From AI Policy to AI Experimentation: How Governments Are Learning Before They Scale

Governments around the world have spent the last few years producing artificial intelligence strategies, ethical principles, responsible AI frameworks, procurement rules, and guidance for public employees.

The harder phase has now arrived.

Public institutions need to use AI in real work.

That creates a management problem very different from writing policy.

Generative AI is already entering government through document drafting, translation, research, citizen services, policy analysis, administrative support, knowledge management, and many informal employee-led experiments. In many cases, adoption is decentralized and faster than the governance systems surrounding it.

A major OECD working paper published in July 2026 examined this exact gap. Drawing on official guidance across 14 countries, academic research, and government case studies, the OECD found that experimentation is becoming an important bridge between high-level AI principles and operational deployment. It also found that government guidance remains fragmented and that monitoring and evaluation are among the weakest areas.

This suggests an important shift in public-sector AI strategy.

Governments increasingly need to manage experimentation as an institutional capability.

Why Conventional Pilots Are Not Enough

Organizations have always used pilots.

A department tests a system with a small group, gathers feedback, and decides whether to expand it.

Generative AI makes this more complicated because the technology is probabilistic and highly dependent on context.

The same underlying model can perform differently depending on the prompt, data, retrieval system, user behaviour, task, language, and surrounding controls.

A pilot that asks whether employees “liked the tool” therefore provides limited evidence.

Government experimentation needs to answer much more rigorous questions.

Does the AI improve the actual outcome?

Does it save time after verification is included?

Where does it fail?

What types of information can safely be entered?

How often do employees accept incorrect output?

Does performance change across languages or user groups?

What happens when the system encounters an exceptional case?

Can the organization demonstrate accountability for the result?

These are operational questions, not technology demonstrations.

The OECD’s Five-Part Lens

The 2026 OECD work proposes evaluating government GenAI experiments across five broad areas: performance, public value, feasibility, usability, and risk management.

That framework is useful because it moves evaluation beyond technical accuracy.

Performance asks whether the system actually performs the intended task effectively.

Public value examines whether the use case creates meaningful value for citizens, employees, or society.

Feasibility considers whether the solution can realistically operate within institutional, technical, legal, financial, and data constraints.

Usability examines how people interact with the system in practice.

Risk management addresses privacy, security, bias, accountability, reliability, and other potential harms.

A system can score well technically and still fail as a government solution.

It may be too expensive to operate, difficult for employees to use, incompatible with legal requirements, or incapable of producing measurable public value.

Governments Have Strategies, but Execution Still Lags

The OECD’s Digital Government Outlook 2026 shows how rapidly AI has moved into public-sector strategy. It reports that 92% of OECD countries have an AI-in-government strategy.

Yet the report also identifies a familiar execution gap.

Governments are significantly better at establishing strategies than at implementing and monitoring them. Tools such as risk assessments, transparency registers, procurement support, and systematic performance measurement are still not widespread enough.

This is where structured experimentation becomes strategically important.

A controlled experiment can translate abstract principles into practical rules.

A policy might say that humans must remain accountable.

An experiment determines exactly where human review is needed.

A strategy might call for transparency.

An experiment determines what should be logged and disclosed.

A responsible AI principle might require fairness.

An experiment identifies the groups across which performance needs to be measured.

Operational learning makes governance concrete.

Sandboxes Need Exit Criteria

One risk is that experimentation becomes permanent.

Organizations sometimes accumulate dozens of AI pilots without deciding which ones should scale, change, or stop.

A mature experimentation system therefore needs explicit gates.

Before the experiment begins, the institution should identify its hypothesis, expected value, target users, permitted data, evaluation criteria, risk threshold, and decision rule.

At the end, there should be only a few possible outcomes.

Scale it.

Redesign it.

Restrict it.

Or stop it.

Stopping an experiment should not be considered failure. Discovering that an AI use case produces insufficient value or excessive risk is exactly what controlled experimentation is supposed to reveal.

Employees Are Part of the Experiment

Generative AI creates another challenge because employees can often adopt it before the institution formally deploys anything.

A public servant with access to a general-purpose AI tool can immediately begin using it for summaries, drafting, analysis, or translation.

Governance therefore cannot rely exclusively on centralized technology projects.

Institutions need safe environments in which employees can experiment without exposing confidential information or accidentally integrating unvalidated AI into consequential decisions.

Training also needs to evolve.

Prompting skills matter, but public servants also need AI judgment.

They need to recognize when outputs require verification, which data should never be entered, when an AI system is inappropriate, and when a human expert needs to take control.

From Innovation Theatre to Institutional Learning

Generative AI creates pressure for governments to move quickly.

The opposite pressure comes from accountability, regulation, public trust, and the consequences of error.

Structured experimentation provides a way to manage both.

It allows institutions to innovate while generating evidence before large-scale commitment.

The most advanced governments may therefore be distinguished less by how many AI pilots they announce and more by how systematically they learn from them.

The critical capability is not experimentation alone.

It is the organizational mechanism that converts experiments into standards, procurement requirements, workforce practices, architecture decisions, measurable outcomes, and clear decisions about what should scale.

The public sector has already entered the generative AI era.

The next question is whether governments can build a disciplined learning system fast enough to govern the transformation intelligently.

Sources and Further Reading

OECD, “Generative AI experimentation in government: Learning from emerging guidelines”, July 2026. DOI: 10.1787/42815683-en.

OECD, “Digital Government Outlook 2026: From Foundations to Transformational Impact”, June 2026. DOI: 10.1787/0496b2bc-en.

OECD, “Trustworthy artificial intelligence in the public sector”, June 2026.

The Hidden Language Gap in AI: Multilingual Does Not Mean Equally Intelligent

Large language models are increasingly presented as global technologies.

A single interface may communicate in English, Arabic, Spanish, Mandarin, French, Hindi, Japanese, Turkish, and dozens of other languages. For users, this creates the impression that the same intelligence simply changes language.

Research is revealing a more complicated reality.

In September 2026, a study published in Scientific Reports tested three widely used language models across 12 languages representing several language families. The researchers found substantial differences in comprehension across languages. Interestingly, English was not consistently the strongest language, and several Romance languages performed better in the experimental setting. The study pointed toward factors such as tokenization, linguistic distance, the quantity of training data, and the origins of that data as contributors to performance differences.

The implications extend far beyond translation.

If AI increasingly mediates access to education, public services, healthcare, business information, and digital knowledge, unequal linguistic capability can become a new form of digital inequality.

Speaking a Language Is Different From Understanding It

An AI model can produce grammatically plausible text in a language while still having weaker underlying competence.

This distinction matters.

A model may translate a simple sentence correctly but struggle with culturally embedded concepts.

It may write fluent prose but misunderstand instructions involving complex grammatical structures.

It may perform reasoning more effectively in one language than another.

It may recognize terminology used by urban professionals while failing on dialects, regional expressions, or locally specific concepts.

Fluency can therefore create false confidence.

Users may assume that a natural-sounding answer reflects equally strong comprehension. In reality, linguistic smoothness and cognitive reliability are separate dimensions.

Training Data Creates an Uneven Map of Human Knowledge

Large language models learn from data.

That data is not distributed equally across languages.

Some languages have enormous digital ecosystems containing books, news, academic research, software documentation, government documents, online discussions, and professionally edited reference material.

Others have much smaller digital footprints.

Even when significant amounts of text exist, the material may not represent the full cultural or linguistic diversity of the population.

The result is an uneven knowledge landscape.

A model may have abundant high-quality training material for formal English and relatively little for a minority language, local dialect, or specialized regional terminology.

This affects more than vocabulary.

It influences what knowledge the AI can access, how confidently it reasons, and which cultural assumptions appear normal inside its outputs.

Bigger Multilingual Models Are Not Automatically Better

Research from the 2026 ACL ecosystem provides another important insight.

The Goldfish project examined language modelling across hundreds of languages and released a suite of more than 1,000 small monolingual models covering 350 languages. The researchers reported that relatively small monolingual models could outperform much larger multilingual models on basic language modelling measures for many low-resource languages.

The finding challenges an intuitive assumption.

Putting more languages into one enormous model does not guarantee that every language receives equal representation or capability.

Multilingual models face competition for model capacity, unequal data availability, and technical choices such as tokenization that can affect languages very differently.

For organizations deploying AI globally, this creates an important governance lesson.

A successful English evaluation does not validate the same system for every other language.

Translation Is Not Always the Answer

One common strategy is to translate a user’s query into English, process it, and translate the response back.

Sometimes this works extremely well.

Sometimes native-language prompting performs better.

Recent ACL research comparing multilingual prompting strategies found that no single approach is optimal across all languages and tasks. Translation was particularly valuable for some low-resource languages, while high-resource languages often gained much less from it. Researchers have begun exploring learned routing systems that choose the appropriate strategy based on the specific request.

This points toward a more sophisticated future for multilingual AI.

Instead of one model pretending every language is equivalent, systems may dynamically decide how a request should be processed.

Cultural Grounding Is the Next Challenge

Language carries culture.

An AI system can translate the words of a request accurately and still miss the social context behind it.

This is particularly significant in areas such as healthcare communication, education, law, government services, financial advice, and customer engagement.

Culturally grounded AI requires more than translated English datasets.

Researchers behind the 2026 UPDESH project, for example, explored synthetic instruction data grounded in language-specific Wikipedia content across 13 Indian languages rather than relying exclusively on top-down translation from English. The resulting dataset contains millions of instruction-following examples designed to improve multilingual and culturally grounded capabilities.

This represents an important methodological shift.

Instead of asking how English-language AI can be exported to the rest of the world, developers can ask how AI should be built from knowledge originating within different linguistic communities.

What Organizations Should Test

A global AI deployment should evaluate performance by language, not simply by product.

Testing should cover comprehension, factual accuracy, instruction following, cultural appropriateness, specialized terminology, safety behaviour, and escalation.

Dialects and regional usage matter.

So do mixed-language conversations, which are common in multilingual societies.

Organizations also need to examine whether safety systems perform consistently across languages. A guardrail that works effectively in English but becomes weaker when a request is made in another language creates a serious control gap.

Performance metrics should therefore be disaggregated.

A single global accuracy score can hide substantial linguistic inequality.

The Next AI Divide Could Be Linguistic

Generative AI promises to make advanced knowledge available through natural conversation.

That promise becomes much more significant when people can use their own language rather than adapting themselves to the language of the technology.

Yet multilingual capability should not be confused with multilingual equality.

AI systems increasingly need to be evaluated as collections of linguistic capabilities rather than as one intelligence that happens to speak multiple languages.

The next frontier is therefore larger than translation.

It involves ensuring that the quality of intelligence, safety, cultural understanding, and access does not depend excessively on which language a person happens to speak.

For a technology positioned as global infrastructure, that is not a secondary technical issue.

It is part of what responsible AI must mean.

Sources and Further Reading

Scientific Reports, “Multilingual large language models do not comprehend all natural languages to equal degrees”, September 2026. DOI: 10.1038/s41598-026-69546-8.

ACL 2026, “Goldfish: Monolingual Language Models for 350 Languages”.

ACL 2026, “No One Fits All: From Fixed Prompting to Learned Routing in Multilingual LLMs”.

ACL 2026, “UPDESH: Synthesizing Grounded Instruction Tuning Data for 13 Indic Languages”.

Smaller May Be Smarter: Why the Future of AI Will Not Belong Only to Giant Models

Artificial intelligence has spent years moving in one direction: bigger.

More parameters. More training data. Larger computing clusters. Larger context windows. More expensive training runs.

That trajectory produced extraordinary capabilities. It also created an assumption that better AI inevitably requires a larger model.

A parallel movement is now challenging that assumption.

Small language models, commonly called SLMs, are becoming increasingly relevant for applications where cost, privacy, speed, specialization, and local deployment matter as much as raw general-purpose capability.

In July 2026, researchers writing in Nature Biomedical Engineering argued that small language models could provide a more scalable, practical, and efficient alternative to LLMs in medical applications. Research on edge computing is pointing in the same direction, particularly for environments with limited processing power, memory, connectivity, or strict privacy requirements.

This may signal a broader change in AI architecture.

The future may involve fewer situations where every task is sent to the largest available model and many more situations where intelligence is distributed across models of different sizes.

What Makes a Language Model “Small”?

There is no universal parameter threshold separating an SLM from an LLM.

The practical distinction is more useful.

Small language models are designed to deliver useful language and reasoning capabilities with substantially lower computational requirements. They can often be specialized for a particular domain or task and may be able to operate on local hardware rather than depending continuously on large cloud infrastructure.

This changes the economics of AI deployment.

A frontier model may know an extraordinary amount about thousands of domains. An organization processing one narrow category of repetitive technical requests may not need all of that capability.

The strategic question therefore changes from:

What is the most powerful model available?

to:

What is the smallest model that can perform this task reliably?

That is a very different procurement and engineering philosophy.

Privacy Changes the Equation

Local AI has one particularly important advantage.

Data may not need to leave the device or institutional environment.

This matters in healthcare, financial services, industrial operations, government, legal services, personal devices, and any environment involving confidential information.

A cloud-hosted AI system can be designed with strong privacy protections. Nevertheless, transmitting data to external computing infrastructure creates governance requirements that may disappear or become substantially simpler when inference occurs locally.

Research published in 2026 has examined privacy-preserving small language models for clinical assistance, reflecting growing interest in deploying task-specific language intelligence without relying entirely on large external models.

For a hospital, for example, a specialized local model that handles a defined documentation task reliably may be more valuable than a vastly more capable general-purpose model that introduces additional infrastructure, privacy, cost, or integration requirements.

Latency Matters Too

Cloud AI creates a physical dependency.

A request travels from the user to remote infrastructure, is processed, and returns.

For document drafting, that delay may be insignificant.

For real-time industrial equipment, mobile applications, assistive systems, vehicles, or devices operating in areas with limited connectivity, milliseconds and network availability matter.

Running intelligence locally reduces that dependency.

It also creates opportunities for AI systems that continue functioning when connectivity is limited or unavailable.

This is one reason edge AI is strategically important. Intelligence can move closer to where data is created and where decisions need to happen.

Small Models Also Change the Economics

Frontier intelligence can be expensive.

For occasional complex reasoning, that cost can be justified. For millions of routine transactions, model selection becomes an operational economics problem.

Imagine an organization handling ten million AI interactions per month.

Perhaps five percent require sophisticated reasoning.

The remaining 95 percent might involve classification, extraction, rewriting, structured responses, or narrow domain questions.

Routing every interaction to the most capable model may represent significant over-engineering.

A more efficient architecture can use smaller models for predictable work and escalate difficult cases to stronger systems.

In this environment, model routing becomes as important as model selection.

Specialization Can Compete With Scale

Small does not automatically mean weak.

A large general model needs broad competence. A smaller model can be trained, fine-tuned, or grounded around a narrower operational domain.

This creates an interesting performance dynamic.

For some specialized tasks, domain adaptation and good system design may matter more than raw parameter count.

A compact model supported by relevant retrieval, deterministic tools, validation rules, and carefully curated domain data can outperform a larger model operating with weak context.

This reinforces an important principle emerging across applied AI: the model is only one component of the system.

The Limitations Are Real

Small models should not be romanticized.

They generally have narrower knowledge, lower reasoning capacity, and less flexibility when confronted with unfamiliar tasks.

Compression can degrade performance.

Specialization can make a model brittle outside its intended domain.

On-device hardware imposes memory, battery, thermal, and processing constraints.

A model that works well on a workstation may behave very differently on a smartphone or embedded device.

These limitations mean the goal should not be to replace all LLMs with SLMs.

The more credible architecture is heterogeneous.

Different intelligence for different work.

AI Architecture Is Becoming a Portfolio Decision

Organizations spent the early generative AI period choosing platforms.

The next phase may require them to design model portfolios.

Some tasks will use frontier reasoning models.

Others will use fast general-purpose models.

Sensitive workloads may use private models.

High-volume repetitive tasks may use specialized SLMs.

Mobile and edge applications may use on-device models.

Deterministic systems will continue to handle tasks that do not need generative AI at all.

This is a more mature architecture because it aligns computational capability with actual operational requirements.

The future of AI therefore may not be a race toward one enormous model handling everything.

It may look more like an intelligent network, where different models operate at different layers and the system chooses how much intelligence each problem actually requires.

In that environment, the smartest AI strategy may sometimes begin with a counterintuitive question:

How small can we go without compromising the outcome?

Sources and Further Reading

Nature Biomedical Engineering, “Small language models in medicine”, July 2026. DOI: 10.1038/s41551-026-01734-3.

Journal of Imaging Informatics in Medicine, “Adaptive, Privacy-Preserving Small Language Models for Multi-Task Clinical Assistance”, March 2026. DOI: 10.1007/s10278-026-01912-4.

ICISCoIS 2026, “The Role of Small Language Models in Edge Computing and On-Device AI”.

When AI Learns the Physical World: Why World Models Could Be the Next Major Leap

For most of the generative AI era, artificial intelligence has operated inside a largely digital universe. It predicts the next word, generates an image from a description, writes software, summarizes documents, and reasons over information presented to it.

The physical world is different.

A robot moving through a warehouse must understand that an object continues to exist when it disappears behind a shelf. A vehicle must anticipate how pedestrians, bicycles, and other vehicles might move. A manufacturing system needs to understand how an action will alter the state of a machine. An intelligent assistant operating in a physical environment must recognize that actions have consequences.

This is why one of the most important emerging directions in AI is the development of world models.

Nature highlighted world models in 2026 as a rapidly growing frontier in artificial intelligence, particularly because models trained on physical environments could strengthen real-world applications such as robotics. At the same time, Nature Machine Intelligence has described a wider transition from embodied intelligence toward “physical AI”, systems capable of perceiving, acting, and adapting inside the real world rather than merely processing information about it.

What Is a World Model?

A world model attempts to build an internal representation of how an environment works.

Instead of generating the next word in a sentence, imagine an AI system predicting the next state of a physical scene.

If a cup is pushed toward the edge of a table, what is likely to happen?

If a robot rotates an object before gripping it, how will its position change?

If a vehicle approaches an intersection while another vehicle accelerates, what future situations are possible?

The model does not need to recreate every law of physics mathematically. It learns patterns about how environments evolve from large amounts of visual, video, sensor, simulation, or interaction data.

This allows the system to perform something increasingly important in AI: simulate before acting.

A robot could consider several possible actions internally, estimate their consequences, and choose a promising option before physically moving.

That is a major step beyond pattern recognition.

Why Generative AI Is Central to the Idea

World models are closely connected to generative AI.

Modern generative systems have become remarkably capable at creating images and videos that represent plausible scenes. The next challenge is to make those generated environments sufficiently coherent that they can be used for prediction, planning, and training.

Generating a visually convincing video is one problem.

Generating a world in which objects behave consistently, actions create predictable consequences, spatial relationships persist, and cause-and-effect relationships remain stable is substantially harder.

A system may generate a realistic-looking ball rolling across a floor while still failing to understand why the ball moved or what would happen if an obstacle were placed in its path.

For physical AI, visual plausibility is therefore insufficient. The model needs operational consistency.

Robotics Could Be the First Major Beneficiary

Traditional robotics requires enormous amounts of task-specific engineering.

Engineers define environments, create control policies, train robots on demonstrations, and anticipate possible failure conditions. The approach can work extremely well in controlled settings, such as industrial production lines.

General-purpose robots face a different challenge.

Homes, hospitals, construction sites, streets, farms, and warehouses contain situations that developers cannot enumerate in advance.

A world model could allow robots to learn more general representations of physical environments and rehearse actions through simulation.

Instead of experiencing every dangerous failure physically, the system could learn partly from simulated futures.

This creates the possibility of a new training architecture.

A robot observes the environment.

The world model predicts multiple possible future states.

A planning system evaluates them.

The robot selects an action.

The real outcome then provides additional evidence for improving future predictions.

The resulting cycle resembles the way intelligent organisms continuously learn from interaction with their surroundings.

The Same Principle Extends Beyond Robots

Physical AI may have applications across multiple sectors.

Manufacturing systems could simulate machine behaviour before changing operating parameters.

Autonomous vehicles could model alternative traffic developments.

Smart infrastructure could anticipate how physical systems respond to interventions.

Agricultural systems could combine environmental observations with action planning.

Scientific AI could use learned representations of physical systems to explore experimental possibilities.

Digital twins could become less static and more predictive, continuously generating alternative future states rather than merely representing the current state of an asset.

The underlying capability is similar across these applications: AI develops an internal environment in which possible futures can be explored computationally.

The Reality Gap Remains a Serious Problem

The opportunity is substantial, but so are the limitations.

A generated world can be internally convincing while still being physically wrong.

Small prediction errors can accumulate over long sequences. Rare events may be poorly represented. Models can learn correlations without understanding causal mechanisms. Safety-critical environments demand a level of reliability far beyond visually plausible simulation.

There is also a fundamental problem known broadly as the simulation-to-reality gap. A system may perform exceptionally well inside its learned or simulated environment and fail when reality contains conditions that the model did not represent correctly.

That makes validation essential.

World models cannot simply become imaginative engines for robots. They need calibrated uncertainty, physical constraints, continuous real-world feedback, and mechanisms that recognize when the model does not know enough to act safely.

From Language Intelligence to Situated Intelligence

Generative AI’s first major breakthrough was linguistic.

The next may be spatial and physical.

Language models learned patterns in enormous collections of human communication. World models attempt something more ambitious: learning regularities in environments where actions have consequences.

If this research trajectory succeeds, AI systems will increasingly move from describing the world to operating within it.

The competitive frontier will then shift again.

Model intelligence will no longer be judged only by how well an AI answers a difficult question. Increasingly, we will ask whether it can understand a situation, imagine possible futures, choose an appropriate action, and adapt when reality develops differently from its prediction.

That is the transition from generative AI as a creator of content to AI as an intelligent participant in the physical world.

Sources and Further Reading

Nature, “World models are AI’s latest sensation: what are they and what can they do?”, April 2026. DOI: 10.1038/d41586-026-00820-5.

Nature Machine Intelligence, “From embodied intelligence to physical AI”, April 2026.