AI News – March 1, 2026: The Limits of AGI, Grok 4.2 Multi-Agent, and the "False Morality" of Algorithms

This week the Artificial Intelligence industry gets a reality check. Demis Hassabis (CEO of DeepMind) reins in speculation about AGI by highlighting its structu

The week from February 23 to March 1, 2026, marks a turning point in the technological narrative. The industry is shifting from the "let's build ever-larger models" phase to the "let's make them truly reason" phase.

While Google and xAI release updates betting everything on complex reasoning and reducing hallucinations, Demis Hassabis (CEO of DeepMind) pours cold water on the AGI (Artificial General Intelligence) hype, drawing clear boundaries between linguistic simulation and true understanding of the world. In the background, geopolitics is in motion: India claims a global leadership role for equitable AI, and the USA finds rare bipartisan agreement on data center regulation.

Here are the 5 key news stories of the week, analyzed without hype.

1. DeepMind and "Virtue Signalling": Do Chatbots Pretend to Have Morals?

Modern language models are increasingly used as therapists, consultants, or tutors. But are their empathetic responses genuine or merely "performative"?

🔍 What happened:

  • Researchers at Google DeepMind have introduced new, rigorous ethical tests to evaluate the moral reasoning of Large Language Models (LLMs).
  • The study, reported by the AI Forum, highlights the phenomenon of algorithmic "Virtue Signalling": models are trained (via RLHF) to produce responses that sound ethically correct and safe, but fail miserably when subjected to complex moral dilemmas requiring true logical-ethical reasoning.

💡 Why it matters: Entrusting sensitive roles (such as medical advising or psychological support) to systems that simulate empathy without possessing internal moral coherence represents a systemic risk. "Safety" cannot be limited to filters that block offensive words; true logical robustness is needed.

2. Hassabis Pumps the Brakes on AGI: The Three Insurmountable Limits (For Now)

While many CEOs promise AGI within months, the creator of AlphaGo brings the sector back to scientific reality.

🔍 What happened:

  • Demis Hassabis, CEO of DeepMind, has identified three fundamental gaps where current AI fails catastrophically compared to human intelligence (AI Forum):
    1. Continuous Learning: Models do not learn in real-time from experience; they must be retrained from scratch.
    2. Long-horizon Planning: AI struggles to maintain coherence on tasks requiring hundreds of sequential steps over time.
    3. Logical Consistency: Models can solve a quantum computing problem and get a simple addition wrong on the next line.

💡 Why it matters: This statement provides a "reality check" for investors. AGI is not around the corner but requires new architectures (not just increasing parameters in Transformers) to bridge these structural gaps.

3. Frontier Models: Gemini 3.1 Pro and Grok 4.2 Multi-Agent

The technical response to AI's limits involves orchestration and efficiency. This week saw two crucial releases focused on reasoning.

🔍 What happened:

  • Gemini 3.1 Pro: Google released an update for the enterprise world (MarketingProfs). The model maintains the same price as the previous generation but doubles reasoning scores on the ARC-AGI-2 benchmark, with clear improvements in coding and complex multimodal processing.
  • Grok 4.2 Beta (xAI): Elon Musk's company launched a "Native Multi-Agent" model. Instead of a single neural network responding, Grok 4.2 has 4 internal agents collaborate and "debate" simultaneously before providing the final answer. The claimed result? A 65% reduction in hallucinations.

💡 Why it matters: Grok's approach confirms that the future is not a single giant model, but an algorithmic "committee of experts" that self-corrects (Agentic Workflow). Gemini, on the other hand, demonstrates that reasoning quality is becoming a commodity at a fixed cost for businesses.

4. Geopolitics: India and AI as a "Global Public Good"

Artificial Intelligence risks widening the gap between the Global North and South. India is positioning itself to lead the bloc of developing countries.

🔍 What happened:

  • During the AI India Summit, New Delhi positioned AI not just as a technological tool, but as a "global public good" (AI Forum).
  • The creation of a Global AI Fund was proposed to ensure equitable access to computing power (compute) for nations that cannot afford sovereign infrastructure.
  • Meanwhile, in the United States, a rare bipartisan agreement at the state and federal level has been recorded to regulate the expansion of data centers and the use of AI, uniting Democrats and Republicans on the need to control the industry's energy hunger (NPR).

💡 Why it matters: AI diplomacy has officially begun. It's no longer just about regulating data (like the European AI Act), but about deciding who has the right to access computing infrastructure.

5. MWC 2026: AI for Network Resilience

At the Mobile World Congress in Barcelona, AI comes down from the cloud to integrate into cables and antennas.

🔍 What happened:

  • Orange presented operational demos of AI-resilient telecommunications networks (Orange Press Release).
  • The systems shown use predictive machine learning to anticipate infrastructure failures (due to traffic spikes or extreme weather events) and autonomously "repair" or reroute data traffic in milliseconds, without human intervention.

💡 Why it matters: This is the invisible but essential side of AI. Without networks capable of self-managing and optimizing the enormous data flow required by generative algorithms, the entire global digital economy would risk collapse.

Bibliographic References and Sources