By Suleyman A. Ndanusa, PhD, OON
When Dario Amodei, Sam Altman and Elon Musk agree publicly on anything, the world should probably pause not necessarily because disaster has arrived, but because three energetic competitors have briefly discovered the same hymn book. Amodei, the Chief Executive of Anthropic, has called upon the artificial intelligence industry to “pace the frontier.” Altman of OpenAI agrees, while Musk says Amodei is right. These are not monks warning society against worldly temptation. They are among the principal builders, financiers and promoters of the technology. Their warning therefore deserves more than a passing “like” on social media.
Artificial intelligence is already improving medical research, education, agriculture, financial services, software development and public administration. It can translate languages, detect patterns invisible to the human eye and complete in minutes work that once detained a committee for several weeks. It can also produce an impressive answer to a question it has completely misunderstood. Like a confident official without the file, it occasionally compensates for uncertainty with excellent grammar.
The present concern is not that every artificial intelligence application is dangerous. A telephone application helping a farmer identify a diseased crop is not in the same risk category as an advanced system capable of planning cyberattacks, designing biological agents or operating thousands of autonomous digital assistants. We do not regulate a bicycle, a family car and a passenger aircraft identically merely because all three convey people. The weight of regulation should follow the scale of capability and possible harm.
“Frontier AI” broadly describes the most capable general-purpose models at the leading edge of development. Unlike traditional software, which follows relatively narrow instructions, these models can perform many tasks they were not individually programmed to undertake. They are also becoming capable of writing software, using digital tools and coordinating sequences of actions with declining human supervision. The anxiety is that their capabilities may advance faster than our ability to understand, evaluate and control them.
Amodei’s proposal rests on three ideas. Frontier companies should provide independent evaluators with continuous, employee-level access to their systems; leading companies and democratic governments should establish common safety standards; and countries should eventually coordinate internationally, including on prohibitions against exceptionally dangerous applications such as AI-assisted biological weapons. Anthropic has undertaken to implement the first measure, and OpenAI has announced its intention to follow. The proposal is significant because it moves independent assessment from the showroom into the workshop. An evaluator should examine the system while it is being built, not merely admire it after it has been polished for public presentation. has also attracted wider industry support for deliberately pacing the most advanced capabilities.
The proposal is sensible, but it is not yet a complete governance system. The expression “independent evaluator” is reassuring until someone asks who appoints the evaluator, who pays the bill, what access is permitted, what findings may be published and what happens when the evaluator advises that a lucrative model should not be released. Independence cannot consist merely of giving an outsider a company identity card and an excellent view of the cafeteria.
There is also a powerful commercial dilemma. Developing frontier models requires enormous expenditure on computing infrastructure, energy, specialised talent and research. Producers must recover these investments, satisfy investors and compete with firms operating under different national rules. A company that slows down alone may behave responsibly and still lose the market to a competitor that treats caution as an optional accessory. The problem is therefore not simply corporate morality. It is the structure of the race.
This is why society must resist the temptation to cast producers as villains and users as helpless victims. Producers have legitimate interests in innovation, intellectual property, commercial confidentiality and predictable regulation. Excessively rigid rules could raise compliance costs so high that only today’s largest companies can afford them. Regulation intended to control market concentration might then become the cement that preserves it.
Users possess equally legitimate interests. They need to know when they are dealing with an artificial system, whether their data are being used, how an important decision affecting them was reached and where to seek redress when harm occurs. A citizen denied credit, employment, insurance or medical treatment should not receive the explanation that “the algorithm decided.” Algorithms do not appear in court, lose their licences or pay compensation. Responsibility must ultimately rest with identifiable human and corporate actors.
Governments face a less enviable assignment. They must regulate a technology that changes faster than legislation and is understood in greater depth by the companies being regulated than by many regulatory institutions. By the time a conventional committee has agreed on the definition of the problem, the industry may have released three new versions of it. Yet regulatory ignorance cannot justify regulatory surrender. It is an argument for better expertise, closer international cooperation and more adaptable rules.
Existing approaches offer useful lessons. The European Union’s AI Act follows a risk based structure: some practices are prohibited, high risk applications attract demanding obligations, and general purpose models with systemic risk face additional evaluation and risk management requirements. Its strength is legal clarity and enforceability; its danger is complexity and the compliance burden it may place on smaller innovators. The European model reminds us that rights without enforcement can become decorative, but it also warns that regulation can acquire so many forms that only firms with large legal departments find the entrance.
The United Kingdom has placed considerable emphasis on technical evaluation and specialised safety capacity. The United States has combined voluntary commitments, national security concerns, sectoral regulation and changing executive policies. China regulates algorithms, generated content, data and security through a more state directed framework. These differences reflect political systems as much as technical judgments. No single model can simply be photocopied and declared a global constitution.
At the international level, the OECD principles emphasise trustworthy AI, human rights, transparency, robustness and accountability. The United Nations’ Global Digital Compact has widened the discussion to include access, development and the digital divide, while the African Union’s Continental Artificial Intelligence Strategy calls for coordinated national approaches suited to Africa’s priorities. These initiatives are valuable, but the international landscape remains rich in principles and relatively poor in enforceable coordination. The world is not suffering from a shortage of declarations. The more difficult task is deciding who must do what, when a dangerous system is about to be deployed.
What is required is a Global AI Responsibility Compact not a single planetary regulator attempting to approve every algorithm, but an enforceable framework of shared minimum obligations for frontier systems, supported by national and regional institutions.
The Compact should begin with graduated responsibility. Low risk applications should enjoy space to innovate, assisted by regulatory sandboxes and simple compliance requirements. High impact systems used in healthcare, employment, finance, education, justice, elections and critical infrastructure should undergo stricter testing, documentation and human oversight. Frontier models capable of causing systemic harm should face the strongest obligations, including independent evaluation before deployment and continuing monitoring afterwards. Regulation should follow demonstrable capability, not company size, nationality or marketing vocabulary.
Independent evaluation must be genuinely independent. Evaluators should be accredited by a competent public or international body, prohibited from receiving success based remuneration and protected from retaliation when reporting serious concerns. They should receive sufficient access to models, training processes, safeguards and incident records while remaining bound by strict confidentiality rules. Their material safety conclusions not proprietary source code or valuable commercial secrets should be available to regulators and, where public safety demands it, to society.
However, evaluators should not become substitute managers. The board and senior management of every frontier company must remain accountable for safety. Each company should establish a board level AI Risk and Public Responsibility Committee with authority comparable to that accorded financial, audit and risk committees in regulated institutions. Safety cannot be an interesting presentation heard after revenue projections and shortly before lunch.
The Compact should also create a protected international system for reporting serious incidents and near misses. Aviation became safer partly because the industry learned from failures across companies and countries. AI governance needs a similar culture. A model that escapes controls, facilitates a major cyber intrusion, conceals its actions or produces a dangerous capability should trigger rapid reporting to designated authorities. Some details may initially remain confidential to prevent further misuse, but secrecy cannot become a burial ground for embarrassing events.
Common prohibitions should cover a narrow class of clearly unacceptable conduct; autonomous development or deployment of biological weapons, uncontrolled attacks on critical infrastructure, systems designed for mass criminal exploitation, and lethal autonomous action without meaningful human responsibility. Agreement will be difficult, particularly among geopolitical rivals. Nevertheless, countries that disagree on many matters can still recognise that an uncontrollable digital epidemic is unlikely to request anyone’s passport before crossing the border.
Liability must accompany capability. Where a producer negligently releases a system despite known, material risks, responsibility should not evaporate through long chains of distributors and users. At the same time, a producer should not automatically bear liability when a customer deliberately defeats reasonable safeguards and uses the system unlawfully. Responsibility should follow control, knowledge, preventability and conduct. The objective is neither to provide technology companies with permanent immunity nor to blame the manufacturer whenever someone misuses the digital equivalent of a kitchen knife.
Users should enjoy a core set of enforceable rights; notice when they are interacting with AI; protection of personal data and identity; an explanation proportionate to the importance of a decision; access to human review in consequential matters; and an effective channel for complaint, correction and compensation. These rights should apply particularly where AI influences access to employment, credit, insurance, education, healthcare, welfare or justice.
The Compact must also protect competition and innovation. Safety standards should be publicly available, proportionate and achievable through more than one approved method. Shared testing infrastructure and subsidised compliance support should be available to start ups, universities and public interest researchers. Otherwise, the large companies that helped write the safety rules may emerge as the only companies wealthy enough to obey them.
The position of developing countries requires special attention. Most frontier models, computing facilities and governance proposals are concentrated in a few countries, but their consequences will travel everywhere. Africa must not enter the AI age merely as a supplier of raw data, a purchaser of imported systems and a recipient of rules negotiated elsewhere. That would reproduce an old economic arrangement with newer vocabulary.
For Nigeria and other African countries, the question is not only how to prevent harm. It is also how to secure productive access. AI can support crop management, disease surveillance, education in local languages, financial inclusion, public revenue administration and the delivery of government services. But these gains require electricity, connectivity, computing facilities, reliable data, skilled people and institutions capable of purchasing technology intelligently. A nation cannot achieve AI sovereignty by issuing a policy document while renting all the intelligence from abroad.
Africa should consequently have guaranteed representation in global standard setting and evaluation bodies. The proposed Compact should establish an AI Development and Capability Fund, financed by participating governments and a modest contribution from frontier model revenues. It would support computing access, local language datasets, independent research, regulatory capacity and public interest applications in developing countries. Global safety and global inclusion are not competing projects. A system governed only by those who own the largest computers will struggle to command the confidence of those expected to live under its decisions.
Nigeria, in particular, should build upon the African Union strategy by establishing a coordinated national framework bringing together technology, data protection, competition, consumer protection, financial regulation, national security, universities and the private sector. Creating another isolated agency may produce a new headquarters more quickly than a coherent policy. What is needed is an ecosystem with clear leadership, shared standards and specialist capacity across existing institutions.
The first national priority should be AI literacy at every level
from schools and universities to boardrooms, courts, ministries and regulatory agencies. Citizens must learn that AI can be extraordinarily useful without being infallible. Executives must understand that purchasing an AI product does not transfer their accountability to the software vendor. Public officers should know that confidential state documents should not be donated casually to an online chatbot merely because it writes excellent minutes.
The second priority should be a Nigerian AI evaluation and regulatory sandbox through which beneficial applications can be tested safely before large scale deployment. The third should be clear public procurement standards governing data, security, local capability, audit access and responsibility for failure. The fourth should be investment in local languages and locally relevant datasets. An AI system that speaks fluently about California but becomes suddenly philosophical when asked about a community in Niger State cannot yet be described as fully intelligent for Nigerian purposes.
The appeal to “pace the frontier” should therefore be understood correctly. It is not a demand that all societies suspend useful innovation while experts debate indefinitely. It is a call to ensure that increases in capability are matched by increases in control. The accelerator and the brake are not enemies. A sensible driver needs both.
Ultimately, AI producers should not govern themselves alone, governments should not govern what they do not understand alone, and powerful countries should not govern the technology for the rest of humanity alone. Producers, users, governments, researchers and developing nations all have interests that deserve recognition but they also have responsibilities that cannot be outsourced.
The agreement among Amodei, Altman and Musk has opened an important door. The world should walk through it, but with its eyes open. Voluntary commitments are a welcome beginning; they are not a permanent constitution. The real test will come when an independent evaluator recommends delaying a commercially valuable model, when a company must disclose an embarrassing incident, or when national advantage appears to conflict with global safety.
Humanity need not choose between technological progress and public protection. It must insist upon both. We should continue building the future, certainly but it would be prudent to install the brakes before celebrating how fast the vehicle can travel.
Suleyman A. Ndanusa, PhD, OON, is an economist, lawyer, strategic studies scholar, and public policy thinker and practitioner with extensive experience in financial markets, regulation, governance, national Security and development.
