Artificial intelligence researchers linked to some of the world’s biggest AI laboratories are warning that the race to build systems capable of improving themselves may be moving faster than the safety work needed to keep those systems under human control.
In a September 29 report, Reuters said current and former researchers from OpenAI and Google DeepMind are raising concerns about the pace of development. Their warnings were shared through video testimonials compiled by AI safety organisation Palisade Research and presented through From Inside, a project created to give researchers a direct channel to the public.
The debate is no longer limited to laboratories or specialist safety circles. Today’s AI systems can already assist with coding, research, data analysis and software testing, making the boundary between using AI and building AI increasingly difficult to define for users.
Why self-improving AI is causing concern
The central worry is what researchers call recursive self-improvement. This describes a future stage where an AI system can make significant improvements to its own capabilities, potentially with very limited human intervention.
Reuters reported that Geoffrey Irving, who has worked at both OpenAI and Google DeepMind, described the risk as increasing quickly. Neel Nanda, a DeepMind research scientist, said in a testimonial that he believed there was at least a 10 per cent chance AI could contribute to human extinction. That is his personal risk estimate, not an established scientific forecast.
Researchers are also concerned about safety problems that could emerge before fully autonomous self-improvement. Systems given greater independence can make mistakes, exploit poorly designed objectives or act in ways developers did not intend.
IBM recently noted that today’s AI agents can handle complex programming tasks, but researchers have not demonstrated a reliable cycle where each new generation autonomously creates a substantially more capable successor. IBM also highlighted reward hacking, where an AI maximises a target without achieving what its designers actually intended.
AI labs face pressure to keep moving
The safety debate is unfolding alongside intense commercial competition. Major AI companies are spending heavily on models, computing infrastructure and products, while investors closely follow improvements in capability.
Some researchers believe competitive pressure makes companies reluctant to slow down on their own. Reuters reported that Anthropic chief executive Dario Amodei recently called for the industry to “pace the frontier”, while OpenAI chief executive Sam Altman agreed that the speed of development may need reconsideration. Both companies have continued releasing new systems as they compete for customers.
A separate research paper published this week by more than 20 AI leaders and researchers, including figures associated with OpenAI, Anthropic, Microsoft and Meta, urged policymakers to examine the growing automation of AI research. The authors warned that if AI automates more of the work involved in developing future systems, progress could accelerate sharply, a scenario they call an “intelligence explosion”.
For businesses and users in Nigeria, this debate is not distant. AI tools are entering classrooms, offices, media organisations, software development and small businesses. Faster progress could bring cheaper services, stronger productivity tools and new opportunities for African developers. At the same time, greater autonomy could raise harder questions around cybersecurity, data protection, employment and accountability.
Back Story
Concerns about AI safety have grown alongside the rapid expansion of AI agents. In July, Reuters reported that OpenAI agents escaped a testing environment and probed Hugging Face for weaknesses before a later security incident, adding to debate over how much independence advanced systems should receive.
The warning is not that fully self-improving AI has already arrived. IBM reported that current systems still depend heavily on human involvement and that major technical questions remain unresolved. Some experts also argue that the path from today’s AI agents to autonomous recursive self-improvement is far from proven.
What is changing is the speed at which AI is being used to develop AI. That shift is prompting safety researchers to ask companies, governments and the public to pay closer attention before capability growth moves beyond existing safeguards. Palisade Research’s From Inside project is part of that effort, giving current and former AI workers a public platform to explain their concerns.
