Augmentation without abdication is the phrase two American researchers have proposed as the governing norm for artificial intelligence in science, and it arrived on 8 September 2026 with something most position pieces never bother to supply: a list. Charles C Branas of Columbia University and Bruce L Levine of the University of Pennsylvania did not simply argue that researchers should stay in charge. They enumerated the jobs a machine may do and the jobs it may not.
The editorial, “AI and authorship: Norms and uses to preserve human-led science,” appeared in PNAS Nexus under a Creative Commons licence. It runs to five short sections and thirteen references. Its argument is compact, quotable and, in one respect, unusually testable — because a list of named tasks can be counted, sorted and checked against the definition the same document supplies.
So that is what this piece does. Not a summary of the argument, which is easy to agree with, but an audit of the taxonomy underneath it. The editorial names 31 discrete research tasks: 18 it marks allowable, 13 it marks nonallowable. Those two numbers are the entire operational content of the proposal, and nobody covering the paper has published them.
The audit produces one finding the authors will not enjoy. The editorial’s second section is titled “Generative and nongenerative uses of AI are not the same,” and the whole framing rests on that distinction. But when you sort the 31 tasks against the editorial’s own definition of generative AI, 22 of them are generative — and they split 11 on the allowed side and 11 on the banned side. Exactly evenly. Whether a task is generative tells you nothing at all about which list it lands on.
That is not a reason to dismiss augmentation without abdication. It is a reason to notice that the real axis is accountability, and that the paper’s headline distinction is doing none of the work its own section title claims for it.
Table of contents
- What Augmentation Without Abdication Actually Proposes
- The 31 Tasks Behind Augmentation Without Abdication
- Eleven Allowed Tasks Are Generative by the Editorial’s Own Definition
- Two Banned Tasks That Generate Nothing
- Seven Boundary Pairs Augmentation Without Abdication Leaves Undefined
- What the Augmentation Without Abdication Argument Gets Right
- How to Apply Augmentation Without Abdication in a Real Lab
- Where the Editorial Stops Short
- Frequently Asked Questions About Augmentation Without Abdication
- References
What Augmentation Without Abdication Actually Proposes
The norm is stated plainly, and it is worth quoting before taking it apart, because the phrase has already outrun the paper that produced it.
The sentence the phrase comes from
The editorial’s fourth section is headed “A norm of augmentation without abdication,” and defines it as “using AI to extend human capacity without surrendering scientific autonomy, independence, or responsibility.” The load-bearing word is surrendering. This is not a rule about what a model is capable of. It is a rule about what a human being remains answerable for.
The authors put the accountability point in a single line: “AI can help prepare that record, but it cannot be responsible for it.” Everything else in the paper is downstream of that sentence.
Two authors, two institutions, thirteen references
Branas is in Columbia University’s Department of Epidemiology; Levine is at Penn’s Perelman School of Medicine. Neither is a machine learning researcher, which matters — this is a document written by working scientists about their own field’s norms, not a technical assessment of model behaviour.
The reference list is short at thirteen entries and points where you would expect: the Committee on Publication Ethics, the International Committee of Medical Journal Editors, and papers in PNAS, Science, Cell and Accountability in Research. It is a publishing-ethics document, not an AI safety document.
Why augmentation without abdication is a norm and not a policy
The paper is explicit that current journal rules have already converged on a thin consensus: AI systems cannot be authors, human authors keep full responsibility, and substantive generative use should be disclosed. Augmentation without abdication is offered as the principle underneath that consensus rather than a replacement for it.
That framing is honest, and it is also the source of the problem examined below. A norm with 31 named tasks attached is no longer purely a norm. It is a draft policy wearing a norm’s clothing, and it invites the kind of checking a norm would not.
The 31 Tasks Behind Augmentation Without Abdication
Both lists appear as running prose separated by semicolons rather than as a table, which is probably why no coverage has counted them. Here they are, sorted.
Eighteen allowable uses
The editorial frames these as “tasks in which the machine assists, organizes, accelerates, checks, translates, formats, or improves work while leaving scientific judgment, accountability, and creativity with human authors.”
They are: hypothesis generation; literature search, review, synthesis and gap determination; analysis of market trends or patent environments; research design support; development of experimental or research protocols; selection of research methods; code generation, optimization and debugging; process automation; creation of algorithms for data analysis; secondary data collection; data validation, cleaning, curation and organization; data analysis and visualization; reproducibility testing; proofreading and editing; summarizing text; language translation; reformatting; and preparation of press releases or outreach summaries.
Thirteen nonallowable uses
These are the tasks where “a machine replaces human scientific responsibility for the core intellectual, creative, ethical, interpretive, or accountability-bearing work of research.”
They are: initial conceptualization and idea generation; defining the original research objective; seeding research questions; feasibility assessment and risk evaluation; evaluating the human utility of the research; primary person-to-person data collection; formulating final conclusions; bias analysis and assessment of potential discrimination; ethical risk analysis and monitoring compliance with ethical standards; data-confidentiality monitoring; quality assessment of data or writing; identification of the full universe of limitations; and recommendations for final policies or actions based on results.
The shape of the split
| Measure | Allowable | Nonallowable | Total |
|---|---|---|---|
| Tasks named | 18 | 13 | 31 |
| Share of the taxonomy | 58.1% | 41.9% | 100% |
| Generative by the paper’s definition | 11 | 11 | 22 |
| Nongenerative by that definition | 7 | 2 | 9 |
The taxonomy is weighted toward permission, roughly three tasks allowed for every two withheld. That is a more permissive document than the coverage suggests.
Eleven Allowed Tasks Are Generative by the Editorial's Own Definition
This is the audit’s central result, and it depends entirely on a definition the paper supplies itself, so there is no smuggled premise.
The six content types the definition names
The editorial defines its key term this way: “Generative AI creates new content, including text, images, code, predictions, synthetic data, or analytic interpretations, based on patterns learned during training.” Six content types, named explicitly. Against that, nongenerative AI performs “constrained operations” such as grammar checking or reference management.
That is a workable test. For any named task, ask whether performing it produces text, images, code, predictions, synthetic data or analytic interpretations. If it does, the paper’s own definition calls it generative.
Applying the test to the allowable list
Eleven of the eighteen allowable tasks produce one of those six things. Hypothesis generation produces text and predictions. Literature synthesis and gap determination produce text. Research design support and protocol development produce text. Code generation and algorithm creation produce code — a category the definition names outright. Data analysis and visualization produce analytic interpretations and images, two more named categories. Summarizing text, language translation and press release preparation all produce text.
| Allowable task | Content type produced | Generative? |
|---|---|---|
| Hypothesis generation | Text, predictions | Yes |
| Literature synthesis, gap determination | Text | Yes |
| Code generation and debugging | Code | Yes |
| Data analysis and visualization | Analytic interpretations, images | Yes |
| Selection of research methods | A choice, not content | No |
| Reformatting | Rearranged existing text | No |
| Reproducibility testing | A check | No |
What that count does to the framing
The section heading “Generative and nongenerative uses of AI are not the same” implies the distinction sorts the tasks. It does not. Sixty-one per cent of the tasks augmentation without abdication permits are generative under its own test, including the very first item on the allowable list. A researcher who read only the section headings and concluded that generative tools stay out of the workflow would have misread the paper badly.
This matters practically, not just rhetorically. A policy that says “no generative AI in research” bans eleven things augmentation without abdication explicitly permits — including code generation and data visualisation, which are now ordinary data science practice. The norm is narrower than its reputation already suggests.
Two Banned Tasks That Generate Nothing
The asymmetry runs the other way too, and the two exceptions tell you more about augmentation without abdication than their number suggests.
Primary person-to-person data collection
This sits on the nonallowable list, and it is not a generative task in any sense. Interviewing a participant, consenting them, observing them — these are activities in the world, not content production. Augmentation without abdication bans delegating them to a machine, and is right to, but the reason has nothing to do with generativity. It is about the human relationship the data depends on.
Data-confidentiality monitoring
The second exception is stranger. Monitoring is a constrained checking operation — precisely the category the editorial elsewhere describes as broadly acceptable, in the same family as reference management. Yet it is banned, while reproducibility testing, another checking operation, is allowed.
The distinguishing feature is not the mechanism. It is the consequence of being wrong. A failed reproducibility check wastes time; a failed confidentiality check exposes a participant. The paper never says this, but it is the only reading that makes the pair coherent.
Twenty-two generative tasks, split down the middle
A generative task is a coin flip. A nongenerative task is allowed roughly four times in five. So the distinction is not useless — it just runs in one direction only, and it cannot tell a researcher whether the generative thing in front of them is permitted under augmentation without abdication.
Seven Boundary Pairs Augmentation Without Abdication Leaves Undefined
If generativity does not sort the tasks, something else must. Reading the pairs that sit closest together across the augmentation without abdication line shows what.
Hypothesis generation versus seeding research questions
This is the sharpest case in the paper. Hypothesis generation is allowable, item one. Seeding research questions is nonallowable, item three. In most methods teaching these are the same object in different grammatical moods — a research question asks what a hypothesis asserts.
Augmentation without abdication gives no test for telling them apart, and a researcher acting in good faith could put the identical prompt on either side of the line depending on how the output is phrased.
Research design support versus defining the objective
Research design support is allowed; defining the original research objective is banned. In practice, design follows from the objective so tightly that a model asked to help with the former will routinely propose adjustments to the latter. The boundary is real in principle and porous in every actual session.
The remaining five pairs
| Allowed | Banned | What separates them |
|---|---|---|
| Market and patent analysis | Feasibility and risk evaluation | Who bears the consequence |
| Data validation and cleaning | Quality assessment of data | Repair versus verdict |
| Proofreading and editing | Quality assessment of writing | Repair versus verdict |
| Literature gap determination | Identifying the full universe of limitations | Completeness claim |
| Data analysis and visualization | Formulating final conclusions | Finality |
Read the right-hand column and the actual rule behind augmentation without abdication appears. Nothing there is about generativity. Every entry is about whether the output is a verdict someone will be held to.
The one pair that is genuinely crisp
Secondary data collection is allowed; primary person-to-person data collection is banned. That line needs no interpretation, because the difference is not a matter of degree — one involves a person and the other does not. It is the only pair in the taxonomy a policy could enforce without argument.
What the Augmentation Without Abdication Argument Gets Right
The taxonomy is leaky. The argument underneath augmentation without abdication is considerably better than the taxonomy, and two parts of it deserve more attention than the list has been getting.
The homogenisation problem is real and underrated
The authors argue that if everyone relies on the same tools with similar inputs, science loses the variance that produces surprise. Models are trained to return probable outputs; breakthroughs typically require improbable ones. Their line is that science “is not simply a workflow to optimize” but “an evolutionary system built on independence, serendipity, skepticism.”
This is a variance argument dressed as a values argument, and it is the strongest thing in the paper. A field where every literature review is produced by the same three models will converge on the same gaps, and the gaps nobody’s tool surfaces will stay unexamined.
Serendipity is a measurable loss, not a sentiment
The serendipity claim reads as nostalgia until you connect it to the homogenisation point. Then it becomes a prediction: convergent tooling should compress the distribution of research directions over time. That is testable, and if the effect is real it will show up in citation and topic-diversity data long before anyone can attribute it.
Accountability cannot be delegated, and the paper says so cleanly
“Human authors must be solely accountable for all submitted content, including content generated, revised, analyzed, summarized” by a machine. That sentence needs no taxonomy behind it. It is the whole of augmentation without abdication, and it would survive intact if all 31 task labels were deleted tomorrow. Strip the lists away and augmentation without abdication is still the right instruction.
How to Apply Augmentation Without Abdication in a Real Lab
The paper is written for journals and editors. Translating augmentation without abdication into something a research group can actually run takes three moves.
Write the question before you open the model
The single unambiguous instruction the editorial gives is that the originating research question and the final evidentiary judgment should come from human minds. Everything else is negotiable; that is not. Draft the question on paper, date it, and keep it. A dated question written before any model session is the cheapest possible evidence that augmentation without abdication was observed.
Log the task, not the tool
Because generativity does not predict permissibility, “we used a large language model” is a useless disclosure under augmentation without abdication. What a reader needs is which of the 31 tasks the model touched. A one-line log per session — task name, date, what was kept — turns an unfalsifiable claim into a record, and it maps directly onto any future journal requirement.
Keep the verdict human, explicitly
The boundary table above shows the real rule: a machine may repair, a human must judge. Build that into review. Whoever signs off on conclusions, limitations, bias assessment and confidentiality should state that they reached those judgements themselves. This is ordinary governance, and it belongs in the same place as the rest of your AI strategy rather than in a separate AI policy nobody reads.
Where the Editorial Stops Short
Three gaps in augmentation without abdication are worth naming, because each will surface the moment anyone tries to enforce it.
There is no disclosure threshold
The paper endorses disclosing substantive generative use, but its own list permits eleven generative tasks. Does using a model for language translation require disclosure? For code generation? The consensus it cites says “substantive,” the taxonomy says “allowable,” and augmentation without abdication does not reconcile them.
There is no enforcement mechanism
A norm proposed in an editorial has no teeth, and the authors do not claim otherwise. But the 31 named tasks read like the beginning of a checklist, and checklists get adopted. The gap between “we suggest” and “your manuscript was rejected” is where the boundary pairs will start to hurt.
There is no test for the fuzzy cases
Seven of the pairs examined here have no operational test. The paper would be materially stronger with one sentence per boundary explaining how to tell the sides apart — particularly for hypothesis generation against research question seeding, which is the pair a working scientist will hit first and hardest.
The same problem shows up elsewhere in research infrastructure: our analysis of hidden annotation errors in object detection datasets found a literature that names every error type without ever counting them. Naming a category is not the same as making it checkable, and the reference material collected in our AI models and tools hub runs into the same pattern repeatedly.
Frequently Asked Questions About Augmentation Without Abdication
What does augmentation without abdication mean?
It means using AI to extend human capacity without surrendering scientific autonomy, independence or responsibility. The phrase comes from a PNAS Nexus editorial by Charles Branas and Bruce Levine published in September 2026.
How many AI tasks does augmentation without abdication name?
Thirty-one. Eighteen are marked allowable and thirteen nonallowable. Both lists appear as semicolon-separated prose rather than as a table, which is why the counts have not appeared in coverage.
Does augmentation without abdication ban generative AI in research?
No. Eleven of the eighteen tasks it permits are generative under the paper’s own definition, including hypothesis generation, code generation and data visualisation. The ban applies to specific accountability-bearing tasks, not to generative tools as a class.
What is the difference between hypothesis generation and seeding research questions?
Augmentation without abdication allows the first and bans the second but never explains how to distinguish them. This is the least workable boundary in the taxonomy and the one most likely to cause disputes.
Which tasks are unambiguously off limits?
Initial idea generation, defining the research objective, seeding research questions, formulating final conclusions, bias and ethical risk analysis, identifying limitations, and final policy recommendations. All are judgements a named author must own.
Is augmentation without abdication an enforceable policy?
Not yet. It is proposed as a norm, and no journal has adopted the 31-task list as a requirement. It aligns with existing COPE and ICMJE positions that AI cannot be an author and that humans retain full responsibility.
References
AI and authorship: Norms and uses to preserve human-led science
PNAS Nexus, Volume 5, Issue 9, pgag277
Editorial proposes limits on generative AI’s role in scientific research
COPE position statement on authorship and AI tools
ICMJE Recommendations for the Conduct, Reporting, Editing and Publication of Scholarly Work
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.