Why is a genomics director pushing back on the AI speed-up story?
Professor Matt Hurles, Director of the Wellcome Sanger Institute, wrote on the institute’s blog on 2 November that the research community should set the optimism about AI accelerating science 10-100x, and “curing all diseases”, alongside a sober look at the bottlenecks in the way. His reason is practical. A scientific project is a chain of tasks, and AI is likely to make some links far quicker than others. Where a fast task hands work to a slow one, the queue simply moves. For anyone running a lab, a funding body or a research office, his piece reads as a list of the places that queue will form.
Strategic Insight: Hurles does not dispute the potential. He calls himself “hugely optimistic” about AI’s long-term effect on science. His argument is that the gains depend on fixing the slow links, and that many of those links are institutions rather than technology.
What is the real constraint on AI-accelerated science?
Hurles sorts the bottlenecks into two kinds. Some are physical, and technology may solve them. Others are psychological, social and institutional: people, teams and organisations each have their own concerns and their own appetite for change. On his account, technical fixes on their own will not be enough.
Why a single speed-up multiplier is the wrong unit
The headline multiplier treats science as one activity. Hurles describes it as many pipelines. Many tasks, such as idea generation, securing funding, data analysis and publication, are shared across kinds of science, but physical experiments, regulatory approval and access to biological material are not shared between, say, biology and mathematics. Within biology itself he separates observational work characterising samples and data gathered from nature from self-contained lab experiments, and he notes that most projects need coordination inside and across institutions.
He then applies the idea of “jagged” AI: tools whose competence varies hugely from task to task, measured against what people expect to be hard. The result, in his view, is uneven acceleration. Some tasks will speed up dramatically, and so will the fields made mostly of those tasks. He cites the current AI revolution in mathematics as a case where many of the tasks lend themselves to AI tools.
| Metric | Value | Strategic Implication |
|---|---|---|
| Claimed AI speed-up of science | 10-100x | Hurles calls this talk optimism and hyperbole, to be weighed against the bottlenecks |
| Near-term bottlenecks he lists for collective attention | 8 | Funding, regulation, contracting, lab automation, sensitive data, peer review, publishing and careers |
| ”Relatively dystopian” scenarios he sets out | 11 | Several describe a process breaking down under increased demand or failing to adapt |
| Suggestions for where effort should go | 6 | He calls the list preliminary and expects it to change over time |
Reality Check: Hurles offers an early sign of the new pressure already visible: the growing volume of written applications, whether for jobs, PhD places or grants, which is putting more strain on review and decision-making.
What happens when one task speeds up and the next does not?
The core of the piece is a prediction. Jagged acceleration will create new bottlenecks and enlarge existing ones, most visibly where an AI-accelerated task feeds a task that has not yet sped up. Hurles thinks the community already has an initial picture of which tasks current AI tools speed up most, so the near-term pinch points can in principle be predicted with some confidence.
Some institutions are already acting. He gives two examples: AI companies investing in automated labs, and funders changing how they run grant funding. His worry is that this effort is fragmented.
Two reasons the bottlenecks may be missed
The first concern is that the community is not looking at the whole system. Each type of institution tackles the bottlenecks that are visible to it, that fit its mission or business model, and that it is equipped to fix. Funders work on funding; technology companies work on technology. The predictable result, he argues, is that some bottlenecks get less attention than others.
The second concern is that the community’s model of how science gets done is thin. He describes current thinking about scientific pipelines as “implicit, simplistic and theoretical” rather than explicit and empirical, and says recent essays on AI and science have often leaned on Popperian ideas, putting generality ahead of the specifics of real projects.
Critical Context: Hurles includes himself in this. He invites scientists to list the task types in their last three substantive projects and how long each took, and admits: “I literally have no idea what my answers to these questions would be.” His point is that, without data on how past projects actually ran, scientists are judging where AI will help by intuition.
Who is affected?
The bottlenecks Hurles names, and his scenarios for how they could go wrong, fall on distinct parts of the research system.
| Group | What changes for them | What is still unknown |
|---|---|---|
| Research funders | His first scenario has funders swamped by applications and making slow, poor decisions | Whether changes to funding processes keep pace with application volumes |
| Regulators | Approval processes could break down under rising demand, or fail to adapt to new AI-enabled risks such as biosecurity | Whether faster approval and adequate safeguards can both be achieved |
| Universities and hospitals | Projects that need contracts between institutions could be badly slowed by contracting delays | How quickly inter-institution agreements, such as those for sharing resources, can be made faster |
| Peer review and publishing | Review could break down under demand, letting “scientific slop” grow or holding up breakthroughs | Whether AI-based review can assure quality at scale |
| Research participants | Trust could fall that the systems opening genetic and health data to researchers are secure | How cutting-edge AI analysis of sensitive data can keep confidentiality and privacy standards high |
None of these five is a laboratory problem. They are processes run by people and institutions, and several of his scenarios turn on those processes meeting a sharp rise in demand.
⚠️ Warning: Hurles’ scenario list is deliberately bleak, but each item is a mechanism, not a mood. Regulatory and peer review breakdowns appear in two directions: dangerous science or low-quality output getting through, or genuine breakthroughs held up.
What this means for leaders of research institutions
Hurles is direct about existing organisations. Existing institutions, and above all those with a large share of core funding, in his words, “will need to adapt their speed of decision-making and flexibility of resource allocation”. He also wants institutions, old and new, to draw on behavioural science, psychology and sociology to think through how research communities adapt.
His first suggestion is to build a far more detailed, empirical picture of the tasks, run in sequence and in parallel, that make up different kinds of science, end to end.
Competitive Reality: Hurles argues that research ecosystems which deal with these bottlenecks and risks early will gain an outsized competitive advantage during this period of rapid change.
The open challenges
Too few kinds of research institution
Hurles thinks the current range of scientific institutions may be too narrow for society to adapt quickly. He is more concerned about non-profits than about companies. There is broad global agreement on what a university is, whilst for-profit organisations keep experimenting with business models. One of his scenarios is the failure of the “bundled” university model, leaving a major imbalance between for-profit and non-profit institutions at the frontier of science, with fewer public goods and more concentrated, siloed knowledge. His answer is more diversity among research institutions, particularly non-profits.
Coordinating with non-human agents
Institutions exist, he notes, to coordinate individual effort and achieve what individuals cannot manage alone. What is different this time is that the “individuals” include non-human agents. He suggests this could drive larger and faster institutional change than earlier technological revolutions, and he expects real difficulty in coordinating work inside institutions and across them.
Trust, sovereignty and unequal access
Several of his scenarios reach beyond the lab. Public trust in institutions could fall because of harmful social effects, concentrations of power without legitimacy or accountability, and institutions whose interests diverge from those they serve. Ideas of AI and data “sovereignty” could silo knowledge and resources within nation states and slow science worldwide. And AI could widen the gap between research systems in richer and poorer countries.
Hidden Cost: In Hurles’ sovereignty scenario, keeping AI, data and resources within national borders slows scientific progress globally.
What to watch
- Funders’ decision times and processes. Hurles names research funders adapting grant processes as one response already under way, and funders overwhelmed by applications as his first scenario. Slower or less reliable decisions as application volumes rise would suggest that scenario is materialising.
- Regulatory approval speed and biosecurity. He lists how regulatory approvals work, and how long they take, as a near-term bottleneck, and failure to adapt to new risks such as biosecurity as a scenario. Changes to approval processes that address both volume and new AI-enabled risks would indicate the regulatory link is being taken seriously.
- Lab automation linked to AI models. He wants advances in lab automation that speed up physical experiments and let AI models and automated labs work in recursive loops. AI company investment here is one of his two examples of institutions acting already; results from such systems would show whether the physical bottleneck is moving.
- AI-assisted peer review. He suggests using AI to scale peer review and to assure the quality of new knowledge. Whether journals and funders adopt this, and whether it reduces or adds to “scientific slop”, would show which way his peer review scenario is heading.
Hurles ends by assuming that whatever is written about AI, his own piece included, will soon look “heroically naive”. The useful part of his reflection is not a forecast. It is the reminder that faster tasks feed slower ones, and that many of the bottlenecks he expects are funding panels, approval queues, contracts and reviewers rather than the analysis AI tools are speeding up.
Source: Matt Hurles, AI acceleration of science – a personal reflection, Wellcome Sanger Institute blog, 2 November 2026. The Wellcome Sanger Institute (Genome Research Limited) is a charity registered in England, based at Hinxton, Cambridgeshire. See also our news report, UK research funders to allow AI in grant assessment.
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching whether UK funders, regulators and research institutions take up the bottlenecks Hurles names. Read more analysis at Insights, or get in touch.