How we used AI
How we used AI
An honest account of where generative AI was and was not used in building this wiki, and how we reviewed it.
We used AI to help build this website and to generate some of its artwork. We did not use it to generate our science, our data, our figures, or our claims.
Why this page exists
iGEM asks every team to be transparent about how they used generative AI. Used well, AI helps communicate science clearly. Using it to author the substance of a wiki, or to generate data, figures or citations, is not allowed and is research misconduct. This page is our account of which side of that line each use falls on.Tools we used
| Tool | Made by | Used for |
|---|---|---|
| Claude Code | Anthropic | Building this wiki: the design system, React components, styling, navigation, page scaffolding, the coded SVG mascot, and the build checks. Also porting our written content from our own documents into pages. |
| Claude | Anthropic | General assistance: thinking through structure, drafting and tightening wording, writing alt text, and checking our pages against each other for contradictions. |
| Runway | Runway AI | Generating the artwork in our promo video, including the eight gut mechanism stills on our home page. Illustration only, never data. |
| Artlist AI Toolkit & Studio | Artlist | Credited in our promo video alongside Runway. The eight stills published on this wiki were confirmed to be Runway output. |
What we used it for
Building the site
This wiki is a React application, and Claude Code wrote most of its infrastructure: components, layout, styling, the navigation and the site map, the decorative artwork that is drawn in code rather than uploaded, and a set of automated checks that fail our build if a page breaks an iGEM rule. None of that is scientific content. It is the container our content sits in.
Putting our writing on the page
Our content was written by us, as documents: the project description, the notebook, the protocols, the engineering record, the dry-lab write-up, the engagement record, and our own biographies. Those documents were then ported onto these pages. The porting was AI-assisted; the writing was not.
We held one rule through that work, and it is the thing we would point a judge at. Where a document contradicted itself, or contradicted another document, the contradiction was reproduced as written and logged, never silently resolved. Our open-questions log runs to dozens of entries for exactly this reason: numbers that disagreed between pages, dates that did not match, a citation with the wrong lead author, a coefficient whose sign disagreed with its own caption. Every one was raised for a human to settle. None was smoothed over.
Artwork
The eight stills in the mechanism story on our home page are frames from our promo video, which was generated with Runway. They are illustration. The mechanism they illustrate is described and cited in our Project Description, and no claim rests on the artwork.
Our mascot is not AI artwork in the usual sense: it is an SVG drawn in code, written by Claude Code, in the same way the rest of the interface was.
How we reviewed the output
Every change to this wiki was reviewed by a team member before it was published. The repository history records who pushed each one, and nothing reaches the live site except by that route.
Three checks ran on everything:
- Scientific content was checked against the source document. Pages reproduce what our own documents say. Where the page and the document disagreed, the document won.
- Every factual claim that a model could have invented was verified against a primary source. Citations were checked against the publisher record, including one where the author order was wrong and one that had no resolvable identifier at all.
- The build refuses to publish certain mistakes. Our compliance script fails the build on an unreferenced image, a broken internal link, a figure hosted off iGEM’s servers, or a page that has drifted out of the navigation. It cannot catch a wrong number, which is why the first two checks exist.
What this does not prove
A review process is not the same as a guarantee. We are confident about the parts that are checkable, the citations, the links, the numbers that appear on more than one page. For anything else, the answer is that a named member of our team is accountable for every page, and we would rather say that plainly than claim the output was verified in a way we cannot evidence.Where AI was not used
None of the following was generated, written, or altered by AI:
- Our experimental data. Every gel, plate and SDS-PAGE image on this wiki is a photograph of our own bench work. Every number in the notebook and the results pages came off an instrument or out of a sequencing read.
- Our dry-lab analysis. The reanalysis was run by us on public data with the pipelines named on the Software page. Every figure is output from that code, and the code is published so it can be re-run.
- Our scientific claims and decisions. What to build, what the results mean, what we got wrong, and what we would do next are ours.
- Our citations. Every reference was checked to resolve to a real publication.
- Anything anyone said to us. Every quotation from an expert, a stakeholder or a member of the public is what that person actually said. No quote on this wiki was generated, paraphrased into quotation marks, or attributed to anyone who did not say it.
- Our biographies. The fourteen student biographies on our Team page are each member’s own words.