When Output Is Cheap, Verification Becomes Scarce
Updated: 3 minutes ago
OpenDeSci Research Insights · Perspective · October 2026 · Version of 26 October 2026
The outlook for AI in science: how AI loops and graphs are reshaping discovery, and why attention, judgment and verification become strategic resources
David Wang¹,²,³ · Prof. Dr. Christian Abegglen³ · Dr. Florian Smeritschnig⁴ · Evan Luthra⁵
¹ OpenDeSci Foundation · ² Wang Advisory GmbH · ³ StGallen imt Business School, St. Gallen, Switzerland · ⁴ StrategyCase.com · ⁵ KOL Capital
In collaboration with StGallen imt Business School
Perspective based on a narrative synthesis of published evidence. This article does not report original empirical research and is not a systematic review.
Artificial intelligence is no longer just a tool that scientists use; it is becoming a participant in how science is done. That changes the economics of discovery. When hypotheses, papers and even peer reviews become cheap to generate, output is no longer scarce, and the bottleneck moves: when output is cheap, verification becomes scarce.
More precisely, scarcity moves both upstream and downstream: upstream to attention and judgment about which questions merit investigation, and downstream to the verification of which answers are reliable. AI can expand the space of hypotheses, but research organisations must still allocate scarce human attention, experimental capacity and institutional commitment. The argument that follows concerns both bottlenecks.
For most of the past decade, AI in science meant better instruments: a model predicted a protein structure, a classifier sorted telescope images, a chatbot summarised a paper. That era is giving way to AI systems that run parts of the scientific method themselves. They read the literature, propose hypotheses, write and execute analysis code, design experiments and decide what to try next. The question is therefore no longer whether AI can answer a scientific question, but whether it can close the loop from question to evidence and back again, and whether anyone can trust what comes out of it.
Two ideas define this next stage. Loops are AI systems that act, observe, learn and iterate, first in software and increasingly in physical laboratories. Graphs are structured, connected knowledge that tells those loops what is already known, what contradicts what, and where each claim came from. Loops give AI agency; graphs give it memory, context and provenance.
The article proceeds in four steps. It first sets out the problem, the evidence and the players. It then describes loops, graphs and a five-stage maturity model for research organisations. It turns next to the trust bottleneck and to what remains scarce beyond verification, using integrated management as a lens for institutional choices and asking where open and decentralised science fit. It closes with dated forecasts and first moves for leaders.
Executive summary
1. Evidence is accumulating
AI has moved from promise to evidence, so far mainly where verification pathways are well specified, and fastest where evaluation can be automated. Two of AlphaFold’s creators shared the 2024 Nobel Prize in Chemistry [9]. AI systems have predicted 2.2 million candidate crystal structures, of which 381,000 remained on the updated convex hull and 736 have so far been matched by materials made independently in laboratories [10]; they have also improved on a 1969 matrix-multiplication algorithm [11] and reached gold-medal standard at the International Mathematical Olympiad [12]. In September 2026, Insilico Medicine reported dosing the first patient in a Phase III trial of rentosertib, an investigational drug with an AI-identified target and a generative-AI-designed molecule [14].
2. The next stage is loops and graphs
The frontier is shifting from models that answer to systems that act. Loops plan, experiment, evaluate and iterate, in software and increasingly in robotic labs. Graphs give those loops memory, context and provenance. On software tasks, the length of tasks AI agents can complete autonomously with 50% reliability has doubled roughly every seven months since 2019, and about every three months since 2024 [26, 27]; that pace says little about experimental science. Across AI as a whole, the arc runs from chatbots (2022) to agents (today) to self-directed loops on shared graphs.
3. A new industry is forming
Frontier labs, AI-scientist startups and governments are converging on automated discovery. OpenAI says it has reached its automated research intern goal [18]. Periodic Labs raised a USD 300 million seed round [21] and Lila Sciences USD 550 million in total [22]. The US Genesis Mission [24] and the EU’s RAISE [25] treat AI for science as strategic infrastructure.
4. The bottleneck moves to trust
When output is cheap, verification becomes scarce. A vendor’s scan reported hallucinated citations in papers accepted at NeurIPS 2025 [34]. An AI detector flagged about 21% of the peer reviews for ICLR 2026 as fully AI-generated [35]. An observational Nature study links AI use to higher individual impact but a narrower science overall [33]. Verification is a central bottleneck, but not the only one: attention and judgment decide what enters the research loop and is worth testing, while responsibility and sovereignty decide who remains accountable and how personal and sensitive data may legitimately be used.
5. An integrated management framework for research organisations
We use the St. Gallen Integrated Management Concept as a lens rather than as a validated answer: research organisations need to align normative responsibility, strategic positioning and operational verification [37]. Open, verifiable infrastructure and trusted public interfaces are likely to decide whether AI makes science faster and more trustworthy, or merely faster.
The problem: science produces more, but gets less per researcher, per dollar and per paper
The world spends more on research than ever. WIPO estimates global R&D expenditure at about USD 2.87 trillion in 2024, in constant 2015 PPP terms, close to triple its real level in 2000 [1].
Yet the return on that effort is falling.
Ideas are getting harder to find. Research effort is rising substantially while research productivity is declining sharply. Keeping Moore’s Law alive now takes more than 18 times as many researchers as in the early 1970s [2].
Breakthroughs are becoming rarer. A study of 45 million papers and 3.9 million patents across six decades found new work increasingly less likely to break with the past and push fields in new directions [3].
Drug discovery is the starkest case. New drugs approved per billion US dollars of R&D halved roughly every nine years from 1950 to 2010, a fall of about 80-fold in inflation-adjusted terms. Scannell and colleagues called it Eroom’s Law, Moore’s Law spelled backwards [4].
At the same time, the system that stores and checks knowledge is under strain.
Volume outruns attention. Articles indexed in Scopus and Web of Science were about 47% higher in 2022 than in 2016, outpacing growth in the number of practising scientists [5].
Reproducibility is fragile. In a Nature survey of 1,576 researchers, more than 70% reported having tried and failed to reproduce another scientist’s experiment [6].
Integrity is under industrial attack. Retractions passed 10,000 in 2023, a record driven by more than 8,000 retractions at Hindawi journals [7]. In a corpus of suspected paper-mill products, annual counts doubled roughly every 1.5 years, against about 15 years for scientific articles overall [8].
This diagnosis matters for everything that follows. Science does not mainly lack intelligence; it lacks throughput, connection and verification. It needs faster ways to search an exploding space of hypotheses, better ways to connect what is already known, and stronger ways to establish what is true. That is exactly where AI is now moving.
Pages from the brochure published with this article summarise the sections in which they appear. Where a page uses data, its sources are named at its foot; page numbers and the numbers in square brackets on the pages refer to the brochure and its own reference list, not to the sources of this article.
The evidence so far: six milestones and the pattern behind them
Six milestones show what AI has already achieved in science, and under which conditions. They do not carry the same evidentiary weight: the sources range from peer-reviewed studies and an official prize announcement to company disclosures, so the source type is part of each claim and is labelled in the table and in the list of sources.
Field | Milestone | Evidence |
|---|---|---|
Structural biology | AlphaFold2 predicted the structures of virtually all 200 million known proteins and has been used by more than two million people in 190 countries. Two of its developers, Demis Hassabis and John Jumper, shared the 2024 Nobel Prize in Chemistry with David Baker [9]. | Official prize announcement |
Materials science | GNoME, a graph neural network, generated 2.2 million candidate structures predicted to be stable relative to existing datasets; 381,000 remained on the updated convex hull, and 736 matched structures independently realised in experiments [10]. That does not validate the full set, and a later perspective found scant evidence of predictions that are at once novel, credible and useful [58]. | Peer-reviewed study; contested |
Algorithms | AlphaEvolve multiplied 4×4 complex-valued matrices with 48 scalar multiplications, improving on Strassen’s 1969 algorithm, and recovers about 0.7% of Google’s worldwide compute [11]. | Company disclosure |
Mathematics | A Gemini Deep Think model reached gold-medal standard at the 2025 International Mathematical Olympiad, scoring 35 of 42 points within the 4.5-hour limit [12]. | Company disclosure; scores certified by IMO coordinators |
Medicine | Rentosertib, an investigational drug with an AI-identified target and a generative-AI-designed molecule, was tested in a 12-week Phase 2a trial in 71 patients, with safety as the primary endpoint; at 60 mg, lung capacity (FVC) rose by a mean 98.4 mL against a 20.3 mL decline on placebo, an encouraging signal from a small trial, not yet conclusive evidence of efficacy [13]. In September 2026, Insilico reported dosing the first patient in a Phase III trial [14]. | Peer-reviewed trial (Phase 2a); company disclosure (Phase III) |
Drug pipelines | As of December 2023, 21 of 24 completed Phase I trials and 4 of 10 completed Phase II trials of AI-discovered molecules were classified as successful: above historical averages in Phase I, in line with them in Phase II. The sample, drawn from AI-native biotechs, was small and non-exhaustive [15]. | Peer-reviewed analysis; small, non-exhaustive sample |
The pattern behind these results is more important than any single headline.
The strongest examples share a clearly specified way to verify an answer, not necessarily a fast one. Structure predictions could be scored against experiments, crystal stability computed, algorithms run and timed, and proofs graded. In mathematics and software, feedback can be near-instant. In materials science and medicine, verification remains experimental, staged and often slow: rentosertib is given for 52 weeks in its Phase III trial, and the company’s release quotes an expected three to four years from the start of Phase III to approval, under favourable conditions [14].
Many of the strongest results still depend on physical or human verification, and much of it is incomplete. Only 736 of GNoME’s predicted structures have so far been matched by materials made independently in laboratories [10], and rentosertib is still being tested in patients [14]. Verification can also shrink a headline: a 2026 author correction cut the number of targets that Berkeley Lab’s autonomous A-Lab had successfully synthesised from 41 to 36, and clarified that its materials were new to the prediction platform, not necessarily new to science [28].
The lesson is precise: AI accelerates science fastest where reliable feedback can be generated, above all where evaluation can be automated, and least where verification is slow, expensive or ambiguous, which describes most of biology, medicine and the social sciences. This is why the next stage of AI in science is not about bigger models alone. It is about loops that connect AI to verification, and graphs that connect it to everything already known.
The players: a new industry is forming around automated discovery
AI for science is no longer a side project of a few research labs. Frontier AI companies, venture-backed AI-scientist startups, drug developers and governments are now building the same thing: systems that close the loop from hypothesis to evidence.
Player | What it is building | Latest signal | Evidence |
|---|---|---|---|
Google DeepMind (frontier lab) | Discovery models and research agents: AlphaFold, GNoME, AlphaEvolve, AI co-scientist | A multi-agent co-scientist generates, debates and evolves hypotheses; some of its leukaemia and liver-fibrosis proposals showed activity in cell-line and organoid tests, a validation its authors call preliminary [16]. In December 2025 it announced its first automated materials lab, to open in the UK in 2026 and aiming at hundreds of materials per day [17]. | Peer-reviewed (Nature, 2026); company announcement |
OpenAI (frontier lab) | Automated research | Says it reached its “automated research intern” goal in September 2026 and targets an automated AI researcher by March 2028 [18]. | Company disclosure, self-reported |
Anthropic (frontier lab) | AI inside scientists’ existing tools | Claude for Life Sciences (October 2025) connects to Benchling, PubMed, 10x Genomics and other platforms, alongside an AI for Science credit programme [19]. | Company announcement |
Edison Scientific (AI-scientist startup; FutureHouse spin-out) | An AI scientist for literature and data | Kosmos runs for up to 12 hours, reads about 1,500 papers and executes about 42,000 lines of code per run. Independent scientists rated 79.4% of its statements accurate [20]. | Preprint |
Periodic Labs (AI-scientist startup) | AI scientists with autonomous labs for the physical sciences | A USD 300 million seed round (September 2025) from a16z, DST, Nvidia, Accel and others. First target: new superconductors [21]. | Press report |
Lila Sciences (AI-scientist startup) | “Scientific superintelligence” with robotic AI Science Factories | USD 550 million raised in total, at a valuation above USD 1.3 billion after an Nvidia-backed extension (October 2025) [22]. | Press report |
Sakana AI (AI-scientist startup) | An end-to-end AI Scientist | A fully AI-generated paper scored above the acceptance threshold in peer review at an ICLR 2025 workshop (average 6.33), then was withdrawn before meta-review, as agreed with the organisers [23]. | Company disclosure |
Insilico Medicine (AI drug discovery) | Generative AI from target to molecule | Reported dosing the first patient in the Phase III GENESIS-IPF-3 trial of rentosertib, September 2026 [14]. | Company disclosure; trial registered (NCT07687459) |
United States (government) | The Genesis Mission | An executive order of 24 November 2025 tasks the Department of Energy with an integrated AI discovery platform; DOE says it will mobilise its 17 national laboratories, industry and academia [24]. | Official documents |
European Union (government) | European Strategy for AI in Science and RAISE | The RAISE pilot launched in November 2025 with EUR 107 million, plus EUR 600 million from Horizon Europe for compute and an aim to double Horizon Europe’s annual AI investment to over EUR 3 billion [25]. | Official documents |
Three shifts stand out.
From models to systems. The frontier is no longer a single model. It is agents that plan, act and check their work, connected to tools, data and increasingly to robots.
From text to experiments. Periodic Labs’ thesis is explicit: the next valuable data will not come from the internet but from autonomous laboratories that generate fresh experimental results, including the negative results that are seldom published [21].
From a private race to public infrastructure. Washington and Brussels now treat AI for science as strategic infrastructure [24, 25].
One gap is just as visible. Judging by the announcements above, most of this capital goes into producing more findings, faster. Far less goes into verifying them, connecting them to what is already known, and making them understandable to the society that pays for science.
The next stage, part one: from answers to loops
A model gives an answer. A loop does the work: it plans, acts, observes the result, evaluates it and decides what to try next. That is the shift now underway in AI for science, and three kinds of loops are already running.
Loop | What the AI does | What closes the loop | Example |
|---|---|---|---|
Reasoning loops | Reads, hypothesises, critiques and refines | Other agents and tournament-style ranking | Google’s AI co-scientist generates, debates and evolves hypotheses [16]. Kosmos coordinates about 200 agent rollouts per objective through a shared world model [20]. Evidence: peer-reviewed (co-scientist, Nature 2026); preprint (Kosmos). |
Evaluator loops | Proposes code, algorithms or proofs | Automated evaluators: code that runs, tests that pass, proofs that check | AlphaEvolve pairs Gemini models with automated evaluators and an evolutionary search over candidate programs [11]. Evidence: company disclosure. |
Physical loops | Designs experiments and runs them with robots | Measurements from the laboratory | In 17 days of autonomous operation, Berkeley Lab’s A-Lab synthesised 36 of 57 target compounds, a count corrected down from 41 in 2026 [28]. DeepMind’s UK lab aims at hundreds of materials per day [17]. Periodic Labs and Lila Sciences are building autonomous labs at scale [21, 22]. Evidence: a peer-reviewed paper corrected in 2026, company plans, press reports. |
The loops are also getting longer. METR, an independent research organisation, measures the length of tasks, in the time human professionals take to do them, that AI agents can complete autonomously with 50% reliability. That length doubled roughly every seven months from 2019 to 2025 [26]. In METR’s January 2026 update, the doubling time since 2024 was about three months (89 days), though METR notes the trend is sensitive to which tasks are measured [27]. These estimates are derived primarily from software tasks and depend on the composition of the benchmark and on uncertain estimates of human completion time. They should not be read as evidence that autonomous scientific research, especially experimental work, will advance at the same rate.
Three consequences follow.
The loop is only as good as its evaluator. In mathematics and software, a candidate can be checked in seconds, so loops spin fast. In materials science, a synthesis takes hours or days. In medicine, a clinical loop closes in years. Where verification is slow, AI speeds up the thinking but not the knowing.
The speed and reliability of feedback, more than the size of the model, set the pace of AI-driven discovery in each field.
Loops create their own data. Autonomous labs do not only test hypotheses. They generate fresh, high-quality experimental data, including negative results that are seldom published [21]. Over time, that data may matter more than any single discovery.
Loops need human governance. Someone must set the objective, the budget, the stop rules and the standard of proof. Even OpenAI, while pursuing automated research, writes that its reasons for doing so do not mean rapid recursive self-improvement is necessarily an outcome to pursue [18].
Loops give AI agency, but agency without memory and provenance produces noise at scale. That is why the second ingredient matters.
The next stage, part two: from documents to graphs
Science is stored as documents, but it works as a graph. Its real structure is a network of entities (genes, molecules, materials, methods), of claims about them, of the evidence behind each claim, and of the citations, replications and contradictions that connect them all. A PDF hides that structure; a graph makes it computable.
In this article, “graphs” refers to three distinct layers: graph-based models, such as GNoME and GraphCast; knowledge graphs, such as TxGNN and OpenAlex; and provenance or research graphs that connect claims to data, code, models, replications and responsible people. These layers can complement one another, but they are not interchangeable.
Layer | What the graph holds | Examples and evidence |
|---|---|---|
Graph-based models | Atoms and bonds, or points on the globe: the graph is the model’s own internal structure | GNoME’s graph networks generated 2.2 million candidate crystal structures predicted to be stable [10]. GraphCast, a graph neural network, outperformed the most accurate operational deterministic forecasting system on 90% of 1,380 verification targets and produces a 10-day forecast in under a minute [29]. Evidence: peer-reviewed. |
Knowledge graphs | Entities and their relations: diseases, drugs and mechanisms; works, authors, institutions and citations | TxGNN learned from a medical knowledge graph spanning 17,080 diseases, 92% of them without an FDA-approved drug; it improved indication accuracy by 49.2% over eight methods and shows the multi-hop reasoning paths behind each prediction [30]. GraphRAG builds an entity knowledge graph from a corpus and answers global questions more comprehensively than standard retrieval [31]. OpenAlex models the research record as an open graph of hundreds of millions of works, free under a CC0 licence [32]; even the decline of disruptive science was measured on the citation graph [3]. Evidence: peer-reviewed (TxGNN), preprint (GraphRAG), documentation (OpenAlex). |
Provenance or research graphs | Claims linked to their data, code, model versions, replications and accountable people | The W3C PROV standard has described provenance as a graph of entities, activities and agents since 2013 [59]. In our assessment, this is the least developed layer in research practice, and the one that decides trust. Evidence: technical standard; authors’ assessment. |
Graphs answer three questions that loops cannot answer alone.
What is already known? A loop without a map of prior work rediscovers the known and calls it new.
What contradicts what? Failed replications, corrections and retractions must travel with a claim, or AI will keep building on work that has already fallen.
Where did this come from? Every result needs a lineage: which data, which code, which model, which human signed off.
The third question, which provenance graphs answer, is the one that will decide trust. A graph without a loop is a library; a loop without a graph is a very fast guesser. Together, they become a discovery engine.
Loops plus graphs: the discovery engine and five stages of maturity
The next operating system of science combines both: loops that act, graphs that remember, and humans who set direction and judge significance.
Each turn of the loop reads from the graph: what is known, what failed, what conflicts. Each turn writes back to it: new evidence, with its lineage attached. Humans sit above the loop. They choose the questions, set the standards of proof and decide what counts as a discovery.
Research organisations will not jump to this model in one step. Applying the maturity logic that two of us developed for enterprise AI [60], informed by the St. Gallen Integrated Management Concept [37], we see five cumulative stages. They are an analytical framework, not an empirically validated classification.
Stage | What it looks like | Value | Risk | You are here if |
|---|---|---|---|---|
1. AI as search | AI finds, summarises and translates papers | Reading speed | Confident summaries of sources that do not exist | Your researchers use AI mainly to search, summarise and translate |
2. AI as assistant | AI writes code, analyses data and drafts text under close human direction | Individual productivity | Unverified errors scale with output | AI drafts code, analyses or text, but a person directs every step and checks every result |
3. AI as in-silico scientist | AI runs closed loops on literature and data, as Kosmos and the AI co-scientist do | Hypothesis throughput | Plausible but wrong findings: even at 79.4% accuracy, roughly one statement in five fails independent checking [20] | At least one AI system generates and tests hypotheses on its own, with people reviewing the output |
4. AI as lab partner | AI closes the loop through robotic experiments, as in A-Lab and the new autonomous labs | Experimental throughput and fresh data | Capital intensity and narrow domains | AI decides which experiments to run next, and robots run them |
5. AI-native discovery systems | Many loops run continuously on a shared knowledge and provenance graph, while humans govern goals, norms and final judgment | Knowledge that compounds | Today a vision, not a reality | Several loops share one provenance-tracked knowledge graph and improve each other’s results |
Available surveys are consistent with widespread use at stages 1 and 2, but they do not directly measure this framework: both are self-reported, and McKinsey’s covers companies in general rather than research organisations. In Wiley’s 2025 survey of 2,430 researchers, 84% used AI tools and 62% used them for research or publication tasks [57]. In McKinsey’s 2026 survey, chatbots were the most widely scaled AI tool: 47% of respondents said their organisations were scaling them across the enterprise, against about two in ten for AI agents [48]. Stage 3 exists today in a handful of frontier systems, and stage 4 in a small number of laboratories.
Where do you stand? A five-question self-assessment
Do your researchers use AI mainly to search, summarise and translate? (Stage 1)
Does AI write code, analyse data or draft text for you, with a person directing each step? (Stage 2)
Does any AI system in your organisation generate and test hypotheses on its own, with people reviewing the output? (Stage 3)
Does AI choose which experiments to run next, and do robots run them? (Stage 4)
Do several AI loops share one provenance-tracked knowledge graph and improve each other’s results? (Stage 5)
Your stage is the number of the highest question you can answer with a yes. If you stop at question 2, you are at the stages where available surveys suggest AI use is already widespread, although no survey measures this framework directly. The decisive transition is not technical: it is the move from stage 2 to stage 3, when AI stops assisting individual scientists and starts running parts of the research process. That move requires a new management system, not just a better model, which is the subject of the sections that follow.
The trust bottleneck: when output is cheap, verification becomes scarce
AI does not only accelerate discovery; it accelerates everything, including error, fabrication and noise. The warning signs are already visible inside science’s most prestigious venues.
Fake references in elite papers. GPTZero, a developer of AI-detection tools, reported that, after scanning 4,841 papers accepted at NeurIPS 2025, it had confirmed 100 hallucinated citations across more than 50 papers, each of which had passed review by several experts; the analysis is the vendor’s own and has not been peer reviewed [34].
AI reviewing the science. Pangram Labs’ AI detector flagged about 21% of the 75,800 peer reviews submitted to ICLR 2026 as fully AI-generated, and more than half as showing signs of AI use, an estimate reported by Nature [35].
AI writing the science. Sakana reported that a fully AI-generated paper scored above the average acceptance threshold in peer review at an ICLR 2025 workshop; as agreed, it was withdrawn before any meta-review. Sakana’s own checks found the system still made occasional citation errors [23].
A landmark claim that collapsed. MIT publicly disavowed a widely discussed preprint on AI-driven materials discovery, stating it had no confidence in the provenance, reliability or validity of its data [36].
The subtler risk is not bad science but narrower science. A Nature study of 41.3 million papers found that scientists using AI publish 3.02 times more papers, receive 4.84 times more citations and become project leaders 1.37 years earlier. Yet AI adoption was associated with a 4.63% contraction in the collective volume of topics studied and 22% less engagement between scientists, as AI-assisted work concentrated in the most data-rich areas [33]. The data are observational, so they show association rather than cause, but the direction is a warning.
If thousands of loops optimise the same well-lit, data-rich corners, science becomes faster and smaller at the same time. Peer review, journals and citation counts were designed for a world in which producing a paper was expensive; that world is ending.
Trust has to be rebuilt as infrastructure, on four principles: provenance by default, so that every claim carries its data, code, model version and the humans accountable for it; verification at machine speed, with automated checks of citations, statistics and code, and real incentives for replication; diversity by design, so that funding and evaluation reward data-poor questions rather than more output; and transparency for the public, with sources shown, uncertainty stated and corrections visible. The organisations that build this layer will matter as much as those that build the models.
Beyond verification: what remains scarce when knowledge becomes abundant
Verification is the most visible bottleneck, but it is not the only one. Research organisations must still decide which questions deserve attention, which results are genuinely original, and which possibilities are worth pursuing. They must make trustworthiness verifiable, assign responsibility for decisions made through human-agent systems, and preserve sovereignty over personal and sensitive data. These six (attention, originality, judgment, trust, responsibility and sovereignty) are not additional technologies. They are the human capabilities and institutional conditions that keep automated science purposeful, accountable and legitimate.
Attention selects what enters the research loop. Originality expands the space of possible questions and explanations. Judgment decides what is true, important and worth pursuing. Trust emerges from evidence and provenance. Responsibility identifies who remains accountable for the outcome. Sovereignty sets the boundaries for legitimate data use.
An integrated management view: AI in science is a management-system question
Most debates about AI in science are about models, benchmarks and compute. The harder question is organisational: how must a university, a research institute or a corporate R&D unit change when AI starts running parts of the research process?
The St. Gallen Integrated Management Concept, developed by Knut Bleicher and continued by Christian Abegglen, offers a practical analytical framework. It connects normative, strategic and operational management, so that technology is not deployed in isolation from purpose, positioning and execution [37]. The framework is used here as a lens for structuring institutional choices; it is not presented as an empirically validated solution to AI in science.
Applied to AI in science, it raises three sets of questions.
Level | The question for research leaders | What maturity looks like |
|---|---|---|
Normative: orientation | Why, and within which limits, do we use AI in research? Who remains accountable for each published claim? | Humans remain accountable for every published claim. AI use is disclosed. Openness and provenance are institutional values, not afterthoughts. |
Strategic: positioning | Where will we build or partner as AI loops spread: data, laboratories, evaluators, graphs or talent? | An explicit value thesis for AI in research: which data assets to build or open, which compute and automated labs to access, which partners to choose, and how to protect a diverse research portfolio. |
Operational: execution | Where do verification gates, disclosure, escalation and human sign-off sit when agents do part of the research? | Provenance is logged in every workflow. Verification gates sit before publication. AI literacy is role-based. Success is measured by validated results and replications, not paper counts. |
Governance cuts across all three levels. It defines who may let an agent run, on which budget, against which standard of proof, and who signs off before a result leaves the building. Abegglen and colleagues call the design of such hybrid management systems, in which people and intelligent agents perform management tasks together, Agentonomics. They present it as the start of a research programme rather than a finished model, and as a lens that makes roles, decision rights, evidence requirements and escalation paths explicit for people and agents [61].
Without this integration, research organisations are likely to repeat what many companies experienced with enterprise AI: pilots everywhere, tools in every lab, and no change in how knowledge is actually produced and verified. In Abegglen’s terms, this is organised standstill [38].
“AI can multiply the output of science. Whether it multiplies understanding is a management question. Research institutions need normative clarity about responsibility, strategic choices about where to build capabilities, and operational discipline in how results are verified. Integrated management keeps people in charge of the loop.” — Prof. Dr. Christian Abegglen, President, StGallen imt Business School
Where open science, DeSci and OpenDeSci fit: verifiable infrastructure and a public interface
If AI loops multiply findings, two kinds of infrastructure become essential: systems that make results verifiable, and interfaces that make them understandable.
This is where open science and decentralised science (DeSci) can matter, if they mature. DeSci is relevant here only where its mechanisms add capabilities that established repositories, persistent identifiers, cryptographic signatures and open research databases do not already provide. The test is functional rather than ideological: each decentralised component should solve a specific provenance, incentive or governance problem better than the established alternative.
DeSci’s first wave was dominated by tokens, and the market data show how far sentiment ran ahead of substance. Bio Protocol’s BIO token, listed on Binance through its Launchpool, peaked at USD 0.92 on its listing day, 3 January 2025, for an estimated circulating market value of roughly USD 1.2 billion (the peak price times the tokens then in circulation) [39, 54]. On 2 October 2026 it traded at about USD 0.03, 96.6% below its peak price, for a market value of about USD 78 million, while circulating supply had grown from about 39% of the total at launch to about 75% [54, 55]. The 58 DeSci tokens tracked by CoinMarketCap were worth about USD 0.3 billion combined on that date, under a third of BIO’s value at its peak [56]. Speculation is not infrastructure.
In our analysis, three lessons stand out. Valuations were set by exchange listings and narratives rather than by research outcomes. A low initial float followed by nearly two years of token unlocks created steady selling pressure. And few projects could point to real-world users, revenue or scientific results. The latest developments point in a more useful direction: Bio Protocol has repositioned around a launchpad for AI research agents, so-called BioAgents [54], and the most valuable token in CoinMarketCap’s DeSci category is now OriginTrail, which runs a decentralised knowledge graph [56].
“DeSci’s first wave priced the dream, not the science. The next wave will be judged on real-world adoption: research that gets funded, results that hold up, and patients and scientists who use the product every day. Decentralisation matters when it makes science more open and more trusted. The real asset is the innovation, not the token.” — Evan Luthra, Co-founder and General Partner, KOL Capital; Advisor, OpenDeSci
DeSci’s underlying primitives may still help with the problems AI creates, as one possible implementation rather than a necessary trust layer. Blockchain can anchor provenance through tamper-evident fingerprints of datasets, code, model versions and claims, so that the lineage in a provenance graph cannot be silently rewritten. Signed records in a well-run repository already achieve much of this; a public ledger adds most where several parties must rely on a record without trusting any single operator. The same test applies to recording credit for verification (portable records of who reviewed, replicated or corrected a result), to transparent funding (grant and replication-bounty flows that funders and the public can follow) and to programmable payouts (regulated stablecoin payments to researchers, reviewers and science creators, across borders and in small amounts). Whether these mechanisms outperform repositories, DOIs, signatures and open databases is an empirical question, and few DeSci projects have yet published evidence at that level of comparison.
Universities do not disappear in this model. As Smeritschnig argued in an earlier essay, what is at stake is their gatekeeping, not their role in education, laboratories and accountability [41].
OpenDeSci, which the authors founded or advise, works on the public interface: the layer where AI-era science meets society. It describes itself as building a modular, AI-native science platform that turns peer-reviewed research into short videos, quizzes and plain-language summaries [40]. Its design principles, as the authors set them out here, follow the argument of this article: every explanation linked to its sources, uncertainty stated plainly, and decentralised technology used only where it strengthens provenance and trust.
The logic is simple. When machines can generate more science than any human can read, the decisive question for society is no longer access to papers. It is access to understanding that can be trusted.
“AI will not replace scientists, and DeSci will not replace universities. What both can replace is gatekeeping that no longer scales: closed review, opaque credit and slow verification. The winners will make verification open, fast and rewarded.” — Dr. Florian Smeritschnig, Founder, StrategyCase.com; Advisor, OpenDeSci
Outlook 2026–2030: five forecasts, and what would prove us wrong
The wider AI arc, in brief. Science is not the only place where AI is moving from answers to action. OpenAI said ChatGPT had reached 900 million weekly users in February 2026 [45] and 1.2 billion at the end of September [53]. The agent era has arrived: the Model Context Protocol, which connects agents to tools, passed to a neutral Linux Foundation body in December 2025, with more than 10,000 public servers according to Anthropic [47], and in McKinsey’s 2026 survey 40% of large organisations said they scale AI agents in at least one function, up from 27% a year earlier, against 22% of smaller ones [48]. OpenAI’s reported five-level scale runs from chatbots and reasoners to agents, innovators and organisations [46]. In our reading, AI for science belongs at the fourth level, innovators: loops connected to graphs.
Self-directed, not self-aware. Autonomy is rising fast, but autonomy is not self-awareness. The closest experimental evidence concerns something narrower: in Anthropic’s October 2025 concept-injection experiments, Claude Opus 4.1 correctly detected and identified injected concepts in about 20% of trials under the best-tested conditions, which points to limited and unreliable introspective access; the study does not establish machine self-awareness [51]. Meanwhile, METR and Epoch AI reported early evidence that AI can complete some coding tasks that would take a human engineer weeks, where the specification can be checked automatically; one run had an AI system working for 19 days without human intervention [50]. If METR’s trend holds, frontier systems could carry out month-long software projects by the end of the decade [26]. Silver and Sutton argue that the next generation of agents will learn mainly from their own streams of experience [52], which, in our view, makes recording that experience as a graph, with provenance, a precondition for trusting what they learn. Gartner expects agentic AI in 33% of enterprise software applications by 2028, and more than 40% of agentic AI projects to be cancelled by the end of 2027 [49]. The lesson for science is simple: the more self-directed the loop, the more the human-set objective, the standard of proof and the record of what happened matter.
One of the boldest public forecasts comes from Anthropic’s Dario Amodei, who argued in 2024 that AI-enabled biology could compress 50 to 100 years of progress into 5 to 10 years, a “compressed 21st century” [43]. He also cautioned that experiments on cells and animals take time that may be irreducible, and that clinical trials remain slow, although AI could reduce the iterations they need [43]. In July 2026 he told STAT: “I don’t think that today we can make progress at a rate of 10 years per year” [44].
Our view is more differentiated: the acceleration is real but uneven, fastest where evaluation can be automated and slowest where verification depends on experiments and clinical trials. On that basis, we make five forecasts. Each has a measurable outcome and a deadline, so that by early 2031 anyone can check whether we were right. They are the authors’ forecasts, not established facts.
Provenance becomes a condition of publication. By the end of 2028, at least three major funders or journals, for example the NIH, Horizon Europe, Nature or Science, will require machine-readable provenance for data, code and AI use with every submission. Measure: published policies.
Self-driving labs scale first in materials and chemistry. By the end of 2030, more than 10,000 new materials synthesised by autonomous laboratories will have been independently validated, and at least one lab will run at hundreds of materials per day [17, 21, 28]. Measure: published counts.
The first regulatory decision on a generative-AI-discovered drug. By the end of 2030, rentosertib will have received a regulatory decision in at least one major market; we expect approval [14, 42]. Measure: regulator announcements.
Verification becomes an industry. By the end of 2028, at least five of the ten largest scholarly publishers will run automated citation and integrity checks on every submission, and NeurIPS, ICML and ICLR will screen reviews for AI generation. Measure: public policies and conference rules.
The public meets science through AI. By the end of 2030, more than half of adults in both the United States and the European Union will report using AI assistants to understand science or health information at least monthly. Measure: Pew Research and Eurobarometer surveys.
What would prove us wrong?
If AI-discovered drugs succeed in Phase II and III at clearly higher rates than history, biology is less of a bottleneck than we assume. In a 2024 analysis of a small, non-exhaustive sample from AI-native biotechs, 4 of 10 Phase II trials completed by December 2023 were classified as successful, in line with historical averages [15].
If autonomous labs absorb billions without producing independently validated discoveries, the physical-loop thesis weakens.
If verification capacity does not scale, trust in science could fall faster than discovery rises. That is the scenario to prevent.
The leadership message
AI in science is not a model race; it is a redesign of how knowledge is produced, verified and shared. The winners will not be the organisations with the most AI pilots, the largest models or the most papers, but those that connect loops to verification, graphs to provenance, and discoveries to public understanding.
The real question is therefore not how much science AI can produce, but how much of it we can trust, connect and understand.
“In the next five years, AI will produce more research than any human can read. The opportunity is not to publish faster. It is to make knowledge verifiable, connected and understandable for everyone, from the scientist in the lab to the student on a phone.” — David Wang, Founder, OpenDeSci
What to do now
Whatever the role, the sequence is the same. Choose one verifiable loop, starting where a credible evaluator already exists. Define the standard of proof: what evidence closes the loop, and what remains uncertain. Log provenance from day one: data, code, model versions, corrections and sign-off. And assign human decision rights: budgets, stop rules, escalation paths and final accountability. The first moves then differ by role.
Who | First moves |
|---|---|
Research leaders and universities | Assess where you stand on the five stages. Name an accountable owner for AI in research. Make AI-use disclosure and provenance logging mandatory. Pilot one closed loop in a data-rich area, with verification gates. |
Funders and policymakers | Fund verification as infrastructure: replication, open data and open scholarly graphs. Reward exploration of data-poor questions. Attach openness and provenance requirements to public AI-for-science programmes. |
Companies and investors | Back loops with real evaluators and proprietary data, not demos. Judge AI-science claims on prospective evidence, such as external synthesis for materials and controlled Phase II results for drugs, and seek independent replication where feasible. Value the verification and provenance layer, not only the model. |
Scientists and science communicators | Use AI agents, but sign only what you can verify. Publish data, code and negative results. Explain findings with their sources and their uncertainty. |
Scope and approach
This perspective draws on a purposive rather than systematic selection of sources, chosen for their relevance to AI-driven discovery, verification and the governance of research, and published mostly between 2023 and 2026. Claims are cited to the original publication wherever possible. Because the sources carry different weight, each entry in the source list is labelled with its evidence type, such as peer-reviewed study, preprint, official record, company disclosure, press report, survey, market data, external estimate or the authors’ own work. Company claims, surveys and vendor estimates are also identified as such in the text.
Declarations
Funding. This work received no external funding.
Use of AI. AI tools, including Claude (Anthropic), were used for literature discovery, drafting support and checking claims and references against the cited sources. The authors reviewed the final text and take full responsibility for it, including all factual claims, quotations and references.
About the authors
David Wang is the founder of OpenDeSci, an AI-native science platform, Managing Director of Wang Advisory GmbH and a Senior Lecturer at StGallen imt Business School. An entrepreneur and management consultant with more than ten years of international consulting experience, he worked at Deloitte and on projects for Bain & Company and Kearney, and has advised private-equity funds and Fortune 500 companies on strategy, value creation and transformation.
Prof. Dr. Christian Abegglen is President of StGallen imt Business School. He is co-author and editor of Das Konzept Integriertes Management, the standard work on the St. Gallen Integrated Management Concept founded by Knut Bleicher, and an advisor and research collaboration partner of OpenDeSci.
Dr. Florian Smeritschnig is the founder of StrategyCase.com and an advisor to OpenDeSci. He holds a PhD from the University of Vienna and a Master’s from the Hong Kong University of Science and Technology, and previously worked at McKinsey and Bitpanda on strategy, growth and operations.
Evan Luthra is a co-founder and General Partner of KOL Capital and an advisor to OpenDeSci. An angel investor since 2014 and an early advocate of Bitcoin, he invests across Web3, blockchain, AI and frontier technology. He was named in the Forbes 30 Under 30 list and is the author of The Millionaire Mindset.
Read the full reading deck or see AI-native science in practice.
Sources
All sources accessed between 30 September and 26 October 2026. Each entry ends with its evidence type.
Bonaglia, D., Rivera León, L. et al. (23 December 2025). End of Year Edition: Despite the Odds, Global R&D Spending Grew Again in 2024, Inching Closer to the USD 3 Trillion Mark. WIPO Global Innovation Index. wipo.int. Type: official statistic (estimate).
Bloom, N., Jones, C. I., Van Reenen, J. and Webb, M. (2020). Are Ideas Getting Harder to Find? American Economic Review 110(4), 1104–1144. aeaweb.org. Type: peer-reviewed.
Park, M., Leahey, E. and Funk, R. J. (2023). Papers and patents are becoming less disruptive over time. Nature 613, 138–144. nature.com. Type: peer-reviewed.
Scannell, J. W., Blanckley, A., Boldon, H. and Warrington, B. (2012). Diagnosing the decline in pharmaceutical R&D efficiency. Nature Reviews Drug Discovery 11, 191–200. nature.com. Type: peer-reviewed.
Hanson, M. A., Gómez Barreiro, P., Crosetto, P. and Brockington, D. (2024). The strain on scientific publishing. Quantitative Science Studies 5(4), 823–843. doi.org. Type: peer-reviewed.
Baker, M. (2016). 1,500 scientists lift the lid on reproducibility. Nature 533, 452–454. doi:10.1038/533452a. doi.org. Type: survey, self-reported (journal news feature).
Van Noorden, R. (2023). More than 10,000 research papers were retracted in 2023: a new record. Nature 624, 479–481. nature.com. Type: journal news analysis.
Richardson, R. A. K., Hong, S. S., Byrne, J. A., Stoeger, T. and Amaral, L. A. N. (2025). The entities enabling scientific fraud at scale are large, resilient, and growing rapidly. PNAS 122(32), e2420092122. doi.org. Type: peer-reviewed.
The Royal Swedish Academy of Sciences (9 October 2024). Press release: The Nobel Prize in Chemistry 2024. nobelprize.org. Type: official announcement.
Merchant, A. et al. (2023). Scaling deep learning for materials discovery. Nature 624, 80–85. doi:10.1038/s41586-023-06735-9. nature.com; summary: Google DeepMind (29 November 2023), deepmind.google. Type: peer-reviewed.
AlphaEvolve team (14 May 2025). AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms. Google DeepMind. deepmind.google. Type: company disclosure.
Google DeepMind (21 July 2025). Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad. deepmind.google. Type: company disclosure; scores certified by IMO coordinators.
Xu, Z. et al. (2025). A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nature Medicine 31(8), 2602–2610. doi:10.1038/s41591-025-03743-2. pubmed.ncbi.nlm.nih.gov. Type: peer-reviewed randomised trial.
Insilico Medicine (10 September 2026). Insilico Medicine doses first patient in GENESIS-IPF-3, the world’s first Phase III trial of a generative AI-driven innovative drug. insilico.com. Trial registration: ClinicalTrials.gov NCT07687459, 320 participants planned at 47 centres in China. Type: company disclosure; trial registry record.
Jayatunga, M. K. P., Ayers, M., Bruens, L., Jayanth, D. and Meier, C. (2024). How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discovery Today 29(6), 104009. doi:10.1016/j.drudis.2024.104009. BCG summary: bcg.com. Type: peer-reviewed analysis, preliminary.
Gottweis, J. et al. (2026). Accelerating scientific discovery with Co-Scientist. Nature 655, 487–496 (published online 19 May 2026). doi:10.1038/s41586-026-10644-y. nature.com; preprint: Towards an AI co-scientist, arXiv:2502.18864 (2025). Type: peer-reviewed.
Hassabis, D., Ibrahim, L., Kohli, P. and Larter, O. (10 December 2025). Strengthening our partnership with the UK government to support prosperity and security in the AI era. Google DeepMind. deepmind.google. Type: company announcement.
OpenAI (6 September 2026). Research acceleration: The view inside OpenAI. openai.com. Type: company disclosure, self-reported.
Anthropic (20 October 2025). Claude for Life Sciences. anthropic.com. Type: company announcement.
Mitchener, L. et al. (2025). Kosmos: An AI Scientist for Autonomous Discovery. arXiv:2511.02824. arxiv.org; Edison Scientific announcement: edisonscientific.com. Type: preprint.
TechCrunch (30 September 2025). Former OpenAI and DeepMind researchers raise whopping $300M seed to automate science. techcrunch.com; Periodic Labs, company website (accessed October 2026): “They generate valuable negative results which are seldom published.” periodic.com. Type: press report; company statement.
Hu, K. (14 October 2025). Exclusive: AI lab Lila Sciences tops $1.3 billion valuation with new Nvidia backing. Reuters; syndicated copy: finance.yahoo.com. Type: press report.
Sakana AI (12 March 2025). The AI Scientist generates its first peer-reviewed scientific publication. sakana.ai. Type: company disclosure.
U.S. Department of Energy, via PNNL (24 November 2025). Energy Department launches Genesis Mission. pnnl.gov; Executive Order 14363, Launching the Genesis Mission, Federal Register, 28 November 2025. Type: official documents.
European Commission (2025). European Strategy for Artificial Intelligence in Science and RAISE: strategy of 8 October 2025; RAISE pilot launched at the AI in Science Summit, 3–4 November 2025. research-and-innovation.ec.europa.eu. Type: official document.
METR (19 March 2025). Measuring AI Ability to Complete Long Software Tasks. metr.org. Type: external estimate by an independent research organisation.
METR (29 January 2026). Time Horizon 1.1. metr.org. Type: external estimate by an independent research organisation.
Szymanski, N. et al. (2023). An autonomous laboratory for the accelerated synthesis of inorganic materials. Nature 624, 86–91. doi:10.1038/s41586-023-06734-w; Author Correction, Nature 650, E1 (published 19 January 2026). nature.com. Type: peer-reviewed, corrected.
Lam, R. et al. (2023). Learning skillful medium-range global weather forecasting. Science 382, 1416–1421. science.org. Type: peer-reviewed.
Huang, K. et al. (2024). A foundation model for clinician-centered drug repurposing. Nature Medicine 30(12), 3601–3613. doi:10.1038/s41591-024-03233-x. nature.com. Type: peer-reviewed.
Edge, D. et al. (2024). From Local to Global: A Graph RAG Approach to Query-Focused Summarization. arXiv:2404.16130. arxiv.org. Type: preprint.
OpenAlex developer documentation. developers.openalex.org; OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts (2022), arXiv:2205.01833. Type: documentation; preprint.
Hao, Q., Xu, F., Li, Y. and Evans, J. (2026). Artificial intelligence tools expand scientists’ impact but contract science’s focus. Nature 649, 1237–1243. doi:10.1038/s41586-025-09922-y. nature.com. Type: peer-reviewed, observational.
GPTZero (21 January 2026). GPTZero finds 100 new hallucinations in NeurIPS 2025 accepted papers. gptzero.me. Type: external estimate by an AI-detection vendor.
Naddaf, M. (2025). Major AI conference flooded with peer reviews written fully by AI. Nature 648, 256–257. doi:10.1038/d41586-025-03506-6. nature.com; Pangram Labs (18 November 2025). Pangram predicts 21% of ICLR reviews are AI-generated. pangram.com. Type: external estimate by an AI-detection vendor, reported by Nature.
MIT Department of Economics (16 May 2025). Assuring an accurate research record. economics.mit.edu. Type: official statement.
Abegglen, C. and Bleicher, K. (2021). Das Konzept Integriertes Management: Visionen – Missionen – Programme. 10th fully updated and expanded edition. Frankfurt/New York: Campus. ISBN 978-3-593-51171-9. Type: book co-authored by one of the authors.
Abegglen, C. (2021). Unternehmen neu erfinden: Das Denk- und Arbeitsbuch gegen organisierten Stillstand. 2nd fully revised edition. Frankfurt am Main: Frankfurter Allgemeine Buch. ISBN 978-3-96251-087-9. Type: book by one of the authors.
Binance (23 December 2024). Introducing BIO Protocol (BIO) on Binance Launchpool! Farm BIO by Locking BNB and FDUSD. Listing on 3 January 2025, 10:00 UTC; initial circulating supply 1,296,529,168 BIO (39.05% of genesis supply). binance.com. Type: exchange notice.
OpenDeSci Foundation website. opendesci.org. Type: the authors’ own organisation.
Smeritschnig, F. (24 January 2026). Decentralized Science Will Not Replace Universities. It Will Replace Their Gatekeeping. OpenDeSci. opendesci.org. Type: essay by a co-author.
Proffitt, A. (8 July 2026). Insilico Medicine launches Phase III trial for AI-developed drug. Bio-IT World. bio-itworld.com. Type: press report.
Amodei, D. (October 2024). Machines of Loving Grace (essay). darioamodei.com. Type: essay.
Herper, M. (6 July 2026). I spoke to Anthropic’s CEO about how AI may affect biotech. Here’s what I learned. STAT. statnews.com. Type: press interview.
TechCrunch (27 February 2026). ChatGPT reaches 900M weekly active users. techcrunch.com. Type: press report of a company figure.
Axios (15 July 2024). OpenAI nears “reasoning”-capable AI, reporting Bloomberg on OpenAI’s five-level scale. axios.com. Type: press report.
Anthropic (9 December 2025). Donating the Model Context Protocol and establishing the Agentic AI Foundation. anthropic.com; Paoli, C. (9 December 2025). Linux Foundation Launches Agentic AI Foundation with Microsoft, OpenAI and Anthropic. Redmond Magazine. redmondmag.com. Type: company announcement; press report.
McKinsey (25 August 2026). The state of AI in 2026: On the road to ROI. Survey of 1,719 participants in 97 countries, fielded 4 May to 8 June 2026. mckinsey.com. Type: survey, self-reported.
Gartner (25 June 2025). Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027. Press release. gartner.com. Type: external forecast.
METR and Epoch AI (10 April 2026). MirrorCode: Evidence that AI can already do some weeks-long coding tasks (preliminary results). metr.org; Epoch AI (2026). MirrorCode: What’s the largest software project AI can complete on its own? Co-developed with METR. epoch.ai. Type: external estimate (preliminary, then full results).
Anthropic (29 October 2025). Signs of introspection in large language models. anthropic.com. Type: company research.
Silver, D. and Sutton, R. S. (2025). Welcome to the Era of Experience. Preprint of a chapter in Designing an Intelligence (MIT Press). storage.googleapis.com. Type: preprint.
OpenAI (29 September 2026). DevDay 2026 recap. openai.com. Type: company disclosure.
CoinMarketCap. Bio Protocol (BIO): price, market capitalisation, supply and all-time high, accessed 2 October 2026. coinmarketcap.com. Type: market data.
1inch. Bio Protocol (BIO) token profile: Binance Launchpool debut on 3 January 2025, YZi Labs investment and supply at launch, citing exchange research. 1inch.com. Type: market data.
CoinMarketCap. Top DeSci tokens by market capitalisation (58 tokens), accessed 2 October 2026. coinmarketcap.com. Type: market data.
Wiley (7 October 2025). ExplanAItions 2025: The Evolution of AI in Research, survey of 2,430 researchers. Press release: newsroom.wiley.com; report: wiley.com. Type: survey, self-reported.
Cheetham, A. K. and Seshadri, R. (2024). Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials Discovery. Chemistry of Materials 36(8), 3490–3495. doi:10.1021/acs.chemmater.4c00643. doi.org. Type: peer-reviewed perspective.
W3C (30 April 2013). PROV-Overview: An Overview of the PROV Family of Documents. W3C Working Group Note; the PROV data model (PROV-DM) is a W3C Recommendation of the same date. w3.org. Type: technical standard.
Wang, D. and Abegglen, C. (26 August 2026). Integrated Management in the AI Era: How the St. Gallen Integrated Management Concept helps organisations move from AI experimentation to AI maturity. White paper, Wang Advisory GmbH, in collaboration with StGallen imt Business School. wangadvisory.com. Type: white paper by two of the authors.
Abegglen, C., Knoll, A. C., Suckfüll, H. C. and von Grone, J. (2026). Agentonomics: Management im Zeitalter intelligenter Agentensysteme. Grundzüge eines managementwissenschaftlichen Forschungs- und Gestaltungsansatzes. Technical Report TUM-I26129, Technical University of Munich, Department of Informatics, August 2026. gimsg.ch. Type: technical report co-authored by one of the authors.
Comments