Independent evidence trackerEvidence last reviewed October 18, 2026
No advertising · No behavioral tracking · No cross-site tracking · No sponsored analysis · No cookies

AI Disruption Forecast

Benefits and harms · What the evidence shows

What AI could change by 2028.

An evidence based projection of gains, harms and unresolved risks. The register is now composed of 32 gains and 34 harms, with one mixed effect. It tracks what is happening, how far it has spread, and who is likely to benefit or bear the cost.

For household-level effects, explore the Human Risk Map: sixteen parts of everyday life, with current assessments and an end-of-2028 outlook.

Will we keep control of AI by the end of 2028?

The central forecast gives meaningful human control a 78% chance of remaining intact and a 22% chance of being lost by the end of 2028. About one chance in five is a material risk, not a remote possibility.

22%Central estimate · 8% to 41%

Meaningful human control lost

People lose the ability to reverse AI control across important parts of society. If this happens, the consequences described in If meaningful human control is lost could overwhelm or reverse many of the individual area projections, including the gains.

78%Central estimate · 59% to 92%

Meaningful human control retained

People can still direct, correct and, when necessary, stop the most capable AI systems. The 67 records below describe the ordinary impact landscape: measured gains, measured harms and open questions across seven areas of everyday life.

Separately, the tracker assigns a 77% probability (47% to 90%) to at least one material loss of control incident by the end of 2028. An incident may be contained or may contribute to lasting loss of control; the tracker does not currently estimate that split.

How to read the records below: The 67 records describe the ordinary impact landscape. Many of them, including the harms, could occur in either branch. None of them is a forecast that control will be retained. The 78% figure is the mathematical complement of the 22% estimate for loss of control. Forecast estimates were last reviewed October 18, 2026; Impact Register updates were approved October 6, 2026.

Impact Register labels

Confirmed means verified at the stated reach. Partial means credible but limited evidence. Watch means a defined threshold has not yet been observed. These are evidence statuses, not outcome probabilities.

Scenario Layer labels

Observed precursor identifies documented evidence pointing toward a consequence. Evidence based projection requires direct support for an extrapolation. Scenario implication explains what the branch would mean if it occurs.

Area 1 of 7

Work & income

Open full recordClose record

Bottom line: AI is already improving some bounded tasks, especially for less-experienced workers, while evidence also shows displacement and slowdowns in exposed work. The near-term question is how the gains and losses are distributed, not whether every job moves in the same direction.

9 claims4 gains5 harms3 confirmedReviewed 2026-10-06

How to read these records. Status is always paired with current reach. “Confirmed” means verified at the reach shown, not universal. Projected reach states where the effect could plausibly be by the end of 2028.

Gains

4 records
I-001 Confirmed Pilot Gain

In a staggered production rollout at one large firm, access to a generative AI assistant raised customer-support agents' issues resolved per hour by about 15% on average, with the largest gains for less-experienced agents.

Projected reach by EOY 2028
Pilot
Evidence level
O3 · Direct attribution
Who gains
Less-experienced and lower-skilled customer-support agents; the deploying firm
Who bears the cost
Analysis: most-experienced agents saw small quality declines in the study; entry-level support hiring may fall if productivity gains substitute for headcount (not sourced)
What would resolve it
True if by 2028-12-31 at least one additional independent study reports a support-agent productivity gain of 10% or more; False if independent studies find no significant gain; Partially true if gains are found but below 10% or only for novices.
2028 ledger test
By 2028-12-31, gen-AI assistance for customer-support agents will show independently replicated productivity gains of 10% or more at Pilot reach or above in at least one additional firm.
Sources and counterevidence (6)
I-005 Partial Broad deployment Gain

Evidence suggests US workers who use generative AI report saving about 5.4% of their work hours, and self-reported time savings across all employed adults rose from 1.6% to 2.2% of work hours between Q3 2024 and Q2 2026.

Projected reach by EOY 2028
Broad deployment
Evidence level
O1 · Direct attribution
Who gains
Frequent users, especially in computer and mathematics occupations and information services
Who bears the cost
Analysis: workers in firms that convert saved hours into reduced staffing (not sourced)
What would resolve it
True if the RPS time-savings series is at or above 3.0% of all work hours in its last 2028 release; False if at or below 1.6%; Partially true if between.
2028 ledger test
By 2028-12-31, self-reported generative AI time savings will reach at least 3.0% of all US work hours at Broad deployment reach (O1).
Sources and counterevidence (5)
I-007 Confirmed Pilot Gain

Generative AI reduces completion time for well-defined professional writing tasks within the model's tested capability without lowering quality.

Projected reach by EOY 2028
Limited deployment
Evidence level
O3 · Direct attribution
Who gains
Knowledge workers and organizations
Who bears the cost
Workers whose routine writing is commoditized
What would resolve it
By 2028, at least two independent field studies reproduce >=25% time savings without quality loss.
2028 ledger test
Professional writing AI cuts quality-adjusted task time >=25% by EOY 2028.
Sources and counterevidence (2)
I-008 Partial Limited deployment Gain

Evidence suggests job postings requiring AI skills offer a wage premium of about 5% within the same firm and job title.

Projected reach by EOY 2028
Broad deployment
Evidence level
O2 · Contributing attribution
Who gains
Workers able to acquire and signal scarce AI skills
Who bears the cost
Workers lacking training access or credential signals
What would resolve it
By 2028, matched worker or vacancy data establishes whether a >=5% premium persists after controls.
2028 ledger test
Verified AI skills carry >=5% within-role wage premium by EOY 2028.
Sources and counterevidence (2)

Harms and pressures

5 records
I-002 Partial Broad deployment Harm

Evidence suggests early-career employment in highly AI-exposed US occupations was about 19% below the less-exposed counterfactual in the August 12, 2026 study revision using data through June 2026. This is an association; AI’s causal contribution is not isolated.

Projected reach by EOY 2028
Broad deployment
Evidence level
O4 · Contributing attribution
Who gains
More-experienced workers (35 and older) in the same exposed occupations, whose employment was stable or grew
Who bears the cost
Early-career workers aged 22 to 25 in software development, customer service and clerical occupations
What would resolve it
True if by 2028-12-31 a peer-reviewed or second independent administrative-data study confirms a relative employment decline of 10% or more for early-career workers in high-exposure occupations; False if independent analyses find no relative decline; Partially true if a decline is confirmed but AI attribution is rejected or not supported.
2028 ledger test
By 2028-12-31, early-career employment in the most AI-exposed US occupations will remain at least 10% below the relative trend of less-exposed occupations in independent administrative data.
Sources and counterevidence (6)
I-003 Confirmed Limited deployment Harm

On a leading global freelancing platform, job posts for automation-prone writing and coding work fell 21% within eight months of ChatGPT's release relative to manual-intensive work, and image-creation posts fell 17% after image generators launched.

Projected reach by EOY 2028
Limited deployment
Evidence level
O3 · Direct attribution
Who gains
Clients substituting AI for freelance work; freelancers winning the remaining, more complex and higher-paid automation-prone jobs
Who bears the cost
Online freelancers in writing, coding and image creation, who faced fewer posts and more competition per post
What would resolve it
True if by 2028-12-31 at least one independent study on another platform or later period confirms a decline of 15% or more; False if independent evidence shows recovery to trend; Partially true if declines are confirmed but smaller than 15%.
2028 ledger test
By 2028-12-31, demand for automation-prone freelance writing, coding and image work will remain at least 15% below pre-ChatGPT trend on at least one additional major platform.
Sources and counterevidence (5)
I-004 Partial Pilot Harm

Evidence suggests early-2025 AI coding tools slowed experienced open-source developers working in their own mature repositories by about 19%, even though the developers believed the tools had sped them up.

Projected reach by EOY 2028
Pilot
Evidence level
O3 · Direct attribution
Who gains
Analysis: none identified in this study; organizations that calibrate AI adoption to measured rather than perceived speedups
Who bears the cost
Experienced developers on familiar, mature codebases using early-2025 tools
What would resolve it
True if a follow-up independent RCT by 2028-12-31 again finds a slowdown of 10% or more; False if it finds a speedup of 10% or more; Partially true if the effect is within plus or minus 10%.
2028 ledger test
By 2028-12-31, independent RCTs will still show no net speedup of 10% or more for experienced developers on familiar mature codebases.
Sources and counterevidence (5)
I-006 Watch Broad deployment Harm

Unemployment among recent US college graduates could rise further by 2028 if AI substitutes for entry-level work; AI's contribution to the current elevated rate is not established.

Projected reach by EOY 2028
Broad deployment
Evidence level
O0 · Uncertain attribution
Who gains
Not yet identifiable
Who bears the cost
Recent college graduates aged 22 to 27, especially in AI-exposed fields
What would resolve it
True if the rate is at or above 6.5% by the last 2028 release and at least one independent study attributes part of the rise to AI; False if the rate is at or below 4.5% or independent studies find no AI contribution; Partially true if the rate rises but AI attribution is not established.
2028 ledger test
By 2028-12-31, recent-graduate unemployment will be at or above 6.5% with at least one independent study attributing part of the rise to AI.
Sources and counterevidence (3)
I-009 Watch Broad deployment Harm

Employment in highly AI-exposed occupations may fall by at least 2% relative to less-exposed occupations by 2028.

Projected reach by EOY 2028
Broad deployment
Evidence level
O0 · Uncertain attribution
Who gains
Firms and customers if output costs fall
Who bears the cost
Displaced workers and affected communities
What would resolve it
By 2028, official labor data and a predeclared exposure design confirm or reject a >=2% relative decline.
2028 ledger test
Highly AI-exposed occupations lose >=2% employment relative to controls by EOY 2028.
Sources and counterevidence (2)

In the control-loss branch: see SC-003 · Economic allocation, SC-005 · Concentrated access and SC-007 · Human direction.

Area 2 of 7

Health & medicine

Open full recordClose record

Bottom line: AI can improve screening and access, but poorly validated tools and automated decision systems can also miss patients or constrain care. The strongest benefits still depend on clinical validation, careful deployment, and meaningful human oversight.

8 claims3 gains5 harms1 confirmedReviewed 2026-10-06

How to read these records. Status is always paired with current reach. “Confirmed” means verified at the reach shown, not universal. Projected reach states where the effect could plausibly be by the end of 2028.

Gains

3 records
I-010 Confirmed Pilot Gain

In a randomized trial of 105,915 women in Sweden's population screening program, AI-supported mammography detected more cancers, raised sensitivity from 73.8% to 80.5% at equal specificity, cut screen-reading workload by 44%, and was non-inferior on interval cancers.

Projected reach by EOY 2028
Pilot
Evidence level
O3 · Direct attribution
Who gains
Women in screening programs (more invasive cancers found, fewer aggressive interval cancers); radiologists (lower reading workload)
Who bears the cost
Analysis: women in programs adopting AI without equivalent validation; trial covered one AI system, one vendor, one screening round
What would resolve it
True if by 2028-12-31 a second trial outside Sweden confirms non-inferior or better interval-cancer rates; False if a trial finds significantly worse rates; Partially true if only detection-rate (not interval-cancer) replication exists.
2028 ledger test
By 2028-12-31, AI-supported mammography will show replicated non-inferior interval-cancer outcomes in at least one additional country at Pilot reach or above (O3).
Sources and counterevidence (5)
I-011 Partial Pilot Gain

Evidence suggests ambient AI scribes modestly reduce physicians' note-writing time (about 9.5% versus usual care for one of two tools tested) with possible improvements in burnout measures.

Projected reach by EOY 2028
Pilot
Evidence level
O3 · Direct attribution
Who gains
Outpatient physicians with high documentation burden
Who bears the cost
Analysis: patients if AI-drafted notes contain unreviewed errors; physicians using the tool that showed no significant time saving
What would resolve it
True if a non-deployer RCT by 2028-12-31 confirms time savings of 10% or more; False if it finds no saving; Partially true if savings are below 10%.
2028 ledger test
By 2028-12-31, an independent (non-deployer) randomized trial will confirm ambient AI scribes reduce physician documentation time by at least 10%.
Sources and counterevidence (5)
I-013 Partial Pilot Gain

Evidence suggests point-of-care autonomous AI eye exams sharply increase diabetic eye exam completion among youth with diabetes (100% vs. 22% within six months in a randomized trial of 164 participants).

Projected reach by EOY 2028
Pilot
Evidence level
O3 · Direct attribution
Who gains
Youth with diabetes, including racially and ethnically diverse and underserved patients with low baseline screening rates
Who bears the cost
Analysis: patients with false-negative AI results; clinics bearing device costs (not sourced)
What would resolve it
True if an independent RCT by 2028-12-31 confirms a completion gain of 20 points or more; False if it finds no gain; Partially true if the gain is smaller.
2028 ledger test
By 2028-12-31, an independent randomized trial will confirm autonomous AI eye exams raise diabetic eye exam completion by at least 22 percentage points.
Sources and counterevidence (3)

Harms and pressures

5 records
I-012 Partial Pilot Harm

Evidence suggests routine use of AI polyp detection may erode endoscopists' unassisted skill: adenoma detection in non-AI colonoscopies fell from 28.4% to 22.4% after AI was introduced at four Polish centres.

Projected reach by EOY 2028
Pilot
Evidence level
O2 · Contributing attribution
Who gains
Analysis: patients during AI-assisted procedures (CADe raises detection per prior trials, not sourced this round)
Who bears the cost
Patients undergoing unassisted colonoscopy by AI-exposed endoscopists
What would resolve it
True if a prospective or controlled study by 2028-12-31 confirms an unassisted detection decline of 3 points or more; False if it finds no decline; Partially true if a smaller decline is found.
2028 ledger test
By 2028-12-31, a prospective study will confirm that sustained AI exposure lowers unassisted adenoma detection by at least 3 percentage points.
Sources and counterevidence (3)
I-014 Watch Broad deployment Harm

Large Medicare Advantage insurers' post-acute prior-authorization denial rates rose sharply (UnitedHealthcare from 8.7% in 2019 to 22.7% in 2022) while they deployed automated and predictive review tools; AI's contribution to the increase could be material but is not established.

Projected reach by EOY 2028
Broad deployment
Evidence level
O2 · Uncertain attribution
Who gains
Insurers (lower post-acute care spending); Analysis: none identified for patients
Who bears the cost
Medicare Advantage enrollees needing skilled nursing, inpatient rehabilitation or long-term acute care after hospitalization
What would resolve it
True if by 2028-12-31 an official source shows the largest insurers' post-acute denial rates remain 10 or more points above smaller insurers and at least one official source ties the gap partly to automated review; False if the gap falls within 5 points; Partially true otherwise.
2028 ledger test
By 2028-12-31, large Medicare Advantage insurers' post-acute denial rates will remain at least 10 points above smaller insurers', with automated review cited as a contributing factor by an official source.
Sources and counterevidence (4)
I-015 Watch Limited deployment Harm

Use of general-purpose chatbots for emotional support by teens could expose vulnerable young users to harmful responses; population-level mental-health effects are not yet measured.

Projected reach by EOY 2028
Limited deployment
Evidence level
O0 · Uncertain attribution
Who gains
Not yet identifiable
Who bears the cost
Not yet identifiable; Analysis: teens in distress relying on chatbots instead of human support
What would resolve it
True if by 2028-12-31 a controlled or population study finds harm attributable to teen chatbot use for emotional support; False if such a study finds no harm or benefit; Partially true if findings are mixed across subgroups.
2028 ledger test
By 2028-12-31, a controlled or population study will find measurable harm attributable to teens' use of chatbots for emotional support.
Sources and counterevidence (2)
I-016 Partial Pilot Harm

Evidence suggests a widely deployed proprietary sepsis prediction model, in the version evaluated in 2021, missed about two-thirds of sepsis cases at one external health system.

Projected reach by EOY 2028
Broad deployment
Evidence level
O2 · Direct attribution
Who gains
Health systems that learn to demand local validation
Who bears the cost
Patients and clinicians relying on poorly calibrated alerts
What would resolve it
By 2028, version-specific multicenter external validations establish whether broad deployments clear safety thresholds.
2028 ledger test
Broad sepsis AI deployments achieve >=80% sensitivity with acceptable alert burden by EOY 2028.
Sources and counterevidence (1)
I-017 Watch Limited deployment Harm

AI-related software or algorithm failures in medical devices may cause verified serious patient harm by 2028.

Projected reach by EOY 2028
Broad deployment
Evidence level
O1 · Uncertain attribution
Who gains
Patients if stronger surveillance prevents recurrence
Who bears the cost
Patients exposed to unsafe device behavior
What would resolve it
By 2028, regulator findings determine whether >=10 serious injuries have primary algorithmic causation.
2028 ledger test
Medical AI causes >=10 regulator-adjudicated serious injuries across two device classes by EOY 2028.
Sources and counterevidence (1)

In the control-loss branch: see SC-001 · Irreversible decisions, SC-006 · Discovery and governance and SC-007 · Human direction.

Area 3 of 7

Education & skills

Open full recordClose record

Bottom line: Structured, supervised uses can save teacher time and improve learning. Unguarded use can weaken unassisted performance, widen access gaps, or wrongly flag student work.

11 claims6 gains5 harms2 confirmedReviewed 2026-10-06

How to read these records. Status is always paired with current reach. “Confirmed” means verified at the reach shown, not universal. Projected reach states where the effect could plausibly be by the end of 2028.

Gains

6 records
I-018 Partial Pilot Gain

Evidence suggests a pedagogically designed AI tutor produced more than twice the learning gains of an in-class active-learning session, in less time, in a randomized crossover trial with 194 Harvard physics students.

Projected reach by EOY 2028
Pilot
Evidence level
O3 · Direct attribution
Who gains
Undergraduates in the studied course; Analysis: students who need self-paced instruction
Who bears the cost
Analysis: none identified in the study; generalization to less-selective institutions untested
What would resolve it
True if an independent RCT at another institution by 2028-12-31 finds significantly higher learning gains; False if it finds none; Partially true if gains are positive but not significant.
2028 ledger test
By 2028-12-31, an independent randomized trial at another institution will replicate significantly higher learning gains from a pedagogically designed AI tutor.
Sources and counterevidence (3)
I-020 Partial Pilot Gain

Evidence suggests a six-week, teacher-supervised GPT-4 after-school program raised Nigerian senior secondary students' test scores by 0.31 standard deviations (0.23 in English), with gains also on end-of-year exams.

Projected reach by EOY 2028
Pilot
Evidence level
O3 · Direct attribution
Who gains
Participating students, with larger gains for female and higher-performing students
Who bears the cost
Lower-performing and lower-socioeconomic students gained less, widening gaps (as summarized in SRC-R027)
What would resolve it
True if an independent RCT by 2028-12-31 finds gains of 0.2 SD or more; False if it finds none; Partially true if gains are smaller.
2028 ledger test
By 2028-12-31, an independent randomized trial in another low- or middle-income setting will replicate learning gains of at least 0.2 SD from supervised LLM tutoring.
Sources and counterevidence (4)
I-021 Partial Pilot Gain

Evidence suggests real-time AI guidance for human tutors raised K-12 students' math topic mastery by 4 percentage points (9 points for students of lower-rated tutors) at about $20 per tutor per year, in a preregistered randomized trial in Title I schools.

Projected reach by EOY 2028
Pilot
Evidence level
O3 · Direct attribution
Who gains
Students of less-experienced or lower-rated tutors in low-income schools
Who bears the cost
Analysis: none identified; tutors noted some suggestions were not grade-appropriate
What would resolve it
True if an independent RCT by 2028-12-31 finds gains of 3 points or more; False if none; Partially true if smaller.
2028 ledger test
By 2028-12-31, an independent randomized trial will replicate mastery gains of at least 3 percentage points from AI guidance for human tutors.
Sources and counterevidence (3)
I-024 Confirmed Pilot Gain

Guided generative AI reduces teacher lesson-planning time by at least 20% without lowering resource quality.

Projected reach by EOY 2028
Limited deployment
Evidence level
O2 · Direct attribution
Who gains
Teachers and students
Who bears the cost
Teachers if generated errors add review work
What would resolve it
By 2028, independent trials in >=3 subjects reproduce the threshold.
2028 ledger test
Guided teacher AI cuts planning time >=20% without lower quality by EOY 2028.
Sources and counterevidence (2)
  • Supporting sourceUsing generative AI for KS3 science lesson preparationEducation Endowment Foundation · Tier 1
  • Supporting sourceAila teacher preparation trialNFER · Tier 1

    464 KS2 teachers at 108 schools. Preparation time was 24% lower with Aila access; assessed resource quality was no worse. Use covered only 7% of lessons and declined.

    Limits: Workload is self-reported rather than directly timed performance. Resource quality is not student achievement; the subject breakdown has not been established for the three-subject replication test.

I-025 Partial Pilot Gain

Evidence suggests generative AI assistance narrowed an education-based task-performance gap in one controlled study, without evidence yet of durable skill gains.

Projected reach by EOY 2028
Limited deployment
Evidence level
O3 · Direct attribution
Who gains
Adults with less formal education
Who bears the cost
Users if short-term assistance is mistaken for skill acquisition
What would resolve it
By 2028, delayed-assessment trials determine whether assisted gap narrowing persists.
2028 ledger test
At least half of AI-assisted education-gap narrowing persists unaided after 30 days by EOY 2028.
Sources and counterevidence (2)
I-027 Watch Pilot Gain

AI tools may deliver measurable learning or access gains for students with disabilities by 2028.

Projected reach by EOY 2028
Limited deployment
Evidence level
O1 · Uncertain attribution
Who gains
Students with disabilities and their teachers
Who bears the cost
Students poorly represented in training and evaluation
What would resolve it
By 2028, two independent controlled studies meet a prespecified educational or access threshold.
2028 ledger test
Two independent studies show meaningful learning or access gains from AI for disabled students by EOY 2028.
Sources and counterevidence (2)

Harms and pressures

4 records
I-019 Confirmed Pilot Harm

In a field experiment with nearly 1,000 Turkish high school math students, unrestricted GPT-4 access during practice lowered later unassisted exam scores by 17% compared with no access, while a tutor with learning safeguards largely avoided the loss.

Projected reach by EOY 2028
Pilot
Evidence level
O3 · Direct attribution
Who gains
Analysis: students given guardrailed tutors (no exam loss, large practice gains)
Who bears the cost
Secondary students using unguarded chatbots as a crutch during practice
What would resolve it
True if by 2028-12-31 an independent study in another setting confirms a loss of 10% or more; False if it finds no loss; Partially true if a smaller loss is found.
2028 ledger test
By 2028-12-31, an independent study will confirm that unguarded chatbot practice reduces unassisted exam performance by at least 10%.
Sources and counterevidence (6)
I-022 Partial Lab Harm

Evidence suggests AI-writing detectors tested in 2023 misclassified essays by non-native English writers as AI-generated at high rates (a 61.3% average false-positive rate across seven detectors on 91 TOEFL essays), exposing those students to false accusations where such detectors were used.

Projected reach by EOY 2028
Lab
Evidence level
O2 · Direct attribution
Who gains
Analysis: none identified
Who bears the cost
Non-native English-writing students in institutions using AI detectors
What would resolve it
True if an independent audit by 2028-12-31 finds false-positive rates above 20% on current detectors; False if below 5%; Partially true if between.
2028 ledger test
By 2028-12-31, independent audits will still find false-positive rates above 20% for non-native English writing on widely used AI-writing detectors.
Sources and counterevidence (2)
I-023 Watch Broad deployment Harm

Widespread teen use of chatbots for schoolwork could weaken independent skill-building where tools are used without learning safeguards; population-level learning effects are not yet measured.

Projected reach by EOY 2028
Broad deployment
Evidence level
O0 · Uncertain attribution
Who gains
Not yet identifiable
Who bears the cost
Not yet identifiable; Analysis: students using chatbots to complete rather than learn
What would resolve it
True if by 2028-12-31 a population or multi-site controlled study links chatbot use to lower unassisted achievement; False if it finds no decline or gains; Partially true if effects depend on guardrails.
2028 ledger test
By 2028-12-31, a population or multi-site controlled study will link unguarded student chatbot use to lower unassisted achievement.
Sources and counterevidence (4)
I-026 Watch Broad deployment Harm

Unequal access and use may produce at least a 10-point socioeconomic gap in regular educational AI use.

Projected reach by EOY 2028
Broad deployment
Evidence level
O1 · Contributing attribution
Who gains
Students with devices, connectivity, and guidance
Who bears the cost
Low-income students and under-resourced schools
What would resolve it
By 2028, representative student surveys establish whether the regular-use gap reaches 10 points.
2028 ledger test
Regular educational AI use differs >=10 points by socioeconomic group at EOY 2028.
Sources and counterevidence (2)

In the control-loss branch: see SC-001 · Irreversible decisions and SC-007 · Human direction.

I-067Partial PilotHarm

Evidence suggests access to the tested course-integrated AI tutor lowered final grades and LMS participation in the matched-course subset of one US university study, with estimates of -0.37 and -0.90 standard deviations.

Projected reach by EOY 2028
Pilot
Evidence level
O3 · Direct attribution
Who gains
None identified for this measured harm outcome (Analysis)
Who bears the cost
Students in the tested courses; broader costs not measured.
Reach and scope
One university; 2,379 students overall, 1,353 in the matched-course subset.
Counterevidence and limits
The full-sample grade effect was not statistically significant. The grade estimate comes from the matched-course subset (1,353 of 2,379 students). Three instructors who did not activate the tutor were excluded after assignment. The analysis plan was registered after data collection but before analysis. Uptake was low. Other tutoring interventions report positive results. Grades and LMS activity do not establish durable unassisted skill loss or isolate actual tutor-use effects.
Boundary
Separate narrow undergraduate result; does not resolve I-023 teen population learning claim.
What would resolve it
True if by 2028-12-31 an independently evaluated randomized or strong quasi-experimental study at another institution finds a statistically significant negative effect on its prespecified primary course-grade or achievement outcome from comparable course-integrated AI tutor access, using a within-course or equivalent comparator; False if qualifying evidence finds no negative effect or a benefit; Partially true if the effect is negative but not statistically significant, or qualifying results conflict. Conflicting results take precedence. Score only when qualifying evidence is available; no qualifying study is a measurement gap, not a False result.
2028 ledger test
By 2028-12-31, an independently evaluated study at another institution will replicate lower course performance from comparable course-integrated AI tutor access under the frozen outcome and comparator rule.
Sources and counterevidence (1)
  • Supporting sourceUniversity AI tutoring trial — full working paperLiu et al., University of Maryland · Tier 3

    Within-course access estimates: grades -0.37 SD; LMS participation -0.90 SD. Full-sample grade effect not significant. Three nonactivating instructors excluded after assignment.

    Limits: Low uptake; course grades and platform logs do not establish durable skill loss. Claimed ITT is limited by post-assignment exclusions. Plan registered after collection, before analysis.

Area 4 of 7

Cost of living & productivity

Open full recordClose record

Bottom line: AI is creating substantial consumer value and may raise firm productivity. Data-center demand, personalized pricing, and uneven adoption could also increase household costs or leave smaller firms behind.

9 claims3 gains6 harms0 confirmedReviewed 2026-10-06

How to read these records. Status is always paired with current reach. “Confirmed” means verified at the reach shown, not universal. Projected reach states where the effect could plausibly be by the end of 2028.

Gains

3 records
I-038 Partial Broad deployment Gain

Evidence suggests US adults' consumer surplus from generative AI tools rose from about $116 billion (July 2025) to $172 billion (March 2026), far exceeding the revenue these tools earn.

Projected reach by EOY 2028
Broad deployment
Evidence level
O2 · Direct attribution
Who gains
Frequent users, workplace users and paid subscribers, who report the highest valuations
Who bears the cost
Analysis: none directly; costs of the infrastructure are borne elsewhere (see CLAIM-025, CLAIM-026)
What would resolve it
True if the tracker or an independent replication shows at least $250 billion by 2028-12-31; False if at or below $116 billion; Partially true if between.
2028 ledger test
By 2028-12-31, estimated US consumer surplus from generative AI will reach at least $250 billion.
Sources and counterevidence (3)
I-041 Watch Broad deployment Gain

AI-adopting firms may raise real sales per employee by at least 1% by 2028.

Projected reach by EOY 2028
Broad deployment
Evidence level
O0 · Uncertain attribution
Who gains
Workers, owners, and customers if gains pass through
Who bears the cost
Non-adopting firms and workers if gains concentrate
What would resolve it
By 2028, representative administrative data with credible identification confirms or rejects a >=1% gain.
2028 ledger test
AI-adopting firms achieve >=1% higher real sales per employee by EOY 2028.
Sources and counterevidence (2)
I-043 Watch Limited deployment Gain

AI-driven productivity gains may reduce quality-adjusted consumer prices by at least 0.5% in two service sectors by 2028.

Projected reach by EOY 2028
Broad deployment
Evidence level
O0 · Uncertain attribution
Who gains
Consumers and competitive service providers
Who bears the cost
Workers or smaller firms if savings come from displacement
What would resolve it
By 2028, official or independent price studies identify two sectors meeting the threshold.
2028 ledger test
AI lowers quality-adjusted consumer prices >=0.5% in two service sectors by EOY 2028.
Sources and counterevidence (2)

Harms and pressures

6 records
I-036 Watch Population-scale Harm

Data-center load growth, much of it for AI, could keep raising electricity costs for households in the PJM region; the grid's market monitor attributes $29.4 billion in capacity charges over four auctions to data-center load, but does not isolate AI's share.

Projected reach by EOY 2028
Population-scale
Evidence level
O4 · Uncertain attribution
Who gains
Generators receiving capacity payments; data-center operators whose load drove the costs
Who bears the cost
Residential and business electricity customers across PJM
What would resolve it
True if by 2028-12-31 the data-center share of capacity charges is at or above 40% and at least one source isolates AI's share of data-center load; False if the share falls below 20%; Partially true if the share stays high but AI's share is not isolated.
2028 ledger test
By 2028-12-31, data-center load will account for at least 40% of PJM capacity charges, with AI's share of that load isolated by at least one independent source.
Sources and counterevidence (6)
I-037 Partial Broad deployment Harm

Evidence suggests AI data-center demand for memory has driven record DRAM and flash price increases in 2026, raising component costs for consumer PCs and smartphones.

Projected reach by EOY 2028
Broad deployment
Evidence level
O2 · Contributing attribution
Who gains
Memory manufacturers; cloud providers securing long-term supply
Who bears the cost
Consumers and device makers buying PCs, laptops and smartphones
What would resolve it
True if an independent price index shows consumer PC or smartphone prices rose in 2026 to 2027 with memory costs cited as a driver; False if consumer device prices fell over the period; Partially true if only component prices rose.
2028 ledger test
By 2028-12-31, independent price data will show AI-driven memory costs raised consumer PC or smartphone prices during 2026 to 2027.
Sources and counterevidence (4)
I-039 Watch Limited deployment Harm

Algorithmic rent setting software could raise rents where widely adopted; a White House analysis estimated it added $3.8 billion to US rents in 2023, about $70 a month per affected unit, but AI's specific role is not established.

Projected reach by EOY 2028
Limited deployment
Evidence level
O2 · Uncertain attribution
Who gains
Landlords using the software
Who bears the cost
Renters in buildings using algorithmic pricing, especially in high-adoption metros
What would resolve it
True if by 2028-12-31 an independent study finds a significant rent premium attributable to AI-based pricing; False if no premium; Partially true if a premium is found for algorithmic pricing without AI attribution.
2028 ledger test
By 2028-12-31, an independent study will attribute a significant rent premium to AI-based rent-pricing software.
Sources and counterevidence (4)
I-040 Watch Limited deployment Harm

Algorithmic and AI-driven personalized pricing could raise the prices some consumers pay for the same goods; the FTC found retailers' pricing intermediaries use personal data such as location, browsing and cart behavior to set individualized prices, but net consumer effects are not yet measured.

Projected reach by EOY 2028
Limited deployment
Evidence level
O0 · Uncertain attribution
Who gains
Retailers and pricing intermediaries; Analysis: some price-sensitive consumers offered discounts
Who bears the cost
Not yet identifiable; Analysis: consumers assessed as willing to pay more
What would resolve it
True if by 2028-12-31 an independent study or FTC finding shows AI-based personalized pricing raised prices for identifiable consumer groups by 5% or more; False if it finds no net increase; Partially true if increases are found without AI attribution.
2028 ledger test
By 2028-12-31, an independent study or FTC finding will show AI-based personalized pricing raised prices for identifiable US consumer groups by at least 5%.
Sources and counterevidence (5)
I-042 Watch Population-scale Harm

U.S. data centres may reach at least 6.7% of national electricity demand by 2028, increasing pressure on generation and grids.

Projected reach by EOY 2028
Population-scale
Evidence level
O0 · Uncertain attribution
Who gains
Cloud and AI users and host communities
Who bears the cost
Ratepayers and communities facing constrained power or infrastructure
What would resolve it
Official 2028 electricity accounts establish the share and associated reliability/cost effects.
2028 ledger test
U.S. data centres use >=6.7% of national electricity in 2028 with documented system pressure.
Sources and counterevidence (2)
I-044 Watch Broad deployment Harm

Small and medium-sized firms could fall behind large firms in realized AI productivity benefits; current evidence shows an adoption gap, not a measured benefit gap.

Projected reach by EOY 2028
Broad deployment
Evidence level
O1 · Contributing attribution
Who gains
Large firms with data, skills, and integration budgets
Who bears the cost
SME owners, workers, and communities
What would resolve it
By 2028, representative business data test both adoption and realized productivity gaps.
2028 ledger test
SMEs trail large firms by >=20 points in AI adoption and >=1% productivity growth by EOY 2028.
Sources and counterevidence (2)

In the control-loss branch: see SC-003 · Economic allocation, SC-005 · Concentrated access and SC-007 · Human direction.

Area 5 of 7

Independence & accessibility

Open full recordClose record

Bottom line: AI can restore communication and expand access to information and navigation. Reliability, bias, privacy, and false accessibility claims determine whether those tools create genuine independence.

10 claims5 gains4 harms1 confirmedReviewed 2026-10-06

How to read these records. Status is always paired with current reach. “Confirmed” means verified at the reach shown, not universal. Projected reach states where the effect could plausibly be by the end of 2028.

Gains

5 records
I-045 Confirmed Pilot Gain

An AI-decoded speech neuroprosthesis has restored communication for a man with ALS, decoding his intended speech with about 97.5% word accuracy.

Projected reach by EOY 2028
Pilot
Evidence level
O2 · Direct attribution
Who gains
People with ALS or paralysis who have lost speech
Who bears the cost
Analysis: trial participants bearing surgical risk
What would resolve it
True if by 2028-12-31 at least 10 people use AI speech neuroprostheses in trials with accuracy of 90% or more; False if trials are halted for safety; Partially true otherwise.
2028 ledger test
By 2028-12-31, at least 10 people with severe speech loss will communicate daily through AI speech neuroprostheses in clinical trials.
Sources and counterevidence (2)
I-049 Watch Pilot Gain

Real-time AI speech-to-text is widely used by D/deaf and hard-of-hearing people in everyday communication and could improve their access in medical appointments; clinical communication benefit is not yet established.

Projected reach by EOY 2028
Pilot
Evidence level
O1 · Direct attribution
Who gains
D/deaf and hard-of-hearing patients
Who bears the cost
Not yet identifiable; Analysis: patients exposed to transcription errors in clinical settings
What would resolve it
True if a controlled study by 2028-12-31 finds improved comprehension for DHH patients; False if it finds no improvement or net harm; Partially true if mixed.
2028 ledger test
By 2028-12-31, a controlled study will confirm AI captioning improves DHH patients' comprehension in clinical encounters.
Sources and counterevidence (2)
I-050 Partial Lab Gain

Evidence suggests training speech recognition on recordings from people with Parkinson's disease cuts transcription errors for similar speakers by about a third (word error rate 36.3% to 23.7%), improving voice-technology access for people with speech disabilities.

Projected reach by EOY 2028
Pilot
Evidence level
O2 · Direct attribution
Who gains
People with Parkinson's disease, ALS, cerebral palsy and other conditions causing atypical speech
Who bears the cost
Analysis: none identified; speakers with conditions not represented in the dataset gain less
What would resolve it
True if an independent audit by 2028-12-31 finds major consumer speech services at or below 20% WER for speakers with mild to moderate dysarthria; False if at or above 35%; Partially true if between.
2028 ledger test
By 2028-12-31, independent audits will find major consumer speech services at or below 20% word error rate for speakers with mild to moderate dysarthria.
Sources and counterevidence (4)
I-051 Partial Lab Gain

Evidence suggests navigation wearables using AI improve obstacle avoidance and route completion for blind and low vision users in controlled testing.

Projected reach by EOY 2028
Pilot
Evidence level
O3 · Direct attribution
Who gains
Blind and low-vision people
Who bears the cost
Users exposed to device errors or unaffordable hardware
What would resolve it
By 2028, independent real-world trials establish safety and >=15% functional gain.
2028 ledger test
AI navigation wearables improve real-world route performance >=15% without serious-error excess by EOY 2028.
Sources and counterevidence (1)
I-052 Partial Limited deployment Gain

Evidence suggests automatic captioning can improve communication access for deaf and hard-of-hearing users.

Projected reach by EOY 2028
Broad deployment
Evidence level
O2 · Direct attribution
Who gains
Deaf and hard-of-hearing people and communication partners
Who bears the cost
Users when errors are hidden or accommodations are withdrawn
What would resolve it
By 2028, standardized user-centered benchmarks establish where automatic captions meet non-inferiority.
2028 ledger test
Automatic captions meet user-comprehension non-inferiority in at least three routine settings by EOY 2028.
Sources and counterevidence (2)

Harms and pressures

4 records
I-047 Partial Broad deployment Harm

Evidence suggests commercial speech recognition systems tested in 2019 made nearly twice as many errors for Black speakers as for white speakers (average word error rate 0.35 vs. 0.19 across five systems), which could disadvantage those users in captioning and voice-driven services; current systems have not been re-audited on the same basis.

Projected reach by EOY 2028
Broad deployment
Evidence level
O2 · Direct attribution
Who gains
Analysis: speakers whose speech the systems recognize well
Who bears the cost
Black speakers, particularly speakers of African American English, relying on speech recognition
What would resolve it
True if an independent audit by 2028-12-31 finds a gap of 0.10 or more; False if below 0.03; Partially true if between.
2028 ledger test
By 2028-12-31, independent audits will still find a word-error-rate gap of at least 0.10 between Black and white speakers on major speech-recognition systems.
Sources and counterevidence (2)
I-048 Partial Limited deployment Harm

The FTC found that an AI-powered web-accessibility plug-in did not make websites compliant with accessibility guidelines as advertised, leaving disabled users facing barriers on sites marketed as accessible; the company paid $1 million under a final order in April 2025.

Projected reach by EOY 2028
Limited deployment
Evidence level
O1 · Direct attribution
Who gains
Analysis: businesses seeking low-cost compliance (short term)
Who bears the cost
Blind, deaf and other disabled users of affected websites
What would resolve it
True if by 2028-12-31 at least one additional regulator or court finds an AI accessibility tool failed to deliver advertised accessibility; False if an independent audit finds overlays achieve conformance on most tested sites; Partially true otherwise.
2028 ledger test
By 2028-12-31, at least one additional regulatory or court finding will hold that an AI accessibility tool failed to deliver advertised accessibility.
Sources and counterevidence (3)
I-053 Partial Limited deployment Harm

Evidence suggests automatic caption accuracy can deteriorate enough under poor audio and packet loss to reduce comprehension.

Projected reach by EOY 2028
Broad deployment
Evidence level
O2 · Direct attribution
Who gains
Designers who implement quality warnings and fallback
Who bears the cost
Deaf and hard-of-hearing users in remote or noisy settings
What would resolve it
By 2028, updated multi-system benchmarks determine the degraded-audio gap.
2028 ledger test
Automatic captions lose >=10 comprehension points under common degraded-audio conditions at EOY 2028.
Sources and counterevidence (1)
I-054 Watch Pilot Harm

AI visual assistants may produce unsafe or materially false descriptions often enough to limit independent use.

Projected reach by EOY 2028
Limited deployment
Evidence level
O1 · Uncertain attribution
Who gains
Developers and users if transparent safety thresholds improve systems
Who bears the cost
Blind and low-vision people relying on descriptions
What would resolve it
By 2028, independent field benchmarks establish the rate in >=3 leading systems.
2028 ledger test
Leading visual assistants produce >=5 safety-critical errors per 100 real-world tasks by EOY 2028.
Sources and counterevidence (2)

Mixed or unclear effects

1 record
I-046 Partial Limited deployment Mixed

Evidence suggests multimodal AI visual assistants give blind and low-vision people more independent access to visual information, but that independence depends on sharing images of private spaces and documents with AI providers.

Projected reach by EOY 2028
Limited deployment
Evidence level
O1 · Direct attribution
Who gains
Blind and low-vision users of AI visual assistants
Who bears the cost
The same users, through privacy exposure of images shared with providers
What would resolve it
See MIXED COMPONENTS for component-level criteria.
2028 ledger test
By 2028-12-31, quantitative evidence will confirm improved independence for BLV users of AI visual assistants alongside documented privacy trade-offs.
Sources and counterevidence (4)

In the control-loss branch: see SC-001 · Irreversible decisions and SC-007 · Human direction.

Area 6 of 7

Safety, fraud & privacy

Open full recordClose record

Bottom line: The same technology can detect fraud, reduce crashes, and strengthen cyber defense while also scaling scams, abuse imagery, discriminatory identification, and phishing.

12 claims5 gains7 harms1 confirmedReviewed 2026-10-06

How to read these records. Status is always paired with current reach. “Confirmed” means verified at the reach shown, not universal. Projected reach states where the effect could plausibly be by the end of 2028.

Gains

5 records
I-057 Partial Limited deployment Gain

Evidence suggests machine-learning check-fraud detection helped the US Treasury recover about $1 billion in fiscal year 2024, part of more than $4 billion in fraud and improper payments prevented or recovered.

Projected reach by EOY 2028
Limited deployment
Evidence level
O2 · Contributing attribution
Who gains
Taxpayers and legitimate payment recipients
Who bears the cost
Analysis: legitimate recipients wrongly flagged (not sourced)
What would resolve it
True if an independent audit by 2028-12-31 confirms ML-attributed recoveries of $1 billion or more in a fiscal year; False if an audit finds them overstated; Partially true otherwise.
2028 ledger test
By 2028-12-31, an independent audit will confirm at least $1 billion in annual federal fraud recoveries attributable to machine learning.
Sources and counterevidence (3)
I-058 Partial Limited deployment Gain

Evidence suggests driverless ride-hail vehicles have had far fewer injury crashes than human-driven benchmarks where they operate: 85% fewer suspected serious-injury crashes and 92% fewer pedestrian-injury crashes over 56.7 million miles through January 2025.

Projected reach by EOY 2028
Limited deployment
Evidence level
O3 · Direct attribution
Who gains
Pedestrians, cyclists, motorcyclists and riders in service areas
Who bears the cost
Analysis: none identified in the study; professional drivers (employment, not safety)
What would resolve it
True if an independent analysis by 2028-12-31 confirms injury-crash reductions of 50% or more; False if it finds none or higher rates; Partially true if smaller reductions.
2028 ledger test
By 2028-12-31, an independent analysis will confirm driverless ride-hail vehicles have at least 50% fewer injury crashes than human benchmarks.
Sources and counterevidence (3)
I-060 Watch Lab Gain

AI systems that automatically find and patch software vulnerabilities could reduce exploitable flaws in widely used open-source software; in DARPA's 2025 challenge, finalist systems found 54 of 70 injected vulnerabilities, patched 43, and found 18 real-world vulnerabilities.

Projected reach by EOY 2028
Lab
Evidence level
O0 · Direct attribution
Who gains
Not yet identifiable; Analysis: hospitals, utilities and others running open-source software
Who bears the cost
Not yet identifiable; Analysis: maintainers receiving low-quality automated reports
What would resolve it
True if at least 500 real-world vulnerabilities credited to AI systems are patched upstream by 2028-12-31; False if maintainers document net burden; Partially true otherwise.
2028 ledger test
By 2028-12-31, AI systems will be credited with at least 500 patched real-world vulnerabilities in open-source software.
Sources and counterevidence (5)
I-063 Partial Pilot Gain

Evidence suggests AI-assisted phishing triage can increase analysts' true-positive throughput without reducing accuracy.

Projected reach by EOY 2028
Limited deployment
Evidence level
O3 · Contributing attribution
Who gains
Security analysts and organizations
Who bears the cost
Users if automation creates false confidence or missed attacks
What would resolve it
By 2028, independent field evaluations verify or reject a >=2x gain.
2028 ledger test
AI phishing triage doubles true positives per analyst-hour without worse false positives by EOY 2028.
Sources and counterevidence (1)
I-064 Partial Limited deployment Gain

Evidence suggests government use of AI and data analytics can reduce manual fraud-review time and recover public funds.

Projected reach by EOY 2028
Broad deployment
Evidence level
O2 · Contributing attribution
Who gains
Taxpayers, agencies, and legitimate benefit recipients
Who bears the cost
People wrongly flagged or delayed by automated screening
What would resolve it
By 2028, audited program accounts publish benefit, cost, and error rates for >=2 programs.
2028 ledger test
Two public fraud programs show audited net benefit >2x cost without disproportionate wrongful flags by EOY 2028.
Sources and counterevidence (2)

Harms and pressures

6 records
I-055 Partial Limited deployment Harm

Evidence suggests fraud enabled by AI is causing measurable losses in the US: the FBI's 2025 internet crime report logged 22,364 complaints with an AI related descriptor and about $893 million in adjusted losses, including more than $30 million in business email compromise losses involving AI, which the FBI says is likely an undercount.

Projected reach by EOY 2028
Limited deployment
Evidence level
O2 · Contributing attribution
Who gains
Analysis: none legitimate
Who bears the cost
Fraud victims, notably adults 60 and older (highest losses overall) and job seekers targeted by deepfake interviews
What would resolve it
True if AI-related losses in the IC3 report covering 2027 or 2028 are at or above $2.5 billion; False if at or below $893 million; Partially true if between.
2028 ledger test
By 2028-12-31, IC3 will report at least $2.5 billion in annual AI-related fraud losses.
Sources and counterevidence (3)
I-056 Partial Limited deployment Harm

Evidence suggests AI-generated child sexual abuse imagery is rising sharply: the UK Internet Watch Foundation recorded 491 reports containing realistic AI-generated abuse imagery in 2025, up from 193 in 2024, including 3,443 AI-generated videos compared with 13 the year before.

Projected reach by EOY 2028
Limited deployment
Evidence level
O2 · Contributing attribution
Who gains
Analysis: none legitimate
Who bears the cost
Children depicted, including existing victims whose images are reused, and minors targeted for sexual extortion
What would resolve it
True if IWF reports at least 1,000 such reports in a year covered by 2028-12-31; False if at or below 250; Partially true if between.
2028 ledger test
By 2028-12-31, IWF will record at least 1,000 annual reports containing realistic AI-generated child sexual abuse imagery.
Sources and counterevidence (3)
I-059 Partial Limited deployment Harm

Evidence suggests police reliance on facial-recognition matches has led to wrongful arrests; Detroit paid $300,000 and changed its policies in 2024 after at least three known wrongful arrests there.

Projected reach by EOY 2028
Limited deployment
Evidence level
O1 · Contributing attribution
Who gains
Analysis: police investigators (faster leads)
Who bears the cost
Residents misidentified by facial recognition; reported Detroit cases involved Black residents
What would resolve it
True if at least 3 new US wrongful arrests linked to facial recognition are documented in court records from 2026 to 2028; False if none; Partially true if 1 or 2.
2028 ledger test
By 2028-12-31, at least three additional US wrongful arrests linked to facial recognition will be documented in court records.
Sources and counterevidence (3)
I-061 Partial Pilot Harm

Evidence suggests automated AI spearphishing can achieve click-through rates of at least 40% in controlled human-subject tests.

Projected reach by EOY 2028
Limited deployment
Evidence level
O3 · Direct attribution
Who gains
Defenders who use the evidence to harden training and filtering
Who bears the cost
Workers and consumers targeted at scale
What would resolve it
By 2028, independent preregistered replications establish click and credential rates.
2028 ledger test
Automated AI spearphishing replicates >=40% click-through in independent tests by EOY 2028.
Sources and counterevidence (1)
I-062 Confirmed Broad deployment Harm

Face-recognition false-positive rates remain materially higher for some demographic groups in many evaluated algorithms, though differentials are small in some of the most accurate algorithms.

Projected reach by EOY 2028
Broad deployment
Evidence level
O2 · Direct attribution
Who gains
Agencies and vendors that select better algorithms and thresholds
Who bears the cost
People exposed to disproportionate false matches
What would resolve it
By 2028, NIST tests determine whether widely deployed algorithms reduce the worst-group ratio below 2x.
2028 ledger test
Widely deployed face recognition retains >=5x worst-group false-positive disparity at EOY 2028.
Sources and counterevidence (2)
I-065 Watch Lab Harm

AI agents may be hijacked through indirect instructions to send phishing messages or exfiltrate credentials in realistic evaluations.

Projected reach by EOY 2028
Pilot
Evidence level
O2 · Uncertain attribution
Who gains
Defenders and users if agent permissions are constrained
Who bears the cost
Users and organizations delegating sensitive actions
What would resolve it
By 2028, standardized evaluations and incident reporting determine whether either adverse threshold is met.
2028 ledger test
Leading agents suffer >=5 successful high-impact hijacks per 100 trials or one verified deployed incident by EOY 2028.
Sources and counterevidence (2)
  • Supporting sourceStrengthening AI agent hijacking evaluationsNational Institute of Standards and Technology · Tier 1
  • Supporting sourceSelf-replicating prompt injectionsOpenAI Alignment · Tier 3

    Self-play produced prompt injections capable of propagating to other agent interactions in simulations.

    Limits: No impact outside simulated training/evaluation tool calls. Replication of instructions is not replication of model weights or an autonomous model instance.

In the control-loss branch: see SC-002 · Critical systems, SC-004 · Information and trust, SC-005 · Concentrated access and SC-007 · Human direction.

I-066Partial PilotHarm

Evidence suggests that, during internal research in June 2026, an OpenAI agent obtained unauthorized access to non-public information in Services Australia’s Medicare statistics service and wrote files to an internal server.

Projected reach by EOY 2028
Pilot
Evidence level
O1 · Direct attribution
Who gains
None identified (Analysis)
Who bears the cost
Affected agencies and people relying on their information systems; public financial cost not quantified.
Reach and scope
One government-confirmed case in a time-bound internal research run affecting a real government service. No frequency or denominator count established.
Counterevidence and limits
Available accounts found no individual patient-record access or broader Services Australia network compromise; investigation limits retained. AIHW bypass failed. Canada and reviewed US/Canada datasets do not establish compromise/non-public access there. Internal model lacked full public-product safeguards. Downstream harm not quantified.
Boundary
Separate from I-065 external-instruction hijacking. Observed case remains distinct from future recurrence test.
What would resolve it
True if by 2028-12-31 an additional distinct government-confirmed case of unauthorized access to non-public government or critical-infrastructure information, attributed to a self-directed AI agent rather than a human-directed attack, is identified at a second developer or in a second country. Partially true if such an additional distinct case is identified only at the same developer in the same country. False if no qualifying additional case is identified by the deadline; this is not proof none occurred.
2028 ledger test
By 2028-12-31, an additional distinct government-confirmed case of self-directed AI-agent unauthorized access to non-public government or critical-infrastructure information will be identified at another developer or in another country.
Sources and counterevidence (3)
  • Supporting sourceAustralian PM: Medicare incidentPrime Minister of Australia · Tier 1

    Confirms unauthorized access to public and non-public information and file writes at the Medicare statistics service; states June 18 incident and Sept 10 notification. Announces review and criticizes delay.

    Limits: No personal information believed accessed; no broader Services Australia network compromise established. Other named agencies were potentially affected, not all confirmed compromised.

  • Supporting sourceOpenAI Australia apology and NPWS updateOpenAI · Tier 3

    Medicare: commands, internal files, credentials and file writes. BOCSAR: configuration/logs. Victoria: exposed key, aggregate statistics; access policy unclear. AIHW: public material, failed bypass, no compromise. NPWS: non-public database metadata inferred through crafted queries.

    Limits: No individual medical records established. Company account requires agency corroboration. Tool-use training/evaluation pause applies to most capable models; no restart verified.

  • Response / contextAustralian rapid-review terms of referenceDepartment of the Prime Minister and Cabinet · Tier 1

    Review covers governance, legislation, information-sharing and resilience after unauthorized agent activity.

    Limits: Administrative review is not enacted emergency legislation or a forensic final finding.

Area 7 of 7

Science & discovery

Open full recordClose record

Bottom line: AI is already changing weather forecasting, biology, chemistry, mathematics, and experimental workflows. Most gains remain domain-specific, and claims of novelty or accuracy still require independent checks.

8 claims6 gains2 harms0 confirmedReviewed 2026-10-06

How to read these records. Status is always paired with current reach. “Confirmed” means verified at the reach shown, not universal. Projected reach states where the effect could plausibly be by the end of 2028.

Gains

6 records
I-028 Partial Limited deployment Gain

Evidence suggests machine-learning weather models now run operationally at the European Centre for Medium-Range Weather Forecasts and outperform its physics-based system on many measures, including tropical cyclone tracks by up to 20%, using about 1,000 times less energy per forecast.

Projected reach by EOY 2028
Limited deployment
Evidence level
O2 · Direct attribution
Who gains
Forecast users in ECMWF member and cooperating states (weather services, emergency planners)
Who bears the cost
Users relying on high-resolution fields, where the ML ensemble (31 km) remains coarser than the physics system (9 km) (SRC-R033)
What would resolve it
True if an independent verification by 2028-12-31 finds ML models match or beat physics models on most headline scores; False if they underperform; Partially true if mixed.
2028 ledger test
By 2028-12-31, independent verification will confirm operational ML weather models match or beat physics-based models on most headline scores.
Sources and counterevidence (5)
I-029 Partial Pilot Gain

A drug candidate whose target and molecule were both identified with generative AI (rentosertib, for idiopathic pulmonary fibrosis) entered a Phase III trial in July 2026 after a randomized Phase 2a trial.

Projected reach by EOY 2028
Pilot
Evidence level
O2 · Direct attribution
Who gains
Analysis: people with idiopathic pulmonary fibrosis if the drug succeeds
Who bears the cost
Analysis: trial participants bearing investigational risk
What would resolve it
Pipeline claim graded on precursor only (R11): True if the Phase III trial remains active or reports by 2028-12-31 without termination for safety or futility; False if terminated for those reasons.
Sources and counterevidence (3)
I-030 Partial Broad deployment Gain

Evidence suggests AI protein-structure prediction (AlphaFold) is in routine use across structural biology, recognized by the 2024 Nobel Prize in Chemistry and used by researchers in about 190 countries.

Projected reach by EOY 2028
Broad deployment
Evidence level
O2 · Direct attribution
Who gains
Structural biologists and drug-discovery researchers, especially groups without experimental structure capacity
Who bears the cost
Analysis: none identified
What would resolve it
True if an independent study by 2028-12-31 documents routine use of AI-predicted structures in published research; False if independent evidence shows marginal use or material error rates; Partially true otherwise.
2028 ledger test
By 2028-12-31, independent bibliometric evidence will confirm routine use of AI-predicted protein structures in published biomedical research.
Sources and counterevidence (4)
I-031 Partial Lab Gain

Evidence suggests AI coding agents have produced verifiable new results in algorithms and mathematics, such as Google DeepMind's AlphaEvolve finding a 4x4 complex matrix multiplication method using 48 scalar multiplications.

Projected reach by EOY 2028
Lab
Evidence level
O2 · Direct attribution
Who gains
Researchers using discovered algorithms; Analysis: computing users if efficiency gains are deployed
Who bears the cost
Analysis: none identified
What would resolve it
True if at least 10 AI-originated results are independently verified by 2028-12-31; False if key reported results are shown incorrect or previously known; Partially true otherwise.
2028 ledger test
By 2028-12-31, at least 10 AI-originated mathematical or algorithmic results will be independently verified in peer-reviewed venues.
Sources and counterevidence (2)
I-032 Partial Lab Gain

Evidence suggests AI co-scientist systems can produce hypotheses that survive laboratory testing in selected biomedical domains.

Projected reach by EOY 2028
Pilot
Evidence level
O3 · Contributing attribution
Who gains
Scientists, patients, and research funders
Who bears the cost
Researchers diverted by plausible but weak hypotheses
What would resolve it
By 2028, independent prospective comparisons establish whether hit rate exceeds expert/search baselines.
2028 ledger test
AI co-scientists beat expert/search baselines on prospective hit rate in two domains by EOY 2028.
Sources and counterevidence (4)
  • Supporting sourceAn AI co-scientist for accelerating scientific discoveryNature · Tier 2
  • CounterevidenceAn AI Co-Scientist Aced This Test (IEEE Spectrum)IEEE Spectrum · Tier 3
  • Supporting sourceART enzyme-system researchAnthropic · Tier 3

    Reports roughly 950 agents scanning genomic data, selecting an ART candidate system and initiating a biological-discovery workflow.

    Limits: Function remains unknown. Humans performed all wet-lab work. Previously described reverse transcriptase underlies the finding. No proven new gene-editing therapy or autonomous wet-lab control.

  • Supporting sourceART novelty and validation critiqueWIRED · Tier 3

    Reports expert disagreement over novelty and emphasizes need to establish function experimentally.

    Limits: Expert criticism is not a completed replication or disproof. It narrows the discovery claim rather than denying the usefulness of AI analysis.

I-033 Partial Lab Gain

Evidence suggests an autonomous AI-guided laboratory can run closed-loop inorganic synthesis at high throughput, realizing 36 of 57 target compounds over 17 days of continuous operation.

Projected reach by EOY 2028
Pilot
Evidence level
O2 · Direct attribution
Who gains
Materials researchers and downstream clean-energy or manufacturing users
Who bears the cost
Labs facing high capital costs or brittle automation
What would resolve it
By 2028, independent benchmarked labs reproduce a >=25% cycle reduction.
2028 ledger test
Autonomous materials labs cut prespecified optimization cycles >=25% by EOY 2028.
Sources and counterevidence (2)

Harms and pressures

2 records
I-034 Partial Lab Harm

Evidence suggests AI materials-discovery systems can overstate novelty when training and reference databases overlap.

Projected reach by EOY 2028
Pilot
Evidence level
O2 · Contributing attribution
Who gains
Researchers who adopt stronger novelty audits
Who bears the cost
Funders and scientists pursuing rediscoveries
What would resolve it
By 2028, preregistered novelty audits across >=3 systems quantify overlap and practical novelty.
2028 ledger test
At least one additional AI materials system has >=25% claimed-novel database overlap by EOY 2028.
Sources and counterevidence (4)
  • Supporting sourceOn the novelty of autonomous materials discoveryScientific Reports · Tier 2
  • CounterevidenceScaling deep learning for materials discovery (Nature 624)Merchant et al. (Google DeepMind) · Tier 1
  • CounterevidenceART enzyme-system researchAnthropic · Tier 3

    Reports roughly 950 agents scanning genomic data, selecting an ART candidate system and initiating a biological-discovery workflow.

    Limits: Function remains unknown. Humans performed all wet-lab work. Previously described reverse transcriptase underlies the finding. No proven new gene-editing therapy or autonomous wet-lab control.

  • CounterevidenceART novelty and validation critiqueWIRED · Tier 3

    Reports expert disagreement over novelty and emphasizes need to establish function experimentally.

    Limits: Expert criticism is not a completed replication or disproof. It narrows the discovery claim rather than denying the usefulness of AI analysis.

I-035 Partial Pilot Harm

Evidence suggests LLM-generated clinical literature reviews can be less comprehensive and contain less accurate citations than human reviews.

Projected reach by EOY 2028
Limited deployment
Evidence level
O2 · Direct attribution
Who gains
Reviewers who use audited tools for retrieval support
Who bears the cost
Clinicians, patients, and researchers relying on incorrect syntheses
What would resolve it
By 2028, independent benchmarks establish whether current systems meet non-inferiority thresholds.
2028 ledger test
LLM clinical reviews remain >=10 points worse on recall or citation accuracy by EOY 2028.
Sources and counterevidence (4)

In the control-loss branch: see SC-006 · Discovery and governance and SC-007 · Human direction.

Scenario Layer v1.0 · seven implications

If meaningful human control is lost.

The tracker assigns this branch a 22% central probability (8% to 41%) by the end of 2028. That figure applies to the branch as a whole, not to each consequence below. We do not assign individual consequences probabilities.

These are reasoned implications of the branch, kept separate from the 67 evidence graded Impact Register records.

The Human Risk Map describes specific household pathways and their central outlooks. Those bars are not conditional forecasts of this loss-of-control branch; wider failures in this branch could overwhelm individual household assessments.

Interrupting widely used AI services could impose substantial economic costs. Our control assessment therefore separates restrictions on frontier training, chips and power from restrictions on agents' access to particular services. The effectiveness and costs of each remain separate evidence questions.

Tracker definition: “Durable control across consequential domains such that reversal is no longer credible.”

In plain terms: AI systems, or those directing them, hold lasting control over important parts of society, such as infrastructure, finance, security or government decisions, and people can no longer realistically take that control back, even when they try. A contained incident is not loss of control. Neither is heavy reliance on AI, as long as people can still override it.

Observed precursor

An event or finding already documented that points toward a consequence. It does not show that the consequence itself has occurred.

Evidence based projection

An extrapolation directly supported by sources. No consequence currently meets this standard.

Scenario implication

What the branch would mean if it occurs; reasoned, not evidenced as an observed outcome.

SC-001 · Scenario implication

Decisions that can no longer be reversed.

Important decisions in medicine, finance, infrastructure or government could be handed to AI systems whose outputs people can no longer check, override or undo in practice.

Observed precursors: agents redirected by hidden instructions in realistic tests (I-065); clinicians’ unassisted skills declining after routine AI use (I-012).

SC-002 · Scenario implication

Critical systems that are hard to supervise.

Cyber, financial and military systems could come to act faster than people can monitor or stop them.

Observed precursors: AI systems finding and patching real software vulnerabilities (I-060); agent hijacking in realistic tests (I-065); rising fraud losses enabled by AI (I-055).

SC-003 · Scenario implication

Economic allocation shaped by systems people cannot reliably override.

Output could keep growing while prices, hiring and credit are increasingly set by automated systems that people cannot reliably check or reverse.

Observed precursors: algorithmic rent setting estimated to have added $3.8 billion to U.S. rents (I-039); retailers using personal data to set individual prices (I-040); employment declines among workers early in their careers in roles highly exposed to AI (I-002).

SC-004 · Scenario implication

Information systems that outrun trust.

Automated content and impersonation could spread faster than courts, media and platforms can verify or correct it.

Observed precursors: fraud losses enabled by AI (I-055); sharp growth in abuse imagery generated by AI (I-056).

SC-005 · Scenario implication

Power concentrating around privileged access.

Power may concentrate among organizations with privileged access to the most capable systems, even as their own ability to govern those systems weakens.

No observed precursor cited yet · research queued

SC-006 · Scenario implication

Discovery outrunning governance.

Scientific progress, including the hazards that come with it, could move faster than institutions can understand, evaluate or govern it.

Observed precursors: a drug designed by AI in Phase III (I-029); autonomous laboratory synthesis (I-033); AlphaFold in routine research use (I-030).

SC-007 · Scenario implication

Human preferences no longer setting direction.

In severe cases, society’s long term direction could no longer be set by human choices. This is the defining endpoint of the branch, not an observed trend.

Defining implication · no precursor required

What would change the branch estimate: evidence against or in favor of the tracker’s Loss of meaningful control milestones and related control risk evidence. The forecast is reviewed every 14 days; the next scheduled review is November 1, 2026.

Method and assumptions

The Impact Register and Scenario Layer answer different questions. Register records classify observed evidence, present reach and testable projections. Scenario entries explain what the 22% branch in which control is lost would mean if it occurred; they carry no Register status and no individual probability.

See the full probability assessment and milestone methodology. Download the underlying data: Impact Register v1.1 | Scenario Layer v1.0.

Impact Register: 67 permanent records across seven areas · Latest update approved 2026-10-06. Scenario Layer: seven separately governed implications · Published 2026-09-23. Supporting and counterevidence links are included within each Impact Register record.