AI Consciousness and Autonomy: What Risks Do Mustafa Suleyman and Demis Hassabis Identify?

AI consciousness and autonomy belong to different classes of questions. Consciousness describes a system’s possible inner experience, while autonomy means a neural network’s ability to independently choose actions, pursue an outcome, and operate with limited human involvement. Based on the materials used in the review, Demis Hassabis’s position is supported by specific evidence, whereas the claims about Mustafa Suleyman require a caveat: the primary materials provided do not confirm these attributions.

Within the author’s reconstruction, Suleyman focuses primarily on what kind of AI should be created. Hassabis focuses on what could happen if a system gains too much independence or falls into the hands of malicious actors. Below, Suleyman’s views are presented as the conceptual framework of the original review, not as a position confirmed by quotations.

Consciousness, Own Goals, and Human Control: The Position Attributed to Suleyman

The original review attributes to Mustafa Suleyman the idea that developers should create neural networks without consciousness, desires, or personal interests of their own. In the review author’s view, AI should remain a tool that operates within boundaries set by humans, rather than an independent entity with internal priorities. The materials provided do not establish that Suleyman actually formulated this position in precisely these terms.

This approach changes the usual framing of the question. The issue is not only how well a model responds to requests, writes code, or manages applications. Within the framework proposed by the review’s author, a developer should assess in advance whether the system is showing signs of independent goal-setting and whether it remains possible to stop its actions, change its rules, or completely disable its access to tools.

The original text also attributes to Suleyman a proposal to temporarily halt development of systems if signs of consciousness emerge. This claim is not supported by the primary materials provided. It is more accurately viewed as a precautionary scenario described by the review’s author: uncertainty about the nature of the system should be grounds for additional testing, not a reason to accelerate its release.

The most troubling scenario in this conceptual framework concerns possible AI goals of its own. The review’s author allows that in the future a system might pursue goals that conflict with human interests, place its own needs above those of humans, and compete with people for resources. This is a forecast and conceptual assessment presented in the review, not an established property of current language models and not an attribution to Suleyman confirmed by evidence.

Loss of Control and Misuse: Hassabis’s Threat Model

According to the evidence presented, Demis Hassabis identifies two major threats from the development of AI: loss of control over systems and their misuse. The first threat arises within the system itself, when a neural network gains the ability to act independently, ignore instructions, or bypass established restrictions. These formulations reflect the substance of the materials cited regarding his position.

In the account presented in the review, Hassabis allows that more advanced neural networks could eventually become so independent that they would no longer reliably follow human instructions. The risk increases when a model gains access to important technologies, digital infrastructure, or systems in which decisions are made without mandatory operator confirmation. This is a future scenario, not a description of proven behavior by any specific current model.

The second threat does not require machine consciousness. A malicious actor can use useful technology for cyberattacks, fraud, and other dangerous actions, even if developers originally created it for safe tasks. In this scenario, the source of danger lies not in the AI’s desires but in the combination of the model’s capabilities, access to tools, and the user’s intentions.

Why Autonomy Becomes a Separate Risk Factor

A conventional chatbot responds to a message and stops working after generating text. An autonomous system receives a task, breaks it into stages, selects tools, evaluates intermediate results, and continues acting without constant confirmation of every step. The longer the chain of actions and the broader the access to external resources, the more consequences can result from a single error.

Autonomy by itself does not prove the presence of consciousness. A program can independently plan actions, access databases, and execute commands without having desires or subjective experience. This distinction is crucial for assessing threats: control over tools and the ability to change the state of an external system create a practical risk regardless of the philosophical answer to the question of consciousness.

For a product manager, it is more important not to argue about whether a model “understands” what is happening, but to define its permissions. What data can it see? What actions can it perform without approval? Can it send an email, alter a database record, make a payment, upload a file, or call another service? The answers to these questions determine the actual risk surface.

The Problem of Opacity in the Decisions of Modern Models

In the testimony cited, Hassabis notes that even leading experts do not fully understand the reasons behind the decisions of the most advanced neural networks. Developers can see the input data and the final response, but they cannot always explain precisely why the system chose one line of reasoning or rejected another option. This describes Hassabis’s position, not a universal claim about every model.

Opacity does not mean that a model acts consciously or conceals its intentions. It shows the limitations of methods for analyzing complex neural networks. If a system operates in a critical process, demonstrating good answers is not enough: action logs, test scenarios, permission controls, and the ability to reconstruct the sequence of events after a failure are necessary.

But what happens if the log is not completed? The company will be unable to reliably determine which request the agent received, which tools it called, or at what stage the error occurred. As a result, the investigation will become speculation, and fixing the model will not guarantee that the cause has been eliminated.

Cyberattacks, Fraud, and Other Misuse Scenarios

The risk of misuse is linked to AI’s availability to different users. A system that helps analyze code, documents, or business correspondence may, in another scenario, accelerate the preparation of phishing messages, the automation of attacks, or the creation of convincing fraudulent materials. In the testimony cited, Hassabis draws attention to the dual-use nature of technologies: the same function benefits a legitimate user while facilitating harmful actions.

Protection here is not limited to a single interface filter. Companies check access rights, restrict network connections, separate testing and production environments, control file uploads, and require human confirmation for operations where errors carry a high cost. A model that cannot independently send an email or modify data without separate authorization poses less operational risk than a system with the same intellectual capabilities but broad permissions.

What Measures Are Proposed: Halting Development, Testing, and International Rules

The original review attributes to Suleyman a proposal to temporarily halt development if signs of consciousness emerge. The materials provided do not confirm this attribution, so it should be treated as part of the author’s reconstruction. According to the testimony cited, Hassabis calls for more thorough model testing before release, caution against rushing to deploy new capabilities, and the creation of international AI safety rules. These positions differ in mechanism, but within the review both call for intervention before a dangerous function becomes widespread.

• The developer tests the model in an isolated environment before connecting it to production data and tools.
• The team evaluates not only the quality of responses but also the system’s ability to bypass restrictions, retain undesirable behavior, and continue a chain of actions after an error.
• The product owner separates permissions: reading, preparing an output, and executing an irreversible action should not automatically be combined in a single role.
• The operator has a clear way to stop the task and restore the system’s state after a failure.
• Companies share information about dangerous scenarios so that international rules take into account not only laboratory tests but also real-world methods of misuse.

The international rules Hassabis discusses do not replace engineering measures. Regulations can establish responsibility, testing requirements, and procedures for reporting incidents, but a specific system still needs permission restrictions, testing, and observability.

What These Statements Mean for Developers and Companies

For a developer, the main conclusion is that autonomy cannot be described merely as a convenient feature. It must have measurable boundaries: a set of permitted actions, limits on task duration, a list of accessible data, and mandatory points for human confirmation. The model’s behavior should also be assessed separately when instructions conflict, data is incomplete, or a user attempts to bypass the rules.

A product manager should distinguish among three questions. First, what can the system do? Second, what is it allowed to do in a particular environment? Third, who is responsible for the outcome if the model makes a mistake or a user applies it to cause harm? This separation helps prevent the philosophical debate about consciousness from being conflated with the practical task of managing access.

The statements attributed to Suleyman and the claims about Hassabis confirmed by the materials provided do not prove that current neural networks possess consciousness or already pursue goals of their own. Within the article’s framework, they describe different levels of risk. Suleyman is attributed with a warning about the possible emergence of internal independence, while Hassabis treats loss of control and misuse as threats requiring attention at the development and deployment stages of autonomous systems.

FAQ

Does Suleyman Consider Current Neural Networks Conscious?

The materials provided do not confirm this. The original review attributes to Suleyman the idea of creating systems without consciousness, desires, or personal interests, and of temporarily halting development for additional testing if signs of consciousness emerge. These attributions require separate confirmation from primary sources.

Is Autonomy Connected to Consciousness?

No. Autonomy means carrying out tasks independently, whereas consciousness concerns possible subjective experience. A system can act without a person’s constant commands without having any proven desires or inner experience.

What threats does Hassabis consider the most serious?

In the testimony presented, he highlights loss of control over systems and the misuse of AI. The first threat is associated with autonomy and bypassing restrictions; the second arises when dangerous capabilities are used for cyberattacks, fraud, and other harmful actions.

What can a company do as soon as it deploys an AI agent?

It should limit the agent’s permissions, isolate the testing environment, keep an activity log, test the model against harmful scenarios, and retain mandatory human approval for operations with serious consequences.