New Delhi: A former Google DeepMind safety researcher has warned that the rapid development of artificial intelligence could eventually create risks serious enough to threaten humanity, arguing that AI safety measures are not advancing at the same pace as the technology itself.
Bilal, who worked as an AI safety researcher at Google DeepMind, recently announced his resignation and shared his concerns about the direction of AI development. In a series of statements reported by India Today, he said the coming period could become increasingly difficult to manage as AI systems become more capable.
His comments come amid a broader debate among AI researchers, technology companies and policymakers over how to manage increasingly capable AI systems while continuing to develop the technology.
Why the researcher resigned
Bilal’s departure from Google DeepMind has drawn attention because of his work in AI safety, an area focused on understanding and reducing potential risks associated with advanced AI systems.
According to India Today, he said he had reached a point where he was no longer comfortable with the direction of AI development and the pace at which safety research was progressing.
His concerns centre on a perceived gap between the rapid improvement of AI capabilities and efforts to ensure that increasingly powerful systems remain controllable and aligned with human objectives.
The researcher has argued that AI development could enter a period of rapid change, making it difficult for existing safety approaches to keep pace.
“Things are about to get crazy”
In comments highlighted by India Today, Bilal warned that the next stage of AI development could be particularly disruptive.
He suggested that the technology is advancing towards systems with substantially greater capabilities and that society may not be sufficiently prepared for what could follow.
His warnings are part of a wider discussion within the AI research community. Researchers disagree about the probability and timing of extreme AI risks, but there is broad interest in questions surrounding AI alignment, misuse, cybersecurity and the control of highly capable systems.
Bilal’s comments should therefore be understood as his assessment of potential future risks rather than a verified prediction that AI will cause a catastrophe.
AI safety has become a major research field
AI safety research covers several areas, including ensuring that AI systems follow instructions reliably, preventing harmful behaviour, evaluating models before deployment and developing safeguards against misuse.
Researchers also examine what could happen if future AI systems become capable of performing increasingly complex tasks with limited human supervision.
One concern is that an AI system could behave in ways that its developers did not anticipate. Another is that humans could deliberately use advanced AI for harmful purposes.
These risks are separate from more immediate problems such as misinformation, fraud, privacy violations and biased automated decisions.
DeepMind has its own AI safety research
Google DeepMind has invested heavily in research into AI safety and responsible development.
The company has teams working on areas including model evaluation, alignment, interpretability and AI governance.
The departure of an individual researcher does not by itself demonstrate that the company’s safety programme is failing. Bilal’s views represent his own assessment of the risks and the direction of AI development.
Google DeepMind and other major AI laboratories continue to publish research aimed at understanding and mitigating potential risks from increasingly capable models.
Why the debate is becoming more urgent
The discussion around AI safety has intensified as companies have developed models capable of handling increasingly sophisticated tasks involving reasoning, coding, scientific research and autonomous workflows.
This has raised a question for researchers: Can safety techniques keep pace with improvements in AI capabilities?
Some researchers believe existing evaluation and alignment methods can continue to improve alongside increasingly capable models. Others argue that more fundamental safety breakthroughs may be needed before AI systems become substantially more autonomous.
Bilal belongs to the group expressing concern that current approaches may not be sufficient for the most advanced systems.
What does “AI can kill us all” mean?
Statements suggesting that AI could “kill us all” refer to a category of hypothetical long-term AI risks rather than a current capability of mainstream AI systems.
The underlying concern is that a sufficiently capable system could potentially pursue objectives in ways that conflict with human interests, particularly if it had access to significant resources or the ability to operate with limited supervision.
There is currently no established evidence that today’s consumer AI systems possess such capabilities.
AI researchers therefore distinguish between current AI risks and hypothetical future risks from advanced AI.
Current risks include hallucinated information, cybersecurity misuse, deepfakes, privacy concerns and automated discrimination. Long-term AI safety research examines scenarios involving systems that could be considerably more capable and autonomous than today’s models.
Former AI researchers raise similar concerns
Bilal’s warnings are not occurring in isolation. Several prominent AI researchers and industry figures have previously raised concerns about the potential risks associated with increasingly capable AI.
At the same time, other researchers and technology companies argue that AI can deliver substantial benefits if development is accompanied by appropriate safeguards, testing and governance.
The disagreement is therefore not simply about whether AI will become more powerful. It also concerns how quickly capabilities will advance, what kinds of systems will emerge and which safety measures will prove effective.
The challenge for AI companies
For companies developing frontier AI systems, the challenge is to balance rapid technological progress with safety testing and risk management.
As models become more capable, developers increasingly conduct evaluations before releasing them publicly. These evaluations can test areas such as cybersecurity, biological risks, autonomous behaviour and the ability of models to resist attempts to circumvent safeguards.
However, the effectiveness of these evaluations remains an active area of research.
Bilal’s resignation highlights this continuing tension between AI development and safety research.
What happens next for AI safety?
The debate over AI’s long-term risks is likely to continue as technology companies develop more capable models.
Governments are also working on AI regulations and safety standards, while researchers are developing techniques to understand how models reach their conclusions and how their behaviour can be controlled.
Bilal’s comments add another warning from within the AI safety community about the pace of technological development.
Whether the extreme scenarios he describes materialise remains uncertain. What is clear is that questions about AI safety, control and responsible deployment are becoming increasingly important as AI systems grow more capable.