37 problems that are still open, each with published evidence — for universities, publishers, journals, editors, authors, reviewers, and researchers in industry and academia. “Solved means” is what an entry must show to win.
for the highest-scoring solution of the round.
of Smart Scholars services for one journal, free, for the highest-scoring entry that solves each problem. For a winner with no journal of their own, the year can go to a journal they choose: one they edit, review for or name. Our services →
to enter. Nobody has to buy or use a Smart Scholars service to enter or to win. Judged on merit, never by chance.
Round 1 closes at 23:59 IST; winners are named by 30 June 2027. Entries sent from the day this page was published count in round 1, which opens on 1 December 2026.
The judges: Rahul Raj Akkali, Authorised Signatory of Smart Scholars, and two experts from outside Smart Scholars, who will be named when round 1 opens on 1 December 2026.
Each box shows the problems of the people who have them.
Showing all 37 problems.
Problem 1Publishers · Journals
Evidence. Fraudulent papers are growing at a rate “far outpacing” legitimate science (Richardson et al. 2025, PNAS).
A screen that flags paper-mill submissions before review, tested on retracted paper-mill papers against genuine ones, with its error rates published.
Problem 2Editors
Evidence. A major publisher retracted more than 100 papers from one cancer journal over faked reviews in 2017 (Retraction Watch).
A check that confirms within minutes that a suggested reviewer is the real person behind the e-mail address.
Problem 3Authors · Universities
Evidence. The Hijacked Journal Checker passed 400 entries in Dec 2025 (Retraction Watch); an agreed definition of a predatory journal exists only since 2019 (Grudniewicz et al., Nature).
An open check that spots a cloned journal site within days and tests any journal against the 2019 definition, explaining each verdict.
Problem 4Journals · Universities
Evidence. 48 of 360 journals accepted a fictitious scientist as an editor (Sorokowski et al. 2017, Nature).
Proof, for every listed board member, that the person agreed to serve, readable by anyone.
Problem 5Journals · Editors
Evidence. 3.8% of 20,621 papers in 40 journals had problematic images (Bik et al. 2016, mBio).
An open figure screen at submission, tested on a public set of known cases.
Problem 6Journals · Editors
Evidence. Rewriting tools turn “artificial intelligence” into “counterfeit consciousness”; such phrases mark machine-rewritten text in published papers (Cabanac et al. 2021).
Detection of machine-paraphrased text in any language, with its error rates published.
Problem 7Journals · Universities
Evidence. LLM use in computer-science papers reached an estimated 17.5% by Feb 2024; 950,965 papers studied (Liang et al. 2024).
A fair check of AI use whose rate of wrongly flagged honest papers is measured and published.
Problem 8Authors · Publishers
Evidence. In biomedicine, only 5.4% of 13,252 citations made after a retraction mention it (Hsiao and Schneider, QSS).
A way to mark, in the citing papers’ own records, that a cited work was retracted, and to tell their authors.
Problem 9Publishers · Libraries
Evidence. 174 open-access journals vanished from the web between 2000 and 2019 (Laakso et al. 2021, JASIST).
Low-cost, automatic preservation that keeps a small journal’s articles readable after the journal closes.
Problem 10Authors · Journals
Evidence. One in five science articles suffers from reference rot; seven in ten of those that cite web pages (Klein et al. 2014, PLOS ONE).
Every web link in a reference list kept readable years later, done automatically at publication.
Problem 11Authors · Indexes
Evidence. For 132,339 bioRxiv and medRxiv preprints, the links to their journal articles are incomplete in the servers’ metadata and in Crossref (Langnickel et al. 2022, JMIR).
An accurate matcher that links each preprint and its published version, both ways, in the registries.
Problem 12Researchers · Industry R&D
Evidence. More than 70% of researchers have tried and failed to reproduce another scientist’s experiments (Baker 2016, Nature); one biotechnology company confirmed the findings of only 6 of 53 papers (Begley and Ellis 2012, Nature).
A low-cost check, before publication, that a paper’s main result can be rerun from its shared data and code.
Problem 13Researchers · Funders
Evidence. 93% of 1,792 authors who promised data did not respond or declined; 6.8% shared it (Gabelica et al. 2022, J Clin Epidemiol).
A way that makes promised data reachable, checked automatically when a paper is accepted.
Problem 14Researchers · Universities
Evidence. Non-native English speakers spend more effort reading, writing and presenting in English; 908 researchers surveyed (Amano et al. 2023, PLOS Biology).
Affordable, accurate publishing in the author’s own language, with every language version linked as one work.
Problem 15Authors · Universities
Evidence. One publisher’s open-access “mirror” journals had no articles by authors in low-income countries (Smith et al. 2021, QSS).
A model that lets authors in low-income countries publish open access at no cost and keeps the journal funded.
Problem 16Reviewers · Editors
Evidence. Reviewers worked over 100 million hours in 2020 (Aczel et al. 2021); agreement to review invitations fell from 0.54 in 2013 to 0.44 in 2017 (Global State of Peer Review 2018, via EASE).
A way to find willing, qualified, independent reviewers faster, or to reward them, that a small journal can afford.
Problem 17Authors · Reviewers · Libraries · Indexes
Evidence. In short literature reviews written by an AI chatbot, 55% of the references from one version and 18% from its newer version were fabricated (Walters and Wilder 2023, Scientific Reports).
A public check that flags ≥95% of nonexistent references in a blind test set of 1,000 submitted reference lists, with <2% false flags, used at ingestion by at least one major index.
Problem 18Researchers · Indexes
Evidence. A considerable number of retracted papers and retraction notices in PubMed are not labelled as such in Web of Science, or cannot be told apart from corrections there (Schmidt 2017, JASIST).
Problem 19Researchers · Public & policy
Evidence. Changes made to published papers without any correction notice, “stealth corrections”, were found and documented (Aquarius et al. 2025, Learned Publishing).
Any change to a published VoR’s text/figures is detectable by a third party (hash/diff) and linked to a notice, demonstrated on ≥3 publishers.
Problem 20Editors · Authors
Evidence. An estimated 6.5% to 16.9% of the peer-review text submitted to four AI conferences could have been substantially modified by LLMs (Liang et al. 2024, arXiv 2403.07183).
A disclosure + verification method that editors can apply per review, validated with <5% false accusation rate on a blind set.
Problem 21Editors · Industry R&D
Evidence. Instructions hidden in preprints, in white text or small font, told AI tools to give the paper a positive peer review (Gibney 2025, Nature news).
A screening step that detects ≥95% of hidden instructions (white text, tiny font, metadata) in a red-team test set, adopted by ≥1 major submission system.
Problem 22Journals · Repositories · Reviewers
Evidence. A large language model’s data-analysis tool was used to create a fake data set for medical research (Taloni et al. 2023, JAMA Ophthalmology).
A method that distinguishes synthetic from real tabular clinical data with AUC ≥0.95 on unseen data, or a provenance chain attached to deposited data.
Problem 23Public & policy
Evidence. Summaries by most of the 10 LLMs tested overgeneralised the conclusions of the scientific texts they summarised (Peters and Chin-Yee 2025, Royal Society Open Science).
Benchmark where AI summaries of 500 papers have overgeneralisation rate below the human-expert baseline, using publisher-supplied claim metadata.
Problem 24Repositories
Evidence. Over 90% of the 66 repositories that answered a 2025 survey said they were encountering aggressive bots (COAR survey 2025).
A shared bulk-access mechanism reducing bot load on participating repositories by ≥50% without reducing human access, measured over 6 months.
Problem 25Repositories · Researchers
Evidence. In October 2025, arXiv’s computer-science category began to require that review articles and position papers have passed peer review at a journal or conference, because of an unmanageable influx of them (arXiv blog, 31 Oct 2025).
A triage method that cuts moderator time per submission by ≥50% while keeping wrongly rejected genuine papers <1%.
Problem 26Researchers · Funders
Evidence. There is little support across the scholarly ecosystem for acknowledging and citing software (Smith et al. 2016, PeerJ Computer Science); in a random sample of 90 biology papers, only 31% to 43% of mentions of software were formal citations (Howison and Bullard 2015, JASIST).
In a sample of 1,000 new papers, ≥70% of used software is cited with a resolvable PID including version.
Problem 27Researchers · Repositories
Evidence. 88.1% of the records in a citation index of research data had received no citation, and citing data was rare in most fields (Robinson-García et al. 2015, JASIST).
Data citations captured by open infrastructure for ≥80% of data reuse in a verified sample.
Problem 28Authors
Evidence. Telling authors apart by their initials distorts measures of coauthorship networks, and Asian names accounted for most of the misidentifications (Kim and Diesner 2015, JASIST).
Disambiguation achieving F1 ≥0.95 on a benchmark of common Chinese/Korean/Indian names.
Problem 29Researchers
Evidence. The coverage of Web of Science and Scopus favours the natural sciences, engineering and biomedical research, and over-represents journals in English (Mongeon and Paul-Hus 2015, Scientometrics).
An evaluation-grade index whose coverage per country is within 10% of a ground-truth national register in ≥10 countries.
Problem 30Researchers · Public & policy
Evidence. Negative results were found to be disappearing from most disciplines and countries (Fanelli 2011, Scientometrics).
Open registry where null results get DOIs and are indexed; measurable increase in published null results in ≥3 fields.
Problem 31Reviewers · Editors
Evidence. Open peer review, where reviewers’ names or reports are published, was among the last parts of open science to be widely taken up: 617 journals were found practising it by 2019, in widely varying ways (Wolfram et al. 2020, Scientometrics).
Peer reviews for ≥20% of articles in a field registered with PIDs linked to the article.
Problem 32Authors · Universities · Funders
Evidence. Many journal editors appear to pressure authors into citing the editors’ own journals (Wilhite and Fong 2012, Science).
An auditable method detecting coercive citation requests in confidential review systems with verified precision.
Problem 33Public & policy
Evidence. Of 7,274 trials on the EU Clinical Trials Register whose results were due, only about half had reported them (Goldacre et al. 2018, BMJ).
Trial registry IDs present in ≥90% of trial publications’ metadata and linked both ways.
Problem 34Universities · Researchers
Evidence. In 2020 there was no globally implementable way to identify a measuring instrument persistently; a working group then proposed one (Stocker et al. 2020, Data Science Journal).
Instruments used in ≥50% of papers of a sample field resolvable via PID.
Problem 35Journals · Researchers
Evidence. An AI system wrote complete research papers end to end, from the idea and the experiments to the write-up, and then ran a simulated review of them (Lu et al. 2024, arXiv 2408.06292).
A verification method (e.g. lab-notebook/provenance attestation) that a journal can use to confirm human accountability for ≥95% of a blind set mixing human and AI-agent papers.
Problem 36Authors · Publishers
Evidence. Research papers have almost certainly been used to train AI models, and publishers have sold access to them in deals made without consulting the authors (Gibney 2024, Nature news).
A method that, for a given paper and model, establishes training inclusion with verified accuracy, plus a credit record linked to the paper’s DOI.
Problem 37Journals · Editors
Evidence. Images made by AI image generators are very difficult to identify, whether by eye or with image-forensic and detection tools (Gu et al. 2022, Patterns).
Raw images in ≥1 journal carry a verifiable chain from instrument capture to figure, or a detector holds AUC ≥0.95 on images from unseen generators.
The sources are linked as evidence only.
Know an open problem that belongs here? Send it with who has it and a published source. A suggested problem goes on this page only after its source is checked.
Free to enter, judged on merit and never by chance.
What we do: DOIs registered from what a journal already has · Crossref records completed — references, abstracts, licences, ORCID iDs, affiliations and ROR IDs, funding · DOIs that resolve · articles preserved · journals rated on the public record · see everything we do →
Start with the audit — ₹4,999 per journal, and the gate to everything else.
Every service on one journal profile — each previews first and acts only on your approval.
Free to use. Nothing you check is shared with the journal.
Mon–Sat, 10:00–19:00 IST. The Ask AI button on every page answers about our services at any hour.
