Speed is the easy part. A model can be confident about living matter and still be wrong.
AI is now part of many kinds of biodesign. Models help to design proteins and enzymes, steer how grown living matter forms, run parts of automated labs, and respond to living systems in living interfaces and artistic practice. Some of this takes minutes. That lowers the barrier to designing with living matter, and studies in biodesign education have begun to test generative AI as a teaching aid (Riffle and Rubery, 2026). But speed also raises a new question: what does a good biodesigner need when the tools are this fast?
This page gives a short answer. It defines one term, AI-native biodesigner, and says what the term asks of a person, whatever kind of living matter they work with. It is written for designers, students and teachers who are new to biodesign. You do not need a biology background.
An AI-native biodesigner is a designer of living matter who works with AI models throughout a project, reads what their outputs cannot show, and asks what a change would mean for the organisms and people involved.
“Throughout” does not mean the models lead. The work often starts with the living matter itself: growing it, handling it, getting to know it. What makes the work native is how it is shaped around the models. They make changing living matter cheap, so you try more. They make it easy to be wrong about whether a change is good, so you build in more ways to check, and keep going back to the living matter.
This term describes doing the work well, not just using AI a lot. That is on purpose. A person can use AI models at every step and still not be AI-native.
The definition has three parts. Judgement is the anchor. Fluency and care support it. Each part is something a person can do, so each can be taught and checked.
They read a model’s output and say what it shows, what it cannot show about the change, and what test with the living matter would check it. A confident prediction is not living matter that works. This part cannot be handed to a tool. It grows from time with living matter: growing it, handling it, watching it fail. A model can shortcut knowing about living matter. It cannot shortcut knowing it.
They can run the tools without needing to code. Running them does not need a lab. They know what a model takes in, what it gives back, and what it costs to run. Fluency gets them started. It does not replace time with the living matter.
Before they make a change, they name who and what it touches: the organism that makes it, the other organisms it meets, the people who handle it, and where it ends up. This connects to work on designing for shared responsibility for safety in emerging technologies, including synthetic biology (Van de Poel and Robaey, 2017).
Judgement and care are two halves of one skill: reading what a model leaves out. Judgement asks what the output cannot show about the change. Care asks what it cannot show about who and what the change touches. A structure predictor gives a shape and a score. Nothing in that output mentions the organism that has to make the protein, or the people who handle it. Models make change cheap, so both questions come up more often, and earlier.
Three questions, one for each part. Ask them of a project, a student, or yourself.
Judgement. Can they say what the output does not show about the change, and what test with the living matter would check it?
Fluency. Can they run the tool again, on a new input, without help?
Care. Before making a change, can they name one organism or person it would touch, and say what it would mean for them?
What the answers mean:
If the answer to the second is no, they are not yet working with the models.
If the answer to the first or the third is no, the work is AI-added or AI-led, depending on where the models sit (see section 03).
Only work that passes all three is AI-native.
Teachers can use these as a learning outcome and as a marking prompt. Section 07 shows weak and strong answers.
The six-week course AI-Native Biodesign: With Proteins uses them this way. Each participant answers the three questions about their own work at the start, and again at the final crit, to see how their work has changed.
The difference between these three comes down to two things: where the models sit in the work, and how their outputs are read.
An AI-added biodesigner keeps the old way of working and uses AI at one step to go faster. The living matter is met only at the end.
An AI-led biodesigner has the models everywhere in the work, and lets them decide. The score becomes the verdict. The best-looking option becomes the choice. The screen stands in for the living matter. The models are native to the work. The judgement is not.
An AI-native biodesigner has the models in the work throughout. That doesn’t mean the models go first. The work often starts with the living matter, and keeps returning to it. They design around what the models make cheap, which is changing living matter. They also design around what the models make easy to get wrong, which is knowing whether the change is any good.
| Models at one step | Models throughout | |
|---|---|---|
| Outputs taken at face value | AI-added | AI-led |
| Outputs read for what they leave out | A careful designer who uses AI a little | AI-native |



Takes the confidence score as the answer.
AI at one step of a workflow that was already set.
Asks what a change touches once it exists, if at all.
Takes every score as the answer, at every step.
Models everywhere, run fast and often.
Rarely asks. There are too many changes to ask about.
Says what the output cannot show, and how to check it.
Models throughout, often after time with the living matter.
Names the organisms and people a change touches before making it.
The term is not tied to one kind of work. Protein design is one context, and it is the one used in the worked example below. Others include:
Enzyme and microbial design, where models suggest changes to what an organism makes.
Grown living matter, such as mycelium, where models predict or steer how it forms.
Automated labs, where models choose and run the next experiment (Zhang et al., 2026).
Living interfaces and artistic practice with living matter, where models respond to living systems, or are shaped by them (Kim et al., 2026).
The three parts stay the same in each context. The tools change. The two questions do not: what does this output not show, and who and what does this change touch?
It is not heavy tool use. Five patterns show the gap between running a tool and reading its result. They are common patterns, not named people. Most of them are AI-led: the tools are used throughout, but their outputs are not read.
Reading the score as a verdict. A designer runs a model and takes its most confident output as the answer. The score says how sure the model is. It does not say whether the result will hold up in a living system, or what it would mean for the organism involved. In protein design, for example, a structure predictor reports how close it expects a predicted shape to be to the true one, not what the protein does (Jumper et al., 2021).
Choosing by looks. A designer generates twenty options and picks the one that looks best on screen. No one asks which organism makes it, or who would handle it.
Repeating inherited claims. A designer uses AI to summarise papers, then repeats a mechanism without tracing it to a measurement. The explanation gets more specific. The evidence stays the same. See mechanism accretion and property capture.
Saying “AI designed” and stopping. A project calls its living matter AI designed. The AI is real. No test sits behind the promise.
Reading the screen in place of the living matter. A designer learns living matter only through model outputs and never grows, handles or tests it. The model becomes a wall between the designer and the living thing. A model is a way in. It does not stand in for the living matter.
It is also not a job title or a certificate. It is not a claim that AI replaces lab work. A model can suggest. Someone still has to test. And it is not a biology degree. The skill is reading, asking and spending time with living matter, and designers can learn it.
“AI-native designer” is already used in product and UX design (Wong, 2026). There it often means a designer who delegates repetitive work to AI, works faster, and spends more time judging what is worth building. Job adverts use it for people who design software products built around AI.
“AI-native” is also used for tools and infrastructure. A 2026 preprint describes an AI-native autonomous biofoundry for enzyme engineering (Zhang et al., 2026). In a 2026 CHI panel paper, the author of this page wrote about “AI-native biological tools”, meaning tools that bring opacity and access barriers for novices and cross-disciplinary users (Kim et al., 2026; see also Kim and Zhou, 2025). In those uses the word describes a tool or a system.
During 2026, biotech news and investment writing also began to call drug companies built around AI “AI-native biotechs”: companies that build their own data, models and labs into one discovery pipeline (BioBuzz, 2026; Corona, 2026). There the word describes how a company is built. This page describes how a person works with living matter. The two can meet, since an AI-native biodesigner might work with such a company or read its claims. The test here stays the same at any speed or scale: judgement about what the output shows and what it cannot show.
A similar name, “AI BioDesign”, belongs to a research programme launched in September 2026 by the Allen Institute, the University of Washington and Fred Hutch Cancer Center. In it, AI models propose biological designs, scientists build and test them at scale, and the results feed back into the models (Allen Institute, 2026). That name describes a research programme. This page describes how a person works.
Some autonomous platforms are built to reduce or remove the need for human judgement and domain expertise (Singh et al., 2025; Zhang et al., 2026). This page takes the opposite position. In biodesign, judgement is the part that stays with the person.
This page uses the term for a person, and in a narrower way than product design. The anchor is judgement about what a model’s output can and cannot show about living matter, and care for the organisms and people a change touches. Speed and delegation are not part of the definition. Where speed and delegation come without judgement, this page calls the work AI-led.
On 4 October 2026 I searched for an earlier definition of “AI-native biodesigner” and did not find one. The term “AI-native” was already used for tools and systems, including in my own 2026 CHI panel paper (Kim et al., 2026). This page extends it to the person. The search was limited: a few web searches, not Scholar, proceedings or job titles. If you know of an earlier use, please send it and I will add it here.
Protein design is a useful example because the gap between a prediction and living matter that works is easy to see.
In the Biodesign Academy masterclass “Design your first protein with AI” (1 October 2026), designers pasted an 86-letter hydrophobin sequence into a free browser tool and got a predicted shape in minutes. Then they changed it. Each change came back with a confidence score. Some scores were not what people expected. Working out why was the main lesson.
Running the tool took minutes. Reading the result took the rest of the session.
Here is what weak and strong answers to the three questions look like for that exercise.
| Weak answer | Strong answer | |
|---|---|---|
| Judgement | “The score doesn’t show function.” | “A high score says the shape is likely. It doesn’t say the changed protein still forms a water-repellent film. You’d have to make it and test the film.” |
| Fluency | Ran it once, with help. | Runs it again, alone, on their own sequence. |
| Care | “It affects the yeast.” | “The yeast has to make more of a protein it doesn’t need, which can slow its growth. And the people who handle the dry powder may breathe it in.” |
The same gap appears in other contexts. A model can predict how grown living matter will form, or which version of an enzyme will work best. In each case, running the model is quick. Reading what it cannot tell you is the work.
The masterclass has since grown into a six-week online course, AI-Native Biodesign: With Proteins. It starts from zero, uses proteins as the way in, and builds judgement, fluency and care week by week on each participant’s own project.
No. Protein design is one context. The term applies to any biodesign work that uses AI models: enzymes and microbes, grown living matter, automated labs, living interfaces, and artistic practice with living matter.
No. AI literacy is a general understanding of what AI is and how it works. An AI-native biodesigner uses that understanding in a specific practice: running models on living matter, reading the outputs, and asking what a change would touch. Literacy helps. It is not enough.
Related, but different. In product design the term often centres on speed and delegation. Here the anchor is judgement about a model’s output on living matter, and care for organisms and people.
No. Someone can use models at every step and be AI-led. AI-native means using them throughout and reading what they leave out.
Work where the models are everywhere and decide everything. The score is the verdict, the best-looking option is the choice, and the screen stands in for the living matter. It is the most common way to use AI heavily and still not be AI-native.
No. It describes how someone works, not their age. A senior designer can be AI-native. A student may not be. It is used here because it is the word people already search for. What matters is the definition.
No. The skill is reading a result, asking good questions, and spending time with living matter. Designers can learn it.
Not to start. You can run the tools in a browser, and that is a real way in. But judgement grows with hands-on time: a kitchen-table mycelium grow, a community lab, a studio. An AI-native biodesigner stays close to the living thing as well as the screen.
No. It means the models are part of the work throughout, not that they lead it. In biodesign, the work often starts with the living matter: growing it, handling it, watching what it does. The models come in alongside that, and the designer keeps going back to the living matter to test what they suggest.
No. It is a way of working, with a short test you can apply.
An AI-added designer uses AI at one step of an old workflow, and meets the living matter at the end. An AI-native designer works with the models throughout a project, often begins with the living matter, and checks what the output cannot show.
Allen Institute (2026, 3 September). AI BioDesign accelerator combines experimental biology and artificial intelligence to learn nature’s design rules. Allen Institute. alleninstitute.org
BioBuzz (2026, 23 March). 10 AI-native biotechs to watch in 2026. BioBuzz. biobuzz.io
Corona, A. (2026). The 2026 AI power shift. Drug Discovery News. drugdiscoverynews.com
Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), 583-589.
Kim, R., Chen, Y., Ramirez-Figueroa, C., Zhou, J., Telhan, O., Nicenboim, I., Dade-Robertson, M., Pevere, M., & Robaey, Z. (2026). Biodesign x AI: Interactions in the Algorithmic Wet Lab. In Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA ’26), April 13-17, 2026, Barcelona, Spain. ACM, New York, NY, USA, 6 pages. doi.org/10.1145/3772363.3790105
Kim, R., & Zhou, J. (2025). Inclusive Molecular Sketching: Accessibility Barriers in AI-Driven Protein Biodesign Workflows. In Proceedings of the 28th International Academic Mindtrek Conference (Mindtrek ’25). ACM, New York, NY, USA, 400-404. doi.org/10.1145/3757980.3762124
Riffle, D., & Rubery, P. (2026). Biodesign Buddy: Integrating Generative Artificial Intelligence in Academic Biodesign. bioRxiv preprint, posted 13 March 2026, not peer reviewed. doi.org/10.64898/2026.03.11.710906
Singh, N., Lane, S., Yu, T., Lu, J., Ramos, A., Cui, H., & Zhao, H. (2025). A generalized platform for artificial intelligence-powered autonomous enzyme engineering. Nature Communications, 16, 5648. doi.org/10.1038/s41467-025-61209-y
Van de Poel, I., & Robaey, Z. (2017). Safe-by-design: From safety to responsibility. NanoEthics, 11(3), 297-306.
Wong, C. K. (2026, 3 June). Being an AI-native designer isn’t what you think it is. The Designer’s Field Guide. thedesignersfieldguide.substack.com
Zhang, C., Yang, L., Qin, Y., Li, D., Dong, S., & Yang, M. (2026). An AI-Native Biofoundry for Autonomous Enzyme Engineering: Integrating Active Learning with Automated Experimentation. bioRxiv preprint, posted 1 February 2026, not peer reviewed. doi.org/10.64898/2026.02.01.703093
Kim, R. (2026). AI-native biodesigner (Version 1.4). Biodesign Academy Glossary. https://www.biodesign.academy/ai-native-biodesigner
Longer version, with sources and reasoning: Kim, R. (2026). AI-native biodesigner: a definition and a three-question test for biodesign education (Version 1.1) [Working paper]. Biodesign Academy Ltd. Zenodo. doi.org/10.5281/zenodo.24177790