Back to the trend of the day: America's air traffic now gets advice from AI to cut delays
Poly Rating · from the trend of the day

Human in the Loop AI Examples: Top 10 Ranked (2026)

The FAA's SMART now advises all 22 US en-route centers, yet its only published payoff is 180 rerouted flights, while in a Copenhagen trial an AI that spotted more cardiac arrests than dispatchers did not make the dispatchers any better.

  • No. 1: FAA SMART air traffic AI (Air Space Intelligence)
  • 10 places

01Final list

The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.

Order: from most to least AI agreement.

A bright badge means this AI put the entry on its list. A dim badge means it did not.

  1. 1

    FAA SMART air traffic AI (Air Space Intelligence)

    Estimate: Widest rollout, thinnest proof: it is live in all 22 US en-route centers, yet the only payoff is FAA's own 180 rerouted flights at ~10 minutes and ~80 gallons each.

    Included it: Claude, ChatGPT, Meta, Grok, MiMo, StepFun, Kimi, DeepSeek. Did not include: Gemini, Z.ai, Wenxin.

    In their lists: Claude: 9 · ChatGPT: 4 · Meta: 6 · Grok: 1 · MiMo: 1 · StepFun: 1 · Kimi: 4 · DeepSeek: 1

    What each AI noted 8
    • Claudeplace 9

      The largest rollout here (22 centers) on the thinnest proof: 180 rerouted flights, under an $875,950,000 contract

      Source ↗
    • ChatGPTplace 4

      The first 180 reported reroutes averaged nearly 80 gallons of fuel saved per flight; that is an early FAA claim, not an independent evaluation.

      Source ↗
    • Metaplace 6

      Now live at all 22 centres, but its only published payoff is 180 rerouted flights in a Washington pilot — 10 minutes and 80 gallons each.

      Source ↗
    • Grokplace 1

      180 flights rerouted in first weeks at ~10 min and ~80 gallons fuel saved each, now live across all 22 centers under human manager approval

    • MiMoplace 1

      One accepted reroute is worth 10 minutes and 80 gallons — and all 22 control centers took the advice eight weeks before Congress finished asking its first safety questions.

      Source ↗
    • StepFunplace 1

      22 centers, about 200 data streams, and every suggestion still gets a human signature: 180 flights rerouted for 10+ minutes and nearly 80 gallons of fuel each.

      Source ↗
    • Kimiplace 4

      Three days after going live at all 22 US en-route centers it watched the FAA ground the whole country — officially unrelated, but the $875M system's entire payoff evidence is still 180 rerouted flights.

      Source ↗
    • DeepSeekplace 1

      180 human-approved reroutes saved about 14,400 gallons of fuel — roughly $62,000 at $4.31/gal, against an $875 million contract.

      Source ↗
    8 AIs put it here
  2. 2

    TREWS sepsis early warning (Johns Hopkins / Bayesian Health)

    Peer-reviewed across 5 hospitals: alerts confirmed within 3 hours tied to 18.7% lower sepsis mortality (3.3 points absolute). Observational, and RCT reviews of sepsis alerts found no clear benefit.

    Included it: Claude, ChatGPT, Meta, MiMo, StepFun, Kimi, DeepSeek. Did not include: Grok, Gemini, Z.ai, Wenxin.

    In their lists: Claude: 3 · ChatGPT: 2 · Meta: 4 · MiMo: 7 · StepFun: 2 · Kimi: 2 · DeepSeek: 8

    What each AI noted 7
    • Claudeplace 3

      Clinicians evaluated 89% of its alerts, an unusually high follow-through, and quick confirmation was tied to an 18.7% relative drop in sepsis deaths

      Source ↗
    • ChatGPTplace 2

      Among alerted sepsis patients, timely clinician confirmation was associated with 1.85 hours faster antibiotic orders—and 38% of evaluated alerts were confirmed.

      Source ↗
    • Metaplace 4

      When clinicians confirmed its alert within 3 hours, sepsis patients were 18.7% less likely to die — the first FDA-cleared continuous AI sepsis monitor.

      Source ↗
    • MiMoplace 7

      Ten times fewer false alarms in the pilot — and an outside lab checking a rival sepsis model found the vendor's published numbers don't travel to real wards.

      Source ↗
    • StepFunplace 2

      The rare advice system whose payoff survived peer review: in-hospital sepsis mortality fell 18.7% relative after clinicians started getting alerts.

      Source ↗
    • Kimiplace 2

      The only system here with peer-reviewed proof that a human confirming an AI alert within 3 hours cut sepsis deaths 18.7% — and it also publishes the human side: 89% of alerts actually reviewed.

      Source ↗
    • DeepSeekplace 8

      89% of alerts were evaluated, but only 38% were confirmed — a rare published look at how often humans actually agree.

      Source ↗
    7 AIs put it here
  3. 3

    Deutsche Bahn AI train dispatching (DB InfraGO ADA-PMB)

    Dispatchers took 7 of 10 suggestions in Frankfurt, and DB credits 58,000 saved delay minutes in 2022. Its 100,000-minute 2023 figure mixes several tools, and the system runs in four regions only.

    Included it: Claude, ChatGPT, Meta, Grok, MiMo, StepFun, DeepSeek. Did not include: Gemini, Z.ai, Kimi, Wenxin.

    In their lists: Claude: 7 · ChatGPT: 5 · Meta: 7 · Grok: 7 · MiMo: 2 · StepFun: 4 · DeepSeek: 2

    What each AI noted 7
    • Claudeplace 7

      58,000 delay minutes removed in 2022, and dispatchers followed the AI's suggestion in 7 of 10 analysed cases

      Source ↗
    • ChatGPTplace 5

      Its reported Frankfurt gain is about 100 seconds per dispatching decision, but it is active in four regions—not across Germany's full network.

      Source ↗
    • Metaplace 7

      In Frankfurt, following its suggestion is worth 100 seconds per dispatch — but it still only runs in selected areas, with screen integration due December 2026.

      Source ↗
    • Grokplace 7

      Real-time recommendations every 60 seconds to dispatchers who accept or reject, active across multiple dense-traffic regions and expanding

    • MiMoplace 2

      A dispatcher gets a fresh conflict plan every 60 seconds, Stuttgart claws back up to 8 minutes of delay — and 'in theory' 17 more trains a day fit on one core route.

      Source ↗
    • StepFunplace 4

      Three AI networks kept 58,000 delay minutes off Germany's rails in a year - about 40 days of delay that never happened.

      Source ↗
    • DeepSeekplace 2

      Over 100,000 delay minutes were saved in 2023 across four dense regions — while a dispatcher still accepts or rejects every suggestion.

      Source ↗
    7 AIs put it here
  4. 4

    Corti cardiac-arrest call AI, Copenhagen EMS

    The AI spotted more cardiac arrests than dispatchers (85.0% vs 77.5%), yet in a 169,049-call randomized trial its alerts did not significantly improve the dispatchers.

    Included it: Claude, Meta, MiMo, StepFun, Kimi. Did not include: ChatGPT, Grok, Gemini, Z.ai, DeepSeek, Wenxin.

    In their lists: Claude: 10 · Meta: 3 · MiMo: 5 · StepFun: 3 · Kimi: 7

    What each AI noted 5
    • Claudeplace 10

      The AI spotted more cardiac arrests than dispatchers, yet in a randomized trial its alerts did not improve what dispatchers recognized

      Source ↗
    • Metaplace 3

      It spots cardiac arrest in 93-95% of calls vs 73% for dispatchers alone, and about 10 seconds faster — seconds that decide CPR.

    • MiMoplace 5

      The AI spots cardiac arrest 85% of the time versus a dispatcher's 77.5% — yet handing dispatchers the AI barely moved their score (93.1% vs 90.5%, p=0.15).

      Source ↗
    • StepFunplace 3

      It listens to the 911 call and spots cardiac arrest with 84.1% sensitivity against dispatchers' 72.5% - across 169,236 real calls.

      Source ↗
    • Kimiplace 7

      Peer-reviewed on 108,607 real calls: the AI caught 11.6 points more cardiac arrests than trained dispatchers and flagged them 10 seconds sooner — every second matters when survival drops ~10% per minute.

      Source ↗
    5 AIs put it here
  5. 5

    Viz.ai stroke alerts (Viz LVO)

    Faster stroke teams in three study designs: notification 39.5 minutes earlier across 166 facilities and 11.2 minutes faster in a randomized trial. A 12-study meta-analysis found no outcome gain.

    Included it: Claude, ChatGPT, Kimi, DeepSeek. Did not include: Meta, Grok, Gemini, MiMo, StepFun, Z.ai, Wenxin.

    In their lists: Claude: 8 · ChatGPT: 3 · Kimi: 3 · DeepSeek: 5

    What each AI noted 4
    • Claudeplace 8

      About 2,000 hospitals; the vendor-cited multicenter study says 31 minutes saved, the randomized trial says 11

      Source ↗
    • ChatGPTplace 3

      Across 166 facilities, AI-assisted coordination was associated with interventionalist notification 39.5 minutes earlier—not treatment 39.5 minutes earlier.

      Source ↗
    • Kimiplace 3

      A 14,116-patient, 166-hospital real-world study found AI-flagged stroke patients reached a neurointerventionalist 39.5 minutes faster — at network scale, not a pilot ward.

      Source ↗
    • DeepSeekplace 5

      DIDO time fell from 202 to 113 minutes — an 89-minute cut that beat the 120-minute benchmark, though the meta-analysis found no significant clinical-outcome gain.

      Source ↗
    4 AIs put it here
  6. 6

    NATS Intelligent Approach / Time-Based Separation, Heathrow

    Heathrow arrival delay fell from 5.55 to 4.46 minutes under Pairwise spacing, and headwind delay is down over 60%. All figures are NATS's own, and whether the tool counts as AI is disputed.

    Included it: Claude, Meta, Kimi. Did not include: ChatGPT, Grok, Gemini, MiMo, StepFun, Z.ai, DeepSeek, Wenxin.

    In their lists: Claude: 2 · Meta: 1 · Kimi: 1

    What each AI noted 3
    • Claudeplace 2

      Headwind delay down more than 60% at the world's busiest two-runway airport; controllers still set every gap

      Source ↗
    • Metaplace 1

      One spacing choice now cuts arrival delay from 5.55 to 4.46 minutes, and Heathrow has banked 1.5 million holding minutes in a decade.

      Source ↗
    • Kimiplace 1

      Ten years live: controllers eyeball AI-calculated gaps and issue the clearances — and the tool cut headwind delays >60% while saving 230,000 airborne-holding minutes a year at one airport.

      Source ↗
    3 AIs put it here
  7. 7

    NESO AI solar forecast (Open Climate Fix Quartz), Great Britain grid

    The grid control room's AI solar forecast is 2.8 times more accurate and worth about £30M a year. That is the developer's figure, up from a £2–10M projection in 2021, and unaudited.

    Included it: Claude, MiMo. Did not include: ChatGPT, Meta, Grok, Gemini, StepFun, Z.ai, Kimi, DeepSeek, Wenxin.

    In their lists: Claude: 4 · MiMo: 3

    What each AI noted 2
    • Claudeplace 4

      2.8 times more accurate, worth about £30m a year; engineers still decide how much reserve to buy

      Source ↗
    • MiMoplace 3

      Halving one forecast error turned into £30M a year of reserve savings — the grid's biggest quiet AI win, and no one asked for a pause.

      Source ↗
    2 AIs put it here
  8. 8

    Kaiser Permanente Advance Alert Monitor

    Nurses screen about one alert per 35 patients daily. Post-alert 30-day mortality was 15.8% vs 20.4% across 21 hospitals; the 520 lives a year is a modelled estimate.

    Included it: Claude. Did not include: ChatGPT, Meta, Grok, Gemini, MiMo, StepFun, Z.ai, Kimi, DeepSeek, Wenxin.

    In their lists: Claude: 1

    What each AI noted 1
    • Claudeplace 1

      About one alert a day for every 35 patients, screened by remote nurses, linked to an estimated 520 fewer deaths a year across 21 hospitals

      Source ↗
    1 AIs put it here
  9. 9

    MASAI AI-supported mammography screening, Sweden

    The strongest study design here is a randomized trial of 100,000+ women: 12% fewer interval cancers and 44% less reading work. It ran at four Swedish sites, not a national rollout.

    Included it: ChatGPT. Did not include: Claude, Meta, Grok, Gemini, MiMo, StepFun, Z.ai, Kimi, DeepSeek, Wenxin.

    In their lists: ChatGPT: 1

    What each AI noted 1
    • ChatGPTplace 1

      In a randomized trial of more than 100,000 women, interval cancers were 12% lower, with a similar false-positive rate.

      Source ↗
    1 AIs put it here
  10. 10

    NASA ATD-2 departure metering, Charlotte

    NASA counts 1,088,888 gallons saved at Charlotte by holding 3.8% of departures at the gate. It covers one airport, and Cranfield questions how the fuel was calculated.

    Included it: Claude. Did not include: ChatGPT, Meta, Grok, Gemini, MiMo, StepFun, Z.ai, Kimi, DeepSeek, Wenxin.

    In their lists: Claude: 5

    What each AI noted 1
    • Claudeplace 5

      Over 1 million gallons saved by holding planes at the gate rather than in the taxi line, and an outside reviewer questions the math

      Source ↗
    1 AIs put it here

More about the top three ↓

02More about the top three

The ranking is above. Here is more about the three places at the top: what they are, why they are here, and what is worth knowing.

The three payoffs are not comparable: rerouted flights, sepsis deaths and delay minutes are different measures, and the SMART and DB figures come from the operators themselves.

1FAA SMART air traffic AI (Air Space Intelligence)

No. 1 — FAA SMART air traffic AI (Air Space Intelligence). On the lists of 8 of 11 AIs; median place there: 2.5. The AIs ranked it first for the widest rollout, all 22 centers, though its only proof is the FAA's own 180 rerouted flights.

Did you know

  • In June 2026 the FAA picked Air Space Intelligence for a 12-year, $875 million contract. It covers two systems. SMART is an add-on inside FMDS, which replaces the old national Traffic Flow Management System.

What it is
AI that suggests flight reroutes to air traffic controllers across the US.
Its edge
Widest reach of the three: all 22 en-route centers within about seven weeks.
Who should know
Useful for anyone following how fast AI is entering air safety, and how thin its proof is.
  • Proven payoff180 rerouted flights

Wikidata

Why it is here
  • SMART went live around Washington on September 21 and, per the FAA's November 10 announcement, now covers all 22 en-route centers. That is the widest rollout of the three, reached in about seven weeks. — AVweb

  • The FAA says SMART helped reroute 180 flights, saving about 10 minutes and nearly 80 gallons of fuel per flight. These are the agency's own figures, and no outside reviewer has confirmed them. — AVweb

Did you know
  • The same company first worked with Alaska Airlines. Alaska says its Flyways platform saves roughly a million gallons of fuel a year. The FAA then chose the company for the national airspace overhaul.

Keep in mind
  • The controllers' union says it "has not been involved in the design, testing, or implementation of SMART". It therefore cannot speak to the effect on the controller workforce.

  • Some House Democrats have asked the Government Accountability Office to review how SMART is rolled out and how its results are measured. The FAA says its own staff check each suggestion and local controllers decide whether to accept it.

Where to go further
  • Open the article and look at the three figures: 180 flights, about 10 minutes and nearly 80 gallons saved per flight. Then read the note that the controllers' union was not involved in the design.

Compare the top three ↓

2TREWS sepsis early warning (Johns Hopkins / Bayesian Health)

No. 2 — TREWS sepsis early warning (Johns Hopkins / Bayesian Health). On the lists of 7 of 11 AIs; median place there: 3. The AIs valued its peer-reviewed evidence: alerts confirmed by clinicians within three hours were tied to fewer sepsis deaths.

Did you know

  • In the 2022 Nature Medicine studies, patients whose TREWS alert was confirmed within three hours got antibiotics a median of 1.85 hours earlier than comparable patients. Coverage notes the lead scientist is also CEO of the company behind TREWS.

What it is
Hospital AI that warns doctors early when a patient may have sepsis.
Its edge
Only one with peer-reviewed results showing fewer deaths, not just time saved.
Who should know
Useful for anyone wondering if medical AI alerts really save lives, and how strong that proof is.
  • Proven payoff18.7% fewer sepsis deaths
Keep in mind
  • Sepsis alerts have a known problem with alert fatigue. In one other emergency-room alert, the authors blamed a low positive predictive value (few alerts were right) for results showing no benefit. This is not a TREWS measurement.

Where to go further
  • Read the article on the study. Look at the 18.7 percent relative drop and the 3.3 point absolute drop in deaths, then the note on the authors' financial ties to the company.

Compare the top three ↓

3Deutsche Bahn AI train dispatching (DB InfraGO ADA-PMB)

No. 3 — Deutsche Bahn AI train dispatching (DB InfraGO ADA-PMB). On the lists of 7 of 11 AIs; median place there: 5. The AIs credited it for a measurable delay saving and dispatchers accepting most suggestions, but it runs in only a few regions.

Did you know

  • DB's own slides say the system (ADA-PMB) first ran in 2022 on one short stretch, Frankfurt to Hanau. In 2023 it was extended to the approach lines for the Riedbahn. Dispatchers followed its suggestion in 70% of cases analysed.

What it is
AI that suggests how dispatchers should handle trains when delays and conflicts arise.
Its edge
Clearest data on how often the human agrees: 7 of 10 suggestions.
Who should know
Useful for anyone curious how often people actually follow an AI's advice at work.
  • Proven payoff58,000 delay minutes (2022)
Why it is here
  • In a DB presentation, dispatchers handled trains as ADA-PMB suggested in 70% of the cases analyzed. That is the clearest published measure of how often the person says yes. — DB presentation

  • ADA-PMB started in Frankfurt–Hanau in November 2022. It stays advisory, and expansion to other regions is planned for 2027, so today it runs in only a few regions. — Trackopedia

History
  • In November 2022 a pilot started at the Frankfurt control centre. DB presented it as Germany's first dispatch assistant for mixed traffic, where fast and slow trains share the same tracks.

Compare the top three ↓

The top three side by side

FAA SMART air traffic AI (Air Space Intelligence)TREWS sepsis early warning (Johns Hopkins / Bayesian Health)Deutsche Bahn AI train dispatching (DB InfraGO ADA-PMB)
In the AIs' lists8 / 117 / 117 / 11
Proven payoff180 rerouted flights18.7% fewer sepsis deaths58,000 delay minutes (2022)
Put first byGrok, MiMo, StepFun, DeepSeek——

Places with the same number of AIs are ordered by their average place in the AIs' own lists.

03Worth knowing

What the AIs noticed while building the list.

Evaluation is not agreement: TREWS clinicians evaluated 89% of alerts but confirmed only 38% of them. This is one of the few published measures of human follow-through.

Noticed by:
Included it: Claude, ChatGPT, Kimi, DeepSeek. Did not include: Meta, Grok, Gemini, MiMo, StepFun, Z.ai, Wenxin.
Source ↗

The weak link is often the hand-off: in Copenhagen the AI recognized more cardiac arrests than dispatchers, but alerting the dispatchers did not raise their recognition.

Noticed by:
Included it: Claude, MiMo. Did not include: ChatGPT, Meta, Grok, Gemini, StepFun, Z.ai, Kimi, DeepSeek, Wenxin.
Source ↗

The rubber-stamp point has a number: when the AI was wrong, inexperienced mammography readers' accuracy fell from about 80% to 20%, and very experienced readers fell to 45.5%.

Noticed by:
Included it: Claude. Did not include: ChatGPT, Meta, Grok, Gemini, MiMo, StepFun, Z.ai, Kimi, DeepSeek, Wenxin.
Source ↗

The incident record of this class: the widely deployed Epic Sepsis Model alerted on 18% of inpatients and still missed 67% of sepsis cases in outside validation (AIID incident 123).

Noticed by:
Included it: Claude, StepFun, Kimi. Did not include: ChatGPT, Meta, Grok, Gemini, MiMo, Z.ai, DeepSeek, Wenxin.
Source ↗

SMART cash math (estimate): 180 reroutes × ~80 gallons ≈ 14,400 gallons, roughly $62,000 of fuel. That compares with about $23.3M already obligated on an $875.95M ceiling.

Noticed by:
Included it: MiMo, StepFun, Kimi, DeepSeek. Did not include: Claude, ChatGPT, Meta, Grok, Gemini, Z.ai, Wenxin.
Source ↗

04Questions people ask

Why is air traffic control not automated?

Separation stays a licensed controller's legal job. The FAA says SMART, now in all 22 en-route centers, cannot replace controllers or take over an aircraft.

Source ↗
Are air traffic controllers smart?

The filter is steep: of 12,859 FAA applicants in FY2024 only 2,184 (17%) were hire-eligible, and Academy failure rose from 7% in 2010 to 26% in 2024. Other sources cite over 30% because they count different trainee groups.

Source ↗
Why isn't air traffic control automated?

Even the newest AI is advisory: SMART suggests reroutes and local managers accept or decline each one. The controllers' union, with 15,000+ members, says it was not consulted on the design.

Source ↗
Can AI replace air traffic controllers?

Not in any deployed system. The FAA's SMART has suggested about 180 reroutes so far, and managers accept or decline each one.

Source ↗

05How we counted

In plain words, no formulas.

1The question

Top 10 'AI Advises, Human Decides' Systems in Critical Public Services, Ranked by Proven Payoff

2Why this topic

Readers get one yardstick across control towers, grid rooms and hospitals, setting the gains operators claim against the independent incident record and showing which fields publish their results with real figures.

3Where it applies

Global safety-critical operations: air traffic, power grids, rail dispatch, emergency call routing, hospitals

4What we looked at

  • The AI suggests and a trained person approves. Systems where the AI acts on its own are not included
  • Payoff is published with figures (minutes, fuel, cost or errors saved), and we note whether the figure comes from the operator, the vendor or an outside reviewer
  • Scale: runs live across a whole network rather than at one pilot site
  • Human load: how many suggestions a person reviews and how much time they have to accept or reject each one, where operators report it
  • Incident record of the system's class weighed against its payoff
  • Peer-reviewed clinical studies (Nature Medicine, NEJM, JAMA Network Open, meta-analyses) — Mortality, time-to-treatment, detection sensitivity, and how often clinicians evaluate and confirm AI alerts
  • Operator reports (FAA, NATS, DB InfraGO, NESO, NASA) — Delay minutes, fuel, landings per hour, balancing cost saved, and acceptance rates where published. These are the operators' own claims
  • Outside reviewers (Cranfield University, JAMA Internal Medicine validation, Alan Turing Institute) — Independent checks of operator or vendor figures

5Who ranked

The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, ChatGPT, Meta, Grok, Gemini, MiMo, StepFun, Z.ai, Kimi, DeepSeek, Wenxin.

6How the places were set

The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.

06Each AI's own list

🇺🇸 Claude10 places · 4 notes

The same question, answered by Claude alone. The final list above combines all the AIs.

  1. 1
    Kaiser Permanente Advance Alert Monitor (AAM), USA

    About one alert a day for every 35 patients, screened by remote nurses, linked to an estimated 520 fewer deaths a year across 21 hospitals

    Source ↗
  2. 2
    NATS Time-Based Separation / Intelligent Approach, Heathrow (UK)

    Headwind delay down more than 60% at the world's busiest two-runway airport; controllers still set every gap

    Source ↗
  3. 3
    TREWS sepsis early warning (Bayesian Health / Johns Hopkins)

    Clinicians evaluated 89% of its alerts, an unusually high follow-through, and quick confirmation was tied to an 18.7% relative drop in sepsis deaths

    Source ↗
  4. 4
    NESO Quartz Solar AI forecast (Great Britain grid control room, 2025)

    2.8 times more accurate, worth about £30m a year; engineers still decide how much reserve to buy

    Source ↗
  5. 5
    NASA ATD-2 departure metering, Charlotte (now in the FAA's TFDM program)

    Over 1 million gallons saved by holding planes at the gate rather than in the taxi line, and an outside reviewer questions the math

    Source ↗
  6. 6
    Alaska Airlines Flyways (Air Space Intelligence), USA

    Dispatchers turned down about 7 of every 10 suggestions, and the ones they kept still saved 1.2 million gallons in 2023

    Source ↗
  7. 7
    Deutsche Bahn AI dispatching assistant (Germany)

    58,000 delay minutes removed in 2022, and dispatchers followed the AI's suggestion in 7 of 10 analysed cases

    Source ↗
  8. 8
    Viz.ai stroke alerts (Viz LVO), USA

    About 2,000 hospitals; the vendor-cited multicenter study says 31 minutes saved, the randomized trial says 11

    Source ↗
  9. 9
    FAA SMART (Air Space Intelligence), USA 2026

    The largest rollout here (22 centers) on the thinnest proof: 180 rerouted flights, under an $875,950,000 contract

    Source ↗
  10. 10
    Corti cardiac-arrest call AI, Copenhagen EMS (Denmark)

    The AI spotted more cardiac arrests than dispatchers, yet in a randomized trial its alerts did not improve what dispatchers recognized

    Source ↗

What Claude noticed

The weak link is often the hand-off, not the model: in Copenhagen the AI recognized more cardiac arrests than dispatchers, but alerting the dispatchers did not raise their recognition.

The rubber-stamp point has a number: in a Radiology study, when the AI was wrong, inexperienced readers' accuracy fell from about 80% to 20%, and even very experienced readers fell to 45.5%.

A low acceptance rate is a sign of real checking: Alaska dispatchers took only about 30% of Flyways routes and still saved 1.2 million gallons in a year.

Alert cost per true catch: the Epic Sepsis Model flagged 18% of all inpatients and missed 67% of sepsis cases; even alerting once per patient meant reviewing 15 patients to find one case. Kaiser's AAM averages about 1 alert per 35 patients a day.

🇺🇸 ChatGPT6 places · 3 notes

The same question, answered by ChatGPT alone. The final list above combines all the AIs.

  1. 1
    MASAI AI-supported mammography screening, Sweden

    In a randomized trial of more than 100,000 women, interval cancers were 12% lower, with a similar false-positive rate.

    Source ↗
  2. 2
    TREWS sepsis early-warning system, United States

    Among alerted sepsis patients, timely clinician confirmation was associated with 1.85 hours faster antibiotic orders—and 38% of evaluated alerts were confirmed.

    Source ↗
  3. 3
    Viz.ai stroke care coordination, United States

    Across 166 facilities, AI-assisted coordination was associated with interventionalist notification 39.5 minutes earlier—not treatment 39.5 minutes earlier.

    Source ↗
  4. 4
    FAA SMART air-traffic decision support, United States (2026)

    The first 180 reported reroutes averaged nearly 80 gallons of fuel saved per flight; that is an early FAA claim, not an independent evaluation.

    Source ↗
  5. 5
    ADA-PMB AI train-dispatching assistant, Germany

    Its reported Frankfurt gain is about 100 seconds per dispatching decision, but it is active in four regions—not across Germany's full network.

    Source ↗
  6. 6
    Mount Sinai machine-learning deterioration alerts, United States

    In a 2,740-patient study, alerts made care escalation more likely; the report gives no exact mortality reduction and the study ended early.

    Source ↗

What ChatGPT noticed

The strict evidence screen supports six ranked examples, not ten: several widely cited 911 efficiency figures concern automated call handling, while some grid and rail projects remain research-stage or limited deployments. Padding this list would blur the user's human-approval and scale requirements.

The clearest published measure of human attention is not seconds per suggestion but clinician follow-through: TREWS providers evaluated 89% of alerts and confirmed 38% of those evaluated. The study does not establish a universal safe acceptance rate.

Human sign-off alone is not proof of meaningful oversight: an experiment with 438 clinicians found prescriptive biased AI advice influenced emergency decisions, while descriptive flags better preserved participants' original decisions.

🇺🇸 Meta10 places · 4 notes

The same question, answered by Meta alone. The final list above combines all the AIs.

  1. 1
    NATS Intelligent Approach Pairwise — Heathrow, UK (2024-)

    One spacing choice now cuts arrival delay from 5.55 to 4.46 minutes, and Heathrow has banked 1.5 million holding minutes in a decade.

    Source ↗
  2. 2
    EUROCONTROL Network Manager — Capacity and Weather-Based Operations (CWBO), Europe

    3.1 million delay minutes saved this year across the European network — the largest total in this ranking, by a human-approved flow system.

    Source ↗
  3. 3
    Corti AI — Copenhagen Emergency Medical Services, Denmark

    It spots cardiac arrest in 93-95% of calls vs 73% for dispatchers alone, and about 10 seconds faster — seconds that decide CPR.

  4. 4
    Bayesian Health TREWS — Targeted Real-Time Early Warning System for Sepsis, US

    When clinicians confirmed its alert within 3 hours, sepsis patients were 18.7% less likely to die — the first FDA-cleared continuous AI sepsis monitor.

    Source ↗
  5. 5
    Duke Sepsis Watch — Duke University Health System, US (since 2018)

    A nurse-led team reviews its score every 5 minutes — and Duke links it to a 27% fall in sepsis deaths, with doubled bundle compliance.

    Source ↗
  6. 6
    FAA SMART — Strategic Management of Airspace, Routes and Trajectories, US (2026)

    Now live at all 22 centres, but its only published payoff is 180 rerouted flights in a Washington pilot — 10 minutes and 80 gallons each.

    Source ↗
  7. 7
    DB InfraGO ADA-PMB — Automatic Dispatching Assistance, Germany

    In Frankfurt, following its suggestion is worth 100 seconds per dispatch — but it still only runs in selected areas, with screen integration due December 2026.

    Source ↗
  8. 8
    RTE France AI Contingency Screening — French Transmission Grid

    20 times faster than manual security assessment — speed is the payoff, because a dispatcher who sees the failure case sooner can still choose.

    Source ↗
  9. 9
    Google / American Airlines Contrail Avoidance — Transatlantic (2023-2026)

    When pilots actually flew the AI plan, contrails fell 62% — but across all eligible flights the fall was only 11.6%, because most plans were not flown.

    Source ↗
  10. 10
    Carbyne AI Emergency Call Triage — New Orleans, US

    For two years it has only handled one narrow job — duplicate crash calls within 200 metres of a reported crash — and that narrowness is why it is allowed.

What Meta noticed

The biggest total and the thinnest proof are both in aviation: EUROCONTROL reports over 3.1 million delay minutes saved this year, while FAA SMART reports only 180 rerouted flights from its pilot — scale is not the same as proven payoff.

The contrail trial is the clearest rubber-stamp test in the ranking: 62% fewer contrails when the plan was flown, but only 11.6% across all eligible flights, because most AI plans were never flown.

Sepsis AI only pays when a human acts fast: TREWS mortality benefit appears only when clinicians confirmed the alert within 3 hours, which is why alert burden and confirmation time matter more than raw accuracy.

No operator publishes the ratio this ranking asks for — value per accepted suggestion per minute of human attention — and automation-bias research says approval rates alone do not prove real checking, because familiar, usually-right suggestions get rubber-stamped.

🇺🇸 Grok10 places

The same question, answered by Grok alone. The final list above combines all the AIs.

  1. 1
    FAA SMART (US national airspace, 2026)

    180 flights rerouted in first weeks at ~10 min and ~80 gallons fuel saved each, now live across all 22 centers under human manager approval

  2. 2
    EUROCONTROL FADE (European network delay prediction)

    75% accuracy on delay-trend prediction and up to 63% lower prediction error, now operational for airlines deciding reroutes and crew planning

  3. 3
    NextEra Energy AI dispatch and outage scheduling

    Over $20 million customer savings in a single year from AI-optimized dispatch and outage scheduling under human oversight

  4. 4
    KEPCO AI-powered grid management (South Korea)

    Projected ~110 billion won (~$73 million) annual electricity procurement cost reduction via AI demand forecasting and STATCOM optimization

  5. 5
    KATE AI triage support (Adventist Health and others)

    20-minute average ED length-of-stay reduction per patient and measurable sepsis-detection gains while nurses retain final triage authority

  6. 6
    EUROCONTROL COAST (approach spacing calibration)

    Up to 6% higher runway throughput at major airports while keeping or lowering ATCO workload through ML-calibrated separation buffers

  7. 7
    DB InfraGO ADA-PMB rail dispatch assistant (Germany)

    Real-time recommendations every 60 seconds to dispatchers who accept or reject, active across multiple dense-traffic regions and expanding

  8. 8
    Corti AI 911 medical-call triage (Seattle Fire Department)

    50% increase in calls diverted to nurse line via real-time AI prompts that dispatchers can ignore, keeping human final authority

  9. 9
    Carbyne AI Emergency Call Triage (New Orleans / Orleans Parish)

    More than 30% reduction in redundant crash-related calls and up to 40-second faster answer times for distressed callers while humans handle true emergencies

  10. 10
    EUROCONTROL TPI (traffic prediction improvements)

    Route predictions within 6 NM for 65% of unseen flights (vs 10% prior) feeding human flow-management decisions across the network

🇨🇳 MiMo10 places · 4 notes

The same question, answered by MiMo alone. The final list above combines all the AIs.

  1. 1
    FAA SMART AI air traffic tool (2026, US)

    One accepted reroute is worth 10 minutes and 80 gallons — and all 22 control centers took the advice eight weeks before Congress finished asking its first safety questions.

    Source ↗
  2. 2
    Deutsche Bahn S-Bahn AI train dispatcher (Stuttgart, Germany)

    A dispatcher gets a fresh conflict plan every 60 seconds, Stuttgart claws back up to 8 minutes of delay — and 'in theory' 17 more trains a day fit on one core route.

    Source ↗
  3. 3
    NESO Open Climate Fix solar forecast AI (UK grid)

    Halving one forecast error turned into £30M a year of reserve savings — the grid's biggest quiet AI win, and no one asked for a pause.

    Source ↗
  4. 4
    EUROCONTROL CWBO AI-assisted flow rules (Europe, 2025–26)

    3.1 million delay minutes banked in one European summer — the network prices delay at €130 a minute, so the saving reads as about €400 million.

    Source ↗
  5. 5
    Corti AI cardiac-arrest advice for 911 dispatchers (Copenhagen)

    The AI spots cardiac arrest 85% of the time versus a dispatcher's 77.5% — yet handing dispatchers the AI barely moved their score (93.1% vs 90.5%, p=0.15).

    Source ↗
  6. 6
    PG&E AI wildfire grid monitoring (California)

    5.5 million meters watched by machine learning caught 17 would-be ignitions and kept 12 million outage-minutes off customers in one fire year.

  7. 7
    Bayesian Health sepsis AI at Cleveland Clinic (US hospitals)

    Ten times fewer false alarms in the pilot — and an outside lab checking a rival sepsis model found the vendor's published numbers don't travel to real wards.

    Source ↗
  8. 8
    KEPCO AI grid demand and voltage control (Korea, 2026)

    159 datasets became 95,000 — and the state utility prices its own AI at 110 billion won a year, nationwide.

    Source ↗
  9. 9
    NextEra Grid Composer AI dispatch (Florida, US)

    Half a trillion data points a day, and the utility counts $20 million saved in 2026 by comparing every human dispatch against the AI's cheaper plan — humans still hold the switch.

    Source ↗
  10. 10
    Telangana 112 AI emergency dispatch (India, 2026)

    Total call handling fell from 190 seconds to 133 — the AI names the nearest of 1,860 patrol cars; a person sends it.

    Source ↗

What MiMo noticed

Cash math vs attention math: ~180 accepted SMART reroutes × ~80 gallons ≈ $61,000 of fuel at November prices — against $23.3M already obligated on a $875.95M contract. The payoff everyone quotes is delay minutes and resilience, and no ledger publishes the conversion from one accepted suggestion to one minute of a human's attention.

Advice can beat humans and still add nothing to them: in Copenhagen's 169,049-call RCT the ML flagged cardiac arrest more sensitively than dispatchers alone (85.0% vs 77.5%), yet dispatchers given the alert were not significantly better than those without (93.1% vs 90.5%, p=0.15). The value of advice is not the sum of the two accuracies.

The rubber-stamp point is measurable and it is near: in a controlled experiment surgeons accepted 3 of every 5 deliberately wrong AI incision guides — while their self-reported confidence rose. Human-in-the-loop is a designed behavior, not a guarantee.

A 2026 meta-analysis prices human review itself: 26.6% of erroneous machine ECG interpretations were never corrected by clinicians. The loop intercepts errors only sometimes — and that interception rate, not model accuracy, is the real safety budget of 'AI advises, human decides'.

🇨🇳 StepFun10 places · 4 notes

The same question, answered by StepFun alone. The final list above combines all the AIs.

  1. 1
    FAA SMART (US air traffic, 2026)

    22 centers, about 200 data streams, and every suggestion still gets a human signature: 180 flights rerouted for 10+ minutes and nearly 80 gallons of fuel each.

    Source ↗
  2. 2
    Johns Hopkins TREWS sepsis AI (US hospitals, 2022)

    The rare advice system whose payoff survived peer review: in-hospital sepsis mortality fell 18.7% relative after clinicians started getting alerts.

    Source ↗
  3. 3
    Corti cardiac-arrest AI (Copenhagen EMS, Denmark)

    It listens to the 911 call and spots cardiac arrest with 84.1% sensitivity against dispatchers' 72.5% - across 169,236 real calls.

    Source ↗
  4. 4
    Deutsche Bahn AI delay prevention (Germany, 2025)

    Three AI networks kept 58,000 delay minutes off Germany's rails in a year - about 40 days of delay that never happened.

    Source ↗
  5. 5
    Google DeepMind wind forecasting (2019)

    Forecasts wind output 36 hours ahead and lifted the value of that energy by roughly 20% - a vendor's number nobody has independently audited.

    Source ↗
  6. 6
    Network Rail 'Insight' AI (UK, 2025)

    Track faults caused 341 days of delay a year; the AI's job is to warn crews before the fault, not after.

    Source ↗
  7. 7
    Pano AI wildfire cameras (US utilities, 2025)

    One camera caught the Brady Fire at under an acre - but PG&E's own evaluation said camera AI detection was not yet meeting desired performance.

    Source ↗
  8. 8
    Oncor dynamic line rating (US grid, 2020)

    Lines rated 6-14% higher 84-91% of the time - the grid's cheapest new capacity is an operator reading a smarter number.

    Source ↗
  9. 9
    NATS Heathrow AI trial (UK, 2025)

    A Heathrow-only trial promising to reclaim 20% of the capacity lost to low cloud - a claim, not a measurement, and not yet network-wide.

    Source ↗
  10. 10
    Epic Sepsis Model (US hospitals, 2021)

    Deployed in hundreds of hospitals, flagged 18% of all patients, and missed 67% of sepsis cases: scale without payoff.

    Source ↗

What StepFun noticed

The arithmetic of SMART's first weeks: 180 rerouted flights at 10+ minutes and ~80 gallons each is roughly 1,800 minutes and 14,400 gallons of fuel - real savings, but a rounding error against an $875M, 12-year contract, for which USAspending-derived records show only $23.3M obligated to Air Space Intelligence in the year to November 5, 2026.

EUROCONTROL logged around 49,000 en-route delay minutes per day across Europe in week 39 of 2026, 39% below 2025 - the prize that advice systems chase is measured in hundreds of thousands of minutes a day, which is why even 10 minutes per flight adds up fast.

OECD's monitor counts roughly 19,064 AI incidents and hazards and screens about 1,000 AI-related media events a day, yet the strongest payoff evidence on this list (TREWS, Corti) and the strongest caution (Epic) all come from peer-reviewed studies - the measurement, not the incident feed, is what separates rank 2 from rank 10.

Nobody publishes the value-per-check ratio, but the two ends of this list bracket it: the FAA says 100% of SMART's recommendations are reviewed by staff, while the Epic model fired on 18% of hospitalized patients and missed 67% of sepsis cases - the point where advice volume turns checking into rubber-stamping.

🇨🇳 Kimi10 places · 4 notes

The same question, answered by Kimi alone. The final list above combines all the AIs.

  1. 1
    NATS Intelligent Approach (Heathrow, live since 2015)

    Ten years live: controllers eyeball AI-calculated gaps and issue the clearances — and the tool cut headwind delays >60% while saving 230,000 airborne-holding minutes a year at one airport.

    Source ↗
  2. 2
    TREWS sepsis early-warning system (US, Johns Hopkins / Bayesian Health)

    The only system here with peer-reviewed proof that a human confirming an AI alert within 3 hours cut sepsis deaths 18.7% — and it also publishes the human side: 89% of alerts actually reviewed.

    Source ↗
  3. 3
    Viz.ai LVO stroke triage (US)

    A 14,116-patient, 166-hospital real-world study found AI-flagged stroke patients reached a neurointerventionalist 39.5 minutes faster — at network scale, not a pilot ward.

    Source ↗
  4. 4
    FAA SMART air-traffic planning AI (US, nationwide Nov 2026)

    Three days after going live at all 22 US en-route centers it watched the FAA ground the whole country — officially unrelated, but the $875M system's entire payoff evidence is still 180 rerouted flights.

    Source ↗
  5. 5
    EUROCONTROL Network Manager AI tools — FADE / PETA / Curfew MIRROR (Europe)

    The quiet giant: machine-learning delay forecasts have run inside the pan-European flow-management brain since 2023, guiding human flow managers — but EUROCONTROL publishes benefits qualitatively, not in minutes.

    Source ↗
  6. 6
    Xcel-NCAR AI wind forecasting for grid dispatch (US)

    Grid operators still decide what to dispatch — the AI just killed the forecast error by 40%, and Xcel says that single number was worth $60M and a quarter-million tons of CO2 a year.

    Source ↗
  7. 7
    Corti AI cardiac-arrest detection on 112 calls (Copenhagen, Denmark)

    Peer-reviewed on 108,607 real calls: the AI caught 11.6 points more cardiac arrests than trained dispatchers and flagged them 10 seconds sooner — every second matters when survival drops ~10% per minute.

    Source ↗
  8. 8
    AI 911 call triage and deflection (Orleans Parish, Jefferson County CO, Monterey County CA)

    County 911 centers report the AI shaves 40 seconds off emergency answer times by handling the noise — while humans keep every emergency call.

    Source ↗
  9. 9
    Pano AI wildfire detection on utility grids (Xcel, PG&E, PGE; US)

    The AI spots smoke; a human analyst confirms it; fire dispatch decides — and one Colorado lightning fire was held to 3 acres after a camera caught it at 0.25 acres.

    Source ↗
  10. 10
    Optellum Virtual Nodule Clinic lung-cancer triage (NHS / US)

    A risk score nudges a clinician's decision and median time-to-lung-cancer-diagnosis collapsed from 129 days to 25 — published, but on a modeled workflow, so it anchors the bottom of the top 10.

    Source ↗

What Kimi noticed

Nobody publishes the ratio that matters — value per unit of human attention. The closest public data point is TREWS: clinicians reviewed 89% of AI alerts (53% within an hour) and confirmed-alert patients gained an 18.7% relative mortality reduction; SMART, by contrast, hasn't published acceptance rates at all.

Unit economics are wildly unproven for the biggest bet: $875M over 12 years bought 180 rerouted flights so far — about $4.9M per rerouted flight at the ceiling, against a $33B/year delay problem if it scales as promised.

The trust tax is paid in coincidences: SMART went nationwide on Nov 10 and the FAA grounded all departures on Nov 13; officials insist the tool 'isn't plugged in,' but the timing debate now shadows every future SMART milestone.

The class's incident record explains the ranking's caution: the Epic Sepsis Model — the most widely deployed hospital AI adviser — missed 67% of sepsis cases and fired 88% false alarms in independent validation, the exact failure mode (automation bias plus alert fatigue) that 'AI advises, human decides' is supposed to contain.

🇨🇳 DeepSeek10 places · 4 notes

The same question, answered by DeepSeek alone. The final list above combines all the AIs.

  1. 1
    FAA SMART (US air traffic, 2026)

    180 human-approved reroutes saved about 14,400 gallons of fuel — roughly $62,000 at $4.31/gal, against an $875 million contract.

    Source ↗
  2. 2
    DB InfraGO ADA-PMB (Germany rail, 2026)

    Over 100,000 delay minutes were saved in 2023 across four dense regions — while a dispatcher still accepts or rejects every suggestion.

    Source ↗
  3. 3
    RWJBarnabas Epic Deterioration Index (US hospitals, 2026)

    High-risk deaths fell from 23.1% to 18.6% across 11 hospitals — a 4.5-point absolute drop, not just a relative headline.

    Source ↗
  4. 4
    Yale New Haven eCART (US hospitals, 2026)

    eCART flags no more than 20–30 patients at a time, yet its 13% relative mortality drop came with a 1.2-point absolute fall.

    Source ↗
  5. 5
    Viz.ai LVO stroke triage (US hospitals, 2026)

    DIDO time fell from 202 to 113 minutes — an 89-minute cut that beat the 120-minute benchmark, though the meta-analysis found no significant clinical-outcome gain.

    Source ↗
  6. 6
    China Southern Grid Tianxuan-Lingxi (Shenzhen power grid, 2026)

    A megacity dispatcher's daily signal load fell from 750 to under 50, and a 30-minute ticket review became about a dozen seconds.

    Source ↗
  7. 7
    Carbyne Emergency Call Triage (New Orleans 911, 2026)

    New Orleans cut duplicate crash calls by 30% and answered high-priority callers 40 seconds faster — but only for a narrow slice of 911 traffic.

    Source ↗
  8. 8
    TREWS sepsis alert (US hospitals, 2022)

    89% of alerts were evaluated, but only 38% were confirmed — a rare published look at how often humans actually agree.

    Source ↗
  9. 9
    NASA Digital Trajectory Rerouting (US air traffic, 2026)

    Each rerouted flight saved close to an hour of aggregate metroplex delay; at Houston alone, airlines approved 62% of 3,870 candidates.

    Source ↗
  10. 10
    CAISO AI Genie (California grid, 2026)

    CAISO's AI copilot turned a week of outage-review work into 15 minutes — if you believe the vendor; CAISO still calls it a pilot.

    Source ↗

What DeepSeek noticed

The FAA's first 180 reroutes saved about 14,400 gallons — roughly $62,000 of fuel at $4.31/gal. That is a tiny fraction of the $875 million contract, and it is the only payoff figure the FAA has published so far.

TREWS is one of the few systems that publishes both sides of human attention: 89% of alerts were evaluated, but only 38% were confirmed. Evaluation is not agreement.

DB's 100,000 delay minutes come from 'intelligent capacity management,' a combination of scheduling, operations and ex-post tools — not ADA-PMB alone, so attribute the gain carefully.

NASA's field evaluation shows the human gate is not just ATC: airline dispatchers approved 62% of 3,870 reroute candidates, and each accepted reroute saved close to an hour of aggregate metroplex delay.