AI Impact Summit: What Actually Changed After Delhi

Search for an AI impact summit and two unrelated things come back under one name. One is a diplomatic process: the fourth in a line of global AI summits, held at Bharat Mandapam in New Delhi from 16 to 20 February 2026 and extended by a day, following Bletchley Park in 2023, Seoul in 2024 and Paris in 2025, and the first of the series hosted by a Global South nation. The next is in Geneva in 2027. The other is a paid enterprise conference. The ISG AI Impact Summit runs in New York on 28 and 29 October 2026, sells an early bird pass at 995 dollars against a 1,295 dollar standard rate, and defines impact as the gap between AI ambition and measurable return on investment, with proof points a chief financial officer will accept.

Neither is hiding anything. They simply have nothing to do with each other, and a first-time searcher cannot tell that from a results page. What they share is a word that makes a testable claim, which is unusual for a conference name and worth taking literally.

What the word impact was meant to fix

The naming was deliberate. The organisers' own framing is that AI's spread is disrupting employment, entrenching bias and driving energy use up sharply, and that this creates an urgent need to move beyond aspirational frameworks toward concrete, measurable impact. The programme was built around three principles called Sutras, People, Planet and Progress, translated into seven working areas called Chakras that ran from human capital and inclusion through safe and trusted AI to democratising access to compute.

Read the series in order and the drift is obvious. Safety at Bletchley, action in Paris, impact in Delhi. Each title concedes something about the last one. Impact is the most exposed of the three, because it is the only one that implies a measurement.

What New Delhi actually produced

The countable outcomes are real and worth stating before any critique. The expo drew more than 300 exhibitors from 30 countries across over 10 thematic pavilions. Delegations came from over 100 countries, including more than 20 heads of state and 60 ministers, alongside the chief executives of Google, OpenAI, Anthropic and DeepMind. India announced plans to add more than 20,000 GPUs to an existing base of 38,000 under its national compute portal, and Microsoft said it was on track to invest 50 billion dollars by the end of the decade to bring AI to lower-income countries. Two significant Indian model families launched at the event. India also set a Guinness World Record with 250,946 pledges collected for an AI responsibility campaign in 24 hours.

Some of it was theatre, and the record includes that too. A university exhibitor presented a robot dog as an indigenous development until observers identified it as a commercially available Chinese product, and the stall was ordered vacated. Delegates were left without food or water during a security lockdown before the Prime Minister's arrival, and the venue closed to the public on the day people had registered to attend.

Those incidents matter less as embarrassments than as a reminder of what an exhibition floor optimises for. A summit that names itself after impact invites this scoring, and gets it.

The gap nobody at the summit owns

The New Delhi Declaration was signed by more than 90 countries, including both China and the United States, which is not a small achievement in the current climate. The problem is what happens on 22 February.

The most careful analysis of the summit series, written by people broadly sympathetic to it, identifies an accountability deficit rather than a bad declaration. The commitments are voluntary and non-binding, and there is no framework to track progress on deliverables from one summit to the next. The model runs ad hoc, with no secretariat and no coordinating body, so continuity depends on whoever hosts next. Nor was participation the same as influence: civil society groups were accredited and given side events, then were not permitted in the room on the day heads of state and chief executives spoke. The same analysis notes that these gatherings often function as trade shows demonstrating industrial capability rather than forums for governance.

That is the whole issue in one line. A summit can produce a signed text, a compute pledge and a world record, and still have no mechanism that would tell anyone, a year later, which of it happened.

The one place impact was measured properly

The instrument existed. It was just pointed at programmes rather than at the summit.

On 17 February, at the same venue, J-PAL ran a day-long research and policy seminar called AI for Social Good: Impact That Works, and the evidence presented there was of a different kind entirely. A randomized evaluation in Tamil Nadu of a low-cost AI tool for identifying people at high risk of heart disease. Machine learning used to improve targeted public service delivery in Togo. Generative AI tested in teacher hiring in Ghana. An early look at an AI tool for reducing flood risk for vulnerable households in Bihar. The event also launched the AI Evidence Playbook, a reference for policymakers, practitioners and donors funding AI-enabled programmes, and the EVAH initiative for generating evidence on AI in health systems across lower and middle income countries. Elsewhere in the accredited programme, sessions examined small AI for the Global South, where the constraint is cost and footprint rather than model scale.

What separates that work from the plenary is a counterfactual. A randomized evaluation asks what would have happened otherwise, which is the only question that distinguishes an effect from a coincidence. No equivalent question was asked of the summit itself.

Four tests before you book any impact summit

The evidence above converts into four questions. They apply to a diplomatic summit, an enterprise conference and everything between.

Who tracks the commitments after everyone flies home

Ask for the mechanism, not the ambition. If nobody can name who publishes what happened to last year's announcements, and on what date, the commitments are marketing with a deadline of the closing ceremony. The absence of that function across four global summits is the single clearest finding in the record.

Is the decision room open or adjacent

Accreditation is not access. Count the events where the people making decisions and the people affected by them are in the same room at the same time, rather than in parallel tracks that never intersect. Programmes that describe a side event as inclusion are telling you where the boundary sits.

Does a euro or a signature move in the building

An announcement is not a transaction. A grant application submitted, a term sheet opened, a partnership agreement initialled, a pilot scheduled with a named regulator: these leave a trace that survives the week. Keynotes do not. This is the same test that separates what actually happens at an AI business summit from what the brochure implies, and it is why which AI summits are worth attending in 2026 is a question about format rather than about speakers.

Would you attend the second edition

The most honest question, and the hardest for a first-time host. Judge an event by what it publishes afterwards, unprompted, when the reporting is unflattering. An organiser willing to state which of its stated outcomes did not materialise has told you more about its standards than any attendee list.

What a summit built for outcomes has to look like

Applying those tests to your own event is uncomfortable, so it is worth being specific about what a structural answer looks like rather than describing an intention.

The EX Future Summit is organised around nine tracks, one of which exists solely to move money: a Global Financial Raise track mapping venture capital, private equity, national grants and Horizon Europe. Eight hosted meet-and-greet sessions run for one purpose, matching universities with private partners across continents, which is the mechanism behind moving research to commercialization rather than discussing it. Policy is not a side panel either. Bali Province's digital residency sandbox opens a cohort at the summit, granting a ninety day residency, a policy pilot lane and access to municipal data, so a founder tests against real regulators instead of a slide. That lane depends on knowing what sovereign AI means in practice for a jurisdiction writing its own rules.

The format follows the same logic: a single continuous thirty hour broadcast from 18 to 20 November 2026 moving between Las Palmas in the Canary Islands and Bali, twelve hours apart, with online attendance free for verified researchers, students and the EX community. Outcomes with numbers attached are the ambition, and one partner already has one, with EnerGaia Institute securing 11 EUIPO SME Fund approvals alongside EX-IX.

This is a first edition with no track record, and by test four above it should be judged on what it publishes in December rather than on what it promises in August. That is the correct standard. It is also the one the whole category has been avoiding.

FAQ

Is the ISG AI Impact Summit the same event as the India one?

No, and the two have no connection. One is a multilateral government process in the summit series that runs to Geneva in 2027. The other is a commercial enterprise conference in New York aimed at executives who need to prove return on AI investment internally. Same phrase, different organiser, audience, price and purpose.

Are AI summit declarations legally binding?

No. The New Delhi text attracted more than 90 signatory countries, which gives it diplomatic weight, but the commitments in it are voluntary and non-binding. There is currently no monitoring or evaluation mechanism attached to the summit series to track whether they are implemented, which is a separate problem from whether the text is good.

What is the difference between an AI safety summit and an AI impact summit?

The emphasis, and it moved deliberately. The first summits concentrated on frontier risk, safety research and building safety institutes. The impact framing shifted attention to deployment, measurable outcomes and access to AI resources for countries outside the small group that builds frontier models. Critics argue the shift traded scrutiny for showcase.

How do you actually measure whether an AI programme had impact?

With a counterfactual rather than a testimonial. The evaluations presented in New Delhi were randomized studies of specific deployments in health, governance, labour, education and climate, designed to isolate what the tool changed against what would have happened anyway. Ask any vendor or organiser what their comparison group was. The answer, or its absence, is the whole assessment.