TODAY’S PAPER | November 23, 2026 | EPAPER

Scientist Two and the new knowledge frontier

.


Naveed R Khan November 23, 2026 3 min read
The writer is an AI and sustainability policy analyst at UCSI University, Malaysia. Reach him at: naveed.r.khan@ gmail.com

Pakistan's next AI policy challenge may be how to govern a research system in which machines can generate hypotheses, run experiments, write code, draft papers and even review one another.

A September 2026 Google Cloud AI Research preprint, ScientistTwo, offers a glimpse of that future. Given a human-defined scientific problem, its multi-agent system identifies weaknesses in existing work, proposes new ideas, runs experiments, performs ablation studies, drafts a paper and sends it through simulated peer review and rebuttal. The authors report that it improved 86 of 107 machine-learning research problems, an 80.4% success rate, with an average 25.2% gain over human state-of-the-art baselines.

The full cycle reportedly took about two to three days and cost roughly $3,800 per task. Those figures require caution. This is a preprint; its benchmarks are predominantly computational AI problems, and much of its evaluation relies on AI reviewers. The authors themselves concede that the system does not yet consistently produce the kind of paradigm-changing work associated with top conference spotlight or oral papers.

Yet the direction is difficult to ignore: the cost of producing plausible scientific work may fall much faster than the cost of verifying it.

That matters enormously for Pakistan.

WIPO's Global Innovation Index 2026 ranks Pakistan 95th overall, but reveals a striking imbalance: 121st in innovation inputs and 72nd in outputs. WIPO again identifies Pakistan as an innovation "overperformer", producing comparatively stronger outputs from constrained inputs. Its strengths include ICT-services exports, ranked 15th globally, and scientific production, ranked 49th. Meanwhile, Pakistan's Economic Survey reports ICT export remittances of $3.38 billion during July-March FY2026, up 19.7% year-on-year.

This is the kind of economy that could benefit disproportionately from autonomous research. Pakistan cannot quickly reproduce the laboratories, endowments or researcher density of richer economies. But it may be able to compress parts of the discovery cycle through shared compute, scientific agents and better access to data.

The danger is equally obvious. If universities continue rewarding publication counts while machines make papers cheaper to produce, Pakistan could automate the wrong thing.

Therefore, a different institutional architecture is needed.

Pakistan should establish a National Verified Research Grid that shares infrastructure connecting the national AI compute grid already envisaged under the National AI Policy with universities, scientific databases, public datasets, software and autonomous research agents. The existing policy already proposes a national network of high-performance computing centers, GPUs and specialised AI hardware. The next step should be to connect that compute to national problems rather than simply "AI projects".

Ministries could publish challenge statements around important national issues, for instance, crop productivity, flood forecasting, disease surveillance, water losses, industrial energy efficiency, logistics and tax compliance. Universities and research teams would compete for compute based on measurable outcomes.

Albeit, the Grid needs an unusual second layer, a National Research Ledger.

Every publicly funded AI-assisted study should carry a machine-readable "research passport": which model generated which hypothesis, which datasets were used, what code was executed, which results were machine-produced, what was independently reproduced and which human investigator accepted responsibility for the final claim.

This is not technological fantasy. ScientistTwo itself employs integrity checks for reproducibility, rule violations, hallucinated references and whether the manuscript accurately represents the underlying code. Pakistan could convert that technical principle into institutional governance.

An AI-generated finding may be published as exploratory research. But it should not automatically qualify for academic promotion, public procurement, clinical guidance, regulation or major government policy until an independent replication or audit crosses a defined verification threshold. Medicine, infrastructure, defence and public finance should face higher thresholds than low-risk computational research.

Universities would also need to rethink incentives. HEC has already adopted a framework that moves away from blanket prohibition of generative AI towards responsible use and AI literacy. The next step is separating authorship from accountability.

A PhD should eventually be judged less by whether three papers were produced and more by whether the researcher formulated an important problem, developed defensible theory, interrogated causality, verified the evidence and can explain why a result deserves to be believed.

Government could push this further by replacing part of publication-based research funding with a Verified National Problems programme. Universities would earn competitive funding when independently reproduced research improves an agreed public metric.

There is an economic opportunity here too. Pakistan's next step beyond software freelancing could be research-as-a-service.

The deeper shift, however, is philosophical.

Science traditionally moves from human curiosity to hypothesis, experiment, criticism and accepted knowledge. Autonomous systems can now occupy several links in that chain. The scarce resource may therefore migrate from producing research to judging it.

Pakistan should neither ban the machine scientist nor worship it. It should build institutions around a harder principle that AI may generate knowledge claims, but accountable institutions must certify that knowledge.

COMMENTS

Replying to X

Comments are moderated and generally will be posted if they are on-topic and not abusive.

For more information, please see our Comments FAQ