An answer with references and polished charts still needs source support, correct analysis, independent reproduction and disciplinary judgment before it becomes a research conclusion. This guide checks materials as of November 15, 2026 and combines current product documentation, university cases and peer-reviewed papers into a traceable workflow. The prompts and four data rows are original teaching examples. The offline script checks synthetic arithmetic.
Read the progress in 2026 by type of evidence
This year's developments connect language models to materials, programs and feedback. Paper2Agent works with a paper and its callable methods. ERA searches over code candidates that can be scored. Research labs are beginning to interpret many existing experimental observations in parallel. Their results depend on different inputs, objectives and checks, so tool selection starts with the task and study design.
| Date and material | What was observed | Scope of the conclusion |
|---|---|---|
| 2026-11-06, BioBasicsBench | Scores on 100 tasks using raw wet-lab evidence | Vendor benchmark. Check the scoring and failure cases in the original |
| 2026-10-21, Rosalind analysis case | Parallel interpretation of existing experimental data | Vendor collaboration and preprint. Follow-up hypotheses require experiments |
| 2026-09-16, Nature Paper2Agent | Papers and code exposed as callable methods | Peer-reviewed computational study, constrained by code and environment availability |
| 2026-08-27, Bocconi/OpenAI | Assignment performance in a classroom randomized experiment | Technical report and institutional account. The grading dimensions affect interpretation |
OpenAI: BioBasicsBench, 2026-11-06 · OpenAI: distributed biological analysis, 2026-10-21 · Nature: Paper2Agent · OpenAI and Bocconi: ChatGPT and causal-thinking training
Record the sample, comparator, execution budget and endpoint as you read. Algorithm scores, expert preferences and experimental validation belong in separate fields even within one study. A model's own confidence label can go into the list of items to check. The researcher still needs to locate the underlying support.
Identify the product, access permissions and storage location
ChatGPT file analysis can help explore tables and inspect analysis code. The API lets a team design its own orchestration and records of tool calls. Specialized scientific models have additional access requirements. Current official help distinguishes Rosalind Research, Discovery and product approvals. The Academic Plan includes no API access or API credits. Labs should check current eligibility before applying.
OpenAI Help: current GPT-Rosalind access · OpenAI Help: data analysis with ChatGPT
| Surface | Useful first tasks | Checks before formal research |
|---|---|---|
| ChatGPT and data analysis | Organize public material, explore small tables, draft code | Exportable inputs, analysis code, results and human edits |
| Your own API workflow | Defined retrieval scope, versioned tasks, batch records | Institutional approval, usage costs, external tool permissions and logs |
| Specialized access such as Rosalind | Connect scientific tools within the approved scope | Organization and user approval, the specific product and separate API eligibility |
| Paper-specific agents | Reuse the program and method from a paper | Version, license, dependencies, test data and applicable field |
OpenAI Help: current GPT-Rosalind access · Nature: Paper2Agent · OpenAI API: web search sources and citations · OpenAI Help: data analysis with ChatGPT
Workspaces have a lifecycle too. The current Prism announcement says it will close on January 1, 2027 and asks users to export in advance. Teams using it can move source files, references and figures into materials they can retain and open themselves. Check that a new tool can export research artifacts before committing a project to it.
Turn the research question into a checkable task
Before searching, write one page describing the population or object, intervention or comparison, outcome measures, time range and exclusion rules. An exploratory project can retain unknowns while marking choices made after seeing the data. Ask the model to identify conflicting definitions before generating subquestions. This helps prevent related but distinct topics from merging into one conclusion.
For example, a graduate student comparing two forms of teaching support should define assignment quality, immediate testing and delayed unaided testing separately. Each measures a different activity. Agree on the endpoint first so you can assess what a polished assignment says about learning. Assign people to check literature, analysis code and the interpretation in the relevant field.
Task: Create a checkable analysis plan using only my question and supplied materials.
List the population or object, comparison, primary outcome, time range and exclusion rules.
Identify incomplete definitions and retain unknowns.
Distinguish prespecified analyses from exploration added after seeing results.
For each step, list inputs, artifacts, verification method and human owner.
List papers whose original text you have not read as candidates to check.
Do not invent DOIs, sample sizes, measurements or experimental results.Turn search results into a record of claims and sources
A search tool can return citations with URLs and the sources it consulted. Open the original, locate the methods, sample and results, and check whether they support the sentence you are writing. A title, abstract and full text provide different levels of evidence. When full text is unavailable, record what you actually read and look for an author archive or other accessible material.
Keep fields for source_id, stable link, version date, passage or table, support, limitations and reviewer. Connect a preprint to its later version and retain negative and conflicting findings. A systematic review also needs the discipline's records for databases, search strings, deduplication and inclusion decisions. Completeness requires its own check.
Using only original material actually read in this round, record every important claim.
Fields: claim_id, source_id, URL/DOI, version date, page or table,
study design, sample/denominator, direct support, limitations and unresolved checks.
Separate reported results, authors' interpretations and your own inferences.
Keep conflicting findings side by side. Do not vote by citation count.
Identify the evidence most likely to overturn the conclusion and the next search question.
List sources for which you saw only an abstract or could not access the full text.Let the model write code while retaining reproducible inputs
Before analysis, retain a read-only original and its checksum. Define missing values, units, duplicate records and the sampling unit. Separate cleaning scripts from result tables and keep the actual plotting data. ChatGPT can show the analysis code it generated. Read the important transformations and recalculate in an environment your team controls.
The four rows below demonstrate records and arithmetic checks. The script verifies unique row IDs, group labels and integer types, then calculates means of 3 for A and 5 for B, with a difference of 2. It includes no random assignment, error model or external validation, and explicitly keeps the causal-claim field false. For real research, add appropriate statistical analysis and independent data.
import json
from fractions import Fraction
# Synthetic classroom fixture. No model call or external data.
rows = [
{"id": "s1", "group": "A", "value": 2},
{"id": "s2", "group": "A", "value": 4},
{"id": "s3", "group": "B", "value": 4},
{"id": "s4", "group": "B", "value": 6},
]
assert len({row["id"] for row in rows}) == len(rows)
assert {row["group"] for row in rows} == {"A", "B"}
assert all(type(row["value"]) is int for row in rows)
means = {}
for group in ("A", "B"):
values = [row["value"] for row in rows if row["group"] == group]
assert len(values) == 2
means[group] = Fraction(sum(values), len(values))
difference = means["B"] - means["A"]
assert difference == 2
result = {
"schema": "research-check-fixture/v1",
"synthetic": True,
"n": len(rows),
"mean_A": str(means["A"]),
"mean_B": str(means["B"]),
"difference_B_minus_A": str(difference),
"causal_claim_supported": False,
}
print(json.dumps(result, sort_keys=True))
When execution fails, record the input version, error, repair and new output together. Once a program runs, inspect leakage, grouping units, train/test splits and dimensions. Changes to the metric or removal of unusual records should produce a reviewable difference for the research lead to accept or reject.
The Stanford case shows how a paper becomes callable
Paper2Agent packages papers, data and programs as methods an agent can call. In Nature's computational biology corpus, 74 of 100 papers were converted successfully. Missing code, data or a working environment were important failure causes. This supports reusable research engineering and suggests checking material availability before choosing a project.
Stanford's account explains that failed attempts and experimental judgments beyond the paper still need the authors' input. Before connecting a paper agent, reproduce the paper's own example, then test new held-out inputs and mark assumptions the authors have clarified. When several paper agents collaborate, record units, input constraints and uncertainty at each handoff.
The Rosalind case turns existing data into follow-up questions
On October 21, OpenAI reported that it, PerturbAI and Scripps Research used thousands of agents to interpret existing mouse brain-cell experimental data in 32 minutes and generate follow-up leads. The linked study is a preprint. The timing belongs to that analysis run. Proposed biological mechanisms need further experiments.
A transferable approach is to divide work into independently checkable objects, then combine reports in one record. Retain inputs, citations, candidate explanations and counterevidence for each object. The consolidating reviewer checks for repeated sources being counted as independent evidence and shared faulty assumptions across agents. More agents also require more capacity for deduplication and review.
Use raw experimental evidence to expose model blind spots
BioBasicsBench, published on November 6, contains 100 raw wet-lab tasks. Astra at xhigh earned 66.7% of available credit on average. This is rubric-based mean reward. The official failures include an assumption overriding experimental controls, showing how a correct local image observation can still produce a wrong conclusion.
An experimental agent can therefore report observations, interpretations and proposed actions in separate fields. Observations identify image regions, measurements and controls. Interpretations list competing hypotheses. Actions specify the evidence still needed for a domain researcher to decide. Ambiguous images, missing controls and contradictory readings should allow an insufficient-evidence outcome and retain the originals for review.
Compare models through the task and validation design
Nature's Co-Scientist paper describes Gemini agents generating and refining hypotheses, with validation in defined biomedical applications including in vitro experiments. ERA searches code candidates around scorable tasks and reports held-out evaluations. These approaches emphasize candidate-question quality and performance after a method is executed.
Nature: Empirical Research Assistance · Nature: Co-Scientist
| Research object | Main model output | Independent check to retain |
|---|---|---|
| Search and synthesis | Literature set, evidence table, draft interpretation | Original support, inclusion decisions and contradictory results |
| Computational method and program | Cleaning code, fitting, simulation and prediction | Retained inputs, test split, recalculation and external data |
| Experimental hypothesis | Candidate mechanism and reasons for a proposed experiment | Controls, replication and disciplinary review |
| Multi-agent synthesis | Reports across objects and cross-disciplinary connections | Deduplication, shared assumptions, review queue and stop conditions |
When selecting a system locally, provide the same material, tools and budget to each candidate. Hide reference results and include unfinished tasks and human rework in the denominator. A better hypothesis score still needs a separate check on your images, programs or reproduction tasks.
Education research needs both assignments and unaided learning
In OpenAI's March study of over 300 students, the primary ITT result improved in microeconomics while the neuroscience difference was not distinguishable. August's Bocconi randomized experiment found ChatGPT support and causal-thinking training improved different assignment dimensions. These institutional studies retain their task and assessment boundaries.
OpenAI: understanding AI and learning outcomes · OpenAI and Bocconi: ChatGPT and causal-thinking training
| Goal | AI-assisted activity | Independent measure |
|---|---|---|
| Master a method | Ask for steps and counterexamples | Unaided transfer problems and delayed tests |
| Judge literature | Organize supporting and conflicting materials | Student explains orally why a citation supports a claim |
| Programming capability | Draft code and suggested tests | Adapt code to new data and explain failures |
| Original questions | Generate candidate directions and challenge assumptions | Assess novelty and testability beyond the preset rubric |
Set the primary endpoint and grader blinding before the experiment, and record actual use and attrition. Distinguish the effect of being offered access from intensity of use. Schedule delayed testing separately. Language, accessibility and device conditions also matter so that comparison groups can access the basic learning resources.
Attach confidentiality, responsibility and disclosure to the artifacts
ICMJE requires disclosure of AI use under the applicable journal's rules and keeps authorship responsibility with humans. Nature Computational Science's policy also restricts uploading confidential review material to public or unsecured AI systems. Before submission, the corresponding author checks the target journal, institution and project requirements.
ICMJE: authors' responsibilities when using AI · Nature Computational Science: transparent use of AI
OpenAI documents default training use for business, education and API data. Connected apps, organizational access and retention still need separate checks. A research team can maintain a list of uploadable materials, with institutional review of the handling route for unpublished data, personal information, licensed material and participant consent.
- Keep the research question, data provenance, licenses and applicable approvals.
- Record the model, date, tool permissions, prompt version and human edits.
- Package original inputs, code, outputs and failure logs for export.
- Assign different roles to check citation support, numerical analysis and disciplinary conclusions.
- Record the actual AI contribution and disclose it under journal and institutional rules.
- Assign a maintainer for source corrections, environment changes and broken links.
Measure value through accepted artifacts and review effort
Start with ten small tasks with known answers or feasible human verification. Define acceptance criteria before the agents run. Count original sources found, code reproduced, usable charts and appropriate insufficient-evidence responses separately. Record failures, waiting and repairs too, so the next round can address the tool or the task definition.
| Measure | Calculation | Context to retain |
|---|---|---|
| Accepted delivery rate | Tasks passing preset checks ÷ all assigned tasks | Difficulty, attrition and failures |
| Total human time | Preparation + supervision + review + repairs | Compare with a manual process held to the same quality standard |
| Cost per accepted item | Total pilot expense ÷ accepted deliveries | If none qualify, report that the ratio is unavailable |
| External reproduction | Results repeated by an independent person or on independent data | Replication design, materials and versions |
| Error recovery | Steps required to withdraw a wrong output and rerun | Whether it already entered a report or external decision |
Extra speed can matter, especially for repeated organization and script drafts. Check whether review queues consume the time saved before accepting a claimed gain. Summarize error types weekly and fix repeated causes such as missing inputs, inconsistent units and unreproducible environments first.
Give every result a path back to its source
- Prespecify the question, outcomes and permitted materials
- Read originals and connect claims to exact source locations
- Fix the inputs and environment, retain model and code versions
- Generate candidate explanations, programs or experiments
- Check controls, held-out data and independent review
- Get the research lead's sign-off and disclose the actual AI contribution
- Retain failures and uncertainty to plan the next validation
Trends to watch over the next 12 to 36 months
These are conditional judgments based on the materials above, with time measured from November 2026. The nearest opportunities connect checkable artifacts, including search to code, papers to tools, and analysis to candidate explanations. Longer-term gains also depend on data permission, external validation, experimental scheduling and team maintenance. One model update cannot establish their timetable.
| Horizon and trend | Conditions needed | Counterevidence to keep watching |
|---|---|---|
| Next 0–12 months, more papers with callable methods | Authors publish maintainable code, data and environments | Frequent dependency failures make reproduction costlier than manual adaptation |
| Next 0–12 months, more specialized scientific tools | Disciplinary evaluation, access approval and integration with existing tools | A general model suffices while specialized access adds administrative effort |
| Next 12–36 months, larger multi-step computation and experiment loops | Instrument interfaces, approvals, recovery and validation advance together | Too many candidates for available experiments, with review becoming the bottleneck |
| Next 12–36 months, courses assess more of the research process | Credible measures of unaided understanding, transfer and retention | Only final text improves and learning gains do not persist |
A lab can start now with one repeatable task using low-sensitivity material. Record the existing process's time and quality, then run a comparable pilot. Keep citations, code and review decisions and expand according to the results. As models improve, the team will already have a research record it can compare and migrate.