A Bigger Ambition for Artificial Intelligence
Artificial intelligence has become very good at attracting attention, but Jack Clark wants policymakers thinking beyond the next impressive chatbot demonstration.
Speaking at the IMF–World Bank meetings in Bangkok on October 17, Anthropic’s co-founder presented AI as a potential accelerator of scientific discovery, alongside an urgent need for safeguards.
The appeal is easy to understand.
Better research could eventually mean better medicines, more effective engineering, and discoveries that open entirely new industries.
Those possibilities explain why a conversation about AI belongs in a room full of economic policymakers.
Science helps determine what economies can produce, which problems societies can solve, and how much progress costs.
Clark’s remarks raise a question that reaches well beyond Silicon Valley: how should countries prepare for machines that could help people discover things faster?
Why Scientific Discovery Changes the Conversation
Much of the public discussion around AI focuses on familiar office tasks.
Can it write an email, summarize a meeting, or rescue a spreadsheet that looks like it survived a small earthquake?
Scientific research introduces a different ambition.
Instead of helping people communicate existing knowledge, AI could help them investigate questions whose answers remain unknown.
Consider the difference between summarizing a published experiment and identifying an overlooked pattern worth testing.
The first saves time.
The second might redirect an entire research project.
That possibility makes AI-assisted science especially interesting, although identifying a promising direction remains only one step toward a discovery.
Researchers still need convincing evidence, suitable methods, and results that withstand scrutiny.
The opportunity lies in helping more worthwhile ideas reach that demanding process.
What a Research Multiplier Could Look Like
The phrase “research assistant” can sound modest until you consider how much supporting work science requires.
A project may involve reading papers, cleaning datasets, writing analysis code, comparing methods, and documenting decisions before an experiment produces anything useful.
As an illustrative workflow, a scientist could ask AI tools to examine several analytical approaches while personally investigating the most promising result.
Another tool might help check whether an apparent finding disappears when the assumptions change.
These examples describe possible uses rather than a universal account of what current systems reliably accomplish.
Their attraction is parallel effort.
A researcher with limited time could explore more possibilities without personally executing every preliminary step.
However, producing more suggestions only helps if researchers can evaluate them efficiently.
Otherwise, the assistant becomes a very enthusiastic supplier of homework.
An Earlier Biology Project Offers Context
Anthropic’s argument has some concrete background, although that background should remain separate from Clark’s latest remarks.
In September, the company described an AI-assisted biology project involving a previously uncharacterized enzyme system.
According to Anthropic, Claude examined DNA data and identified distinctive features associated with a reverse transcriptase, an enzyme that copies RNA into DNA.
The company described the system as having characteristics reminiscent of CRISPR-associated arrangements.
Crucially, it also said the system’s function remained unknown.
That distinction keeps the announcement in perspective.
An intriguing biological observation can justify further investigation without establishing a practical gene-editing tool or medical treatment.
The example illustrates a plausible contribution from AI: drawing researchers’ attention to something worth examining more closely.
What happens afterward determines its scientific importance.
Discovery Still Needs an Evidence Trail
An AI-generated explanation can sound convincing even when its underlying argument falls apart.
Scientific work therefore needs more than fluent answers and attractive charts.
Researchers must be able to trace an analysis back to its data, inspect the code, understand the assumptions, and reproduce the result.
Imagine a system flagging a surprising relationship in a large dataset.
Before celebrating, the team needs to rule out missing values, measurement errors, accidental data leakage, and ordinary coincidence.
That verification takes effort, but it gives the finding meaning.
A useful AI research workflow should preserve the steps that produced a conclusion.
If another scientist cannot follow those steps, the apparent acceleration may disappear during review.
Fast exploration becomes valuable when it produces evidence that other people can examine and trust.
The Independent Picture Is Encouraging—and Uneven
Stanford’s 2026 AI Index provides a broader view of AI’s expanding scientific role.
Its science chapter reports approximately 80,150 AI-related natural-science publications in 2025, an increase of 26% over the previous year.
That suggests substantial research activity, though publication volume alone cannot establish the quality or usefulness of the work.
The same report highlights important weaknesses.
It describes models performing strongly on chemistry questions while struggling with research replication and more complete scientific tasks.
The practical lesson is about choosing the right responsibility.
A system that answers difficult questions successfully may still struggle to carry an investigation from beginning to end.
Scientists need evaluations that resemble the work they actually intend to delegate.
A demanding quiz and a dependable laboratory colleague require different kinds of competence.
Where Faster Research Meets the Physical World
Even excellent digital assistance cannot remove every constraint on scientific progress.
An engineering team may generate a promising design quickly, then spend months obtaining materials, building prototypes, and measuring performance.
A biology team may narrow its experimental options while still needing equipment, laboratory time, and careful interpretation.
These are illustrative examples of why research speed depends on the whole process.
Accelerating one stage can simply move the queue somewhere else.
That does not make the improvement unimportant.
It changes where organizations should invest next.
If analysis becomes faster, experimental capacity may become more valuable.
If researchers test more ideas, they may need better systems for tracking failures and comparing results.
The strongest gains could come from redesigning workflows around useful AI capabilities, rather than attaching a chatbot to an unchanged process.
Understanding the $44 Trillion Headline
The economic figure attached to Clark’s remarks deserves careful handling.
Anthropic’s published scenario explorer includes an extreme case in which US GDP reaches $44.4 trillion in 2030, measured at 2025 prices.
In that model, the figure stands 32.4% above the corresponding economy without AI.
Comparing it with today’s smaller economy answers a different question, so the two percentage comparisons should not be treated as interchangeable.
The explorer also presents modest and substantial scenarios with smaller economic effects.
Its purpose is to examine what different assumptions might imply.
It does not establish that the largest outcome will happen.
The useful discussion therefore concerns those assumptions: future capabilities, adoption, autonomy, productivity, and workers’ ability to adjust.
A large number attracts attention; understanding its foundations makes that attention productive.
Capability Must Become Useful Adoption
A capable system creates possibilities, but an organization must still turn those possibilities into dependable work.
For a research institution, adoption might require secure data access, suitable software, staff training, and agreement about which outputs need expert review.
These practical questions can decide whether an impressive demonstration becomes a daily tool.
Consider a hypothetical university team evaluating AI assistance for data analysis.
The relevant comparison includes the time spent preparing inputs, checking results, correcting mistakes, and documenting the final work.
A faster first draft of an analysis may help considerably.
It may also save little if verification becomes much harder.
Measuring the complete workflow gives institutions a sounder basis for deciding where to expand.
Successful adoption should make good research easier to accomplish, with benefits visible beyond the demonstration.
More Output Does Not Guarantee Better Science
AI makes it easier to generate text, code, and candidate explanations.
That creates a temptation to measure progress by how much material a research team produces.
More output can be useful, but it can also bury the important result.
Imagine receiving several hundred proposed hypotheses with little indication of which deserve experimental attention.
The immediate problem becomes prioritization.
Researchers need systems that help assess relevance, uncertainty, and testability, alongside tools that generate possibilities.
Institutions also need incentives that reward dependable findings.
Otherwise, faster production could increase the burden on reviewers without improving knowledge at the same pace.
The meaningful measure is what survives examination and helps subsequent work.
A shorter list of well-supported ideas may contribute more than an enormous folder of plausible suggestions awaiting someone else’s weekend.
Human Expertise Changes Shape
AI assistance could alter where researchers spend their time.
Some may devote fewer hours to routine coding or preliminary literature searches and more to experimental design, interpretation, and choosing worthwhile questions.
That is a possible direction, rather than a guaranteed outcome for every discipline.
It also creates a training challenge.
Early-career researchers often develop judgment by performing the same supporting tasks that institutions may want to automate.
If those tasks change, supervisors need other ways to teach the underlying skills.
A student should understand why an analysis works, where it can fail, and how to challenge its conclusion.
Operating the software is only part of that education.
The long-term opportunity depends on developing people who can direct powerful tools intelligently and recognize when those tools have gone wrong.
Economic Growth Has a Distribution Problem
Anthropic’s scenario explorer also makes clear that a richer economy can produce difficult outcomes for workers.
Its most transformative scenario combines substantial growth with worsening prospects for many people in knowledge occupations.
Those are modeled possibilities, not established predictions.
Nevertheless, they expose a weakness in treating GDP as a complete account of success.
Households experience economic change through wages, job security, prices, and access to services.
An impressive national total cannot answer every question about those experiences.
This matters when discussing AI-assisted science because the eventual benefits may travel through patents, companies, public institutions, and healthcare systems.
Who owns a discovery and who can afford its application can shape its social value.
Economic preparation therefore includes decisions about access and opportunity, alongside efforts to increase productivity.
Participation Requires More Than Buying Software
For countries outside the largest AI markets, the scientific opportunity raises practical questions about participation.
Could local universities use these tools to investigate locally important problems?
Would researchers have affordable access, appropriate datasets, and enough computing support?
For the Philippines, possible areas of interest include agriculture, disaster preparedness, public health, and coastal research.
These are potential applications, rather than projects announced in Clark’s speech.
Their value would depend on local expertise and evidence.
A tool developed elsewhere may need careful evaluation against local conditions before researchers can rely on it.
Training, institutional partnerships, and shared research infrastructure could help smaller teams participate.
The goal should be to strengthen their ability to ask and investigate important questions, while giving them the resources to judge the answers.
Safeguards Belong Inside the Workflow
Clark paired his optimism with a call for oversight, particularly where scientific capabilities could be misused.
That concern has a practical counterpart inside research institutions.
Teams need clear responsibility for approving sensitive activities, controlling access, and reviewing what automated systems are allowed to do.
Oversight works best when it follows the actual risk of a task.
Summarizing public literature and initiating consequential laboratory activity deserve different levels of authority.
Organizations can build that distinction into their workflows through permissions, records, and review requirements.
Planning these arrangements early may also make useful adoption easier.
Researchers should know which activities they can pursue independently and which require additional scrutiny.
Clear rules reduce uncertainty for responsible users while preserving accountability as systems become more capable and take on more complex work.
What Would Demonstrate Meaningful Progress?
The next stage of this story needs evidence that reaches beyond ambitious speeches.
Useful indicators would include independently reproduced discoveries, measurable improvements in complete research workflows, and reliable performance across institutions with different resources.
Costs matter, too.
A system that saves time but requires expensive computing and extensive expert correction may offer a narrower benefit than its headline suggests.
Researchers should also report where AI assistance failed.
Those accounts help other teams avoid repeating mistakes and reveal which capabilities still need improvement.
For policymakers, broader access and tangible public benefits deserve attention alongside commercial adoption.
The strongest case for AI-assisted science will emerge from results that other people can verify, use, and build upon.
That is how an appealing possibility becomes a dependable contribution to progress.
A Scientific Opportunity Worth Building Carefully
Clark’s remarks put a consequential idea before economic leaders: AI could help expand the capacity to discover, with effects that eventually reach far beyond the laboratory.
The encouraging part is the possibility of giving researchers more room to explore difficult questions.
The demanding part is turning that possibility into trustworthy results and broadly useful applications.
Governments, universities, companies, and scientists each have work to do.
They must improve tools, train people, strengthen experimental capacity, and decide how responsibility should follow increasing capability.
No economic scenario can complete those tasks for them.
But preparation can begin before the largest predictions are resolved.
If AI helps researchers test better ideas and produce reliable knowledge more efficiently, the benefits could be substantial—and they will have something firmer beneath them than a spectacular headline.
SOURCES
- Reuters — Jack Clark on AI, scientific discovery, economic growth, and safeguards. (Reuters)
- Anthropic — Scenarios for Our Economic Future. (anthropic.com)
- Anthropic — Claude Discovers a Novel Enzyme System with CRISPR-Like Repeats. (anthropic.com)
- Stanford HAI — The 2026 AI Index Report: Science. (hai.stanford.edu)
The Kingy Brief
Get The Kingy Brief.
AI changes, original tests and one practical thing to try. Fridays at 09:00 Vancouver time.
Free · Double opt-in · Unsubscribe anytime
Sign up for The Kingy Brief on Beehiiv.
Signup help and newsletter schedule
Signup form provided by Beehiiv. After submitting, check your inbox for "Confirm your subscription to The Kingy Brief" and open its confirmation link. Check Spam or Promotions if you cannot find it.
Fridays at 09:00 Vancouver time: source-checked AI changes, original tests and one practical thing to try. We skip a week when there is not enough verified material. Free. Unsubscribe anytime.