BIOEE 7600-103 · Week 5
An agent helps you write
9 October 2026
Public copy of the slides as
shown in class, with class-specific detail removed. Course page:
benninglab.org/teaching/agentic-ai
Trending · this week
Hot goss
6 Oct · OpenAI dumps 719 math manuscripts on GitHub
- From an unreleased internal model, grouped into 372 result families. About 4,000 open problems posed; ~3 hours of compute per result on average.
- ~42% of top-line results are formally verified in Lean, a proof assistant: software that checks every logical step of a proof mechanically. A Lean-checked proof is correct, provided the theorem was stated correctly. The rest are unchecked, and none are peer-reviewed.
- Day 2: three manuscripts withdrawn after a sign error broke a chain of dependent proofs; 14 more revised.
- Snapshot of the field: "100+ reactions to 100+ solutions" (Proofs & Prompts, 8 Oct; shared by Daniel Litt).
"Quasi-Riemann is a result that I thought was complete science fiction."
Tim Santens"My feelings on recent developments are very mixed and complex."
Terence Tao"The results OpenAI announced are amazing … That said, I am sad about how these questions were treated."
Inhyeok Choi"If OpenAI wanted to destroy the mathematical community, this would be a great way to go about it."
Henry Wilton"They can simply move on and slop-bulldoze the next conjecture. And, of course, the credit remains theirs."
Enrico Fatighenti"I’m worried how big corporates are controlling academic research."
Barna Saha
Today
Three exercises, ordered by how much of the writing the agent does
1 · Friendly reviewer
- It edits one paragraph of your draft.
2 · Methods from code
- It describes your analysis script.
3 · Proposal from a review
- It originates a project summary from last week's review.
benninglab.org/teaching/agentic-ai/writing
Start now · runs take 2–5 minutes
Pick one
Open the course page and copy the prompt for whichever exercise interests you most.
- 1 · Friendly reviewer: a whole paper or draft for context, and the one paragraph to edit.
- 2 · Methods from code: an analysis script of yours.
- 3 · Proposal from a review: your Week 4 review (
review_original.html is best; it carries the DOIs).
- Already ran one? Open your output.
Your materialUse something you wrote, or have the right to share.
Exercise 1 · what we ran
One abstract, ten rewrites
Benning, Eckhart, Geber & Moeller 2019, Am Nat. Five runs each of "make it clearer" and "make it more compelling", no checker.
| Change from the original | Clear (of 5) | Compelling (of 5) |
|---|
| "could potentially persist" loses "potentially" | 4 | 4 |
| Simulations "showed" softened to "indicated" / "suggested" | 3 | 4 |
| "estimate contributions of herbivory" becomes "test whether herbivory sets the range limit" | 0 | 4 |
| A claim added that is not in the original | 0 | 4 |
| "strongly associated" turned causal | 0 | 2 |
| Opening line rescoped ("many", "most", "every" species) | 2 | 3 |
| Any change to what is claimed, or how strongly | 5 | 5 |
Exercise 1 · "make it more compelling", run 4
"Compelling" is a request to change claims
AimWe usedcombined field experiments andwith demographic simulations to estimatetest contributionswhether mammalian herbivores set the range limit of mammalian herbivory to a range boundary in the CalifornianCalifornia annual plant…
Last sentence…empirical support to recent developmentstheory inon range limits, and they suggest that forecasts of range shifts that consider only climate may miss important constraints.
Exercise 1 · the guarded prompt
A separate checker, claim by claim
10 / 10plain rewrites changed a claim or its strength
0 / 3guarded rewrites did
Guarded rewrite vs. originalSpecies' geographic distributions have already shifted during the Anthropocene. However, we often do not know what aspects of the environment drive range dynamics, much less which traits mediate organisms' responses to these environmental gradients. Most studies focus on possible climatic limits to species' distributions and have ignored the role of biotic interactions, despite theoretical support for their importance in setting distributional limits. We used field experiments and simulations to estimate contributionsthe contribution of mammalian herbivory to a range boundary in the Californian annual plant Clarkia xantiana ssp. xantiana. A steep gradient of increasingThe probability of herbivory occurredincreased steeply across the boundary,. and aA reanalysis of prior transplant experiments revealed that herbivory drove severalfold declines in lifetime fitness at and beyond the boundary. Simulations showed that populations could potentially persist beyond the range margin in the absence of herbivory. Using data from a narrowly sympatric subspecies, ClarkiaC. xantiana ssp. parviflora, we also showed that delayed phenology is strongly associated with the susceptibility of C. xantiana ssp. xantiana's susceptibilityxantiana to herbivory and its low fitness beyond its border. Overall, our results provide some of the most comprehensive evidence to date of how the interplay of demography, traits, and spatial gradients in species interactions can produce a geographic range limit, and they lend empirical support to recent developments in range limits theory.
Your runsDid Claude change any claims meaningfully? Did it help with clarity?
Exercise 2 · methods from code
The paper's analysis script, citation removed
The R Markdown from the 2019 paper's Dryad archive. Author line and citation stripped, web search off.
- Two plain runs: a one-line prompt.
- One guarded run: asks first, then a separate verifier ties every statement to a line of code.
- Each output scored, statement by statement, against the code and the published methods.
Exercise 2 · Claude's methods vs. the paper's
More detailed than the paper
More than the paper
- Length: ~1,500 words (plain), ~2,200 (guarded), all on the statistics.
- 23–40 statements true of the code that the paper's main text omits or words differently.
- Transforms, spans, filters, constants:
log(x + 1), loess span = 1, dropped stems, hard-coded residual SD.
- Packages for each step, and robustness refits the paper puts in appendices.
- Sometimes more correct: the paper says "log(fitness)"; the code, and all three runs, say
log(fitness + 1).
Where it fell short
- Missed small details: R² as a model check, Nagelkerke's pseudo-R².
- Errors: 3 per plain run, mostly following the code's comments instead of what it runs (as the paper does). Guarded: 0, after its verifier fixed 19 points in its own draft.
Exercise 3 · proposal from a review
A Project Summary from a Week 4 review
Rehearsal topic: gene flow and adaptation at range edges. Prior work only from references the review cites.
Exercise 3 · what it wrote
Why does adaptation stop at a range edge?
- The question: swamping by gene flow from central populations (Kirkpatrick & Barton 1997), or drift eroding the variance adaptation needs (Polechová & Barton 2015)? Or is the edge not a limit to adaptation at all?
- Aim 1: replicated lab landscapes with an imposed gradient, crossing dispersal rate with deme size. Swamping predicts range contraction with dispersal at every deme size; drift limitation predicts an interaction.
- Aim 2: at replicate field edges, transplant beyond each edge, measure the drift-threshold parameters, and carry crosses from climate-matched and central sources to F2 and backcross over three years.
- Left for me: the study organism, number of edges, preliminary data, partners, trainees.
What the rules say now
Current guidelines from journals and agencies
| Source | Rule |
|---|
| NSF notice, Dec 2023 (updated May 2026) | Disclosure encouraged; proposer responsible for accuracy. Reviewers may not upload proposals to non-approved tools. |
| NSF PAPPG 24-1 Suppl. 1, Dec 2025 | Fabrication, falsification or plagiarism done with AI tools is research misconduct. Disclosing AI use in proposals is encouraged, not required. |
| NIH NOT-OD-25-132, Jul 2025 | Applications "substantially developed by AI" not considered original; AI detection; 6 applications per PI per year. |
| NIH NOT-OD-23-149, Jun 2023 | Reviewers may not use GenAI on applications or critiques. |
| Science journals | AI-generated text allowed if disclosed: cover letter, plus Methods or Acknowledgments depending on use. AI not an author; AI images need editor permission. |
| Wiley (Ecol Lett, BES journals), Jul 2026 | Disclose at submission and in Methods / Acknowledgements; spelling and grammar exempt. |
| ESA journals | Not an author; disclose; spelling and grammar exempt; peer review not delegable. |
| Evolution (SSE / OUP) | Disclose in cover letter and Methods; report model, version and "query structure, syntax". |
| Springer Nature | LLM not an author; document use in Methods; copy-editing exempt; no GenAI images. |
Discussion
What do you think?
- Did your reviewer give helpful edits? How does it compare to asking a human for a friendly review?
- Was the generated proposal novel, milquetoast, wrong, or something else? Would you use genAI to produce a first draft of a proposal?
- Evolution asks you to report the "query structure" of AI use. How would you report a 40-step agent session?