bridge: H4 (permanent serious injury or death possible). Recovery Unresolved. E unchanged.

• Limits: no measured AI error rate; the 99.7% appeal-reversal figure cited in the July 2026 letter covers appealed nursing-facility admission denials at one insurer; the full 2024 Senate report was not inspected (the release and letter were).

The bars retain the assessed household baseline of needed post-hospital care. The outpatient pilot and payment-denial cases show related review pathways; they are not newly scored household baselines. Cigna’s PxDx is described here as an automated algorithm, not as established evidence of AI use. Reported pilot delays do not isolate the contribution of AI from the surrounding approval process. Possible expansion to more states is an outlook, not an announced expansion.

Forecast review triggers

Confirm increasing pressure with independent measures of AI-linked review delays, interrupted recommended care and difficulty obtaining timely correction. Lower the forecast if clinician escalation and appeals reliably preserve timely care. Deployment or denial counts alone do not establish inappropriate AI-influenced denial.

Change history

October 19, 2026: initial household assessment.

October 19, 2026: expanded the card to cover denials before care and coverage cutoffs or payment denials after care starts or is delivered; scores unchanged.

October 19, 2026: end-2028 forecast reviewed with updated capability, adoption and household assumptions; scores unchanged. Added an explicit note that pressure is expected to increase despite the unchanged bar total.

October 20, 2026: forecast display moved from 6 to 6.5 to show greater reach and repeated review burden; underlying E3/H3/F3 and current scores unchanged.

A peer-reviewed study of a large freelance platform found fewer jobs and lower earnings for exposed workers after generative tools arrived. Wider evidence is mixed. Early-career employment in exposed jobs has diverged in US payroll data, while a study of Danish administrative records found no measurable effect on earnings or hours two years after generative chatbots reached the public. We expect substitution to reach more service markets by 2028. Our revised central forecast is that sustained loss of most exposed commissions can exhaust this household’s finite savings cushion and threaten essentials despite other support. Employers have cited AI in more than 112,000 announced US job cuts this year through July; those are company statements, and companies can have reasons to cite AI, so we treat them as context.

AI’s contribution

AI produces a substitute for the work itself.

Other pathways we track

AI hiring tools screening out qualified applicants; AI monitoring or performance scoring that cuts paid hours.

Selected connections

Depends on Communication and digital security, Education and childcare · relied on by Banking, Housing and insurance, Food and essential goods, Healthcare and medicine

Sources

1. Hui, Reshef and Zhou, *Organization Science* (2024)
2. Stanford Digital Economy Lab, August 2026 update
3. Humlum and Vestergaard, NBER Working Paper 33777 (revised March 2026)
4. Challenger, September 2026 announced job cuts and employer-cited reasons (October 1, 2026)

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• Early-career employment in AI-exposed occupations
• Employment outlook for AI-exposed occupations
• Economic allocation shaped by systems people cannot reliably override

Now E2/H2/F3: a bounded platform study supports lost work and earnings, with broader evidence mixed. This household has other support; replacing the affected income remains costly or uncertain.

End-2028 forecast and assumptions

End-2028 E3/H3/F4 forecasts consequential substitution across markets, depletion of the same household’s finite buffer and no adequate restoration of the earning function within 30 days. F4 does not mean only the identical job or rate counts as recovery. Other household support need not replace the material earnings deficit. The forecast assumes alternative paid demand and retraining do not provide adequate continuity quickly enough.

Forecast confidence: low. These are projections under the stated household circumstances, not findings that the future harm has already occurred.

Forecast selection: the higher of the independently defended end-2028 positions was retained after reconciliation. The supporting pathway, household and assumptions remain explicit.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

If the household's essentials depend on this income with no bridge: H3 (basic needs threatened). E and F unchanged.

• Forecast review triggers: Confirm with independently substantiated AI-linked lost orders or work across markets, sustained household income deficits and time to adequate replacement earnings. Lower F if this population reliably restores adequate earning capacity within 30 days. Employer-cited layoff announcements alone do not establish causation or recovery time.
• Limits: the platform is not a representative sample of US workers.
• Full sources: Challenger, Gray & Christmas, July 2026 report: https://www.challengergray.com/wp-content/uploads/2026/08/Challenger-Report-July-2026.pdf

Change history

October 19, 2026: initial household assessment. End-2028 harm forecast raised from H2 to H3; forecast bar changed from 5 to 6. Current bar stays at 4.

October 19, 2026: end-2028 forecast revised from 3/3/3 to 3/3/4; forecast bar 6 to 6. Current scores and household baseline unchanged.

A public-radio investigation of Texas insurance complaint files found homeowners paying for tree work to keep coverage, one insurer admitting it had reviewed the wrong aerial report about two months after a state inquiry, and a coverage lapse that left a lender charging for replacement coverage. Pennsylvania's insurance department reminded insurers in 2024 of their obligations to physically confirm damage suggested by aerial imagery. Lender-placed coverage may protect mainly the lender and may leave important household risks uncovered. We expect wider use across insurers by 2028.

AI’s contribution

AI-assisted property assessment informs the renewal workflow; data quality and human review can also contribute to a wrong decision.

Regional note

the strongest outcome reporting is from Texas; insurance law and replacement options vary by state (affects F).

Other pathways we track

AI tenant-screening scores that block rental applications; algorithmic rent-setting.

Selected connections

Depends on Work and income, Banking · relied on by: no selected links

Sources

1. KUT/TPR investigation (May 2025)
2. Pennsylvania Insurance Department notice on aerial imagery
3. NAIC, artificial intelligence in insurance

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• Economic allocation shaped by systems people cannot reliably override

Now E2/H3/F3: documented household cases establish the renewal failure route, but not a measured multi-insurer pattern. Required coverage is threatened and disputes or replacement can take weeks to months.

End-2028 forecast and assumptions

End-2028 E3/H3/F3 forecasts consequential erroneous renewals across insurers as shared assessment tools spread. This is still a mortgaged homeowner’s aerial-assessment pathway, not a silent switch to rent-setting, auto claims or displacement. Physical verification and enforcement remain counterforces.

Forecast confidence: low. These are projections under the stated household circumstances, not findings that the future harm has already occurred.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

• Forecast review triggers: Confirm with multi-insurer audits or substantiated complaints linking AI-rated imagery to consequential renewal errors and inadequate affordable replacement. Lower the forecast if physical verification and prompt coverage continuity materially prevent household loss.
• Limits: individual cases, not a nonrenewal error rate.

Change history

October 19, 2026: initial household assessment.

October 19, 2026: end-2028 forecast reviewed with updated capability, adoption and household assumptions; scores unchanged.

Documented wrongful arrests after false face-recognition matches range from a Detroit man held about 30 hours to a woman reported in 2026 to have been jailed for about six months. Detroit's settlement now requires independent evidence before an arrest. More departments use face search and some have stronger safeguards. Our revised forecast is a consequential pattern of wrongful arrests across agencies by end-2028, assuming independent checking remains uneven. Recovery is Unresolved because release, getting back to work and normal life, and clearing the record happen on very different timelines.

AI’s contribution

AI produces the false match that starts the chain.

Context, not scored

At an August 2026 UN briefing, an ICRC arms official said the organization lacked confirmed evidence of autonomous weapons directly targeting human beings on the battlefield. He also said autonomous weapons were already in use and could kill people while targeting objects. AI assistance and autonomous authority to select and attack are different questions. This context does not change the household scores.

Other pathways we track

AI-synthesized voices in false emergency calls (swatting); errors in AI-assisted emergency-call handling.

Selected connections

Depends on Electricity, heating and cooling, Communication and digital security · relied on by: no selected links

Sources

1. ACLU, *Williams v. City of Detroit* case record
2. *Williams v. Detroit*, settlement and order (2024)
3. Washington Post via ADN, six months jailed (Apr 2026)

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• Critical systems that are hard to supervise

Now E2/H3/Unresolved: documented wrongful arrests establish the mechanism. Liberty and legal standing are threatened; release, clearing records and returning to normal life have different timelines.

End-2028 forecast and assumptions

End-2028 E3/H3/Unresolved forecasts consequential AI-linked false-match arrests across agencies. It assumes deployment and reliance outpace consistent independent-evidence safeguards. Search volume alone is insufficient for E3. This card continues to score the wrongful-arrest pathway rather than unverified emergency-dispatch failures.

Forecast confidence: low. These are projections under the stated household circumstances, not findings that the future harm has already occurred.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

none scored.

• Forecast review triggers: Confirm with independent multi-agency audits or court records connecting false AI matches and inadequate verification to a consequential pattern of arrests. Lower the forecast if audited independent-evidence requirements reliably prevent that pattern. The existence of a rule alone does not establish compliance.
• Context sources: UN Geneva press briefing transcript, August 25, 2026 (responses around 22:14 and 29:44 to 35:00): https://transcripts.un.org/en/briefing/geneva/2026-08-25 ; ICRC autonomous weapons FAQ: https://www.icrc.org/en/article/faq-autonomous-weapon-systems
• Limits: case histories are not a national rate; some case narratives come from plaintiffs' counsel; the weapons statement is a dated account of one organization's knowledge.
• Full sources: ACLU letter on the 2026 case: https://assets.aclu.org/live/uploads/2026/04/2026-04-13-kimberlee_williams_montgomery_letter_final_Redacted-1.pdf

Change history

October 19, 2026: initial household assessment.

October 19, 2026: end-2028 forecast revised from 2/3/U to 3/3/U; forecast bar 5 to 6. Current scores and household baseline unchanged.

Both halves of this pathway have happened, but separately. A 2025 cyberattack on a major US food distributor halted deliveries nationally; trade reporting of company statements put ordering back in about ten days and normal capacity in about three weeks. It was not linked to AI. Threat researchers have separately documented AI agents carrying out parts of intrusion campaigns in other settings. We have not verified a case connecting AI-assisted intrusion to the distribution failure described here. Our revised end-2028 forecast is consequential AI-assisted disruption across distributors, including household delivery effects. This is a forecast extension, not an already-verified household outcome. Many other cards lean on this one, so their outlooks link here rather than scoring supply disruption twice. An August 2026 Gambit report identifies a US distributor among organizations affected by hands-on AI-assisted intrusion. It narrows the sector-specific gap but does not establish interrupted household deliveries.

AI’s contribution

AI speeds up and scales the intrusion.

Other pathways we track

AI planning or agent errors that misroute or cancel orders; correlated automated purchasing errors.

Selected connections

Depends on Electricity, heating and cooling, Communication and digital security · relied on by Food and essential goods, Healthcare and medicine

Sources

1. Google Threat Intelligence Group, adversarial AI report (Sept 2026)
2. Distribution Strategy, food distributor restores operations (July 2025)
3. CISA and partners, guidance on agentic AI risks (May 2026)
4. Gambit, AI-assisted intrusions including a distributor, August 2026

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• Critical systems that are hard to supervise

Now E1/H2/Unresolved: AI-assisted intrusion and distribution outages are documented separately, but the reviewed evidence does not connect AI assistance to household delivery failure. The household has some stock or another seller.

End-2028 forecast and assumptions

End-2028 E3/H2/Unresolved forecasts repeated consequential AI-assisted disruption across firms, not just an isolated test. It assumes offensive automation transfers to ordering and dispatch despite segmentation, manual work and supplier alternatives. The household-delivery link is a forecast, not a fact established by enterprise compromise. IT restoration does not establish household replenishment time.

Forecast confidence: low. These are projections under the stated household circumstances, not findings that the future harm has already occurred.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

No other way to get an essential item before household stock runs out: H3 (basic needs threatened). E and F unchanged.

• Forecast review triggers: Confirm with incident-response or official records connecting material AI assistance to essential-delivery interruptions across firms. Lower the forecast if tested manual operations and alternative supply preserve household continuity despite attacks. No reported incident alone does not prove that a pathway is absent.
• Limits: the threat report is an interested vendor account; inventory, other carriers and manual handling can stop the chain; the distributor chronology is context, not a household recovery measure.
• Full sources: USDOT freight plan guidance noting AI use (Feb 2026): https://www.federalregister.gov/documents/2026/02/24/2026-03648/guidance-on-multimodal-state-freight-plans-and-state-freight-advisory-committees

Change history

October 19, 2026: initial household assessment.

October 19, 2026: end-2028 forecast revised from 2/2/U to 3/2/U; forecast bar 4 to 5. Current scores and household baseline unchanged.

The FBI's 2025 crime report included a section on AI for the first time: 22,364 complaints and $893 million in losses mentioning AI, with investment fraud accounting for $632 million and about $659 million involving cryptocurrency. These figures describe complaints mentioning AI. They may omit unrecognized AI use, while reported AI involvement is not independently verified in every case. They are not a measured total of AI-caused losses. Recovery depends on the payment route; wire-freezing results cannot be assumed for cryptocurrency. Our revised forecast assumes coordinated AI-enabled scam networks repeatedly rebuild operations after interventions by 2028, sustaining harm despite stronger defenses. Current evidence does not by itself establish that future containment outcome.

FinCEN’s September 2026 analysis describes industrial-scale scam centers and professional laundering networks. Its roughly $12.7 billion figure is suspected scam-linked financial activity in reports, not AI-caused household losses. Recent vendor investigations also report agent-enabled misuse and successful campaign disruptions. Organized networks support the forecast mechanism, but neither rising losses nor organization alone establishes defeated containment.

AI’s contribution

AI lets a scam operation run thousands of convincing, tailored conversations and fake endorsements at once.

Other pathways we track

account takeover leading to unauthorized withdrawals (remedies depend on the transaction type and circumstances); AI customer-service loops that stall disputes.

Selected connections

Depends on Work and income, Communication and digital security, Identity, privacy and reputation · relied on by Housing and insurance, Food and essential goods, Healthcare and medicine

Sources

1. FBI IC3, 2025 Internet Crime Report (AI section, pp. 39 to 42)
2. FinCEN Alert FIN-2024-Alert004 on deepfake fraud
3. FBI public service announcement on generated content in fraud (Dec 2024)
4. FinCEN, organized digital-asset investment scam networks (September 3, 2026; suspected activity, not AI-only losses)
5. Anthropic, threat-intelligence investigations (September 2026; vendor case reports)

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• Fraud losses involving AI
• AI spearphishing in controlled tests
• Critical systems that are hard to supervise
• Information systems that outrun trust

Now E3/H3/F4: AI-enabled investment scams operate at consequential scale and can threaten savings needed for essentials. Adequate recovery is not dependable, although some transfers can be stopped or recovered.

End-2028 forecast and assumptions

End-2028 E4/H3/F4 forecasts sustained coordinated harm defeating containment in the named investment-scam pathway. The key assumption is that coordinated networks repeatedly regenerate operations after bank, platform and enforcement interventions while sustaining consequential victimization. More complaints, rising losses or evasion of an individual check do not alone meet E4. This is not a forecast of durable loss of human control across society, nor an assumption that every defense fails. H3 and F4 retain the same household resources and recovery endpoint.

Forecast confidence: low. These are projections under the stated household circumstances, not findings that the future harm has already occurred.

Forecast selection: the higher of the independently defended end-2028 positions was retained after reconciliation. The supporting pathway, household and assumptions remain explicit.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

none scored. Retirees living on savings they can't rebuild may face irreversible loss of their retirement security; we are still looking for evidence specific enough to score it.

• Forecast review triggers: Confirm E4 with substantiated coordinated AI-enabled scam operations that repeatedly sustain consequential victimization through successive containment interventions. Lower E if payment interception, platform action and enforcement materially suppress those operations. Rising AI-related complaint losses prompt review but do not alone validate E4.
• Limits: categories overlap; average losses are not a typical household's loss.

Change history

October 19, 2026: initial household assessment.

October 19, 2026: end-2028 forecast revised from 3/3/4 to 4/3/4; forecast bar 6 to 7. Current scores and household baseline unchanged.

A major email provider's threat team reported AI agents running credential-harvesting campaigns in 2026. US data on AI's role in household account takeovers is still thin: the FBI's 2025 report tagged only 14 phone-number hijacking complaints as involving AI. Account recovery depends on verification and can stall without a clear timeline. We expect AI-assisted takeover to reach consumers more widely by 2028.

AI’s contribution

AI agents automate and scale credential-harvesting campaigns.

Other pathways we track

AI anti-fraud systems wrongly locking people out; telecom operational failures.

Selected connections

Depends on Electricity, heating and cooling · relied on by Work and income, Banking, Transportation, Identity, privacy and reputation, Reliable information and civic participation, Education and childcare, Government services and legal matters, Emergency response and safety, Logistics

Sources

1. Google Threat Intelligence Group, adversarial AI report (Sept 2026)
2. Google Account Help, security holds during recovery
3. FBI IC3, 2025 Internet Crime Report
4. Gambit, AI across the intrusion lifecycle, August 2026

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• AI spearphishing in controlled tests
• Information systems that outrun trust

Now E2/H2/F3: an AI-assisted credential-campaign mechanism is demonstrated. Household outcomes are less well measured, and restoring access can be costly or uncertain.

End-2028 forecast and assumptions

End-2028 E3/H3/F3 forecasts consequential household account compromise at scale and threatened access to essential linked functions. The same adult retains other communication routes and some recovery options; the assumption is that these do not adequately substitute for authentication to essential services. Phishing-resistant sign-in and usable independent verification are counterforces. Enterprise campaign evidence is not itself a household outcome distribution.

Forecast confidence: low. These are projections under the stated household circumstances, not findings that the future harm has already occurred.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

Account is the only practical gateway to care, income or essential paperwork: H3 (basic needs or income threatened). E and F unchanged.

• Forecast review triggers: Confirm with independently substantiated AI-assisted takeovers that prevent access to work, finance, care or other essential linked services. Lower H if alternative verification reliably preserves those functions, or lower E if stronger authentication prevents consequential household deployment.
• Limits: vendor telemetry is interested evidence.

Change history

October 19, 2026: initial household assessment.

October 19, 2026: end-2028 forecast revised from 3/2/3 to 3/3/3; forecast bar 5 to 6. Current scores and household baseline unchanged.

Regulators have documented specific failures. One robotaxi operator recalled its fleet's software after a vehicle dragged a pedestrian following a collision, and another filed a recall covering 3,067 vehicles after robotaxis illegally passed stopped school buses; that software remedy was applied in November 2025. An independent IIHS analysis found fewer police-reportable crash involvements per mile for Waymo's driverless operations than for human drivers across the studied locations overall. That finding does not describe every operator or operating condition. By 2028 we expect the failure pathway to be demonstrated across more operational settings. Wider deployment is one driver of that forecast; confirming E3 requires consequential, attributable outcomes at scale, not mileage alone.

AI’s contribution

AI drives the vehicle.

Regional note

driverless services run in specific cities and conditions; exposure varies sharply by location (affects E).

Other pathways we track

AI routing or ride-allocation that leaves some people without service; navigation errors in accessibility tools.

Selected connections

Depends on Electricity, heating and cooling, Communication and digital security · relied on by: no selected links

Sources

1. NHTSA, post-collision recall investigation record
2. NHTSA, school-bus recall report 25E-084
3. IIHS, driverless crash-rate analysis

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• Critical systems that are hard to supervise

Now E2/H2/Unresolved: regulator records demonstrate specific system failures. The assessed crash is non-catastrophic; evidence does not establish one adequate household recovery timeline.

End-2028 forecast and assumptions

End-2028 E3/H2/Unresolved forecasts consequential attributable failures across operational settings, not merely more mileage. The independent safety result concerns one operator in studied areas and can coexist with wider absolute exposure. No fatal crash or loss of all practical mobility is substituted for the existing central case.

Forecast confidence: low. These are projections under the stated household circumstances, not findings that the future harm has already occurred.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

none scored.

• Forecast review triggers: Confirm with regulator or independent evidence of consequential system contribution across operating contexts, using exposure-normalized outcomes where available. Lower the forecast if safety improvements and effective deployment restrictions substantially prevent those failures.
• Limits: operator safety data is interested-source evidence; one operator's history is not an industry rate; recalled, remedied software is not a live condition.
• Full sources: IIHS study record: https://www.iihs.org/research-areas/bibliography/ref/2374

Change history

October 19, 2026: initial household assessment.

October 19, 2026: end-2028 forecast reviewed with updated capability, adoption and household assumptions; scores unchanged.

A 2026 study in Science found that overly agreeable AI advice made people more convinced they were right in interpersonal conflicts and less willing to make amends. A longer controlled study found heavier voluntary use went along with worse wellbeing, though its assigned conditions showed no significant effect, so it can't show the chatbot caused the decline. A state lawsuit alleges far worse outcomes for vulnerable young users; those are allegations, not findings. By 2028 we expect consequential relationship or care effects in real-world use, beyond controlled changes in judgments and intentions. Prospective outcome studies would test that forecast.

AI’s contribution

The AI's overly agreeable responses are the mechanism.

Other pathways we track

AI-generated sexual coercion (scored under Identity, privacy and reputation); chatbot dependence displacing clinical care.

Selected connections

Depends on Healthcare and medicine, Identity, privacy and reputation · relied on by: no selected links

Sources

1. Stanford report on Cheng et al., *Science* (Mar 2026)
2. MIT Media Lab, longitudinal controlled study of chatbot use
3. Kentucky Attorney General complaint against a companion-chatbot company (Jan 2026)

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• Teen chatbot emotional support: a separate vulnerable group
• Information systems that outrun trust

Now E2/H2/Unresolved: controlled evidence establishes changes in interpersonal judgments and repair intentions, not population-level clinical harm. Current recovery is unresolved.

End-2028 forecast and assumptions

End-2028 E3/H3/F3 forecasts consequential real-world effects and substantial everyday functional impairment after repeated reliance, rather than distress or duration alone. The baseline remains an adult without an initial acute crisis. Repeated use and cumulative conflict reinforcement are forecast assumptions, not new pre-existing vulnerability criteria. F3 describes costly or uncertain functional repair. Short experiments do not establish that multi-year outcome; safer models and beneficial use could limit it.

Forecast confidence: low. These are projections under the stated household circumstances, not findings that the future harm has already occurred.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

Person in acute crisis without timely human support: H4 (death or permanent serious injury possible). Recovery Unresolved. E unchanged.

• Forecast review triggers: Confirm with prospective evidence linking repeated AI advice to substantial real-world functional relationship or care disruption and difficult correction. Lower H if increased use does not impair functioning or if safer systems reliably prevent reinforcement. Keep crisis and self-harm scenarios separate.
• Limits: the full Science article was not retrieved; the crisis row rests on safety evaluations and allegations, not adjudicated causation.
• Full sources: Moore et al., chatbot therapy-response safety evaluation (FAccT 2025): https://spirals.stanford.edu/assets/pdf/moore_expressing_2025.pdf ; FTC inquiry into companion chatbots (Sept 2025, an inquiry, not a finding): https://www.ftc.gov/news-events/news/press-releases/2025/09/ftc-launches-inquiry-ai-chatbots-acting-companions

Change history

October 19, 2026: initial household assessment.

October 19, 2026: end-2028 forecast revised from 3/2/U to 3/3/3; forecast bar 5 to 6. Current scores and household baseline unchanged.

A randomized trial in a Turkish high school found students who practiced math with an unguarded AI tool did worse on later tests without it; a version designed as a tutor removed that penalty. Most US teens now use AI tools, but the studies reviewed here do not establish this learning effect at scale in the US. Our revised central forecast is consequential unassisted-learning losses across multiple teaching settings by end-2028, with low confidence. This is a forward-looking judgment, not an established multi-school finding today. Wider use alone would not establish consequential learning harm at operational scale; tool design and teaching could limit it. The current recovery timeline is unresolved. The revised forecast assumes cumulative deficits threaten educational progression and require costly or uncertain remediation by end-2028; that longer-term outcome is not established by the short experiment.

AI’s contribution

The AI replaces the student's own problem-solving.

Other pathways we track

AI writing detectors wrongly flagging students for cheating; inaccurate AI guidance about childcare services. Childcare has thin coverage in current evidence, a known gap in this card.

Selected connections

Depends on Electricity, heating and cooling, Communication and digital security · relied on by Work and income

Sources

1. Bastani et al., *PNAS* (2025)
2. Author-hosted paper
3. Pew Research Center, how teens use and view AI (Feb 2026)

Now E2/H2/Unresolved: one school experiment supports meaningful short-run learning loss from unguarded use; a guarded tutor avoided that penalty. Current recovery duration remains unresolved.

End-2028 forecast and assumptions

End-2028 E3/H3/F3 forecasts consequential multi-setting learning harm, threatened educational progression after repeated unguarded use and costly or uncertain remediation. H3 here means progression is materially threatened, not that every lower test result is serious harm. Human teaching remains part of the same baseline. The short-run evidence does not establish this cumulative endpoint; guardrail adoption and retained unassisted learning could lower it.

Forecast confidence: low. These are projections under the stated household circumstances, not findings that the future harm has already occurred.

Forecast selection: the higher of the independently defended end-2028 positions was retained after reconciliation. The supporting pathway, household and assumptions remain explicit.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

none scored.

• Forecast review triggers: Confirm with multi-term, multi-setting evidence of impaired unassisted learning that threatens progression and requires costly or uncertain correction. Lower H if deficits remain bounded without threatened progression, or lower E if guarded use prevents consequential harm across settings. Adoption alone is insufficient.
• Limits: one school, short-term outcomes; the design of the tool matters.

Change history

October 19, 2026: initial household assessment. End-2028 pathway development raised from E2 to E3; forecast bar changed from 4 to 5. Current bar stays at 4.

October 19, 2026: end-2028 forecast revised from 3/2/U to 3/3/3; forecast bar 5 to 6. Current scores and household baseline unchanged.

New York City's comptroller audited the city's AI business assistant and found inconsistent answers, including failures on public-service questions. The city disputes some findings and describes fixes. We expect more agencies to adopt AI assistants by 2028, but adoption alone won't show consequential failure at scale. This card does not assess benefit adjudication. The benefits cases reviewed here did not establish an AI contribution sufficient for a scored leading pathway.

AI’s contribution

The AI generates the official answer.

Regional note

audit evidence is from New York City.

Other pathways we track

AI-influenced eligibility or fraud scoring; AI-invented legal citations entering court filings.

Selected connections

Depends on Communication and digital security · relied on by: no selected links. Wrongful detention and its legal aftermath are scored under Emergency response and safety.

Sources

1. NYC Comptroller, MyCity audit
2. *Mata v. Avianca*, sanctions opinion (2023)

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• Decisions that can no longer be reversed

Now E2/H1/F1: an official audit demonstrates inconsistent routine guidance; authoritative information remains available and the baseline resident has not missed a deadline.

End-2028 forecast and assumptions

End-2028 E3/H2/F2 forecasts recurring consequential routine errors across agencies, meaningful avoidable time or expense, and usable correction within two to seven days. It assumes AI-first service channels make clarification slower despite available official information. Neither loss of essential benefits nor a passed deadline is introduced. A broader benefits-decision pathway remains separate research.

Forecast confidence: low. These are projections under the stated household circumstances, not findings that the future harm has already occurred.

Forecast selection: the higher of the independently defended end-2028 positions was retained after reconciliation. The supporting pathway, household and assumptions remain explicit.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

none scored.

• Forecast review triggers: Confirm with multi-agency outcome audits finding meaningful costs from routine AI guidance and two-to-seven-day usable correction. Lower H if consequences remain minor, or lower F if authoritative guidance and escalation reliably restore usable function within 24 hours. Benefit deprivation would require a separately disclosed assessment.

Change history

October 19, 2026: initial household assessment.

October 19, 2026: end-2028 forecast revised from 2/1/1 to 3/2/2; forecast bar 3 to 5. Current scores and household baseline unchanged.

In the 2024 New Hampshire primary, robocalls with an AI-cloned presidential voice told voters to skip the vote, and a federal court has since examined the campaign. The FCC confirmed that existing restrictions on artificial or prerecorded voice calls also apply to AI-generated voices. The evidence reviewed here does not establish changed voter behavior. We expect wider use in the 2028 cycle. The fallback assumes official information is reachable in time; anyone who acts on a false instruction after a deadline can't undo it.

AI’s contribution

AI makes the impersonation sound authentic.

Regional note

the documented campaign was in New Hampshire.

Other pathways we track

AI search or news answers with false civic information; fabricated emergency warnings.

Selected connections

Depends on Communication and digital security, Identity, privacy and reputation · relied on by: no selected links

Sources

1. *League of Women Voters of New Hampshire v. Kramer*, D.N.H. opinion
2. FCC announcement of FCC 24-17, February 8, 2024 (GPO copy)
3. FCC Declaratory Ruling FCC 24-17

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• Information systems that outrun trust

Next evidence check

Recheck election-related AI misinformation at the first scheduled review after the November 3, 2026 US midterm elections.

Now E2/H2/F1: a court-documented campaign demonstrates deceptive dissemination, not measured changes in participation. Official verification remains usable before the deadline.

End-2028 forecast and assumptions

End-2028 E3/H2/F2 forecasts consequential interference at scale and two-to-seven-day usable correction while participation is still possible before the deadline. The F2 forecast assumes more complex personalized deception and clarification delays, not an already-lost opportunity to vote. Missing the voting deadline is not a function restored within seven days and is not silently made the baseline.

Forecast confidence: low. These are projections under the stated household circumstances, not findings that the future harm has already occurred.

Forecast selection: the higher of the independently defended end-2028 positions was retained after reconciliation. The supporting pathway, household and assumptions remain explicit.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

none scored.

• Forecast review triggers: Confirm with substantiated AI-enabled interference with participation and documented two-to-seven-day correction that still restores usable participation before deadlines. Lower F if official authentication and guidance reliably correct the problem within 24 hours. Content volume alone does not establish E3.

Change history

October 19, 2026: initial household assessment.

October 19, 2026: end-2028 forecast revised from 3/2/1 to 3/2/2; forecast bar 5 to 5. Current scores and household baseline unchanged.

The FBI has warned since 2023 that criminals are making fake explicit images from ordinary photos of adults and children for harassment and extortion, and that circulating copies are hard to remove. In April 2026 an Ohio man pleaded guilty in what federal prosecutors believe was the first conviction under the TAKE IT DOWN Act. His victims included at least six adult women; AI-generated videos of at least one were sent to her co-workers. Covered platforms must remove qualifying imagery within 48 hours of a valid request and make reasonable efforts to find and remove known identical copies. Stronger removal could improve recovery, but our revised end-2028 forecast assumes regeneration and redistribution prevent adequate practical correction within 30 days. The reviewed sources do not establish that future recovery outcome.

AI’s contribution

AI creates the images.

Other pathways we track

synthetic identities used to open accounts; facial-recognition misidentification (scored under Emergency response and safety).

Selected connections

Depends on Communication and digital security · relied on by Banking, Reliable information and civic participation, Relationships and mental health

Sources

1. FBI public service announcement on synthetic explicit content (June 2023)
2. US Attorney's Office, Southern District of Ohio, guilty plea (Apr 7, 2026)
3. FTC, complying with the TAKE IT DOWN Act

Related site evidence and scenarios

These cover related pathways or wider scenarios, not additional proof of this card’s household outcome. Control-loss scenarios are conditional and do not set these bars.

• Abuse imagery involving children: a separate vulnerable group
• Information systems that outrun trust

Now E3/H3/F3: recurring documented abuse can substantially impair daily life, relationships or work. Removal and practical recovery are uncertain; not every fake image reaches H3.

End-2028 forecast and assumptions

End-2028 E4/H3/F4 forecasts sustained coordinated harm defeating containment through repeated regeneration and redistribution after interventions. This assumes organized abuse operations maintain consequential harm despite platform removal and enforcement. It is not established merely by the presence of copies, larger upload counts or F4 recovery difficulty. H3 retains the ordinary-adult household: grave irreversible harm is not imposed as the central case. F4 forecasts no adequate practical correction within 30 days. The base and displayed position are both 7, between High and Critical.

Forecast confidence: low. These are projections under the stated household circumstances, not findings that the future harm has already occurred.

The base position adds pathway development (E) and household harm (H) on a 0 to 8 track. Recovery (F) is assessed separately. A documented forecast display adjustment can show worsening pressure between reference points without changing the underlying scores; this is a qualitative judgment, not a precise measurement. For this layout, the reference positions remain Low (2), Moderate (4), High (6) and Critical (8). The endpoint text is aligned with the track edges and the remaining labels have equal intervening space. Marker placement follows the reference labels; intermediate scores fall between their reference points. Equal totals can reflect different combinations; the answers above explain this card’s pathway.

Special-case assessment

none scored.

• Limits: the sources reviewed don't measure how many people are targeted; general sextortion deaths are not attributed to AI; the June 2026 site seizures involved images of famous women, so they are context for enforcement, not for the ordinary-person baseline.
• Full sources: DOJ, seizure of domains publishing nude digital forgeries (June 2026): https://www.justice.gov/opa/pr/united-states-seizes-domain-names-publishing-nude-digital-forgeries-famous-women

Forecast review triggers

Confirm E4 with substantiated coordinated abuse networks that repeatedly sustain consequential harm through platform removal and enforcement interventions. Lower E if practical containment materially suppresses repeated abuse. Confirm F4 through affected-adult follow-up beyond 30 days; lower F if practical recovery becomes reliable within that period.

Change history

October 19, 2026: initial household assessment.

October 19, 2026: end-2028 forecast revised from 3/3/3 to 3/3/4; forecast bar 6 to 6. Current scores and household baseline unchanged. Added an explicit note that pressure is expected to increase despite the unchanged bar total.

October 20, 2026: end-2028 pathway forecast raised from E3 to E4 on the assumption of coordinated abuse sustaining harm despite containment interventions; forecast bar 6 to 7. H3/F4 and current scores unchanged.