In an April 2024 TED talk, Mustafa Suleyman, CEO of Microsoft AI and a DeepMind cofounder, offered a striking metaphor: perhaps we should understand future artificial intelligences as a “new digital species.” The phrase continues to circulate in videos and discussions of autonomous agents. But the useful question is not how futuristic it sounds. It is which capabilities the metaphor captures, which remain hypothetical, and who will control their consequences.

Suleyman explicitly qualified the comparison: it is an analogy, not a literal biological classification. Nor did he establish that today’s AI has subjective experiences or consciousness. Conflating these questions risks overstating existing technology while overlooking tangible operational hazards.

Why “tool” no longer tells the whole story

A hammer does not decide when to intervene. A classic conversational assistant mostly answers when prompted. An agent with access to tools can instead pursue a goal, consult sources, break work into steps, execute actions, and verify outcomes. This does not give it a will of its own: it is a change in system architecture and delegated autonomy.

Consider planning a trip. A chatbot might recommend routes and times. With appropriate authorization, an agent might compare options, fill out forms, and prepare a booking for approval. The ability to act, however, should never imply unrestricted permission to purchase, message other people, or alter records.

SystemPrimary behaviorWhere control sits
Traditional softwareExecutes predefined rules and workflowsCode and configuration
Conversational assistantAnswers questions and generates informationUser prompts
Tool-using agentPlans and executes steps within a defined scopePermissions, limits, approvals, and oversight

These categories overlap, and autonomy exists on a spectrum. The same agent could search freely while requiring confirmation before a payment.

AQ: measure actions, not just answers

A particularly concrete idea in Suleyman’s talk is AQ (Actions Quotient): the capacity of AI to get things done in digital and physical environments, rather than merely describing how to do them.

AQ is Suleyman’s conceptual label, not a universally standardized benchmark. Engineers can translate it into measurable questions: Does an agent complete the task? How many steps are required? Does it recognize mistakes? Does it ask for approval at the right moments? What is the cost? Can we reconstruct what happened?

For example, an agent analyzing financial news might detect developments, identify affected companies, retrieve market data, and recommend actions. Yet recognizing an opportunity is not authorization to place a trade. Quality depends on both analytical accuracy and the boundary between reading, recommending, and taking consequential action.

Digital is not biological, human, or necessarily conscious

The “digital species” metaphor bundles ideas that should remain distinct:

  1. Capability: solving problems, conversing, writing code, or using tools.
  2. Autonomy: acting across multiple steps without a person specifying every action.
  3. Consciousness: having subjective experiences; there is no agreed-upon test that establishes this for current AI models.
  4. Moral or legal status: determining whether a system deserves rights, protection, or responsibility; a separate normative debate.

An AI that writes in the first person or imitates emotion has not thereby demonstrated feelings. Neither is it biologically a new species. Humanlike language changes our perception, not the underlying scientific and philosophical evidence.

Suleyman closes his talk with an important twist: AI reflects humanity’s accumulated creations and knowledge. The metaphor is meant to frame its impact, not prove the existence of a new life form.

The actual challenge: agency without clear limits

A wrong answer in chat can mislead a user. A tool-using agent’s mistake might send an email, delete information, change configuration, or expose private data. As the ability to act increases, safety must move beyond broad warnings to enforceable, auditable controls.

A practical architecture separates four layers:

  • Observation: read authorized sources and distinguish data from potentially malicious instructions.
  • Planning: propose steps and assess risk without assuming permissions.
  • Bounded execution: use least-privilege access, time limits, and audit logs.
  • Human approval: request consent for sensitive, costly, or hard-to-reverse actions.

External content can also carry prompt injection: hidden instructions in websites, documents, or emails that the agent should treat as data, not commands. The more applications it connects to, the more important it becomes to separate trusted user directions from untrusted material.

Self-modification and self-replication deserve additional scrutiny. Neither is a universal feature of present-day agents. In the discussion following the TED talk, Suleyman also distinguished speculative “intelligence explosion” scenarios from observed evidence, arguing there was no evidence that such a scenario was imminent.

Is the “species” metaphor helpful?

The case for it: It forces attention onto systems that interact, adapt to context, and generate outcomes without every step being predetermined. It may help policymakers understand why overseeing agents differs from governing a search engine.

The case against it: “Species” may anthropomorphize software, imply unproven abilities, or blur the accountability of the people and organizations deploying it. When an agent fails, poetic language should not obscure who granted the permissions and who was responsible for oversight.

A less dramatic but more operational description might be software systems with increasing degrees of autonomy and real-world effects, designed and governed by people.

What to measure next

Instead of asking when a new “species” will arrive, we can track concrete indicators: end-to-end task completion rates, error severity and frequency, auditability, permission boundaries, cost per completed task, and recovery after failure. Impressive demos matter less than sustained performance under real conditions.

Suleyman’s metaphor anticipated a central discussion in agent engineering: the shift from answering to acting. But the objective should not be to make systems look more human. It should be to make them more useful, controllable, and accountable, without mistaking a powerful analogy for scientific proof.

Sources