Tag: AI

The Replication Crisis Is a Market Failure (And We Designed It That Way)

Also published on my newsletter

The replication crisis isn’t a mystery. After presiding over the review for thousands of grants at NSF’s Biological Sciences Directorate, I can tell you exactly why science struggles to reproduce its own findings: we built incentives that reward novelty and punish verification.

A 2016 Nature survey found that over 70% of scientists have failed to reproduce another researcher’s experiments. But this isn’t about sloppy science or bad actors. It’s straightforward economics.

Thanks for reading sciencepolicyinsider! Subscribe for free to receive new posts and support my work.

The Researcher’s Optimization Problem

You have limited time and resources. You can either:

  1. Pursue novel findings → potential Nature paper, grant funding, tenure
  2. Replicate someone’s work → maybe a minor publication, minimal funding, colleagues questioning your creativity

The expected value calculation is obvious. Replication is a public good with privatized costs.

How NSF Review Panels Work

At NSF, I watched this play out in every review panel. Proposals to replicate existing work faced an uphill battle. Reviewers—themselves successful researchers who got there by publishing novel findings—naturally favor creative, untested ideas over verification work.

We tried various fixes. Some programs explicitly funded replication studies. Some review criteria emphasized robustness over novelty. But the core incentive remained: breakthrough science gets you the next grant; careful verification doesn’t.

The problem runs deeper than any single agency. Universities want prestigious publications. Journals want citations. Researchers want tenure. Nobody’s optimization function includes “produces reliable knowledge that someone else can build on.”

The Information Market Is Broken

Even when researchers try to replicate, they’re working with incomplete information. Methods sections in papers are sanitized versions of what actually happened in the lab. “Cells were cultured under standard conditions” means something different in every lab. One researcher’s gentle mixing is another’s vigorous shaking.

This information asymmetry makes replication attempts inherently inefficient. You’re trying to reproduce a result while missing critical details that the original researcher might not even realize mattered.

The Time Horizon Problem

NSF grants run 3-5 years. Tenure clocks run 6-7 years. But scientific truth emerges over decades. We’re optimizing for the wrong timescale.

During my time at NSF, I saw brilliant researchers make pragmatic choices: publish something surprising now (even if it might not hold up) rather than spend two years carefully verifying it. That’s not a moral failing—it’s responding rationally to the incentives we created.

What Would Actually Fix This

Make replication profitable:

  • Count verification studies equally with novel findings in grant review and tenure decisions
  • Fund researchers whose job is rigorous replication—make it a legitimate career path
  • Require data and detailed methods sharing as a funding condition, not an afterthought
  • Make failed replications as publishable as successful ones

The challenge isn’t technical. It’s institutional. We designed a market that overproduces flashy results and underproduces reliable knowledge. Until we fix the incentives, we’ll keep getting exactly what we’re paying for.

The Unsung Hero: Why Exploratory Science Deserves Equal Billing with Hypothesis-Driven Research

For decades, the scientific method taught in classrooms has followed a neat, linear path: observe, hypothesize, test, conclude. This hypothesis-driven approach has become so deeply embedded in our understanding of “real science” that research proposals without clear hypotheses often struggle to secure funding. Yet some of the most transformative discoveries in history emerged not from testing predictions, but from simply looking carefully at what nature had to show us.

It’s time we recognize exploratory science—sometimes called discovery science or descriptive science—as equally valuable to its hypothesis-testing counterpart.

What Makes Exploratory Science Different?

Hypothesis-driven science starts with a specific question and a predicted answer. You think protein X causes disease Y, so you design experiments to prove or disprove that relationship. It’s focused, efficient, and satisfyingly definitive when it works.

Exploratory science takes a different approach. It asks “what’s out there?” rather than “is this specific thing true?” Researchers might sequence every gene in an organism, catalog every species in an ecosystem, or map every neuron in a brain region. They’re generating data and looking for patterns without knowing exactly what they’ll find.

The Case for Exploration

The history of science is filled with examples where exploration led to revolutionary breakthroughs. One of my lab chiefs at NIH was Craig Venter, famous for his exploratory project: sequencing the human genome. The Human Genome Project didn’t test a hypothesis—it mapped our entire genetic code, creating a foundation for countless subsequent discoveries. Darwin’s theory of evolution emerged from years of cataloging specimens and observing patterns, not from testing a pre-formed hypothesis. The periodic table organized elements based on exploratory observations before anyone understood atomic structure.

More recently, large-scale exploratory efforts have transformed entire fields. The Sloan Digital Sky Survey mapped millions of galaxies, revealing unexpected structures in the universe. CRISPR technology was discovered through exploratory studies of bacterial immune systems, not because anyone was looking for a gene-editing tool. The explosive growth of machine learning has been fueled by massive exploratory datasets that revealed patterns no human could have hypothesized in advance.

Why Exploration Matters Now More Than Ever

We’re living in an era of unprecedented technological capability. We can sequence genomes for hundreds of dollars, image living brains in real time, and collect environmental data from every corner of the planet. These tools make exploration more powerful and more necessary than ever.

Exploratory science excels at revealing what we don’t know we don’t know. When you’re testing a hypothesis, you’re limited by your current understanding. You can only ask questions you’re smart enough to think of. Exploratory approaches let the data surprise you, pointing toward phenomena you never imagined.

This is particularly crucial in complex systems—ecosystems, brains, economies, climate—where interactions are so intricate that predicting specific outcomes is nearly impossible. In these domains, careful observation and pattern recognition often outperform narrow hypothesis testing.

The Complementary Relationship

None of this diminishes the importance of hypothesis-driven science. Testing specific predictions remains essential for establishing causation, validating mechanisms, and building reliable knowledge. The most powerful scientific progress often comes from the interplay between exploration and hypothesis testing.

Exploratory work generates observations and patterns that inspire hypotheses. Hypothesis testing validates or refutes these ideas, often raising new questions that require more exploration. It’s a virtuous cycle, not a competition.

Overcoming the Bias

Despite its value, exploratory science often faces skepticism. It’s sometimes dismissed as “fishing expeditions” or “stamp collecting”—mere data gathering without intellectual rigor. This bias shows up in grant reviews, promotion decisions, and journal publications.

This prejudice is both unfair and counterproductive. Good exploratory science requires tremendous rigor in experimental design, data quality, and analysis. It demands sophisticated statistical approaches to avoid false patterns and careful validation of findings. The difference isn’t in rigor but in starting point.

We need funding mechanisms that support high-quality exploratory work without forcing researchers to shoehorn discovery-oriented projects into hypothesis-testing frameworks. We need to train scientists who can move fluidly between both modes. And we need to celebrate exploratory breakthroughs with the same enthusiasm we reserve for hypothesis confirmation.

Looking Forward

As science tackles increasingly complex challenges—understanding consciousness, predicting climate change, curing cancer—we’ll need every tool in our methodological toolkit. Exploratory science helps us map unknown territory, revealing features of reality we didn’t know existed. Hypothesis-driven science helps us understand the mechanisms behind what we’ve discovered.

Both approaches are essential. Both require creativity, rigor, and insight. And both deserve recognition as legitimate, valuable paths to understanding our world.

The next time you hear about a massive dataset, a comprehensive catalog, or a systematic survey, don’t dismiss it as “just descriptive.” Remember that today’s exploration creates the foundation for tomorrow’s breakthroughs. In science, as in geography, you can’t know where you’re going until you know where you are.

Bold Ventures in Science: NSF’s NEON and NIH’s BRAIN Initiative

My favorite projects…

As loyal readers know, these are my two favorite science initiatives. They stand out as beacons of progress: the National Science Foundation’s National Ecological Observatory Network (NEON) and the National Institutes of Health’s Brain Research through Advancing Innovative Neurotechnologies (BRAIN) Initiative. These groundbreaking endeavors showcase the commitment of U.S. science agencies to tackling complex, large-scale challenges that could revolutionize our understanding of the world around us and within us.

NSF’s NEON: A Continental-Scale View of Ecology

Imagine having a window into the ecological processes of an entire continent. That’s precisely what NEON aims to provide. Initiated in 2011, this audacious project is creating a network of ecological observatories spanning the United States, including Alaska, Hawaii, and Puerto Rico.

Yes, NEON has faced its share of challenges. The project’s timeline and budget have been adjusted since its inception, growing from an initial estimate of $434 million to around $469 million, with completion delayed from 2016 to 2019. But let’s be honest: when did you last try to build a comprehensive ecological monitoring system covering an entire continent? These adjustments reflected the project’s complexity and the learning curve in such a pioneering endeavor.

The payoff? NEON is now collecting standardized ecological data across 81 field sites from Hawaii to Puerto Rico and in between. This massive time series in some 200 dimensions will allow scientists to analyze and forecast ecological changes over decades. From tracking the impacts of climate change to understanding biodiversity shifts, NEON provides invaluable insights that could shape environmental policy and conservation efforts for future generations.

NIH’s BRAIN Initiative: Decoding Our Most Complex Organ

Meanwhile, the NIH’s BRAIN Initiative is taking on an equally monumental task: mapping the human brain. Launched in 2013, this project is aptly named, as it requires a lot of brains to understand… well, brains.

With annual funding that has grown from an initial $100 million to over $500 million, the BRAIN Initiative is a testament to the NIH’s commitment to unraveling the mysteries of neuroscience. Mapping all 86 billion neurons in the human brain by 2026 might seem a tad optimistic. But I’m increasingly impressed with our progress, and I am hopeful we’ll be able to get some meaningful statistics about variability across individuals.

The initiative has already led to the development of new technologies for studying brain activity, potential treatments for conditions like Parkinson’s disease, and insights into how our brains process information. It’s like a real-life adventure into the final frontier, except instead of outer space, we’re exploring the inner space of our skulls.

The Challenges: More Feature Than Bug

Both NEON and the BRAIN Initiative have faced obstacles, from budget adjustments to timeline extensions. But in the world of cutting-edge science, these challenges are often where the real learning happens. They’ve pushed scientists to innovate, collaborate, and think outside the box (or skull, in the case of BRAIN).

These projects have also created unique opportunities for researchers to develop new skills. Grant writing for these initiatives isn’t just an administrative hurdle; it’s a chance to think big and connect individual research to grand, overarching goals. It’s turning scientists into visionaries, and isn’t that worth a few late nights and extra cups of coffee?

Conclusion: Big Science, Bigger Possibilities

NEON and the BRAIN Initiative represent more than just large-scale scientific projects. They’re bold statements about the value of basic research and the importance of tackling complex, long-term challenges. They remind us that some questions are too big for any single lab or institution to answer alone.

As these projects evolve and produce data, they’re not just advancing our understanding of ecology and neuroscience. They’re also creating models for conducting science at a grand scale, paving the way for future ambitious endeavors.

So here’s to the scientists, administrators, and visionaries behind NEON and the BRAIN Initiative. They’re showing us that with enough creativity, persistence, and, yes, funding, we can tackle some of the biggest questions in science. And who knows? The next breakthrough in saving our planet or understanding consciousness could be hidden in the data they’re collecting right now.

Automated hypothesis generation: an AI role in science

When I was getting my PhD in Ann Arbor during the 1980’s, just staying up to date with the relevant literature to my own thesis project was a constant challenge. There was a paper magazine back then called Current Contents (CC). CC contained just that: the tables of content for all of the relevant journals (in Life Sciences). It was a critical resource because there was no other way—even then—to keep tabs on the collective scientific output.

Keeping tabs was not just for general knowledge about the field. Or even about properly giving credit to others. Rather, it was critical to the hypothesis creation. Asking the right question (at the right time) is what determines scientific success in many cases. But you can’t ask the right question without understanding whether it’s been already asked. And really you can’t ask the right question without a full understanding of what the current state of scientific knowledge is.

At the time, it was the habit, in many high impact papers, to have the last figure in the paper be a cartoon schematic that represented the author’s view of where the field was—at the moment of the paper’s acceptance into the journal. In my field of molecular neuroscience, this often was a series of shapes and arrows representing key biomolecules and pathways. It was often amusing to go from one paper to the very next that a particular group put out and see that some of the arrows would mysteriously reverse directions from the cartoon in the previous paper. This was presumably because the paper’s results along with other results had changed the thinking of the author.

In any case, that cartoon figure was always a clue into what the next hypothesis to be tested would be for a particular research group. So in a sense, you could predict the trajectory of scientific inquiry from that cartoon figure at the end of a paper.

That was the 1980’s. Our scientific knowledge base has expanded exponentially since then. One of the current versions of Current Contents is called Faculty of 1000 (F-1000). It’s on-line of course. The idea is that leaders in the field curate the papers that you should read based on your profile. It’s a great idea I guess, although science being as competitive as it is, I have doubts that the elect would give up some brilliant and undiscovered insight of a paper to the unwashed, if it really might supercharge some scientific inquiry. However, as a scientist, you have many other choices. Google Scholar comes to mind—it’s both comprehensive and I’m pretty sure it uses AI extensively to tailor its results. So machine-driven instead of human-driven (as in the case of F-1000).

However, the cartoon figure at the end of papers has become pretty obsolete (although it does still make appearances). That’s because pretty much all of science—certainly life sciences—has become incredibly complex. In my field, you can’t make a cartoon big enough to represent all the relevant biomolecules and pathways and the arrows have become incredibly intertwined because of the multiplicity of feedback loops and cross-talk links.

So not only is it difficult to glean the next hypothesis for the clever reader (even when there is a cartoon). It’s impossible for the author to do the same.

This has pushed much of science from the paradigm of Popper to exploratory research. In such science, I might read the data stream from some set of sensors, correlate that data with some other external variable (like seasonality) and publish a correlation that is intriguing. Correlation of course is not causation—we all know that.

And yet, science has the tools to do excellent hypothesis-based research. In neuroscience, optogenetics methods allow us to turn on and off neural circuits to understand their effects upon behavior. In molecular biology, CRISPR does the same for genetic circuits and networks.

The problem is not executing the research. It’s the ability to ask the right question. For biology, generating a hypothesis that is parsimonious with all of the current knowledge in a scientific discipline is challenging for human scientific superstars and downright impossible for your typical graduate student coming up with a thesis project. I believe that the same is true for any area of science where the volume of knowledge and relevant data has expanded exponentially.

But all is not lost. I think this is a perfect domain for AI as it exists today. Keeping tabs of many disparate but relevant data points and then coming up with a next move? That’s how AI’s beat humans in chess right now. So… AI in collaboration with human scientists might be a very fruitful collaboration going forward. And it may yet save hypothesis-based research.

Hubble Telescope

I know that we are waiting on the James Webb Space Telescope, but disturbingly, the Hubble Space Telescope is sitting in safe mode after a gyro failure this past weekend (hat tip to NASA watch). This is the telescope that has been the workhorse of NASA’s astronomy program.

In a sense this is the future. If we continue to send very complicated gadgets to technological edge environments, particularly in the near future with AI on-board, they are going to have to be much more resilient. Space and deep ocean are examples of such environments. There are implications for big science and DOD.

Drone attacks

The Atlantic has an excellent piece about the use of drones by non-state actors for bad purposes, here. My own view is that edge computing and AI will render these technologies vastly more destructive in the future. Not in terms of mass destruction, but in terms of targeted destruction. The key is how to defend against AI-enabled swarms. Could the new 5G networks somehow be deployed in an emergency to do just that?

Spoofing AI

This an interesting new scientific meme. It made it into Science on the basis of a presentation at the International Conference on Machine Learning here. The idea is that hackers can easily defeat AI’s (think “social engineering” used on a machine).

Meanwhile there is the contrasting meme of us getting spoofed by AI, in the FT, here. In this case AI’s are able to make videos of people doing things that they did not do.

All of this gets to the cybersecurity aspects of AI that potentially put society at risk.