AI Risks: Why DeepMind CEO Demis Hassabis Says Urgent Research Cannot Wait
AI risks moved back to the centre of the global technology debate at the India AI Impact Summit in New Delhi, where Google DeepMind co-founder and chief executive Sir Demis Hassabis argued that safety research must keep pace with rapidly advancing systems. His warning was not a rejection of artificial intelligence. It was a call to treat a transformative technology with the seriousness required when its benefits and dangers are developing at the same time.
Speaking at Bharat Mandapam in February 2026, Hassabis described the summit as an important forum for international dialogue at a pivotal moment for AI. He highlighted two broad dangers: malicious use by bad actors and the possibility that increasingly autonomous systems could behave in ways their designers did not intend. Those concerns remain relevant beyond a single conference because AI systems are becoming more capable, more widely deployed and more deeply connected to everyday digital services.
The debate about AI risks is often presented as a choice between innovation and caution. That framing is too simplistic. Healthcare, education, scientific discovery and business productivity can benefit from advanced tools. However, meaningful progress depends on confidence that powerful systems are tested, monitored and governed responsibly. The question is not whether the world should use AI. The question is how quickly institutions can build safeguards while the technology continues to improve.
Why Hassabis Put AI Risks at the Centre of the Discussion
Hassabis has spent years arguing that AI could become one of the most consequential technologies in human history. At the Delhi summit, his message combined optimism with urgency. The official DD News summary of his address reported that he described the gathering as a critical platform for international cooperation and said global dialogue was becoming increasingly important.
Reporting on Hassabis’s warning identified two different problems. The two AI-risk categories highlighted during the summit were malicious use and loss of control. The first is misuse. A general-purpose model designed for beneficial tasks can also lower barriers for cyberattacks, fraud, manipulation or harmful technical activity. The second is loss of control. As systems gain more autonomy, they may take actions that were not anticipated by developers or users.
| Area of concern | What it means | Why it matters |
|---|---|---|
| Malicious use | People deliberately use AI tools for harmful purposes | Useful capabilities can be redirected towards scams, cyberattacks or other abuse |
| Unintended behaviour | A system produces harmful results or acts outside the intended objective | Greater autonomy can make failures harder to predict and reverse |
| Systemic disruption | Widespread adoption changes jobs, information systems or public trust | Harm may emerge across society even without one dramatic failure |
These categories should not be treated as identical. Understanding AI risks requires precision. A scammer using an AI tool is a different problem from an autonomous system acting unpredictably. Workforce disruption is different again. Effective policy begins by distinguishing the risks clearly enough to match each problem with an appropriate response.
AI Risks Are Already More Than a Future Scenario
Some discussions about AI risks focus almost entirely on hypothetical future systems. Long-term questions matter, but existing evidence should not be ignored. The International AI Safety Report 2026 assesses general-purpose systems, the harms already being documented and the risks that remain uncertain but potentially severe.
The report groups major AI risks into three categories: malicious use, malfunctions and systemic risks. It notes that capable models can assist with cyber activity, generate misleading information and produce unreliable answers. It also explains that AI capabilities remain uneven. A model may perform impressively on difficult evaluations while still failing at simpler tasks or producing confident factual errors.
That combination is important. The public debate about AI risks should not begin only when systems become extraordinarily powerful. AI risks can also appear when people place too much trust in imperfect tools. A fluent answer can sound authoritative even when it contains a fabricated citation, a faulty conclusion or a misleading recommendation.
The same report describes an evidence dilemma for policymakers. Waiting for complete certainty can leave society exposed to serious harm. Acting too early with poorly designed rules can create ineffective regulation or lock in weak approaches. The challenge is to build proportionate safeguards that can adapt as evidence improves.
The Delhi Summit Was About Impact as Well as Safety
The India AI Impact Summit was held at Bharat Mandapam from 16 to 20 February 2026. Its official materials framed the event around three pillars: People, Planet and Progress. The summit press-release page also records the New Delhi Frontier AI Impact Commitments, announced as an effort to advance inclusive and responsible AI.
The location mattered. Reuters reported that the Delhi gathering marked the first time this series of global AI summits had been held in the developing world. It brought technology leaders and policymakers together while India positioned itself as a major centre for AI adoption and infrastructure investment. Readers can explore that wider summit context through the Reuters overview.
AI risks formed one part of a wider agenda. The summit also addressed access, skills, public services, infrastructure and inclusion. This broader frame is useful because responsible development cannot be reduced to preventing extreme outcomes. It should also consider who benefits from AI, who bears the costs and whether lower-income countries can influence the rules shaping the technology.
Responsible Innovation Requires More Than a Slogan
Calls for responsible AI are common. The harder task is translating them into practical systems. AI risks cannot be managed only through speeches, general principles or voluntary promises. Organisations need clear testing procedures, security controls and decision-making structures.
A serious approach should include:
- Pre-deployment evaluations: Test advanced models for dangerous capabilities, reliability problems and predictable misuse routes before release.
- Security protections: Reduce the chance that sensitive models, data or tools can be stolen, manipulated or accessed improperly.
- Human oversight: Keep meaningful human control over high-impact decisions and autonomous actions.
- Incident reporting: Create processes for documenting failures, sharing lessons and responding quickly when harms appear.
- Independent research: Support researchers who can study safety questions without depending entirely on the companies building the systems.
- International coordination: Develop common minimum expectations for risks that cross borders.
These steps do not eliminate AI risks. They create a stronger foundation for managing AI risks responsibly. The details will vary by sector. A writing assistant does not require the same safeguards as an autonomous system connected to critical infrastructure. A medical tool requires different evidence from an entertainment app.
Risk-based regulation is therefore more useful than treating every AI system as though it poses the same danger.
Cybersecurity Is One of the Clearest Pressure Points
Cybersecurity illustrates why AI risks require careful attention. AI tools can help defenders review code, detect suspicious behaviour and respond to incidents. The same capabilities can also assist attackers. The International AI Safety Report states that AI systems can identify software vulnerabilities and generate malicious code, while the overall balance between offensive and defensive advantage remains uncertain.
This is a dual-use problem. AI risks become more complicated when the same capability can be valuable in the hands of a security team and dangerous in the hands of a criminal group. The answer is not to assume that every use is harmful. It is to strengthen safeguards and improve defensive readiness.
The News Ink’s verified cybersecurity guide explains practical protections for everyday users, including stronger passwords, multifactor authentication and caution around scams. Those habits remain important as AI-generated impersonation and automated fraud become more convincing.
AI risks in cybersecurity also reinforce the need for cooperation. Digital threats cross national borders quickly. A vulnerability discovered in one country can affect users elsewhere. Effective responses require developers, companies, governments and researchers to share information without exposing sensitive details unnecessarily.
Greater Autonomy Changes the Safety Question
The rise of AI agents has made the discussion more urgent. Traditional chatbots generally respond to user prompts. More capable agents can plan tasks, use tools and act across multiple steps with less direct supervision. This can make software more useful, but it can also increase the consequences of a mistake.
In May 2026, Hassabis told Axios that the next wave of AI agents could act as a societal stress test for more powerful systems. He said he still broadly expected artificial general intelligence around 2030, while viewing 2029 as a possibility. The Axios report also recorded his call for faster safety preparation.
Forecasts about artificial general intelligence remain uncertain. There is no universally agreed definition, and timelines vary widely. However, the policy lesson does not depend on accepting one precise prediction. AI risks deserve attention because more autonomous systems are already emerging and because institutions need time to build safeguards.
The important questions are practical:
- What permissions should an AI agent receive?
- Can its actions be reviewed and reversed?
- What happens when the system encounters an unexpected situation?
- Who is accountable when an automated action causes harm?
- How should high-risk tools be tested before deployment?
These are governance questions as much as technical questions.
India’s Human-Centric Message Added an Important Perspective
Prime Minister Narendra Modi used the summit to argue that AI should remain human-centric. In his official address, he said AI should have an open sky while command remains in human hands. He also presented the M.A.N.A.V. vision, emphasising moral systems, accountable governance, national sovereignty, accessibility and legitimacy.
That framing connects directly with AI risks. Governance should not exist only to slow technology down. It should help ensure that systems serve people rather than treating users merely as data sources or passive recipients of automated decisions.
The Global South perspective is especially important. Countries differ in infrastructure, labour markets, languages and public-service needs. A safety framework designed only around the priorities of a small number of wealthy economies may miss significant risks and opportunities elsewhere.
The strongest international approach will need to account for both frontier-model safety and everyday deployment. A country may face immediate concerns involving fraud, education, labour displacement or access long before it confronts the most advanced autonomous systems.
AI Risks Should Not Obscure the Benefits
Responsible analysis of AI risks must avoid two extremes. One extreme treats AI as an automatic solution to complex social problems. The other treats every advance as evidence of inevitable disaster. Neither position is useful.
Google DeepMind announced new collaborations in India focused on science and education during the summit. Its official partnership announcement argues that frontier labs, governments, academia and civil society need deep collaboration if AI is to benefit humanity.
There are real reasons for optimism. AI tools can assist researchers, improve access to knowledge, support translation and help professionals work more efficiently. The News Ink’s article on AI trends in 2026 examines how agents, automation and generative tools are already reshaping digital life.
However, benefits do not cancel AI risks. They make responsible governance more important. Public trust is difficult to build and easy to lose. A system that saves time in one setting can cause harm in another if it is deployed without adequate testing or human review.
The Workforce Question Needs Honest Answers
Employment is one of the most immediate concerns surrounding AI risks. New tools can increase productivity, automate repetitive work and create new roles. They can also reduce demand for some tasks, particularly when companies adopt automation quickly.
Predictions should be treated cautiously. No one can state with certainty how many jobs will disappear, how many new roles will emerge or how quickly labour markets will adapt. The effects will differ by industry, country and skill level.
The News Ink’s verified article on whether AI will replace jobs explores that debate in more detail. The responsible response is not panic, but preparation. Governments, schools and employers should invest in reskilling, technical literacy and adaptable education.
Hassabis also emphasised the continuing value of technical knowledge. That does not mean every worker must become a computer scientist. It means people will benefit from understanding how to use tools critically, verify outputs and recognise limitations.
Workforce policy belongs inside the discussion of AI risks because disruption can be serious even when no dramatic technical failure occurs.
What Effective AI-Risk Research Should Prioritise
The phrase AI safety can become vague unless it is connected to concrete questions. Research should cover both immediate harms and longer-term uncertainty.
| Research priority | Key question |
|---|---|
| Misuse prevention | How can developers reduce harmful use without blocking legitimate applications? |
| Agent oversight | How should autonomous systems be monitored, limited and stopped when needed? |
| Reliability | How can users identify hallucinations, errors and misleading outputs? |
| Cybersecurity | Will new tools advantage defenders, attackers or both? |
| Social impact | How will AI affect work, inequality, education and access to information? |
| Governance | Which rules should be international, and which should remain sector-specific? |
AI risks require technical research, but technical research alone is not enough. Economists, educators, legal experts, cybersecurity professionals and social scientists all have a role. Decisions about deployment involve values, trade-offs and institutional capacity.
Companies also need incentives to share evidence. Safety research becomes less useful when independent experts cannot examine important systems or when incident reporting remains inconsistent. Transparency should be designed carefully, because publishing every detail can create new security problems. The goal is accountable disclosure, not careless exposure.
International Cooperation Will Remain Difficult but Necessary
AI risks must be considered even when AI development is shaped by commercial competition and national strategy. Countries want investment, infrastructure and access to advanced tools. Companies want to release useful products quickly. These incentives can make cooperation difficult.
Yet AI risks do not respect national borders. Cyber threats, misinformation and economic disruption can spread internationally. A serious accident involving a widely deployed system could affect people far beyond the country where the model was developed.
The most realistic path may be a layered approach. Governments can establish domestic rules for high-impact uses while supporting international minimum standards for frontier safety, security testing and incident response. Companies can compete on products while still cooperating on shared risks.
This will not remove political disagreement. It can reduce the chance that safety becomes an afterthought in a race for speed.
Frequently Asked Questions About AI Risks
What warning did Demis Hassabis give at the Delhi summit?
Hassabis highlighted two broad AI risks: malicious use by bad actors and unintended behaviour from increasingly autonomous systems. He also called for international dialogue and cooperation.
Are AI risks only about future superintelligence?
No. AI risks already include scams, cyber misuse, unreliable outputs, misinformation and workforce disruption. Longer-term concerns about highly autonomous systems add another layer of uncertainty.
Did the India AI Impact Summit focus only on safety?
No. The summit addressed People, Planet and Progress, with discussions involving inclusion, infrastructure, public services, skills and responsible development.
Can AI still provide major benefits?
Yes. AI can support science, education, healthcare and productivity. The central challenge is ensuring that deployment is safe, accountable and useful.
Why is international cooperation important?
AI systems and digital threats cross borders. Shared standards, research collaboration and incident-response processes can improve safety while allowing countries to develop their own domestic policies.
AI Risks Demand Preparation, Not Panic
AI risks deserve urgent research because powerful technologies are rarely governed well through last-minute reactions. Hassabis’s warning at the Delhi summit was not an argument against progress. It was a reminder that progress becomes more sustainable when institutions prepare for misuse, malfunctions and disruption before the most serious harms appear.
The completed summit also showed that the debate is broader than one company or one country. India emphasised inclusive development and human control. Researchers highlighted evidence gaps. Technology leaders discussed global cooperation. The next stage will depend on whether those ideas produce durable systems for testing, oversight and accountability.
AI risks should neither be exaggerated for attention nor dismissed as obstacles to innovation. The responsible path is more demanding: invest in research, strengthen safeguards, educate users and keep human judgement at the centre of deployment.
For more technology analysis and AI coverage, follow The News Ink on X.
