AI Won't Slow Down. Now What?

When the people building the most powerful AI systems in the world start warning that development may be moving too quickly, it is worth paying attention.

But what does “slow down” actually mean? Should model development stop altogether? Is the real risk the technology itself, or the fact that organisations are adopting it without understanding their data, responsibilities and consequences?

In the 50th episode of Impact of AI: Explored, we explored these questions with four excellent guests: Paul Slater, Louise Humpington, Dan Boyles and Femke Cornelissen.

This was a particularly special episode for us. James O'Regan and Gerjon Kunst have now reached episode 50, and we used the occasion to bring together a group of people with different perspectives on AI strategy, governance, technology, adoption and the future of work.

Our central question was simple:

“If the people building frontier AI systems say the race is moving too fast, should we listen?”

The answer from everyone on the panel was yes—but with some important qualifications.

Meet the Guests

Paul Slater

Paul Slater is an AI enablement advisor, author, speaker and host of the Humanity Working podcast. He spent more than a decade at Microsoft, where he led global digital transformation initiatives and worked on strategy for a major life sciences business.

Today, Paul focuses on helping organisations prepare for an AI-shaped future by combining technology adoption with human transformation. His work covers AI readiness, workforce enablement, leadership, governance and organisational change.

Paul is also the author of The AI-Ready Human, a book about the human capabilities people need to remain relevant as technology transforms work. He argues that adaptability, judgment, resilience and human creativity will become increasingly important as AI becomes more capable.

Louise Humpington

Louise Humpington brings a governance and legal perspective to the conversation. In the episode, she described herself as a “cynical former lawyer”—which turned out to be a useful perspective when discussing responsibility, liability and the motivations behind calls to slow AI development.

Her contribution focused on what safe AI should mean in practice. That includes assurance frameworks, ethical benchmarks, independent testing, human oversight and clear accountability when systems cause harm.

Louise's broader work sits at the intersection of AI, business, governance and organisational responsibility. Her perspective is particularly valuable because she looks beyond what AI can do and asks what organisations are legally, ethically and socially allowed to do with it.

Dan Boyles

Dan Boyles is Head of AI at Hello:AI Collective and a global AI consultant and Microsoft Copilot expert. He has worked on more than 50 AI and Copilot deployments across the UK and Europe, with a particular focus on financial services.

Dan specialises in the practical intersection of AI, compliance, security and business value. He is also a former Microsoft Global Master Trainer and is known for explaining AI adoption in a direct, jargon-free way.

During the episode, Dan repeatedly brought the conversation back to what happens in real organisations. His key concern was not necessarily whether the next frontier model would be released tomorrow, but whether businesses understand where their data is going, who owns the outcomes and what happens when something goes wrong.

Femke Cornelissen

Femke Cornelissen works in AI transformation and adoption and brings extensive experience from inside large organisations. She is particularly interested in how companies redesign their operating models, roles and governance structures for an AI-enabled future.

Femke's contribution focused on the gap between the speed of technological development and the speed at which organisations develop the skills, culture, leadership and governance needed to use AI responsibly.

She also highlighted an emerging challenge: the transition from personal AI to team AI and, ultimately, enterprise AI. As people create more agents and automated workflows, organisations need to know who owns them, what data they use and what happens when the person who built them leaves the business.

Setting the Stage

AI development is moving quickly. Frontier labs are competing to build more capable models, while businesses are being encouraged to adopt AI tools, agents and copilots at an equally rapid pace.

At the same time, the people closest to the technology are warning that the current speed may not be sustainable. That warning could be driven by genuine safety concerns, concerns about cybersecurity, worries about uncontrolled capability growth—or more commercial and legal considerations.

As Louise pointed out, we should always ask what sits behind these statements:

• Why is this being said now?
• Who benefits from the message?
• Is the motivation safety, regulation, liability or competitive positioning?
• What exactly do we mean when we say “slow down”?
• Who would actually be slowed down?

Those questions matter because “pause AI development” can mean many different things. It could mean stopping the training of new frontier models, delaying public releases, introducing more testing before deployment, or creating time for governments and organisations to establish better safeguards.

For most businesses, however, a pause in frontier model development would not immediately change their daily work. Many organisations are still trying to understand their existing data, implement sensible security controls, configure Copilot properly and identify use cases that create genuine business value.

That creates an important distinction between the AI conversation taking place online and the reality inside most organisations.

In this blogpost, we explore the main themes from the episode: why a pause might be useful, why governance must become preventative rather than reactive, why AI adoption should begin with business challenges, and why humans—not models—may become the most important source of competitive advantage.

Episode Highlights

“Zero risk” is not a serious position

Paul's first major point was that claiming there is a zero per cent chance of serious harm from advanced AI says more about the person making the claim than it does about the technology.

He was not arguing that AI will definitely destroy humanity. Instead, he challenged the certainty with which some people dismiss the possibility of serious consequences.

His position was more nuanced:

“We should listen to warnings about AI risk, but we should also remain sceptical about the motivations behind those warnings.”

The issue is not whether we can predict the precise probability of catastrophe. The issue is whether we are willing to acknowledge uncertainty and act responsibly despite it.

The most important question may be: “Who owns the mistake?”

Dan brought the discussion back to a question that applies to every AI deployment:

“Who owns the mistake?”

It is easy to become excited about what an AI system can do. It is much harder to answer who is responsible when the system makes a bad recommendation, exposes sensitive information, discriminates against someone or produces an incorrect result that is acted upon.

That question applies at several levels:

• Who owns the data?
• Who owns the AI agent?
• Who approves the output?
• Who monitors the system?
• Who is responsible when the system behaves unexpectedly?
• Who can turn it off?

For Dan, this is why responsible adoption is not about chasing the newest model. It is about understanding the technology already available and ensuring that it can be used safely.

“The human makes the AI work”

One of the strongest moments in the conversation came when Paul argued that humans may become the key differentiator in an AI-enabled economy.

If every organisation has access to broadly similar models, tools and agents, the technology itself becomes less of a competitive advantage. The difference comes from how well people understand the business, make decisions, manage risk and apply judgment.

Dan immediately recognised the value of the idea and joked that he was going to borrow it for a presentation.

The irony of AI is that the human remains central. The quality of the outcome depends not only on the model, but also on the person defining the problem, selecting the data, evaluating the output and deciding what to do next.

Deep Dive: Why Slowing Down Could Help Organisations Catch Up

The debate about slowing down AI often focuses on frontier models, geopolitical competition and the risk of damaging economic growth.

But there is another perspective that received less attention in the wider public debate: slowing down could give organisations time to catch up.

AI companies are increasingly using AI to help build the next generation of AI systems. That creates a development cycle that is difficult for most organisations to follow. Businesses are still learning how to govern one generation of tools while the market is already promoting the next.

In practice, many companies have not yet completed the basics:

• They do not know which teams are using AI.
• They have not mapped where business data is being sent.
• They lack clear ownership for agents and automated workflows.
• Their employees have not received adequate AI literacy training.
• They have not defined acceptable use policies.
• They have not established review processes for AI-generated content.
• They are unsure who is accountable for errors.
• They have not connected AI adoption to their actual business strategy.

This is the gap between AI capability and organisational readiness.

Femke described a progression from personal AI to team AI and enterprise AI. A person may create an AI assistant to help with their own work. The next step is for a team to share agents and workflows. Eventually, those systems may become important enterprise capabilities.

That progression introduces new governance questions. If an employee builds an agent and then leaves the organisation, who owns it? If an agent uses sensitive information, who approved that access? If several teams create similar solutions, who decides which one becomes the standard?

These questions are not theoretical. They are already appearing inside organisations that have moved from experimentation to scaled adoption.

From Reactive Governance to Preventative Governance

Louise made a particularly important distinction between reactive and preventative governance.

At the moment, governance often functions as a clean-up exercise. Something goes wrong, and organisations then investigate what happened, who was responsible and how to prevent a repeat.

That is not enough for AI systems that can operate at scale and speed.

A stronger approach would require models and AI-enabled systems to meet agreed safety, security, cybersecurity and ethical benchmarks before they reach the market or are deployed in high-impact situations.

That could include:

• Independent testing.
• Documented assurance processes.
• Clear human oversight.
• Defined limits and operating conditions.
• Transparent ownership.
• Monitoring after deployment.
• The ability to suspend or shut down the system.
• Evidence that employees understand both the benefits and risks.

The EU AI Act was discussed as one example of the direction of travel. Louise highlighted the principle that just because a system can perform a task does not mean an organisation should use it without proper oversight.

In some cases, the law may increasingly move from “you should not do this” to “you are not allowed to do this unless you can demonstrate that it is safe, fair and properly controlled.”

Real-Life Stories & Examples

The Business That Wanted to Deploy an Untested Model

Dan described working with organisations that were still hesitant to enable Copilot because they were concerned about data, security and governance.

At the same time, some businesses were considering deploying brand-new open-source models directly into their work environments without fully understanding the implications.

His response was straightforward: stop and assess the situation first.

The fact that a model is technically impressive does not make it appropriate for a business environment. Before deploying it, an organisation needs to understand where the data goes, what retention policies apply, how the model is secured and what regulatory requirements affect the use case.

The Recruitment System That Screened Out Older Candidates

Louise shared an example involving AI-assisted recruitment.

A talent acquisition team instructed an AI system to look for “youthful, energetic enthusiasm” and a “dynamic profile.” The system interpreted those instructions in a way that effectively favoured younger candidates. It screened out people whose university or college education occurred before a particular year.

The organisation then had to review thousands of CVs manually to determine whether it had exposed itself to discrimination claims.

The lesson was clear:

“Discrimination is still discrimination, even when it is deployed at scale.”

An AI system does not remove an organisation's responsibility. In fact, automation can increase the scale of the harm and make it more difficult to identify where the original judgment went wrong.

The Robot Dog at the Dog Show

Louise also told a much lighter story. During a family camping trip, a robot dog appeared at a real dog show.

The actual dogs reacted with complete suspicion, barking at the robot and refusing to accept it as one of their own. The robot dog was eventually disqualified because it was not, in fact, a real dog.

The joke that followed was whether this could be the answer to AI governance: simply disqualify AI because it is not human.

Of course, the problem is not that simple. But the story captured an important human reaction to AI. People do not automatically trust systems merely because they appear intelligent or capable.

AI Can Amplify Stupidity as Well as Intelligence

James shared an example of a company responding to a customer complaint by accidentally publishing the prompt used to generate the response.

Instead of producing a carefully reviewed message, the organisation effectively exposed something like: “Write a sympathetic email to the customer.”

That prompted a wider observation: AI amplifies whatever is already present. It can amplify expertise, but it can also amplify poor thinking, weak processes and bad decisions.

Paul extended this idea to the growing volume of AI-generated content. As more people use the same models, online writing, presentations and websites can begin to look increasingly similar.

He described this as the “claudification” of content: a world where AI-generated patterns become so common that people either copy them or deliberately try to avoid them.

The risk is not only that the internet becomes full of generic content. It is also that people stop developing the human qualities that make creative work valuable: taste, judgment, wisdom and originality.

Those qualities develop over time. They cannot simply be generated on demand.

The AI Expert Problem

The panel also discussed the sudden growth in people presenting themselves as AI experts.

Some of these people are genuinely helping their organisations experiment and solve problems. Others may be making business-critical decisions without sufficient experience in security, data, compliance, architecture or organisational change.

Dan compared this with earlier waves such as cloud, security and GDPR. In those areas, organisations typically looked for people with established expertise in the relevant disciplines.

AI should not be treated as completely separate from those disciplines. Effective AI adoption requires a combination of expertise:

• AI and machine learning.
• Data architecture.
• Cybersecurity.
• Privacy and compliance.
• Software engineering.
• Cloud architecture.
• Business strategy.
• Organisational change.
• Human resources.
• Marketing and communication.
• Governance and risk management.

Louise described this as a “mycelium network of governance”: a connected system in which knowledge and responsibility flow between different specialists rather than remaining trapped in organisational silos.

That approach matters because AI risks often appear between disciplines. A model may be technically sound but legally problematic. A use case may be commercially attractive but impossible to govern. A deployment may be secure but rejected by employees because they were not involved in the change.

The right people need to be in the room together.

Key Takeaways

• We should listen when frontier AI developers warn that development is moving too quickly, but we should also question their motivations and commercial interests.
• “Slow down” needs to be defined clearly. It could mean pausing training, delaying releases, increasing testing or giving organisations time to establish proper governance.
• Most organisations are not waiting for the next frontier model. They are still trying to understand and use the tools they already have.
• The biggest immediate risk is often not artificial general intelligence. It is poor implementation, weak data governance and unclear accountability.
• Every AI deployment should answer the question: who owns the mistake?
• AI adoption should start with a business challenge, not with a desire to use the newest model.
• Productivity alone is not a sustainable competitive advantage if every organisation has access to the same tools.
• Human judgment, strategy, creativity and wisdom may become more important as AI becomes more capable.
• Governance needs to become preventative rather than merely reactive.
• Organisations need clear owners for AI agents, workflows, data and decisions.
• AI literacy is essential. Employees need to understand both what AI can do and where it can fail.
• Just because AI can perform a task does not mean an organisation should deploy it without human oversight.
• AI can amplify intelligence, but it can also amplify poor thinking, bias and organisational dysfunction.
• Responsible AI requires a combination of technical, legal, security, data, business and human expertise.
• Organisations should be able to turn AI off and continue operating. If they cannot, they may be becoming dependent on a system they do not control.

Closing Thoughts

The most striking thing about this conversation was that every guest said they would support some form of pause in AI development.

That does not mean they want to stop innovation permanently. It means they want to create the conditions for innovation to continue responsibly.

Louise would want agreed benchmarks and assurance frameworks before models reach the market. Femke would want evidence that the technology creates real value before organisations scale it. Paul would want international cooperation and meaningful external oversight. Dan would want proof that organisations can turn AI off and continue operating without collapsing.

Those answers are different, but they point in the same direction.

We do not necessarily need less innovation. We need better foundations around the innovation we already have.

For us, the biggest lesson from this episode is that AI strategy should not begin with the question, “Which model should we use?” It should begin with:

• What problem are we trying to solve?
• What data is involved?
• Who is affected?
• Who owns the outcome?
• What could go wrong?
• How will we know whether it is working?
• Can we explain the decision?
• Can we turn it off?

This was the 50th episode of Impact of AI: Explored, and we could not have asked for a better discussion to mark the occasion.

Thank you to Paul, Louise, Dan and Femke for joining us, and thank you to everyone who has listened, watched, shared and challenged us over the first 50 episodes.

What do you think? Should AI development be paused, slowed down or simply governed more effectively? And if you would pause it, what would need to happen before pressing “unpause”?

Join the conversation—and we will see you in episode 51.