Daniele de Rigo

23/ From the report's [7] summary (https://web.archive.org/web/20260912032811/https://www.oecd.org/en/publications/pisa-2025-results-volume-i_73451bc5-en.html) Besides the complex interplay of covariates "PISA data show that students who use [ #genAI] for specific schoolwork tasks, such as summarising texts, drafting or conducting research, attain lower scores in science than students who do not use [ #genAI]. In these scenarios, students who do not use [ #genAI] outperform those who do, on average, by around 20 points – equivalent to being around one year of schooling ahead of their peers"

3/ [A] also notes how now a scenario exists in which an autonomous #generativeAI "harness, backed by an enormous amount of computational resources, performs this entire iteration internally, and ends up producing the final ansatz [...] while the [ #genAI] company running the harness keeps the process to arrive at that ansatz almost completely out of public view. Technically, [a key problem ...] would now be solved; but there would be almost no value added to #mathematics as a consequence"

22/ On learning gaps even during school age, an #OECD report (Programme for International Student Assessment, PISA) [7] may offer some clues:

• Premise: "In #science, #reading and #mathematics, 20 points represents the average annual pace of learning of 15-year-olds in countries that participate in PISA"

• Although a "complex" relationship, on #science learing a clear point: "non-users of [ #genAI] chatbots for specific schoolwork tasks outperform their peers in most countries and economies"

18/ The commentary cites another study (randomized controlled trial involving #SoftwareEngineers, still preprint ahead of #PeerReview) [6] to "investigate whether skills are being lost in the field of #ComputerScience" [4] This early study appears to suggest how #genAI-assisted "participants did particularly poorly on questions that required them to diagnose #errors in the code, which suggests that they had failed to learn the concepts behind the #code that they had just produced"

16/ #References [3] Avila, A., Bhargava, M., Birkar, C., Deligne, P., Deng, Y., Donaldson, S., Duminil-Copin, H., Figalli, A., Hairer, M., Huh, J., Kontsevich, M., Lindenstrauss, E., Lions, P.-L., Maynard, J., McMullen, C., Mori, S., Ngô Bảo Châu, Okounkov, A., Scholze, P., Smirnov, S., Tao, T., Viazovska, M., Villani, C., Werner, W., Zelmanov, E., 2026. A severe misalignment of AI in mathematics. Math and AI. https://mathandai.org (digitally preserved at: https://purl.org/INRMM-MiD/z-9I86UHEZ )

15/ The declaration continues [3] "The issues the #mathematical community faces now are similar to issues that other #scientific and #creative professions are facing, and indicate issues that all of humanity might face: how to make sure that, as AI changes the way work is done, we do not lose sight of what that work was meant to achieve in the first place" Clues pile up in other fields, with "reliance on" #genAI tools allegedly degrading "the abilities of physicians and software engineers" [4]

9/ As anecdotal evidence, after recent exchanges with somewhat exasperated colleagues (note: the fields I'm referring to here are not #mathematics, but mainly #ComputationalModelling, e.g. in #EarthSciences, #ecology, #ClimateChange, ... or just simply their #PeerReview !)

• There seems to be a growing prevalence of those who claim that it is obsolete (usual "left-behind syndrome"?) to do scientific/technical work without depending more and more on #generativeAI (allegedly without upper limits)

7/ On the omission and invisibility of #NegativeResults, [1] cites "an elegant, comic paper" [2] showing how even a most improbable thesis can be defended this way: "how to craft analyses to prove listening to the Beatles made undergraduates younger" [1] Which kind of "meta"-thesis on #genAI itself might be defended by actors investing "enormous amount of computational resources" with "the refusal [...] to disclose their #NegativeResults or reveal the process towards obtaining their solutions"?

4/ in [B] @tao@mathstodon.xyz also notes how "there are no clear frontiers that are separating the "«[ #genAI]-feasible» problems from the «[ #genAI]-hard» problems (which certainly still exist, given that the difficulty level of problems are unbounded, and can even be undecidable). This is [...] also compounded by the refusal of [ #genAI] companies to disclose their #NegativeResults or reveal the process towards obtaining their solutions" which brings to mind #CherryPicking and other biases in #science [1]

5/ An issue of too vast a scope, @deevybee@mastodon.social notes [1]: "It’s not that [scientists] are unconcerned about doing #science well; it's just that many of them don’t recognise that there are serious problems with current practices" and asks "How can that be?" "Many researchers persist in doing research in a way virtually guaranteed not to deliver meaningful results" incurring insidious biases: "#PublicationBias, low statistical power, #pHacking and #HARKing (hypothesising after results are known)" [1]

24/ Figure I.4.13 in the report [7] (p. 239) shows the equivalent schooling time gap vs. the mean score in #science (recalling how the scores are designed so that "20 points represents the average annual pace of learning of 15-year-olds"). The only non-negative gap appears to be among students who use #genAI "for schoolwork for a general purpose («to help me learn»)" and only for "those who report moderate use" ("about once or twice a week") (See "scores adjusted for socio-economic status")

19/ #CognitiveOffloading vs learning, as in [A] "Other technologies have made particular skills obsolete in the past, notes Tapani Rinta-Kahila, an information-systems researcher at the University of Queensland in Brisbane, Australia. For example, GPS navigation systems have eroded people’s navigation skills. #GenerativeAI tools, however, are “the first technology that automates various #cognitive faculties around thinking and interpretation, which were long considered unique human skills”" [4]

17/ In June, a commentary [4] wrote "Evidence suggests that [ #genAI]-driven ‘deskilling’ is starting to happen in medicine, computer science and other fields" citing a study [5] "of physicians in Poland who specialize in #endoscopy [...which] shows how quickly AI tools can erode human abilities" While a co-author of the study [5] warns that "more studies are needed to confirm the phenomenon", he also notes "There is no established solution against #deskilling right now"

13/ Blurred spectrum between two refusal ways to disclose #NegativeResults and the process towards obtaining any given "successful" solution

• #generativeAI companies using these "opaque" solutions as advertising and with goals "severely misaligned" [3] vs. the goals of the #ScientificCommunity

• #scientific/ #technical experts increasingly adopting an "opaque" process to deliver "successful" solutions, hiding (or simply being more and more unable to keep up with) even the more severe pitfalls

12/

• This peculiar behaviour appears not new: several colleagues vividly remember past "epidemics", the peaks of contagion in the form of (sort of) "religious" or "sectarian" exaltation when academy/industry advertising campaigns succeeded in persuading of an imminent and irreversible technical/scientific revolution - e.g. #Bayesian "semantic" networks, #SupportVectorMachines, #RandomForest classifiers/regressors, #blockchain everywhere, ... "new species" which with time found their niche

8/ #References [1] Bishop, D., 2019. Rein in the four horsemen of irreproducibility. Nature 568, 435+. https://doi.org/10.1038/d41586-019-01307-2 (free access version: https://ora.ox.ac.uk/objects/uuid:59a5ddb4-407c-4cd5-9964-b611b15d47c4/files/me1715fd9a834e2208349f6c754de87e3) [2] Simmons, J.P., Nelson, L.D., Simonsohn, U., 2011. False-positive psychology: undisclosed flexibility in data collection and analysis allows presenting anything as significant. Psychological Science 22 (11), 1359–1366. https://doi.org/10.1177/0956797611417632 #DOI

6/ She notes [1]: "Back in 1975, psychologist Tony Greenwald noted that science is prejudiced against #NullHypotheses: We even refer to sound work supporting a null hypothesis as “failed experiments.” That prejudice results in #PublicationBias: researchers are less likely to write up studies showing no effect; editors less likely to accept them, and so no one can learn from them" [1] cites the #pHacking bias: "the practice of trying many analyses and only reporting the 'significant' results"

4/ The commentary by @hildabast@mastodon.online (https://web.archive.org/web/20250623150019/https://absolutelymaybe.plos.org/2024/05/05/scientific-uncertainty-isnt-a-justification-for-getting-things-seriously-wrong/) reflected on how "#ScientificUncertainty enables people to lead themselves, and others, astray" providing "a lot of space for bias to flourish" Controlled experiments differ from the "messier" real life, but the commentary noted: "Someone being prematurely very sure may convince others that they really know what they’re talking about. But they are at the mercy, then, of their biases. Pretending uncertainty is not there is a minefield"

2/ This attitude reminds the first, perhaps most obvious, of the four "coping strategies" on #uncertainty in the #SciencePolicyInterface, as discussed by @Jeroen_van_der_Sluijs@mstdn.science [2] "Monster-exorcists want to expel the monster. Uncertainty simply does not fit within symbolical order where #science is seen as the producer of authoritative objective #knowledge" [2] Variant: "keeping the uncertainties in knowledge claims deliberately under the table because they do not fit a political agenda" [2]

2/ From min. 16: #LLM training data "basically grabbing everything" on the internet. "there's a lot of garbage there, right? And there's no documentation. So if we don't know what's in the data, we are in no way positioned to actually mitigate the harms. [...] the synthetic text that you can create using a language model recreates the #SystemsOfOppression that are in the training data and can also mislead people if it gets fluent enough. We think that maybe there's a thinking mind behind it"

7/ A long-lasting "tension" exists between communicating #science with citizens and policy

• "effectively" (i.e. with an "advertising" mindset where uncertainties and doubts may be "safely" omitted not to "confuse" the audience);

• rather than "honestly" (i.e. following the essence of science so that results are important steps in a path which is not predetermined by our beliefs, but instead is subordinated to the facts of reality; and reality does not owe our beliefs and gambles any respect)

6/

#References (2)

[3] Dries, C., McDowell, M., Rebitschek, F.G., Leuker, C., 2024. When evidence changes: communicating uncertainty protects against a loss of trust. Public Understanding of Science 33 (6), 777–794. https://doi.org/10.1177/09636625241228449

#DOI

(on #uncertainty vs #hubris in #science, in the literature there are several paths of reflection of course; among the many e.g.

• https://hostux.social/@dderigo/110524816915476662

• https://hostux.social/@dderigo/110620576244087429 )

3/ However, "findings can be overturned when new evidence arises" [3]: "how communicating and explaining #uncertainty around scientific findings affect trust in the communicator when findings change" was found [3] not to be so obvious as in the "default" guess [1]. Sometimes, "communicating uncertainty buffers against a loss of trust when evidence changes" and "explaining uncertainty does not appear to harm trust" [3] A commentary by @hildabast@mastodon.online on the study [3] offers a broader perspective

5/ #References [1] Stavrova, O., Kleinberg, B., Evans, A.M., Ivanović, M., 2024. Expressions of uncertainty in online science communication hinder information diffusion. PNAS Nexus 3 (10), pgae439+. https://doi.org/10.1093/pnasnexus/pgae439 [2] van der Sluijs, J.P., 2005. Uncertainty as a monster in the science-policy interface: four coping strategies. Water Science & Technology 52 (6), 87–92. https://doi.org/10.2166/wst.2005.0155 (free access version: https://web.archive.org/web/20251129182336/https://dspace.library.uu.nl/bitstream/handle/1874/386038/Uncertainty_as_a_monster.pdf?sequence=1 - thanks @Jeroen_van_der_Sluijs@mstdn.science !) #DOI

3/ Curiously, #Ramanujan [2] noted another approx. [B] (longer) with the same accuracy as in the case [A] for f(163): "31 places of decimals". In #GNUlinux Bash, easy to compare the two even with bc: echo 'scale=60; l( 640320^3 + 744 ) / sqrt(163)' | bc -l; echo 'scale=60; 4/sqrt(522)*l( ((5+sqrt(29))/sqrt(2))^3 * (5*sqrt(29) + 11*sqrt(6) ) * ( sqrt((9+3*sqrt(6))/4 ) + sqrt((5+3*sqrt(6))/4 ) )^6 )' | bc -l; To show π in bc, use π = 4 arctan(1): echo 'scale=60; 4*a(1)' | bc -l; #PiDay

10/

• Surprisingly often associated: a refusal "to disclose their #NegativeResults or reveal the process towards obtaining their solutions", to reuse the words in [B] in another context

• #NegativeResults sometimes take the form of clear conceptual mistakes or serious #ScientificCode errors, easy for experts to "smell" even if undisclosed and "secretly" fixed (later on)

• The "refusal to disclose" sometimes appears to be unconscious, as if the ability to understand its gravity were lacking

20/ #References [4] Lenharo, M., 2026. Is AI ruining our skills? Early results are in — and they’re not good. Nature d41586-026-01947–1+. https://doi.org/10.1038/d41586-026-01947-1 (free access: https://www.removepaywall.com/search?url=https://www.nature.com/articles/d41586-026-01947-1 ) [5] Budzyń, K., Romańczyk, M., Kitala, D., Kołodziej, P., Bugajski, M., Adami, H.O., et al., 2025. Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. The Lancet Gastroenterology & Hepatology. https://doi.org/10.1016/s2468-1253(25)00133-5 #DOI

11/

• Those who opt for this extremism in delegating and giving up control of key #technical / #scientific tasks no longer do so timidly, but rather with increasing ostentation, an apparent arrogance which is perhaps instead stentorian theatricality (an analogy made me think a lot, as if it echoed something else deep: in the classic fight-or-flight dichotomy, now the mode might be an overly ostentatious fight ceremony, a colleague observed, almost like in a Batesian mimicry: but fearing what?)

2/ Gardner cited #Ramanujan [2] for the general form [A] : f(n) = exp( π √n ) which is almost integer for some values of n (22, 37, 58, see https://www.imsc.res.in/~rao/ramanujan/CamUnivCpapers/Cpaper6/page4.htm) But n=163 was not explicitly mentioned in [2]. #References [2] Ramanujan Iyengar, S., 1914. Modular equations and approximations to π. The Quarterly Journal of Pure and Applied Mathematics 45, 350-375. (https://zbmath.org/45.1249.01 - note: eq. [B] here is wrong! ; original: http://www.imsc.res.in/~rao/ramanujan/CamUnivCpapers/Cpaper6/page1.htm ; PDF: https://ramanujan.sirinudi.org/Volumes/published/ram06.pdf )

5/

Min. 18: #dehumanisation, "three parts" working #definition on "what it is in the #mind of the person doing #dehumanization and what it is to the person who's on the receiving end of that. [...]

• It is the #CognitiveState of failing to perceive another human as fully human.

• It is the acts that express that cognitive state or otherwise entail the assertion that another human is not fully human.

• And then finally it's the #experience of being subjected to those acts by someone else."

14/ The declaration [3] notes "The most precious resources of our profession are students and ideas, and these we nurture with great care" "We are witnessing a general threat to intellectual work, with misalignment between the outcome of the use of AI and its initial purpose. In many fields and activities, years of training have traditionally served not only to produce a final answer or product, but also to develop understanding and the ability to formulate new questions and ideas"

9/ On steady doubt vs prejudice: https://archive.org/details/the-pleasure-of-finding-things-out-the-best-short-works-of-robbins-jeffrey-feynman-richard-phillips/page/199/mode/2up "to work hard on something, you have to get yourself believing that the answer's over there, so you'll dig hard there, right? So you temporarily prejudice" yourself: "Forget what you hear about science without prejudice" Still, doubt defines scientists: while "doing whatever they're doing, they're not so sure of themselves" "They can live with steady doubt, think “maybe it's so” and act on that, all the time knowing it's only “maybe”"

3/ #StochasticParrot [2]: make vivid "difference between how we use language and what's coming out of what I now like to call #SyntheticTextExtrudingMachines. So once that paper was out in the world [...] I was frequently asked «How do I know that you're not just a stocastic parrot?» And I after being asked this question a couple times decided that I'm not going to be in conversation with people who won't posit my humanity as a basic axiom of the discussion. This is a #dehumanizing question"

8/ Here, scientific "honesty"/ "integrity" is not anything mystical, but rather a cultural achievement of centuries of scientific experience: e.g. Feynman insisted on underlining how researchers "learned from experience that the truth will out" so that "this type of integrity" is just "care not to fool yourself" "Other experimenters will repeat your experiment and find out whether you were wrong or right" [...] "although you may gain some temporary fame and excitement" https://archive.org/details/the-pleasure-of-finding-things-out-the-best-short-works-of-robbins-jeffrey-feynman-richard-phillips/page/209/mode/2up

4/ #References

[2] Bender, E.M., Gebru, T., McMillan-Major, A., Shmitchell, S., 2021. On the dangers of stochastic parrots: can language models be too big? In: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. Presented at the FAccT ’21: 2021 ACM Conference on Fairness, Accountability, and Transparency, ACM, Virtual Event Canada, pp. 610–623. ISBN:978-1-4503-8309-7 https://doi.org/10.1145/3442188.3445922

On dehumanisation

• https://philarchive.org/archive/KROMDS
• https://www.google.it/books/edition/Habeas_Viscus/lePTBAAAQBAJ?hl=it&gbpv=1&dq=%22racialization%20is%20understood%20not%20as%20a%20biological%20or%20cultural%22&pg=PT13&printsec=frontcover

21/ #References [6] Shen, J.H., Tamkin, A., 2026. How AI impacts skill formation. arXiv. https://doi.org/10.48550/ARXIV.2601.20245 #DOI

2/ The comment [A] continues: "Crucially, this iteration would only work well at producing such insights if the iterator did not have access to the final [idea] in advance, as this naturally inhibits the exploration of alternate routes to the [idea] that are superficially «dead ends», but in fact end up being highly instructive in the nature of their failure" Actually, the term #ansatz is used in [A][B] rather than my approximative simplification "[idea]" (e.g. see https://en.wikipedia.org/wiki/Ansatz)