Why Real Science Still Breaks Automated Labs

Introduction

Self-driving laboratories are one of the most important tests of the wider AI bloom idea. If advanced AI can help humanity accelerate science, then it must eventually do more than generate theories. It must help turn ideas into reliable discoveries in the physical world.

In controlled settings, autonomous laboratories have already produced impressive results. AI systems can choose experiments, robotic equipment can run them, and software can analyse outcomes and decide what to test next. In some materials science and chemistry applications, this can compress research cycles from weeks to hours. Yet the most ambitious vision — laboratories that can autonomously navigate the full messiness of real scientific work — remains much harder than many headlines imply. The central obstacle is not computation. It is reality. Scientific environments contain hidden variables, tacit human knowledge, fragile procedures and biological complexity that resist standardisation. The closer automation gets to the real world, the more these difficulties matter.[PubMed]pubmed.ncbi.nlm.nih.govAutonomous 'self-driving' laboratories: a review of…by AV Tobias · 2025 · Cited by 65 — This article reviews and provides perspe…

Where closed-loop systems succeed today

The strongest successes in self-driving laboratories tend to appear in domains where experiments are highly structured, variables are tightly controlled and outcomes can be measured automatically.

Materials science has become a leading example. Researchers have built systems that repeatedly synthesise materials, measure their properties and use machine-learning models to select the next experiment. Closed-loop platforms have explored vast combinations of processing conditions, compositions and structures far faster than conventional trial-and-error approaches.[arXiv]arxiv.orgOn-the-fly Closed-loop Autonomous Materials Discovery via Bayesian Active LearningJune 11, 2020…Published: June 11, 2020[ACS]pubs.acs.orgACS PublicationsSelf-Driving Laboratories for Chemistry and Materials Scienceby G Tom · 2024 · Cited by 726 — More advanced SDLs combinin…

This works particularly well when:

Many materials problems fit this pattern. Researchers may be trying to maximise conductivity, stability, catalytic efficiency or another measurable property. Once the optimisation target is defined, active-learning systems can efficiently search through thousands of possibilities.[arXiv]arxiv.orgOn-the-fly Closed-loop Autonomous Materials Discovery via Bayesian Active LearningJune 11, 2020…Published: June 11, 2020

This is why many demonstrations of autonomous science focus on batteries, catalysts, thin films, semiconductors and related areas. The laboratory environment is still complex, but the experiment itself is often more repeatable than in biological research. As a result, AI-guided experimentation can generate genuine throughput gains.[Collegium Helveticum]collegium.ethz.chCollegium HelveticumThe Rise of Self-Driving Labs in Chemistry and Materials…This lecture will explore how the convergence of automati…[sheffield]sheffield.ac.ukself driving labs making chemical research faster and smarterSheffield UniversitySelf-driving labs: making chemical research faster and…9 May 2025 — Researchers have built an automated platform…Published: May 2025 The success of these systems matters for the broader scientific-acceleration story. They show that parts of discovery can already be transformed into high-speed feedback loops between algorithms and physical experiments. But they also reveal an important limitation: many of the easiest domains to automate are not representative of science as a whole.

Tacit knowledge is harder to automate than procedures

One of the most underestimated obstacles is tacit knowledge.

Scientific papers often describe experiments as though they are fully specified recipes. In practice, much laboratory work depends on skills, judgments and contextual knowledge that researchers rarely write down completely.

A protocol may say that a sample should be mixed gently. An experienced scientist knows what “gently” means in a particular context. A robotic system does not. A paper may describe a culture preparation method that technically reproduces a procedure while missing dozens of small practical details that affect results. Experienced researchers often detect problems through subtle observations that are difficult to formalise into software rules.[University of Bristol]research-information.bris.ac.ukTwo Kinds of Science D24 for Po SUniversity of BristolCollins, H., Shrager, J., Bartlett, A., Conley, S., Hale, R., &…by H Collins — “Can Robots Help Solve the Reprodu…

This problem appears repeatedly in discussions of scientific reproducibility. Even when laboratories attempt to follow identical protocols, results can diverge because crucial information never entered the formal record. Automation can improve consistency once a process is fully specified, but many scientific processes are not fully specified to begin with.[crukcambridgecentre.org.uk]crukcambridgecentre.org.uk‘robot scientist’ eve finds less one third scientific results are reproducibleRobot scientist' Eve finds that less than one third of…6 Apr 2022 — Statistically significant evidence for repeatability was found fo…[PMC]pmc.ncbi.nlm.nih.govPMCTesting the reproducibility and robustness of the cancerby K Roper · 2022 · Cited by 27 — Automation makes experimental replication technically easier, as laboratory robotics are more accura…

Human researchers routinely make adjustments based on:

These interventions are often invisible in published methods sections. A robotic platform can only automate what has been captured and encoded. When key knowledge exists mainly in people’s heads, automation encounters a hidden wall.[NLR]docs.nlr.govPerspectives for self-driving labs in synthetic biologyPerspectives for self-driving labs in synthetic biologyJanuary 23, 2023 — by D Arnold · 2023 · Cited by 123 — Self-driving labs (SDLs)…Published: January 23, 2023

This creates a paradox. The scientific fields that appear most ripe for automation are often those that have already become highly standardised. The more exploratory and uncertain a field becomes, the more tacit knowledge tends to matter.

Fragile experiments break closed loops

Another challenge is that real laboratories are filled with failure modes that are difficult to anticipate.

A closed-loop system assumes that experimental results can be trusted as feedback for future decisions. But many experiments fail for reasons unrelated to the scientific question being studied.

A pipette may clog. A reagent may degrade. A sensor may drift. A culture may become contaminated. A robotic arm may slightly misalign a sample. A measurement may appear valid while quietly incorporating systematic error.

Human scientists often recognise these failures because they understand the broader context of the experiment. They notice unusual smells, colours, textures, timing irregularities or equipment behaviour. Many of these signals remain difficult to capture through automated monitoring systems.[PubMed]pubmed.ncbi.nlm.nih.govAutonomous 'self-driving' laboratories: a review of…by AV Tobias · 2025 · Cited by 65 — This article reviews and provides perspe…

This becomes especially problematic in autonomous systems because errors can propagate. If the AI interprets a faulty measurement as genuine scientific information, it may choose subsequent experiments based on a false signal. The laboratory can then optimise towards artefacts rather than discoveries.

The problem resembles challenges seen in autonomous vehicles. Driving works well on carefully mapped roads under predictable conditions. Edge cases create disproportionate difficulty. Scientific laboratories have their own version of edge cases, except they occur constantly because experimentation is fundamentally about probing the unknown.

Why biology remains harder than materials science

The gap between materials science and biology illustrates the challenge particularly clearly.

Materials systems are often complicated, but they can be relatively stable. Researchers can repeatedly create similar samples and measure similar properties under controlled conditions.

Biology is different.

Living systems are noisy, adaptive and context-dependent. Cells change behaviour over time. Genetic pathways interact with one another. Small environmental differences can produce large effects. Experimental outcomes often depend on factors that researchers do not yet fully understand.[NLR]docs.nlr.govPerspectives for self-driving labs in synthetic biologyPerspectives for self-driving labs in synthetic biologyJanuary 23, 2023 — by D Arnold · 2023 · Cited by 123 — Self-driving labs (SDLs)…Published: January 23, 2023

Even supposedly standard biological materials vary. Cell lines drift genetically. Reagents age. Organisms respond differently to environmental conditions. Biological systems contain layers of feedback that make outcomes difficult to predict and difficult to reproduce.[PMC]pmc.ncbi.nlm.nih.govPMCTesting the reproducibility and robustness of the cancerby K Roper · 2022 · Cited by 27 — Automation makes experimental replication technically easier, as laboratory robotics are more accura…

This creates several difficulties for autonomous laboratories:

Measurement is often ambiguous

In materials science, success may be measured through a clear property such as conductivity or strength.

In biology, researchers often care about complex phenomena such as toxicity, immune response, disease progression or cellular differentiation. These outcomes may not have a single straightforward measurement. Multiple competing interpretations can fit the same data.[LinkedIn]linkedin.comAutonomous Labs in Nature Magazine: A New Frontier for…Autonomous systems explore combinatorial space at industrial scale. Sha…

Experimental timescales are longer

Many biological processes unfold over days, weeks or months.

A materials experiment might complete in minutes. A biological experiment may require cell growth, incubation, sequencing, imaging and multiple validation stages. This reduces the speed advantage of closed-loop optimisation.[NLR]docs.nlr.govPerspectives for self-driving labs in synthetic biologyPerspectives for self-driving labs in synthetic biologyJanuary 23, 2023 — by D Arnold · 2023 · Cited by 123 — Self-driving labs (SDLs)…Published: January 23, 2023

Translation to the real world is harder

Drug discovery illustrates the problem.

A system may autonomously optimise compounds in cell cultures, yet success in a controlled laboratory environment does not guarantee success in animals or humans. Toxicity, metabolism, side effects and regulatory constraints introduce layers of complexity that cannot be fully captured by a narrow experimental loop.[LinkedIn]linkedin.comAutonomous Labs in Nature Magazine: A New Frontier for…Autonomous systems explore combinatorial space at industrial scale. Sha…

As a result, many impressive autonomous demonstrations solve only a small part of the broader scientific challenge.

The reproducibility problem cuts both ways

Automation is often presented as a solution to science’s reproducibility crisis.

There is truth in this claim. Robots can execute procedures more consistently than humans. Automated systems can log actions in greater detail and reduce variation introduced by individual researchers.[PMC]pmc.ncbi.nlm.nih.govPMCTesting the reproducibility and robustness of the cancerby K Roper · 2022 · Cited by 27 — Automation makes experimental replication technically easier, as laboratory robotics are more accura…

Yet automation also exposes a deeper problem.

If published findings cannot reliably be reproduced even under highly controlled conditions, then autonomous systems inherit a noisy and imperfect scientific record. Models trained on flawed literature may pursue unproductive directions. Closed-loop systems can optimise experimental throughput without necessarily improving scientific validity.[crukcambridgecentre.org.uk]crukcambridgecentre.org.uk‘robot scientist’ eve finds less one third scientific results are reproducibleRobot scientist' Eve finds that less than one third of…6 Apr 2022 — Statistically significant evidence for repeatability was found fo…

The challenge is not merely performing experiments faster. It is ensuring that the information generated remains trustworthy.

This distinction matters for AI-driven scientific acceleration. A future with millions of autonomous experiments per day only improves discovery if those experiments produce reliable knowledge rather than automated noise.

The hardware bottleneck remains stubbornly physical

Much AI progress comes from software scaling.

Laboratories do not scale that way.

Every experiment still requires physical infrastructure: instruments, chemicals, sensors, maintenance, calibration and safety systems. Expanding experimental throughput often means purchasing additional hardware rather than simply allocating more computation.[ScienceDirect]sciencedirect.comDemocratizing self-driving labs: advances in low-cost 3D…by S Doloi · 2025 · Cited by 32 — Laboratory automation through…

Researchers reviewing self-driving laboratories consistently identify hardware integration as one of the hardest engineering challenges. Scientific instruments are frequently designed as standalone devices rather than components in a unified autonomous system. Connecting them into reliable workflows requires substantial custom engineering.[PubMed]pubmed.ncbi.nlm.nih.govAutonomous 'self-driving' laboratories: a review of…by AV Tobias · 2025 · Cited by 65 — This article reviews and provides perspe…

Even highly capable AI systems cannot eliminate these constraints.

A model may generate ten thousand promising hypotheses overnight. The laboratory may still only be capable of physically testing a few hundred. The bottleneck shifts from thinking to acting.

This is one reason many analysts see self-driving laboratories not as a replacement for scientists but as an attempt to rebalance an increasingly uneven system. AI can already generate candidate ideas faster than many laboratories can evaluate them. The challenge is making experimentation scale without assuming that the physical world behaves like software.

What this means for the larger AI bloom vision

The limitations of self-driving laboratories do not invalidate the case for scientific acceleration. In some ways they clarify it.

The strongest version of the AI bloom argument is not that science becomes fully automated overnight. It is that intelligence, robotics, instrumentation and experimental infrastructure gradually become more tightly integrated, allowing civilisation to learn from the physical world much faster than before.

Current autonomous laboratories demonstrate part of that future. They show that closed-loop experimentation can substantially increase throughput in specific domains. They also reveal that real science is not simply a search problem waiting for a larger model. It is a physical process shaped by messy environments, incomplete knowledge, hidden variables and living systems that resist simplification.[ACS Publications]pubs.acs.orgACS PublicationsSelf-Driving Laboratories for Chemistry and Materials Scienceby G Tom · 2024 · Cited by 726 — More advanced SDLs combinin…

The near-term lesson is therefore double-edged. Scientific discovery may become dramatically faster in areas that can be standardised and automated. But the most important breakthroughs in medicine, biology and complex real-world systems may continue to require deep human involvement for longer than many optimistic forecasts assume.

For advocates of long-run human flourishing, that distinction matters. The future may depend less on replacing scientists than on creating richer partnerships between AI systems, robotic laboratories and human researchers who remain uniquely good at navigating uncertainty when reality refuses to follow the protocol.

Amazon book picks

Further Reading

Books and field guides related to Why Real Science Still Breaks Automated Labs. Use these as the next step if you want deeper reading beyond the article.

Book

The Genesis Machine

By Amy Webb, Andrew Hessel

Relevant to why biology is harder to automate than clean digital workflows.

See on Amazon

Book

The Knowledge Machine

By Michael Strevens

Directly supports the page’s theme that automated science must still obey rigorous methods.

See on Amazon

Book

A Crack in Creation

By Jennifer A. Doudna, Samuel H. Sternberg

Gives real-world context for the complexity and stakes of experimental biology.

See on Amazon

Book

Failure

By Stuart Firestein

Explains why failed, messy and ambiguous experiments are central to real science.

See on Amazon

Browse more on Amazon:The Genesis MachineThe Knowledge MachineA Crack in Creation

As an Amazon Associate I earn from qualifying purchases.

eBay marketplace picks

Marketplace Samples

Example marketplace items related to this page. Use the search link to explore similar finds on eBay.

Shop location

UsingUSA USA USAUKCanadaAustraliaIreland

Example eBay listing

Laboratory Area OSHA Notice Sign Metal, Plastic, Or Decal - 5 or 10 Pack

SearcheBay.co.uk: laboratory sign

Browse similar oneBay.co.uk

Example eBay listing

Laboratory 5x10 Hanging Plus Wall or Door Sign | Medicine Signage

SearcheBay.co.uk: laboratory sign

Browse similar oneBay.co.uk

Browse more oneBay.co.uk

Example items shown for inspiration; availability and pricing can change. Branchoria may earn a commission if you purchase through outbound eBay links.

Endnotes

1. Source: pubs.acs.org
Link:https://pubs.acs.org/doi/10.1021/acs.chemrev.4c00055

2. Source: arxiv.org
Link:https://arxiv.org/abs/2006.06141

Published: June 11, 2020

3. Source: pmc.ncbi.nlm.nih.gov
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC12290927/

4. Source: crukcambridgecentre.org.uk
Title: ‘robot scientist’ eve finds less one third scientific results are reproducible
Link:https://crukcambridgecentre.org.uk/news/%E2%80%98robot-scientist%E2%80%99-eve-finds-less-one-third-scientific-results-are-reproducible

5. Source: pmc.ncbi.nlm.nih.gov
Title: PMCTesting the reproducibility and robustness of the cancer
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC8984295/

6. Source: docs.nlr.gov
Title: Perspectives for self-driving labs in synthetic biology
Link:https://docs.nlr.gov/docs/fy23osti/84830.pdf

Published: January 23, 2023

7. Source: linkedin.com
Link:https://www.linkedin.com/posts/tarselli_will-self-driving-robot-labs-replace-biologists-activity-7430570231183347712-5Cis

8. Source: linkedin.com
Link:https://www.linkedin.com/pulse/self-driving-laboratories-drug-discovery-ai-automation-nagesh-nama-lvube

9. Source: sciencedirect.com
Link:https://www.sciencedirect.com/org/science/article/pii/S2635098X25000968

10. Source: linkedin.com
Link:https://www.linkedin.com/pulse/has-our-industrial-scientific-communitybeen-acting-little-qyqje

11. Source: youtube.com
Title: An autonomous laboratory for the accelerated synthesis of novel materials
Link:https://www.youtube.com/watch?v=blooHbR5_eM

12. Source: pubmed.ncbi.nlm.nih.gov
Link:https://pubmed.ncbi.nlm.nih.gov/40852582/

13. Source: collegium.ethz.ch
Link:https://collegium.ethz.ch/events/fellow-year-2024-2025/the-rise-of-self-driving-labs-in-chemistry-and-materials-science

14. Source: sheffield.ac.uk
Title: self driving labs making chemical research faster and smarter
Link:https://sheffield.ac.uk/cmbe/news/self-driving-labs-making-chemical-research-faster-and-smarter

Published: May 2025

15. Source: research-information.bris.ac.uk
Title: Two Kinds of Science D24 for Po S
Link:https://research-information.bris.ac.uk/ws/portalfiles/portal/450758294/Two_Kinds_of_Science_D24-for_PoS.pdf

16. Source: nature.com
Link:https://www.nature.com/articles/d41586-026-00974-2

Additional References

17. Source: researchgate.net
Link:https://www.researchgate.net/publication/397193999_The_Bright_Future_of_Materials_Science_with_AI_Self-Driving_Laboratories_and_Closed-Loop_Discovery

18. Source: institute.global
Link:https://institute.global/insights/tech-and-digitalisation/a-new-national-purpose-reimagining-uk-science-and-technology-through-lovelace-disruptive-invention-laboratories

19. Source: cam.ac.uk
Title: opinion the science reproducibility crisis and what can be done about it
Link:https://www.cam.ac.uk/research/news/opinion-the-science-reproducibility-crisis-and-what-can-be-done-about-it

20. Source: viterbischool.usc.edu
Title: Viterbi | School of Engineering The Reproducibility Crisis in Science
Link:https://viterbischool.usc.edu/news/2022/11/the-reproducibility-crisis-in-science-these-researchers-have-a-fix/

21. Source: synbiobeta.com
Title: briefly bio launches to tackle reproducibility crisis in science
Link:https://www.synbiobeta.com/read/briefly-bio-launches-to-tackle-reproducibility-crisis-in-science

22. Source: flyingpenguins.io
Title: This isn’t a daring prediction so much as a foregone
Link:https://www.flyingpenguins.io/p/will-ai-make-sciences-reproducibility

23. Source: ifp.org
Title: scaling materials discovery with self driving labs
Link:https://ifp.org/scaling-materials-discovery-with-self-driving-labs/

24. Source: preprints.org
Title: These systems integrate machine learning, robotic
Link:https://www.preprints.org/manuscript/202509.1369

25. Source: royalsocietypublishing.org
Link:https://royalsocietypublishing.org/rsos/article/12/7/250646/235354/Autonomous-self-driving-laboratories-a-review-of

26. Source: facebook.com
Link:https://www.facebook.com/NaturePortfolioJournals/posts/a-technology-feature-in-nature-outlines-the-advances-made-to-self-driving-labs-w/1431959458960316/