Designing the experiment
88Deciding which hypothesis is worth testing, and which experimental design could actually falsify it, requires scientific judgment no current AI system can reliably originate on its own.
Biology looks automatable from the outside in a specific way: much of the raw work — pipetting, sequencing, image analysis, literature search — is repetitive enough that software and robots already do parts of it faster and more consistently than a person. What is harder to automate is the judgment that decides which experiment is worth running in the first place.
AI tools like protein-structure predictors and lab-automation robots are real and expanding fast, changing how much of the routine bench and analysis work gets done. What they have not changed is who decides what question is worth asking, interprets an ambiguous or contradictory result in context, and takes responsibility for a published claim being right.
Moderate
Share of the work a machine could do
A substantial share of biology's routine execution — sample processing, sequencing, literature search, basic image and data analysis — is already partly automated or heading that way. The core of the job that survives is choosing which question is worth asking, designing an experiment that can actually answer it, and interpreting results a machine cannot yet reliably tell apart from noise.
Scored from the tasks, not the job title. Lower is safer.
Jobs AI cannot take →What machines cannot take
Interpreting ambiguous or conflicting data
85Deciding whether an odd result is a real biological finding, a technical artifact, or noise draws on contextual scientific judgment that current AI pattern-matching does not reliably replicate.
Delicate wet-lab and field technique
80Living tissue, cell cultures and field organisms behave unpredictably; today's lab-automation robots execute pre-defined protocols precisely, but still cannot improvise when a sample does not behave as expected.
Fieldwork in unpredictable environments
74Tracking animals, surveying remote habitats or collecting samples under changing weather and terrain still depends on physically present, adaptable human judgment machines are not close to matching.
Scientific accountability and peer review
66A named researcher stands behind a published claim's accuracy and stakes their reputation on it; no journal or funding body has worked out how, or whether, to put an AI system's name there instead.
What they already take
Sample processing and sequencing
78Automated liquid-handling robots and high-throughput sequencers now process thousands of samples with far less human hands-on time than a decade ago, especially in genomics-heavy labs.
Protein structure prediction
72Deep-learning models like AlphaFold can now predict a protein's three-dimensional shape from its sequence in minutes, a task that used to take years of laboratory crystallography per protein.
Routine image analysis and cell counting
66AI-based image recognition now counts cells, flags abnormalities and measures structures in microscopy images with a consistency that reduces, though does not eliminate, tedious manual scoring.
Literature search and summarization
58AI tools can now scan and summarize thousands of papers far faster than a researcher reading manually, though verifying the summaries against the original findings still falls to a human.
How the work is changing
AI predicts and designs molecules, not just analyzes them
Following AlphaFold's 2020 breakthrough in predicting protein structure — work that won its creators a share of the 2024 Nobel Prize in Chemistry alongside protein-design pioneer David Baker — AI tools increasingly help design new proteins and molecules, not just interpret existing ones.
Cloud labs let biologists run experiments by writing code
Remote-operated robotic laboratories now let researchers submit experiments as code and receive results without ever touching a pipette themselves, shifting some bench work toward a more software-engineering-like workflow.
Wet-lab and computational skills increasingly sit in the same person
Biology projects increasingly require comfort with both hands-on experimental technique and computational data analysis, and researchers who can do both are becoming more valuable than narrow specialists in either alone.
Preprints are speeding up a traditionally slow field
Rapid-sharing platforms like bioRxiv, launched in 2013, let biologists post findings before formal peer review, compressing a publishing cycle that once took a year or more into weeks, at some cost to pre-publication vetting.
New jobs branching off
Computational biologist / bioinformatician
Analyzes large genomic, proteomic or ecological datasets computationally, a role that barely existed as a distinct career before the genomic era and now sits at the center of most large biology projects.
Synthetic biology engineer
Designs and builds novel biological systems — engineered microbes, genetic circuits, lab-grown tissue — treating living cells as programmable components rather than only objects of study.
Lab-automation scientist
Designs and manages robotic, remote-operated laboratory workflows for cloud labs and high-throughput facilities, a hybrid role between traditional bench science and automation engineering.
Genetic counselor
Translates increasingly detailed genomic test results into terms patients and families can actually use to make medical decisions, a fast-growing role sitting between clinical genetics and direct patient care.
Outlook
The tasks most exposed to automation are the ones that can be fully specified in advance: a routine sequencing run, a standard image-analysis pipeline, a protein-structure prediction a model has been trained on thousands of similar examples for. The tasks most protected are the ones that cannot be pre-specified: which question is worth years of a career, what an odd result actually means, and how to interpret a living system that keeps refusing to behave like the model predicted.
That split is unlikely to shrink the number of people called biologists so much as change what the job actually involves day to day — more time spent designing experiments and interpreting results, less time spent on the repetitive execution now increasingly handled by robots and software. A biologist working in 2040 will very likely spend less time at the pipette and more time at the keyboard than one does today.
None of that changes who is accountable when a published biological claim turns out to be wrong, or who decides which experiment gets funded and run in the first place. Until an AI system can be trusted with that responsibility, and no regulator or funding body currently intends to hand it over, the center of the job stays human.