Loading the Elevenlabs Text to Speech AudioNative Player…

Artificial intelligence is accelerating, the companies building it are rewriting their own safety rules, and some researchers believe an intelligence explosion could begin as early as 2027. The technology is impressive, the money involved is extraordinary, and the governance looks increasingly uncomfortable. But how close are we really, and what should business leaders be doing about it?

We may be approaching an AI tipping point, but it’s not the one everyone expected

The AI industry has developed an extraordinary ability to make the future sound both inevitable and approximately six months away.

Every major announcement seems to bring us closer to artificial general intelligence, autonomous businesses or machines that can outperform human researchers. Meanwhile, plenty of organisations are still struggling to connect their CRM properly to their website.

Both realities can exist simultaneously.

What makes late 2026 particularly interesting isn’t another chatbot passing another examination. It’s the convergence of three developments that could fundamentally alter how artificial intelligence evolves.

First, AI is becoming embedded in real commercial operations. Second, the systems themselves are getting better at carrying out increasingly complex tasks, including helping researchers develop better AI. Third, the voluntary safety commitments designed to control these advances are proving increasingly vulnerable to competitive pressure.

That last point deserves considerably more attention than it receives.

The people building the world’s most powerful AI systems are also, to a considerable extent, deciding how dangerous those systems are allowed to become.

It’s an interesting governance arrangement. Imagine the aviation industry announcing that aircraft manufacturers could independently decide when their safety standards had become commercially inconvenient.

Yet something remarkably similar is playing out in frontier AI.

What does an AI tipping point actually mean?

An AI tipping point occurs when technological improvements begin creating structural changes in business, investment, capability or competitive behaviour.

It doesn’t necessarily mean artificial general intelligence has arrived. It means the consequences of AI development are beginning to extend beyond individual tools.

Research cited in Broadcom’s Private Cloud Outlook 2026 suggests that organisations are already reassessing where AI workloads should operate. The survey reported that the proportion primarily using public cloud for production AI inference fell from 56% in 2025 to 41% in 2026.

Cost has become a major concern.

That shouldn’t surprise anyone who has spent time implementing AI commercially. Experimenting with a model is relatively inexpensive. Running thousands of automated processes continuously, with large context windows, repeated reasoning calls and multiple agents, can become a different financial proposition.

The boardroom conversation is changing from “What can we do with AI?” to “What are we actually spending, and what is it producing?”

A sensible question, albeit one that would have been equally useful before signing the enterprise licence agreement.

Marketing faces its own tipping point.

Research and modelling cited in the supplied material estimate that AI-generated answers displaced approximately 17% of organic search clicks by the second quarter of 2026, with substantially greater disruption forecast.

These are model-based estimates, not a definitive measurement of search traffic across the entire internet. Nevertheless, the direction of the challenge is commercially significant.

Businesses have spent decades optimising websites to appear in search results. Increasingly, they must also consider whether their expertise, products and brands are represented accurately within AI-generated answers.

The importance of credible, structured and authoritative content is unlikely to diminish.

The uncomfortable part is that being visible to an AI assistant doesn’t necessarily mean receiving a website visit.

Bottom Line: The commercial tipping point isn’t simply more AI adoption. It’s the moment organisations must redesign costs, processes, customer acquisition and governance around technology that is improving faster than many businesses can adapt.

How quickly are frontier AI systems improving?

One of the more useful measures of AI progress comes from METR, the independent organisation studying model evaluation and threats.

Instead of relying entirely on traditional benchmarks, METR examines how long a task would take a human expert and measures whether an AI agent can complete it successfully.

Its research suggests that the length of tasks frontier systems can complete has been increasing rapidly.

By May 2026, the research cited in the source material placed leading models near a 12-hour task horizon at a 50% success rate.

That sounds extraordinary, and in some respects it is.

But there’s a rather important distinction between completing a twelve-hour task half the time and employing an autonomous worker who can be trusted to operate for twelve hours.

Businesses generally prefer their systems to work more often than they fail.

METR’s research indicates that task horizons at an 80% success rate are substantially shorter. The organisation also highlights limitations in evaluating the longest tasks.

A separate study involving experienced open-source developers found that participants using early-2025 AI tools took 19% longer to complete assigned work, despite believing the technology had made them faster.

This doesn’t mean AI coding tools are ineffective. It demonstrates how easily perceived productivity and actual productivity can diverge.

And that distinction matters enormously when companies start making staffing, investment and operational decisions based on benchmark headlines.

A tool that writes code quickly is useful. A system that independently understands requirements, writes dependable software, tests it, identifies unexpected problems and takes responsibility for the result is something else entirely.

We are progressing towards the latter, but the evidence doesn’t justify pretending we have already arrived.

What are Responsible Scaling Policies, and why are they under pressure?

Responsible Scaling Policies, or RSPs, were introduced to help frontier AI developers manage potentially catastrophic risks as their models became more capable.

Anthropic introduced its framework in 2023. OpenAI developed its Preparedness Framework, while Google DeepMind established its Frontier Safety Framework.

The basic principle was sensible.

If an AI system approached capabilities that could create severe risks, developers would need to demonstrate adequate safeguards before continuing development or deployment.

In some circumstances, that originally meant pausing.

The difficulty is what happens when a safety commitment encounters commercial competition.

In February 2026, Anthropic revised its policy, removing its original unilateral pause commitment. Its reasoning reflected a genuine collective-action problem: if one developer stops while less cautious competitors continue, the cautious developer may lose influence without reducing the overall threat.

That argument isn’t entirely unreasonable.

It’s also precisely why voluntary commitments become fragile when commercial pressure increases.

A September 2026 research preprint, cited in the supplied material, examined revisions to frontier AI safety frameworks. It reported that 75% of traced commitment edits weakened or removed previous commitments, with many material changes insufficiently disclosed in public summaries.

These findings concern the revisions examined, not necessarily every commitment made by every AI developer. Nevertheless, they raise an important question.

Can a safety framework meaningfully constrain an organisation if that organisation can rewrite the framework when the constraint becomes inconvenient?

The source research also identifies concerns about limited independent scrutiny, redacted safety assessments and the use of competitor-relative risk thresholds.

Under such approaches, acceptable risk can become partly dependent on what competing systems can already do.

That risks turning safety into a moving target.

If everyone accelerates together, nobody necessarily considers themselves unusually reckless.

For business leaders, this should sound familiar. A company can have an impressive governance policy and still discover that the policy becomes remarkably flexible when a major contract is at stake.

The difference here is the potential scale of the consequences.

What is an intelligence explosion, and could it happen in 2027?

An intelligence explosion describes a hypothetical process in which artificial intelligence becomes capable of improving the technology used to create more intelligent AI systems.

The concept dates back to British mathematician I. J. Good in 1965.

The underlying argument is relatively straightforward.

If developing smarter machines requires intelligence, and machines become capable of performing that research, they could begin improving the next generation of systems themselves.

Those improved systems could then accelerate further research.

In theory, the process creates a self-reinforcing cycle of rapidly increasing capability.

It’s important to understand that this doesn’t require AI to become better than humans at every job simultaneously.

It may only require AI to become sufficiently capable at the specialised work involved in designing, training, testing and improving AI models.

This is why autonomous AI research has become such an important milestone.

The influential Situational Awareness essay by Leopold Aschenbrenner and the subsequent AI 2027 scenario developed by Daniel Kokotajlo and colleagues brought particular attention to 2027 as a potential turning point.

Their arguments focus on accelerating compute investment, algorithmic improvements and the possibility of large numbers of AI research agents operating in parallel.

These are scenarios and forecasts, not established outcomes.

The distinction matters because the term “intelligence explosion” has a remarkable ability to turn speculative timelines into apparent calendar appointments.

Is the evidence supporting an AI intelligence explosion in 2027?

The evidence is mixed, which is considerably more interesting than either blanket dismissal or breathless certainty.

On the supporting side, investment in AI infrastructure has grown enormously.

The supplied research places estimates for major technology-company capital expenditure in 2026 between approximately $650 billion and $830 billion.

Algorithmic efficiency has also improved significantly, allowing increasingly capable systems to operate with fewer computational resources than earlier approaches would have required.

Perhaps most importantly, leading AI laboratories are already using AI to accelerate aspects of their own research.

A November 2026 METR assessment cited in the research estimated that Anthropic’s internal benchmark progress had increased to approximately 1.9 times its earlier rate.

That is meaningful evidence of AI-assisted research acceleration.

However, it isn’t proof that autonomous AI researchers have achieved recursive self-improvement.

Much of the reported progress still involves humans directing, evaluating and integrating AI-generated work. Key supporting evidence was also not publicly available in full.

The other side of the argument is equally important.

The AI 2027 scenario authors themselves reportedly assessed progress during 2026 as running at approximately 65% to 75% of their projected pace.

Reliable autonomous research remains a major outstanding challenge.

Physical infrastructure also imposes constraints. More powerful systems still require chips, electricity, cooling, engineering expertise and increasingly complex verification.

Software may improve rapidly, but constructing a new power station remains stubbornly resistant to prompt engineering.

The sensible conclusion is that rapidly accelerating AI-assisted research is increasingly credible, while a full intelligence explosion beginning in 2027 remains uncertain.

An event doesn’t need to happen in 2027 to become strategically important.

If major capability thresholds arrive in 2029 or 2032 instead, businesses and governments could still face extraordinary disruption.

Why weaker AI governance matters to every business

It’s tempting to regard frontier AI safety as something for governments, research laboratories and people who enjoy discussing existential risk over dinner.

That would be a mistake.

The consequences of unreliable or poorly governed AI are already relevant to everyday business decisions.

Organisations are giving AI systems greater access to customer information, internal documents, software repositories, communications and operational workflows.

As agents become more capable, they’ll also be given greater authority to act rather than simply advise.

This creates a responsibility gap.

Who is accountable when an automated system makes an expensive mistake? How is its behaviour monitored? What happens when its capabilities change following a model update?

And, perhaps most importantly, can the organisation actually stop it?

An AI policy that exists only as a document is not operational governance.

Boards should be interested in the answers long before something goes wrong, rather than discovering a passionate new commitment to AI risk management during an emergency meeting.

There is also a competitive question.

If leading AI developers accelerate their research while reducing some voluntary safeguards, businesses may experience more frequent model releases, shifting capability boundaries and less predictable operational risks.

Technology procurement therefore becomes partly a question of resilience.

Choosing an AI supplier isn’t just about finding the model that produces the most impressive demonstration.

It’s about understanding reliability, data access, monitoring, security, contractual responsibilities and the ability to move between providers when circumstances change.

Bottom Line: Organisations can’t control the global AI race, but they can control how much authority they delegate to systems they don’t fully understand.

What business leaders should do before 2027

The sensible response isn’t to predict the precise arrival date of artificial general intelligence. It’s to build an organisation capable of adapting as AI improves.

Five priorities stand out:

  1. Measure real productivity. Evaluate time saved, quality, error rates, supervision requirements and total operating costs rather than relying on demonstrations or employee perceptions.
  2. Introduce autonomy gradually. Start with bounded workflows, clear permissions and human approval for consequential actions.
  3. Develop practical AI governance. Establish ownership, monitoring, escalation processes, vendor assessments and clear rules for sensitive information.
  4. Reassess commercial exposure. Examine how AI-generated search answers, automated customer interactions and new competitors could affect revenue and customer relationships.
  5. Plan for several timelines. Consider what happens if powerful autonomous systems emerge within two years, five years or considerably later. Investments should remain useful across more than one scenario.

This is ultimately a leadership challenge.

The organisations that succeed won’t necessarily be those making the loudest claims about their AI strategy.

They’ll be the ones that understand where the technology creates measurable value, where it introduces unacceptable risk and where existing business models may need to change.

The AI future is arriving unevenly, and that’s what makes it dangerous to underestimate

The debate about an intelligence explosion in 2027 tends to produce two camps.

One believes we’re approaching an unprecedented technological acceleration that could fundamentally alter civilisation within a few years.

The other points, reasonably, to unreliable agents, exaggerated benchmarks and commercial promises that haven’t yet translated into consistent results.

Both positions contain useful observations.

The mistake is assuming that one must be entirely correct.

AI can be genuinely extraordinary at specialised intellectual tasks while remaining surprisingly unreliable in ordinary business situations.

Frontier laboratories can make impressive advances in automated research without having solved recursive self-improvement.

And companies can publish sophisticated safety frameworks while facing incentives that make their commitments harder to maintain.

The more immediate concern isn’t whether an intelligence explosion begins on schedule in 2027.

It’s whether our institutions, businesses and governance systems can keep pace with increasingly powerful technology, particularly when the companies developing it are under pressure to move faster.

We don’t need to know the exact date of the next breakthrough to recognise the challenge.

The real tipping point may arrive when AI becomes powerful enough to change our decisions faster than we’re capable of understanding their consequences.

That’s a much more important conversation than guessing the year a machine becomes smarter than its creator.

And unlike the predicted intelligence explosion, it’s a conversation we can start doing something about today.

Frequently Asked Questions
What does an AI tipping point actually mean?

An AI tipping point occurs when technological improvements begin creating structural changes in business, investment, capability or competitive behaviour.

What are Responsible Scaling Policies?

Responsible Scaling Policies were introduced to help frontier AI developers manage potentially catastrophic risks as their models became more capable.

Is there evidence supporting an AI intelligence explosion in 2027?

The evidence is mixed, with significant investment in AI infrastructure and improvements in algorithmic efficiency, but reliable autonomous research remains a major challenge.

AIG Agents
What is an AI Agent?AIAI Insights

What is an AI Agent?

Damon SegalDamon SegalMarch 25, 2025
AI Hardware
The Interplay of Hardware and Energy in Advancing Artificial IntelligenceAIPhysical AITech

The Interplay of Hardware and Energy in Advancing Artificial Intelligence

Damon SegalDamon SegalJanuary 31, 2025
AI News 31 January 2025
This Week in AI, AGI, and ASI: The Latest DevelopmentsAI News

This Week in AI, AGI, and ASI: The Latest Developments

Damon SegalDamon SegalFebruary 1, 2025
The owner of this website has made a commitment to accessibility and inclusion, please report any problems that you encounter using the contact form on this website. This site uses the WP ADA Compliance Check plugin to enhance accessibility.