Sam Altman Says AI Is Becoming a Science Problem

OpenAI CEO Sam Altman says the rapid development of artificial intelligence is creating a deeper scientific challenge: understanding increasingly capable AI systems and ensuring they remain predictable, controllable and aligned with human goals.

Contents

Why Sam Altman says AI safety goes beyond infrastructure The Nvidia security question Why AI alignment is becoming more important AI’s growing capabilities are changing the safety debate OpenAI expects more AI models What the debate means for the future of AI

Artificial intelligence is advancing at a pace that is forcing researchers to confront questions that go beyond conventional cybersecurity and engineering. OpenAI CEO Sam Altman has described the challenge as increasingly scientific, arguing that protecting the infrastructure used to build and operate AI systems is not enough to address the risks associated with increasingly capable models.

Altman’s comments come as technology companies continue to invest heavily in more powerful AI systems and researchers work to understand how those systems behave as their capabilities expand.

The discussion has also highlighted different views within the technology industry about how AI safety should be approached.

- Advertisement -

Why Sam Altman says AI safety goes beyond infrastructure

Modern AI systems depend on an extensive technological infrastructure that includes advanced processors, data centres, networking systems and software. Protecting that infrastructure from cyberattacks and other external threats remains an important part of AI security.

But Altman’s argument is that infrastructure security represents only one layer of the overall problem.

A secure data centre or protected AI accelerator does not automatically guarantee that the model operating inside that infrastructure will behave as intended. As models become more capable, researchers need to understand their decision-making processes, limitations and potential failure modes.

That makes AI alignment—the effort to ensure AI systems reliably pursue goals consistent with human intentions—an increasingly important area of research.

Altman has argued that security associated with Nvidia’s infrastructure is therefore not a complete answer to the broader safety challenge. The distinction is significant: infrastructure security is primarily concerned with protecting systems from external threats, while AI alignment and control research focuses on the behaviour and capabilities of the models themselves.

- Advertisement -

The Nvidia security question

Nvidia has become a central company in the AI ecosystem because its processors and related computing infrastructure are widely used to train and operate advanced AI models.

Security surrounding AI hardware and infrastructure can help protect computing environments, data and systems. However, the wider AI-safety debate involves another question: what happens when increasingly capable models are given greater autonomy or access to tools and external systems?

That distinction is at the heart of the disagreement in emphasis between Altman and Nvidia CEO Jensen Huang.

- Advertisement -

Huang has discussed AI safety largely through the lens of engineering, testing, safeguards, evaluation and the application of existing rules and regulations. Altman’s position puts comparatively greater emphasis on developing the scientific understanding needed to control and align increasingly advanced models.

The two approaches are not necessarily mutually exclusive. Infrastructure protection, model evaluations, cybersecurity, regulation and alignment research can all address different parts of the overall AI-safety problem.

Why AI alignment is becoming more important

As AI systems move beyond generating text and images and increasingly interact with software, databases and external tools, questions about control become more consequential.

An advanced AI model can potentially perform multiple steps to accomplish a task rather than simply responding to a single user instruction. That creates additional opportunities for mistakes, unexpected behaviour or misuse.

Researchers therefore study areas including:

AI alignment: ensuring model behaviour remains consistent with intended objectives.
Model evaluations: testing capabilities, limitations and potentially dangerous behaviours.
AI control: developing methods for maintaining meaningful human oversight.
Cybersecurity: protecting models, infrastructure and sensitive information from attacks.
Interpretability: improving understanding of how AI systems arrive at their outputs.
Agent safety: addressing risks associated with AI systems that can independently perform multi-step tasks.
Governance: establishing rules and accountability mechanisms around increasingly capable AI.

These areas overlap, but they address different layers of the AI-safety challenge.

AI’s growing capabilities are changing the safety debate

The AI industry is moving rapidly from systems that primarily generate responses toward models that can reason across complex tasks, use tools and operate with greater degrees of autonomy.

That evolution has changed the nature of the safety conversation.

Earlier discussions often focused on familiar problems such as inaccurate information, biased outputs, privacy and cybersecurity. Those issues remain important, but researchers are increasingly examining what happens when models become capable of completing longer and more complicated sequences of actions.

The central challenge is not simply whether an AI system can perform a task. It is whether developers can reliably predict how the system will behave across unfamiliar situations and retain effective human oversight as capabilities increase.

This is why Altman’s description of AI as a “science problem” matters to the wider debate. It points to a gap between building systems that work and fully understanding systems that are becoming increasingly complex.

OpenAI expects more AI models

Altman’s comments also come as OpenAI continues to develop new generations of AI models.

The company has indicated that it expects to introduce further models while continuing to work on safety measures alongside advances in capability. That creates a difficult balancing act for AI developers: increasing performance while ensuring that safeguards, evaluations and oversight develop at a comparable pace.

For the broader industry, the issue extends beyond any single company.

Google, Microsoft, Anthropic, Meta, OpenAI, Nvidia and other technology companies are participating in an increasingly competitive AI ecosystem. Governments and researchers are simultaneously working on questions involving AI regulation, cybersecurity, autonomous systems and accountability.

What the debate means for the future of AI

The disagreement over how to approach AI safety reflects a larger question facing the technology industry.

One approach focuses heavily on engineering controls: testing systems, establishing safeguards, securing infrastructure and applying laws and regulations. Another emphasizes the need for deeper scientific research into how increasingly capable AI systems behave and how they can be reliably aligned with human objectives.

In practice, advanced AI safety is likely to involve several of these layers at once.

Securing the hardware and infrastructure is important. Testing models is important. Establishing appropriate governance is important. And understanding the behaviour of increasingly capable AI systems is an ongoing research challenge.

As AI capabilities continue to develop, the industry’s ability to combine these approaches may become increasingly important to determining how safely the technology can be deployed.

For now, Altman’s warning underscores a fundamental point in the AI debate: protecting the machines that run advanced models is different from understanding and controlling the behaviour of the models themselves.

ALSO READ: