2022–23: Generative AI and the New Governance Problem

2022–23: Generative AI and the New Governance Problem

For much of the 2010s, AI governance was largely concerned with systems designed for relatively specific purposes.

A facial-recognition system.

A recommendation engine.

A fraud-detection model.

An automated hiring tool.

A predictive system used in healthcare.

These systems could certainly create serious governance challenges.

But policymakers could often identify the application, the sector and the organisation responsible for deploying it.

Generative AI changed this structure.

Beginning in 2022, increasingly capable generative AI systems became publicly accessible at an unprecedented scale.

Large language models could write, summarise, translate, explain, code and interact conversationally.

Image-generation systems could create visual content from natural-language prompts.

Other systems could generate audio, video and increasingly multimodal content.

The technology was not simply another AI application.

It represented a different governance problem.

A single underlying model could be used for thousands of different purposes.

The same system could be helpful in one context, unreliable in another and potentially harmful in a third.

This created a fundamental challenge for AI governance.

How do you govern a technology that is general-purpose, rapidly evolving and capable of being integrated into applications that its original developer may never have anticipated?

THE BREAKTHROUGH OF LARGE LANGUAGE MODELS

The technical foundations of generative AI had been developing for years.

Transformer architectures, large-scale datasets, increasingly powerful computing infrastructure and improvements in model training had gradually enabled increasingly capable language models.

By the early 2020s, these developments were producing systems with strikingly broad capabilities.

Large language models could generate coherent text across many subjects.

They could translate between languages.

They could write computer code.

They could answer questions.

They could summarise documents.

They could assist with research and writing.

The significance was not simply that the models were better than previous systems.

It was that one model could perform many different tasks.

This generality became central to the governance challenge.

THE ARRIVAL OF CHATGPT

In November 2022, OpenAI released ChatGPT to the public.

Its rapid adoption demonstrated that generative AI was no longer confined to specialist research communities.

People could interact directly with a highly capable language model through a simple conversational interface.

The governance implications became immediately visible.

AI was entering education.

It was entering workplaces.

It was entering software development.

It was entering media and content production.

It was entering customer service.

It was entering research.

It was entering everyday personal use.

This created a major shift in the social visibility of AI.

Earlier AI systems often operated in the background.

Generative AI interacted directly with people.

The public could see what these systems could do.

And they could also see where they failed.

THE PROBLEM OF GENERAL-PURPOSE AI

The concept of general-purpose AI became increasingly important during this period.

A general-purpose model is not necessarily designed for one narrow task.

It can support a broad range of downstream applications.

This creates a different distribution of responsibility.

Imagine a foundation model developed by one company.

Another organisation integrates it into an educational platform.

A third organisation uses that platform in schools.

A fourth party develops a specialised application on top of the model.

If something goes wrong, who is responsible?

The model developer?

The application developer?

The deploying organisation?

The end user?

Potentially, responsibility may be distributed across all of them.

This challenged governance approaches that focused primarily on the final application.

THE FOUNDATION MODEL PROBLEM

Generative AI also introduced a new layer of governance around foundation models.

A foundation model can serve as the technological base for many downstream systems.

This means that decisions made during model development can influence thousands of later applications.

Training data choices can affect downstream behaviour.

Model architecture can affect capabilities.

Safety testing can influence how risks are understood.

Release decisions can determine how widely capabilities become available.

Documentation can affect whether downstream developers understand limitations.

This means governance may need to begin before a model reaches the final application.

The lifecycle becomes:

Development.

Training.

Evaluation.

Release.

Integration.

Deployment.

Monitoring.

Updating.

Incident response.

This lifecycle perspective became increasingly important in AI governance.

HALLUCINATIONS AND RELIABILITY

One of the most visible problems with generative AI was the production of incorrect information presented in convincing language.

These outputs became widely described as “hallucinations.”

The governance significance of the problem is important.

A traditional software system may fail by producing an obvious error.

Generative AI can fail differently.

It can produce an answer that appears authoritative, coherent and plausible even when it is incorrect.

This creates a new form of reliability risk.

The problem becomes particularly serious when AI is used in high-stakes environments.

Healthcare.

Law.

Finance.

Education.

Scientific research.

Public administration.

In such settings, users may assume that fluent output reflects accurate information.

It does not necessarily do so.

Generative AI therefore demonstrated that language fluency and factual reliability are different properties.

This became a major issue for AI evaluation.

FROM ACCURACY TO EVALUATION

The rise of generative AI changed the question of how AI systems should be evaluated.

For a specialised model, evaluation may focus on a relatively defined task.

For a general-purpose model, the evaluation problem is much broader.

A model may perform well on one benchmark and poorly in another environment.

It may produce harmful content under certain prompts.

It may reveal sensitive information.

It may be vulnerable to adversarial manipulation.

It may generate incorrect information.

It may behave differently across languages and cultural contexts.

It may perform differently after updates.

This created demand for more comprehensive evaluation systems.

The governance question therefore became:

What should be measured before and after a general-purpose AI system is deployed?

This question would become central to frontier AI governance.

THE COPYRIGHT QUESTION

Generative AI also transformed the copyright debate.

Large models are trained using enormous quantities of information.

This raised questions about how copyrighted works can be used in AI training and what rights creators should have.

The governance questions became complex.

Can copyrighted material be used to train models?

Under what conditions?

Does training constitute a legally relevant form of copying?

Should creators have the ability to opt out?

Should AI-generated content be labelled?

What obligations should model developers have regarding training data?

Different jurisdictions began approaching these questions differently.

The result was another major governance challenge.

AI regulation was increasingly interacting with existing intellectual property law.

This demonstrated an important feature of AI governance.

New AI technologies rarely operate inside a completely new legal environment.

They interact with laws that already govern data, copyright, privacy, competition, consumer protection and employment.

THE PRIVACY PROBLEM

Generative AI also created new privacy concerns.

Large language models can process enormous quantities of information.

Users may enter sensitive information into AI systems.

Organisations may use AI to process internal documents.

Models may sometimes reproduce information that appeared in training or other data sources.

This creates questions about data protection and responsible information handling.

The privacy question therefore evolved.

Earlier AI governance had focused heavily on whether data could be collected and processed.

Generative AI introduced another concern:

How should highly capable models interact with sensitive information throughout their lifecycle?

This includes training, fine-tuning, user interaction, storage, retrieval and deployment.

THE MISINFORMATION PROBLEM

Generative AI also changed the information environment.

AI systems could produce large amounts of text, images, audio and video at relatively low cost.

This created legitimate concerns about synthetic media and misinformation.

The problem was not simply that false content could exist.

False information has always existed.

The change was the scale and speed at which synthetic content could potentially be produced.

A single individual could generate large quantities of plausible text or visual material.

Organisations could automate content production.

Synthetic media could potentially make it more difficult to determine whether particular content was authentic.

This created new governance questions around provenance, disclosure, platform responsibility and public resilience.

THE EDUCATION QUESTION

Education became one of the first sectors where generative AI’s governance implications became highly visible.

Students could use AI systems to generate essays.

Teachers could use AI for lesson planning.

Researchers could use AI for summarisation and coding.

Institutions had to rethink assessment.

The central question was not simply whether students should be allowed to use AI.

The deeper issue was whether educational systems were designed for a world in which AI could perform many forms of written and analytical work.

This challenged traditional assumptions about assessment.

If an AI system can generate a polished essay in seconds, what exactly should an educational institution be measuring?

Writing ability?

Critical thinking?

Research skills?

Originality?

Problem-solving?

The answer affects how education systems prepare people for an AI-enabled economy.

THE LABOUR MARKET QUESTION

Generative AI also changed discussions about automation.

Earlier AI debates often focused on automating routine or highly structured tasks.

Generative AI demonstrated that increasingly sophisticated cognitive tasks could also be assisted or partially automated.

Writing.

Coding.

Translation.

Customer support.

Research assistance.

Design.

Marketing.

Analysis.

This did not mean that all these occupations would simply disappear.

The impact of AI on work is more complicated.

Some tasks may be automated.

Others may be augmented.

New tasks may emerge.

Existing jobs may change.

Productivity may increase.

Organisational structures may evolve.

The governance challenge is therefore not simply job replacement.

It is managing the transition.

This brings AI governance into the domain of labour policy, skills development and education.

THE SAFETY QUESTION BECOMES BIGGER

Generative AI also changed the conversation around AI safety.

Earlier discussions often focused on immediate harms such as bias, privacy and discrimination.

These concerns remained important.

But increasingly capable models created additional questions.

Could advanced AI systems be misused for cyber operations?

Could they assist harmful biological research?

Could autonomous systems take actions with limited human supervision?

Could increasingly capable models develop unexpected behaviours?

Could highly capable AI systems create systemic risks across multiple sectors?

These questions helped bring frontier AI safety into mainstream policy discussions.

The governance agenda was expanding again.

FROM RESPONSIBLE AI TO FRONTIER AI

This produced an important distinction.

Responsible AI traditionally focused on ensuring that AI systems were fair, transparent, safe and accountable in practical deployment.

Frontier AI governance increasingly focused on the risks associated with highly capable models at the edge of technological capability.

The two agendas overlap.

But they are not identical.

A governance framework for an automated hiring system may focus on discrimination, explainability and individual rights.

A governance framework for a highly capable foundation model may also need to consider dangerous capabilities, model evaluations, deployment thresholds, security, misuse and systemic risks.

This created a new policy frontier.

THE 2023 GOVERNANCE RESPONSE

By 2023, governments and international organisations were responding rapidly.

The European Union was moving toward the AI Act.

The G7 launched the Hiroshima AI Process.

The United Kingdom hosted the AI Safety Summit at Bletchley Park.

The United States issued new policy measures and executive action around AI.

China continued developing its own regulatory framework for generative AI and algorithmic systems.

International organisations expanded their work on AI governance.

This was a significant change.

AI governance was no longer a relatively specialised policy field.

It had become a major international policy agenda.

THE BLETCHELY MOMENT

The AI Safety Summit held at Bletchley Park in November 2023 represented a particularly visible moment in this transition.

Governments and other stakeholders came together to discuss the risks associated with highly capable AI.

The Bletchley Declaration recognised the need for international cooperation around AI safety.

The importance of this moment was not that governments suddenly agreed on every aspect of AI governance.

They did not.

The significance was that advanced AI safety had become a subject of international diplomacy.

AI governance was moving beyond national regulation.

It was becoming geopolitical.

THE G7 HIROSHIMA AI PROCESS

The G7’s Hiroshima AI Process represented another major development in 2023.

The process focused on advanced AI systems and international principles and a code of conduct for organisations developing advanced AI systems.

This reflected a growing recognition that frontier AI could not be governed only through domestic regulation.

Companies developing highly capable models often operate internationally.

Their systems are deployed across borders.

Their infrastructure and supply chains are global.

Their impacts can cross jurisdictions.

International coordination therefore becomes necessary.

THE EUROPEAN UNION’S RESPONSE

The European Union was also developing a major regulatory response.

The EU AI Act would eventually create a comprehensive risk-based legal framework for AI.

Generative AI created additional complexity for the regulatory design because general-purpose AI systems do not fit neatly into traditional product categories.

The emergence of general-purpose AI therefore influenced the evolution of European AI regulation.

The policy question became not only:

“What risk does this application create?”

but also:

“What responsibilities should apply to the underlying general-purpose model?”

This was a significant conceptual change.

THE UNITED STATES AND A DIFFERENT APPROACH

The United States also intensified its AI governance activity during 2023.

Rather than relying on a single comprehensive AI law at that stage, the US approach involved a combination of executive action, existing sectoral regulation, standards work, agency activity and state-level initiatives.

The important point for this series is not to label one approach as universally superior.

The important point is that different governance models were emerging.

The EU was moving toward comprehensive horizontal regulation.

The United States was combining existing regulatory institutions with new executive and standards-based approaches.

China was developing a state-led framework involving regulation, standards and industrial strategy.

Other countries were developing their own combinations.

The global AI governance landscape was therefore becoming more diverse.

THE GOVERNANCE GAP

Generative AI exposed a major weakness in earlier governance approaches.

Principles could identify desirable values.

But principles alone could not answer operational questions.

How should a model be evaluated?

Who should conduct the evaluation?

What information should companies disclose?

When should a model be considered too risky to deploy?

Who is responsible for downstream harms?

How should incidents be reported?

What happens when a model is updated?

How should governments cooperate across borders?

These were not purely philosophical questions.

They were institutional questions.

And institutional questions require concrete governance mechanisms.

THE RISE OF AI EVALUATION

One of the most important consequences of the generative AI era was therefore the growing importance of evaluation.

If AI systems are increasingly powerful, governments and organisations need ways to measure their capabilities and risks.

This creates a new governance infrastructure.

Benchmarks.

Red-team testing.

Safety evaluations.

Robustness testing.

Bias assessments.

Cybersecurity testing.

Capability evaluations.

Monitoring.

Incident reporting.

Independent assessment.

The governance question becomes increasingly empirical.

Instead of simply asking whether an AI system is trustworthy, institutions increasingly need evidence demonstrating how trustworthy it is under defined conditions.

This is a major shift.

THE EMERGENCE OF AI AUDITING

Generative AI also increased interest in independent auditing.

If companies evaluate their own systems, questions may arise about incentives and credibility.

Independent evaluation can provide additional assurance.

But independent auditing creates its own challenges.

Who qualifies as an auditor?

What standards should auditors use?

How much technical access should they receive?

How should confidential information be protected?

Who pays for the evaluation?

What happens when the evaluator identifies a serious problem?

These questions would become central to the next phase of AI governance.

GOVERNANCE BECOMES A TECHNICAL SYSTEM

This is one of the most important lessons of 2022–23.

AI governance was increasingly becoming technical.

Effective governance could not depend only on legislation and policy statements.

It also required technical expertise.

Governments needed people who understood model evaluations.

Regulators needed technical capacity.

Universities needed AI safety and governance research.

Standards organisations needed technical methodologies.

Civil society organisations needed the ability to independently assess AI claims.

This meant that the boundary between technical AI research and public policy was becoming increasingly blurred.

THE GLOBAL SOUTH CHALLENGE

Generative AI also created new challenges for developing countries.

The technology became globally accessible, but the underlying infrastructure remained highly concentrated.

Many countries could use advanced AI systems without having the capacity to develop them.

This creates both opportunities and dependencies.

Generative AI can provide access to educational resources, translation, software assistance and productivity tools.

But countries may become dependent on foreign providers for critical AI infrastructure.

This raises questions about data sovereignty, digital dependency, local-language representation and long-term technological capacity.

For the Global South, the governance question is therefore not simply:

“How should we regulate generative AI?”

It is also:

“How can we ensure that our societies can meaningfully participate in the AI economy?”

THE BANGLADESH QUESTION

For Bangladesh, the generative AI transition makes institutional readiness increasingly important.

The country does not need to reproduce the exact governance architecture of the EU, United States or China.

Its circumstances are different.

But Bangladesh does need the capacity to understand what these systems can do, where they are being used, what risks they create and which institutions should be responsible for addressing those risks.

That requires more than a national AI policy document.

It requires research capacity.

University readiness.

Public-sector expertise.

Technical standards.

Data governance.

AI literacy.

Evaluation capability.

Industry engagement.

And mechanisms for evidence-based policymaking.

Generative AI therefore makes university and institutional readiness part of national AI governance.

THE BIG GOVERNANCE TRANSITION

The period from 2022 to 2023 can therefore be understood as another major transition in the history of AI governance.

Before generative AI, the central governance question was often:

“How should AI applications be made responsible?”

After generative AI, another question became unavoidable:

“How should increasingly general-purpose AI models themselves be governed?”

This is a much larger problem.

It involves model developers.

Infrastructure providers.

Application developers.

Deployers.

Regulators.

Auditors.

Standards organisations.

International institutions.

Users.

The governance ecosystem has expanded.

AI governance is becoming a system of systems.

THE END OF THE PRINCIPLES-ONLY ERA

The earlier principles remain important.

Human rights still matter.

Fairness still matters.

Transparency still matters.

Safety still matters.

Accountability still matters.

But the generative AI era demonstrated that principles must be operationalised.

A principle without an evaluation mechanism can remain abstract.

A safety commitment without testing can be difficult to verify.

A transparency requirement without documentation standards can be difficult to enforce.

An accountability principle without clearly assigned responsibility can become ambiguous.

This is why the next phase of AI governance is increasingly about implementation.

FROM PRINCIPLES TO POWER

The evolution can now be seen more clearly.

The first phase asked why AI should be governed.

The second developed principles for responsible AI.

The third began translating those principles into governance frameworks.

Generative AI accelerated the fourth phase.

Governance now increasingly concerns who has the power, capability and institutional authority to evaluate, regulate and control increasingly capable AI systems.

This is where AI governance begins to intersect more directly with geopolitics.

WHO CONTROLS FRONTIER AI?

By 2023, a relatively small number of companies possessed the resources necessary to train and deploy the most capable large-scale models.

That concentration created a new governance question.

Can governments effectively govern technologies whose development depends on highly concentrated private-sector expertise and infrastructure?

This is not simply a regulatory problem.

It is also an institutional power problem.

Governments need technical expertise.

Companies possess technical expertise.

Universities contribute independent research.

Standards organisations develop methodologies.

International organisations provide coordination.

Civil society provides external scrutiny.

The future of AI governance will depend partly on how these different sources of power interact.

CONCLUSION

The emergence of generative AI in 2022–23 changed the global AI governance landscape.

The technology demonstrated that AI systems could be general-purpose, widely accessible and rapidly integrated into many sectors.

That created new governance challenges involving reliability, misinformation, copyright, privacy, labour, education, security and safety.

More importantly, generative AI exposed the limitations of governance frameworks designed primarily around individual AI applications.

The governance object was changing.

It was no longer only the application.

Increasingly, it was the underlying model.

This led to a new emphasis on foundation models, general-purpose AI, model evaluation, safety testing, transparency, accountability and lifecycle governance.

It also pushed AI safety into international diplomacy.

The OECD principles and UNESCO Recommendation remained important.

But the governance conversation was becoming more operational.

The question was no longer simply whether AI should be trustworthy.

The question was increasingly:

How can trustworthiness be measured?

Who verifies it?

Who is responsible?

Who has the technical capacity to evaluate it?

Who has the authority to intervene?

And ultimately:

Who governs increasingly powerful AI?

That question takes the global AI governance story into its next phase.

THE NEXT PHASE:

The next article examines the 2023 Hiroshima AI Process and the G7’s attempt to develop international governance mechanisms for advanced AI.

It will explore how AI governance moved from global principles toward international coordination, codes of conduct and the emerging diplomacy of frontier AI safety.

ATLAS AI GOVERNANCE SERIES:

“From Principles to Power: The Evolution of AI Governance”

Part 1 — What Is AI Governance—and Why Has It Become So Important?

Part 2 — Before AI Governance: What Did We Learn from Data Protection, Cybersecurity and Internet Governance?

Part 3 — Why Did AI Governance Become a Global Policy Issue in the 2010s?

Part 4 — 2019: OECD AI Principles and the Start of a New Era in AI Governance

Part 5 — 2021: UNESCO Recommendation and the Global Ethics Dimension of AI Governance

Part 6 — 2022–23: Generative AI and the New Governance Problem

Part 7 — 2023: Hiroshima AI Process and the G7