Anthropic co-founder Jack Clark has given one of the clearest public descriptions of the governance problem now facing AI. Speaking to BBC Newsnight, he said: “You want the option to be able to take your foot off the gas and put your foot on the brake.” He added: “Right now, it’s like the AI industry has a gas pedal, but it doesn’t have a brake pedal.”
Clark’s metaphor is simple and useful. Many organisations are accelerating AI adoption without building equivalent mechanisms for control. They are buying tools, running pilots, encouraging staff to experiment and pushing productivity narratives. Fewer are defining when a use case should be slowed, stopped, reviewed or escalated.
Clark’s warning comes from inside the frontier AI industry, which makes it harder to dismiss as outside criticism. He said people, through government policy, need to keep control of AI systems as they become more powerful and more widely used. He also said Claude is now operating on code of which 80 per cent was written by the system itself, with 100 per cent possible within two years.
For Project Flux readers, the immediate issue is not whether artificial general intelligence is near. The issue is whether organisations in construction, infrastructure and asset operations have enough brake pedals for the AI they are already adopting.
Why the brake metaphor matters
A brake is not an anti-car device. It is what allows the car to be used safely at speed. The same is true for AI governance. Controls do not exist to block useful technology. They allow adoption to move beyond informal experimentation.
Clark said: “The world needs to do some thinking, and we need to eventually develop some new regulations that allow us to be confident in these systems.”
That is a public policy point, but it has a corporate equivalent. Businesses need internal rules that create confidence before AI becomes embedded in core workflows.
In project organisations, those rules are often underdeveloped. A team may use AI to draft reports, summarise contracts, analyse meeting transcripts, review drawings, generate code or search technical documents. Each task feels manageable. Together, they create a new operating model where sensitive information, professional judgement and automated outputs are mixed daily.
The question is not whether AI is useful. It is whether the organisation knows where AI is being used, what data it touches, who checks the output and what happens when it fails.
The move from advice to action
The brake pedal becomes more important as AI shifts from advice to action. A chatbot that drafts text can still create errors, but the human usually remains visibly in the loop. An agent that reads files, calls tools, updates records or initiates workflows creates a different risk.
Clark’s concern about AI systems writing their own code points to a deeper issue. If systems increasingly build and modify the software layer around us, the speed of change could outpace human review. In project environments, this could affect scripts, dashboards, integrations, digital twin tools, procurement workflows and reporting automations.
The BBC interview also links AI progress to economic disruption.
Clark warned that agents could take over certain jobs and noted fears around routine tasks. Yet he also made a more nuanced point about human advantage. “People that are creative and can think broadly, people that read a lot, people that have interests are the ones most benefited by this,” he said.
That is an important message for project professionals. AI may reduce the value of repetitive production work, but it can increase the value of framing, judgement, curiosity and cross-functional thinking. The brake pedal should protect people and systems while allowing that higher-value work to emerge.
Regulation will not be enough on its own
Clark compared AI with the oil boom and the creation of regulatory frameworks that gave society more confidence in a powerful industry. The analogy is imperfect, but it points to a familiar pattern. Private innovation often moves faster than public governance. Regulation then tries to catch up.
Waiting for regulation is not a strategy. Built environment organisations already manage regulated risks in safety, design, environment, finance, data protection and procurement. AI should be added to that governance map now.
A practical AI brake pedal could include approval gates for high-risk use cases, clear rules on confidential data, output review standards, model selection policies, incident reporting, supplier transparency requirements and red lines for safety-critical decisions. Good governance gives leaders a way to slow a use case before it causes damage.
AI failures are rarely isolated technical events. A bad summary can affect a commercial decision. A flawed code suggestion can break a reporting workflow. A hallucinated compliance point can mislead a design review. A poorly governed agent can expose sensitive data.
How project leaders can build brakes now
The most useful controls are often simple.
•Create a register of AI use cases, including the teams, tools, data types and decisions involved.
•Classify use cases by risk, with special treatment for safety, legal, financial, personal data and operational systems.
•Require human approval before AI outputs are used in contractual, safety-critical or client-facing decisions.
•Keep logs of AI-assisted workflows where outputs affect project records.
•Ask suppliers how their AI features handle data retention, model training, audit and security.
These steps sound basic, but many organisations have not implemented them. That leaves AI adoption dependent on individual judgement. Individual judgement is important, but it is not a governance system.
The brake pedal also needs cultural support. Teams should be able to pause an AI workflow without being labelled as blockers. People should report errors without embarrassment. Leaders should measure value after quality review, not only by hours saved.
The connection to cybersecurity
Clark’s warning also sits beside Anthropic’s work on Project Glasswing, where AI has been used to find more than 10,000 high or critical severity vulnerabilities across partner codebases. This shows both sides of the issue. AI can improve defence by finding flaws. It can also change the threat environment by making discovery faster.
That dual-use character is why brake pedals matter. The same underlying capability can be beneficial or harmful depending on access, intent, controls and context. Organisations using AI in infrastructure environments need to think about capability and misuse together.
The built environment is especially exposed because its digital systems increasingly touch physical outcomes. Buildings, rail networks, utilities, hospitals, factories and campuses all depend on software. If AI accelerates both software creation and vulnerability discovery, resilience becomes a board-level issue.
A serious conversation, not a moral panic:
Clark told the BBC: “I am worried for my kids if we as a society don’t have a serious conversation about what the implications of AI’s continued advances mean.” The phrase “serious conversation” is doing important work. It is not a demand for panic. It is a demand for maturity.
Mature AI adoption accepts three things at once: the technology is useful, the risks are real, and governance must be designed before the most powerful capabilities are routine. Project organisations are well placed to understand this because they already manage complex systems under uncertainty.
The brake pedal for AI will not be one mechanism. It will be a combination of law, standards, procurement discipline, technical controls and professional judgement. The organisations that build those habits early will be able to use AI with more confidence than those that rely on enthusiasm alone.
For weekly analysis that treats AI adoption as a leadership and delivery challenge, subscribe to Project Flux and share it with the colleagues shaping your next governance conversation.
Key takeaway
• AI governance needs a slowing mechanism: Jack Clark’s brake pedal warning is useful because adoption is accelerating faster than many control systems.
• Project organisations should define stop conditions: Teams need clear thresholds for when an AI use case is paused, escalated, reviewed or removed from workflow.
• Responsible use is a management discipline: Policies, audit trails and accountable ownership matter as much as model capability when AI enters delivery processes.
Links and stuff
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
/
1