AI systems may soon do more than answer questions or produce text. Mustafa Suleyman, cofounder of DeepMind and founder of Inflection, sees a shift toward interactive AI: systems that can use software and communicate with people to carry out tasks. That prospect could change how people use technology, while raising difficult questions about control and safety.
From generating content to getting things done
Suleyman describes AI’s development in three broad phases. The first focused on classification: training computers to identify patterns in images, video, audio and language. The current phase is generative, producing new material from inputs. He expects the next phase to be interactive.
In that model, people would explain a high-level goal in conversation. An AI could then use available tools, contact other people or interact with other AI systems to pursue it. Instead of navigating buttons and typing instructions into separate programs, a person might make conversation the main way to direct technology.
The change is consequential because today’s tools generally respond to specific instructions. An AI that can act on broader goals has more room to make decisions along the way. Suleyman calls this a shift toward technology with agency, and says people may underestimate how significant that is.
More capability means sharper questions about control
Giving an AI permission to act creates a tension: the system needs enough freedom to complete a task, but people need to remain in command. Suleyman says that balance depends on boundaries an AI cannot cross. Those limits, in his view, should apply to the underlying code, the system’s interactions with people and other AIs, and the incentives of the companies building it.
He also argues that independent institutions and governments should have direct access to check whether those limits hold. Oversight would not happen only at a national or international level. Individuals might also grant an AI limited permission to handle personal data or answer certain kinds of questions.
Some capabilities, he says, call for particular caution. Suleyman points to recursive self-improvement: an AI changing its own code without human oversight. He suggests that this kind of activity might need a license. His comparison is to activities involving hazardous materials and to drone use, which he says is restricted because of privacy risks.
Can regulation keep pace?
Suleyman rejects the idea that AI will be impossible to regulate. He argues that internet governance offers examples of progress, pointing to efforts against spam, revenge porn, radicalization material, and the online sale of weapons and drugs. He sees a combination of public pressure, institutional action and government rules as a way to manage AI, too.
That confidence is contested. The article notes that some of Suleyman’s claims about online regulation are not supported by the numbers. It points to continuing cybercrime, the growth of nonconsensual deepfake pornography, and the marketing of drugs and guns on social media. It also says platforms could do more to filter harmful content.
Those examples complicate the idea that existing approaches can simply be extended to AI. Interactive systems could take actions across tools and conversations, so oversight would need to address not only what a system says but also what it is allowed to do. Suleyman’s proposal is to set clear limits and give outside institutions a way to verify them.
Optimism alongside practical risks
Suleyman presents himself as focused on both benefits and threats, rather than choosing an optimistic or pessimistic label. He says AI models have become more controllable as they have grown, and describes Inflection’s chatbot Pi as designed with safety as a top priority. He also says the company keeps details of its approach private while building a business and paying for computing hardware.
His broader argument is that debate should focus on practical concerns, including privacy, bias, facial recognition and online moderation. He points to aviation and cars as examples of complex technologies that operate alongside safety rules, and calls for both industry responsibility and government action.
Whether interactive AI can be bounded as reliably as Suleyman hopes remains a central question. The systems he describes would act on people’s behalf, so their usefulness would depend on trust: users would need to know what permissions they have granted, and society would need ways to check that companies respect the limits. Suleyman’s vision is not just a more capable chatbot, but a new kind of technology whose authority has to be deliberately constrained.