[TYPES] AI-generated conference submissions

[Ralf Jung]

Hi all,

>> I would like to echo the point that Klaus made: How does knowing that LLMs
>> were used to produce a paper change the way you review it?
The kind of mistakes LLMs make, and the kind of mistakes humans make, are quite 
different. So to the extent that we try to identify technical flaws in papers 
during the review process, knowing that it was AI-generated can be very helpful 
to focus the validation effort on the parts of the formalism that are most 
likely to be wrong. In an ideal world of course we'd fully rigorously check 
everything from scratch, but obviously that is not practical, so it is 
worthwhile to focus this effort where it is most likely to be fruitful.

It would help even more, of course, if such submissions come with mechanized 
proofs, as those make it a lot harder to produce something that is 
plausible-looking but entirely wrong; a failure mode that is also possible 
without LLMs but that has now been fully automated and brought to a whole 
different scale. (And that is what these AIs we are talking about are primarily 
trained to generate, after all -- something that is plausible. Correctness is 
secondary.)

Kind regards,
Ralf

On 17.03.26 18:15, Alcides Fonseca wrote:
> [ The Types Forum, http://lists.seas.upenn.edu/mailman/listinfo/types-list  ]
> 

> 
> Assume I submitted a paper with the following acknowledgement: “LLMs were
> used to generate the code for the experiments, analyze them and produce the
> manuscript.”
> Will I get desk rejected?
> Will I get a different review? Will you trust the claims less because I
> used LLMs? In some cases,  LLMs actually improve the correctness by
> allowing authors to have machine-checked proofs in reasonable time (to
> Jon’s point). So maybe you should trust LLM-generated papers more than pure
> manuscripts.
> 
> Personally, I agree with Klaus: it should make no difference, and we
> shouldn’t require any acknowledgement.
> 
> 
> 
> The relevant impact is already covered by another point in ACM’s policy: *"In
> addition, all persons listed as an author on an ACM submission certify
> that: […] They agree to be held responsible and accountable for any issues
> relating to the correctness or integrity of the Work and compliance with
> all related ACM Publications Policies with the understanding that,
> depending on the circumstances, not all authors will necessarily be held
> equally accountable.”*
> 
> This is much more important IMHO. Personally, I would much rather require
> that you specify which author is responsible for what (as many publications
> in life sciences do) than what tools or whether genAI was used to create
> it. So you have an idea, here is that section from a couple of papers:
> 
> Authors’ Contributions
> 
> Conceptualization: P.B.,R.S.,M.C.-F.,A.F.Funding acquisition: P.B.,
> 
> M.C.-F, A.F. Data curation: P.B. Investigation: P.B. Methodology: P.B.
> 
> Resources:P.B.,M.C.-F.Software:P.B.,A.F.Supervision:R.S.,M.C.-F.,
> 
> A.F.Visualization:P.B.Writing—original draft:P.B.Writing—review
> 
> & editing: P.B.,R.S.,M.C.-F.,A.F.
> 
> 
> 
> Author contributions
> 
> MC-F, AF, and PB conceptualized and designed the study
> 
> and wrote the manuscript. PB and MR collected and assembled
> 
> the datasets. PB, MR, AF, and MC-F analyzed the data and
> 
> interpreted the results. All authors contributed to the article and
> 
> approved the submitted version.
> 
> 
> 
> Each author might use LLMs in different ways, but they are accountable for
> what they did. And that is the basis for trust (and penalties, if you want).
> 
> 
> The main issue with genAI is the economics: you can produce a paper much
> faster than before. There are more incentives in our world to produce new
> papers than to review and reproduce papers. Has anyone have ever received
> an NSF/ERC/other grant for reproducing existing work?
> 
> If the economics do not change, there is no ACM policy that will save this.
> The reciprocal reviewing of ICLR and other ML conferences (and OOPSLA!)
> addresses the problem head-on: they are able to change the economics of
> writing vs reviewing (at the cost of a possible decrease in PC quality).
> 
> 
> I agree things are changing, not necessarily for the best, but we should
> tackle the root of the problem that is the incentives for publishing vs
> reviewing.
> 
> 
> — Alcides
> 
> On Tue, Mar 17, 2026 at 3:36 PM Anitha Gollamudi <anitha.gollamudi at gmail.com>
> wrote:
> 
>> [ The Types Forum, http://lists.seas.upenn.edu/mailman/listinfo/types-list 
>> ]
>>
>> I would like to call out a specific point, also echoed by Klaus. AI tools
>> can specifically help non-native English speakers write better narratives.
>> It takes multiple (re-) tries to present/finesse a technically correct
>> idea---both with the grammar and narrative. I am wondering if the planned
>> AI detection tools are aimed to flag such instances as well? I understand
>> it is a grey area, but I am also afraid that it could end up being a
>> counter-productive policy.
>>
>> -Anitha
>>
>> On Tue, 17 Mar 2026 at 11:22, Jonathan Aldrich <
>> jonathan.aldrich at cs.cmu.edu>
>> wrote:
>>
>>> [ The Types Forum,
>> http://lists.seas.upenn.edu/mailman/listinfo/types-list 
>>> ]
>>>
>>> Penalties for ACM policy violations are covered here:
>>>
>>>
>>>
>> https://urldefense.com/v3/__https://www.acm.org/publications/policies/penalties-for-publication-violations__;!!IBzWLUs!Sjaz8owhfEo9XGIOjdHvYlt4K1i0NrcU2OJwhOfFW-zvlNaateRbntcrFSLLEBufF-BToem_Mxww-uzmXyT4YGGe2ItcCGK8yGCGee6rNw$
>>>
>>> Regarding application to cases of AI use, I'd like to echo Mae's call for
>>> empathy and grace, particularly in cases that might involve students. A
>> lot
>>> of violations are likely unintentional in this context. ACM definitely
>>> tries to take these factors into consideration in addressing ethics and
>>> plagiarism cases.  A harmless use of AI by a student who neglects to
>>> acknowledge AI use, perhaps because they were unaware of the policy,
>> might
>>> receive only a warning ("Level I") and if already published, a
>> corrigendum
>>> to the paper acknowledging the AI use.
>>>
>>> If the AI use compromises the integrity of the paper the paper would of
>>> course be rejected or, if already published, retracted ("Level II").
>> When
>>> caught during reviewing, these kinds of violations can be handled by
>>> conference chairs (e.g. with the approach Mae suggests), with
>> notification
>>> to ACM's Ethics and Plagiarism committee.  I remember how big of a deal
>>> even a single paper rejection was to me as a student; I am sure that this
>>> level of action is enough to motivate care and compliance with disclosure
>>> requirements for the vast majority of our community.
>>>
>>> The penalties do escalate according to the context.  Careless use by
>>> someone who should know better (typically more senior researchers) might
>> be
>>> a Level III violation, with a 1-year publication/participation ban; if it
>>> was intentional or repeated it would be a 2 year ban (Level IV), and if
>> it
>>> is severe, intentional, and repeated it would be Level V (5 year ban).
>> In
>>> practice, bans of 5 years, or more (for stacked Level V penalties) are
>>> typically for things like reviewing rings, severe research misconduct,
>>> severe harassment or abuse in ACM-relevant contexts--in which case it is
>> of
>>> course very well justified.  Without revealing details of confidential
>>> cases, scandals that make the news are likely to be at this level of
>>> severity (to be clear, "makes the news" is not among our
>> criteria--there's
>>> correlation here, not causation).  Keep in mind that a 5 year ban is
>>> already career-ending or at least career-altering for someone whose work
>>> centers on publishing in ACM venues.
>>>
>>> The ACM spends a lot of volunteer time and money (some of these involve
>>> in-depth investigations and lawyers) on such issues, and when they come
>> to
>>> the publications board, the conversations involve careful and substantive
>>> deliberation, with the typical result a strong consensus on the action
>>> taken. Having been a part of the process, I can say while it is
>> imperfect,
>>> costly, and sometimes slow, it is generally fair and its outcomes
>> achieve a
>>> large measure of justice and help to protect our community.
>>>
>>> Best,
>>>
>>> Jonathan
>>>
>>> On Tue, Mar 17, 2026 at 3:30 AM Drossopoulou, Sophia <
>>> s.drossopoulou at imperial.ac.uk> wrote:
>>>
>>>> What happens to authors when such an AI generated paper is
>> encountered? I
>>>> would have thought that if ACM had a strict policy, eg no papers
>>> published
>>>> for the next 5 years, then the problem would be significantly reduced
>>>>
>>>>        Sophia
>>>> ———•••———
>>>> Sent from mobile phone, hence succinct
>>>>
>>>>> On 17 Mar 2026, at 03:54, Jonathan Aldrich <
>>> jonathan.aldrich at cs.cmu.edu>
>>>> wrote:
>>>>>
>>>>> CAUTION: This message came from outside Imperial. Do not click links
>>> or
>>>> open attachments unless you recognise the sender and were expecting
>> this
>>>> email.
>>>>>
>>>>>
>>>>> [ The Types Forum,
>>>> http://lists.seas.upenn.edu/mailman/listinfo/types-list     ]
>>>>>
>>>>> Ugh, I have been fortunate enough not to encounter that in my
>>> reviewing.
>>>>> But I hear it is very common in other subfields so it was probably
>>> just a
>>>>> matter of time before it got to PL.
>>>>>
>>>>> If this is an ACM conference, our policies already require disclosing
>>> the
>>>>> use of AI for anything beyond grammar-checking-style applications:
>>>>>
>>>>>
>>>>
>>>
>> https://urldefense.com/v3/__https://www.acm.org/publications/policies/new-acm-policy-on-authorship__;!!IBzWLUs!RC4FcA0-IHc55AxdQMuWNHIfsDdNqAlbvh2m7TpSk7wCaTZl7_iQsMZfqu-oMv768kp_7vwINcEPMtugQGMRVD8jcVueMjwevkEYu5ZLnw$
>>>>>
>>>>> If there are any non-ACM PL conferences that don't have this policy,
>> I
>>>>> would encourage them to adopt it.
>>>>>
>>>>> ACM is evaluating tools that can (heuristically) detect AI use, as
>> well
>>>> as
>>>>> tools that can identify hallucinated references.  We should start to
>>> see
>>>>> deployment of these within the next year--if anyone on this list is
>> an
>>>> ACM
>>>>> PC chair and wants to do an early trial, let me know and I can
>> connect
>>>> you
>>>>> with people who may be able to arrange that.  It's of course
>> important
>>> to
>>>>> have a human verify any tool reports based on heuristics as they may
>> be
>>>>> incorrect, but the point is that they can save time in identifying
>>>> problems.
>>>>>
>>>>> Regarding reviewing workload, it's very unfortunate.  The tools
>>> mentioned
>>>>> above will eventually help some.  In the meantime, it's my view that
>>> once
>>>>> you determine that a paper is so flawed it cannot be accepted,
>>> especially
>>>>> if that flaw involves misconduct such as undisclosed AI use or
>>> otherwise
>>>>> makes the paper very difficult to read, it's reasonable for the
>>> reviewer
>>>> to
>>>>> stop reading and return a review based on the portion they read.  Of
>>>>> course, I would mention the situation to the PC chair to make sure
>> they
>>>> are
>>>>> OK with this; most probably are.
>>>>>
>>>>> Best,
>>>>>
>>>>> Jonathan
>>>>>
>>>>>
>>>>>> On Mon, Mar 16, 2026 at 9:50 PM Stephanie Balzer <
>>>> stephanie.balzer at gmail.com>
>>>>>> wrote:
>>>>>>
>>>>>> [ The Types Forum,
>>>> http://lists.seas.upenn.edu/mailman/listinfo/types-list 
>>>>>> ]
>>>>>>
>>>>>> Dear all,
>>>>>>
>>>>>> I have now been numerous times on the receiving end on a what it
>>> appears
>>>>>> to me (almost) entirely AI-generated conference submissions that I
>> was
>>>>>> assigned to review.  Of course I have no proof, but to me it was
>>> pretty
>>>>>> obvious.  The submissions in question consist of an amalgamation of
>>>>>> meaningful words (sometimes not entirely from the context the paper
>>>>>> ought to be about), are generally well written, although
>> meaningless,
>>>>>> and even come backed up with some rules with horizontal lines and
>>> proof
>>>>>> sketches (sometimes from various contexts).  That catch, however, is
>>>>>> that the whole composition doesn't make sense.
>>>>>>
>>>>>> What are we going to do about this as a community?
>>>>>>
>>>>>> I have numerous concerns here: My immediate concern is that I do not
>>>>>> like to spend my time on such submissions.  Even though it's quite
>>>>>> obvious immediately that the paper is meaningless, it still takes
>> some
>>>>>> time to make sure and justify the verdict.  Another concern I have
>> is
>>>>>> the risk that, under time pressure, no due diligence is done, and we
>>> may
>>>>>> end up accepting such a paper.
>>>>>>
>>>>>> As a first step we may require authors to declare whether AI was
>> used
>>> in
>>>>>> preparing their submission and what for and we delimit what uses are
>>>>>> permitted.
>>>>>>
>>>>>> Looking forward to your thoughts,
>>>>>>
>>>>>> Stephanie
>>>>>>
>>>>>>
>>>>
>>>
>>

-- 
Website: https://urldefense.com/v3/__https://research.ralfj.de__;!!IBzWLUs!XZggTvlvvx-mlQSm7AVDG2WUgKuFMtkBaFPwr7-KrvXo28uQ85tjIBBz8a4LxNmIJ69Top5UsVs5ljE2nSmYB64NDaL1p4w$