UN Creates Its First AI Scientific Panel: Warning on Governance

⏱️ Lectura: 9 min

The UN’s first AI scientific panel has already issued its initial warning: the window to regulate artificial intelligence effectively is closing. The notice comes as leading labs release models every few months while legal frameworks still debate basic risk definitions.

📑 En este artículo

1. TL;DR
2. What happened: the AI scientific panel’s warning
3. Context and history
4. Technical Details and Performance
5. How to Follow the Process and Get Involved
6. Impact and Analysis
7. What’s Next
8. Frequently Asked Questions
   1. What is the Independent International Scientific Panel on AI?
   2. How is it different from the Global Dialogue on AI Governance?
   3. Why is it compared to the IPCC?
   4. Can the panel ban or restrict AI models?
   5. What tools do AI safety institutes use to evaluate models?
   6. When will the panel publish its first full report?
9. References

The Independent International Scientific Panel on AI, created by the UN General Assembly within the Global Digital Compact, works as a mirror of the IPCC but for artificial intelligence: it gathers technical evidence, it doesn’t dictate laws.

TL;DR

• The UN created the Independent International Scientific Panel on AI within the Global Digital Compact, adopted at the Summit of the Future in September 2024.
• The panel is explicitly modeled on the IPCC, the UN’s scientific body on climate change.
• In parallel, the Global Dialogue on AI Governance was launched, an intergovernmental forum to coordinate rules between countries.
• The panel warns that the gap between the speed of model development and states’ regulatory capacity keeps growing.
• Its mandate includes publishing periodic scientific assessments on AI capabilities, risks, and opportunities.
• The warning comes as the European AI Act and initiatives from the US and China move forward with different, poorly coordinated approaches.

What happened: the AI scientific panel’s warning

The UN’s AI scientific panel sent a direct message to governments: if regulation keeps its current pace, it will permanently lag behind the technology it’s trying to oversee. The warning doesn’t propose a specific law; instead, it demands that states speed up the creation of risk assessment mechanisms before the most advanced systems are deployed at scale.

The warning is based on a pattern the panel has documented since its formation: AI labs release new models with capabilities that exceed safety tests designed just months earlier. Without a permanent independent scientific assessment mechanism, every model update forces regulators and labs to start from scratch.

Context and history

The panel didn’t come out of nowhere. In September 2024, UN member states adopted the Global Digital Compact during the Summit of the Future, the document that for the first time established concrete multilateral commitments on digital governance and artificial intelligence. That compact mandated two parallel mechanisms: the Independent International Scientific Panel on AI and the Global Dialogue on AI Governance.

The design deliberately copies the Intergovernmental Panel on Climate Change (IPCC), the body that since 1988 has centralized scientific consensus on climate so governments can negotiate policy on common ground. The logic is the same: before discussing what to ban or allow, countries need a shared technical diagnosis of what an AI model can actually do and what risks it carries.

That precedent already existed partially in the network of AI Safety Institutes, first launched by the United Kingdom and the United States after the 2023 Bletchley Park AI Safety Summit, and later replicated by Japan, Singapore, Canada, South Korea, and the European Union. The UN’s scientific panel aims to expand that effort to a truly global scale, including countries that have so far been left out of those bilateral conversations.

flowchart TD
A["Global Digital Compact (2024)"] --> B["AI Scientific Panel"]
A --> C["Global Governance Dialogue"]
B --> D["Periodic technical reports"]
C --> E["Negotiation between states"]
D --> E
E --> F["National and regional policies"]

Technical Details and Performance

The technical discussion behind the warning centers on capability evaluations (capability evals): standardized tests that measure whether a model can, for example, help synthesize a pathogen, write a functional exploit, or deceive a human evaluator. These evaluation frameworks are the technical input the AI scientific panel needs to produce its reports.

One of the most widely used open frameworks among AI safety institutes is Inspect, published by the UK’s AI Safety Institute. It allows you to define a battery of evaluations and run them against any model compatible with a standard API:

pip install inspect-ai

With the package installed, a minimal evaluation looks like this:

from inspect_ai import Task, task, eval
from inspect_ai.dataset import Sample
from inspect_ai.scorer import match
from inspect_ai.solver import generate

@task
def prueba_basica():
    return Task(
        dataset=[Sample(input="What is the capital of France?", target="Paris")],
        solver=generate(),
        scorer=match(),
    )

eval(prueba_basica(), model="openai/gpt-4o-mini")

This example runs a single control question, but the same pattern scales to hundreds of safety tasks: jailbreak attempts, malicious code generation, or deceptive reasoning. AI safety institutes use variants of this approach to produce the data that feeds the AI scientific panel.

Besides automated evaluations, AI safety institutes combine human red-teaming tests: experts who actively try to break a model’s safeguards before its public release. The panel aims to standardize common criteria so that these exercises, currently fragmented across labs and countries, produce comparable results.

How to Follow the Process and Get Involved

The work of the panel and the Global Dialogue is public, though scattered across several UN platforms. To follow it closely:

• The official Global Digital Compact page centralizes the founding documents and the session calendar.
• The UN Digital Library indexes resolutions and full-text reports, searchable via API.
• National AI safety institutes (United Kingdom, United States, Japan, Singapore) publish their own technical evaluation reports, which often feed into the panel.

To automate tracking of new documents related to the panel, the UN Digital Library exposes a search accessible via HTTP:

curl "https://digitallibrary.un.org/search?p=scientific+panel+artificial+intelligence&of=xm"

That command returns records in MARC XML format; with a simple parser you can filter by publication date and get notified when a new panel document appears.

Impact and Analysis

The panel’s warning exposes an underlying tension: several AI governance mechanisms already exist, but they operate with different scopes and speeds. The following table compares the main ones:

Mechanism	Scope	Approach	Main Limitation
UN Scientific Panel	Global, all member states	Technical diagnosis, no regulatory power	Can’t impose rules, only evidence
Global Dialogue on AI Governance	Global, intergovernmental	Political coordination between countries	Depends on consensus, moves slowly
Network of AI safety institutes	Bilateral between specific countries	Technical evaluation of frontier models	Excludes most countries
European Union AI Act	Regional, legally binding	Risk-level based regulation	Only applies within the EU

The panel doesn’t compete with these mechanisms: it aims to be the common evidence layer that’s currently missing. Without a shared diagnosis, each regulatory bloc ends up measuring different risks with different criteria, which makes any real coordination between countries difficult.

For teams building products on top of third-party models, this fragmentation is already a practical problem: a safety evaluation approved under the European AI Act doesn’t necessarily meet the criteria of an AI safety institute in another country. A common scientific standard, even if not binding, would reduce the need to repeat full audits for every jurisdiction.

💭 Key point: the panel resembles the IPCC in design, but faces a problem the IPCC never had: its object of study changes capabilities every few months, not every decade.

What’s Next

The original mandate of the Global Digital Compact provides for periodic sessions of the Global Dialogue on AI Governance and the publication of regular scientific reports from the panel, following the model of the IPCC’s assessment reports. The exact schedule for upcoming publications depends on the General Assembly and hasn’t been finalized yet.

Meanwhile, the capacity gap between governments also worries the panel: most countries lack the budget or technical staff to evaluate frontier models on their own. Part of the panel’s mandate is to level that gap, offering assessments that any member state can use as a reference, without needing to build its own AI safety institute from scratch.

What this warning does make clear is the panel’s priority: convincing governments that early coordination is cheaper than reactive regulation after a serious incident. It’s the same argument the IPCC used for decades, with mixed results.

⚠️ Heads up: a scientific panel without binding power can document the risk perfectly and still not change the pace at which labs deploy their systems.

📖 Summary on Telegram: View summary

Try it yourself: install inspect-ai with pip install inspect-ai and run your first safety evaluation against a model in minutes.

Frequently Asked Questions

What is the Independent International Scientific Panel on AI?

It’s a body created by the UN within the Global Digital Compact to produce periodic scientific assessments on the capabilities and risks of artificial intelligence, without regulatory power of its own.

How is it different from the Global Dialogue on AI Governance?

The panel produces technical evidence; the Global Dialogue is the forum where governments negotiate policy using that evidence as common ground.

Why is it compared to the IPCC?

Because it copies its institutional design: a body of experts that centralizes scientific consensus so governments can negotiate policy on a shared diagnosis, instead of data fragmented by country.

Can the panel ban or restrict AI models?

No. Its mandate is diagnostic, not regulatory. Binding decisions remain the responsibility of each state or regional bloc, such as the European Union with its AI Act.

What tools do AI safety institutes use to evaluate models?

Open frameworks like Inspect, which automate batteries of safety and capability tests against any model compatible with a standard API.

When will the panel publish its first full report?

The Global Digital Compact provides for periodic reports following the IPCC model, but the exact schedule for the next publication still depends on the UN General Assembly.

References

• The Next Web: coverage of the UN scientific panel’s warning on the AI governance window.
• Global Digital Compact (UN): founding document that creates the scientific panel and the Global Dialogue on AI Governance.
• IPCC: the UN’s climate panel that served as the institutional model for the new AI panel.
• Inspect (UK AI Security Institute): open framework for model safety evaluation used by AI safety institutes.
• UN Digital Library: searchable archive of resolutions and official documents related to the panel.

📱 Like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.

Imagen destacada: Foto de Siavosh Hosseini en Unsplash