<\/p>
The collection of mathematics papers released by OpenAI<\/mark><\/a> currently contains 719 papers produced by the company’s internal model. OpenAI cautions that the results are at various stages of verification and some may contain errors.<\/p> In an update dated October 7, 2026<\/mark><\/a>, the company announced the withdrawal of three papers. A sign error had been found in one proof. Two other papers relied on the same reasoning, so all three had to be withdrawn. Fourteen other papers were revised, with changes including corrections to proofs and more precise assumptions.<\/p> Revising and withdrawing papers is part of scientific research. In this case, the review process worked: errors were found and the company responded. But the number of corrections doesn’t tell us whether the entire collection is being checked quickly enough or how much work remains.<\/p> The published papers may contain valuable ideas, but using them requires an assessment of their proofs, assumptions and the dependencies between individual results. That raises a practical question: how should this assessment be organized when AI makes it possible to produce more and more proposals?<\/p> Alex Imas of Google DeepMind and James Manyika of Google describe the burden of checking results in an essay published on October 8 on the DeepMind Institute website<\/mark><\/a>. They refer to the study \u201cAI in science: Early insights\u201d<\/mark><\/a>, produced by teams at Google, Google DeepMind and MIT FutureTech.<\/p> The study included a survey of 637 scientists in the US and UK conducted in the summer of 2026. Respondents reported saving an average of 6.9 hours a week through AI. The authors describe this figure as a net time saving.<\/p> A separate question covered checking results. About 46% of respondents who saved time with AI said they spent more than a quarter of the time gained from generating text, code or hypotheses on verification and corrections. This tells us about the workload involved in checking outputs. But that response shouldn’t be used to reduce the reported net savings a second time.<\/p> About 41% of respondents also said they had more untested hypotheses than three years earlier. These are participants’ own reports and the sample isn’t representative of the scientific community as a whole. The findings point to a problem, but they don’t prove that AI is causing widespread research backlogs.<\/p> How might such backlogs develop? If AI identifies many new substances with potentially useful properties, scientists have to decide which ones to investigate. They then have to synthesize them in the lab and take measurements. Generating a longer list of candidates doesn’t increase the number of available lab stations or the time available on equipment.<\/p> Two tasks need to be distinguished here. One is detecting the model’s mistakes. The other is testing whether an experiment supports a proposed hypothesis. Even correct calculations can’t replace that research. More accurate AI may therefore make fewer mistakes while also proposing more solutions that require time-consuming tests.<\/p> When assessing AI productivity, it’s easy to confuse receiving an answer with completing a task. A document may have a clear layout, correct language and persuasive conclusions. But the person responsible for approving it still has to establish whether those conclusions can be trusted.<\/p> In a market analysis, for example, someone has to check whether the compiled data covers the same periods and comparable products. Reviewing recommendations for management requires checking whether they account for the specific company’s circumstances. A developer reviewing code, meanwhile, has to assess its behavior in less typical situations that basic tests don’t cover.<\/p> Accuracy is only the starting point. An analysis may contain reliable data but overlook the issue that matters most to the decision at hand. A program may work but be difficult and expensive to maintain. Therefore, reviewers also have to assess a solution’s usefulness and the consequences of implementing it.<\/p> If a company measures only the time spent preparing a first draft, it overlooks this part of the work. A full accounting has to include revisions, consultations and document approval. Only then can the company assess how much time the team as a whole has saved.<\/p> Imagine a department where several people prepare analyses and one person is responsible for approving them. With AI, the authors can produce more reports. If they all go to the same person, that person’s workload grows even if the quality of individual documents stays the same.<\/p> For the author, submitting an analysis may mark the end of a task. For their manager, receiving it is the start of another.<\/p> The same pattern can affect an editor receiving more articles, a developer reviewing colleagues’ code or a manager assessing proposals. Time saved by authors can mean extra work for the recipient, especially when authors send every generated version and leave the recipient to choose the best one. The time needed for that assessment also determines whether the team as a whole benefits.<\/p> The workload can grow even without an increase in the number of documents. Each document simply has to get longer. AI makes it easier to add supporting arguments, comparisons and alternative scenarios. The recipient still has to read them and decide whether they contribute anything useful.<\/p> As tasks multiply, the person responsible for quality needs more time or help. If deadlines stay the same and no one takes over some of their duties, pressure builds to check less thoroughly. A carefully written document may then receive an overly cursory review because, at first glance, it gives no cause for concern.<\/p> Much depends on the form in which authors hand over their work. They can use AI to flag inconsistencies and unnecessary passages, but that review shouldn’t be treated as confirmation that the data is accurate. They still need to consult sources, check calculations or run appropriate tests. Authors should also select the proposals that merit further assessment themselves. A report prepared this way is easier to review.<\/p> The first step, then, is to decide which proposals actually need further assessment. The ease of generating them isn’t a reason to send every one to a manager. Without an initial selection, ideas will keep piling up with no one available to consider them.<\/p> Anyone making a recommendation should explain its basis, what they’ve already checked and which uncertainties remain. The person approving it also needs to know what’s expected of them: a data assessment, a choice between options or approval for a specific action. That way, they don’t have to reconstruct the author’s entire process.<\/p> Some checks can be automated. Software can recalculate totals, check data completeness or run code tests. These solutions should be introduced alongside tools that prepare documents and write software. But users need to understand the scope of these checking tools. A table with correctly calculated totals can still contain data that isn’t comparable.<\/p> The depth of a review should also depend on the consequences of a potential error. A draft title needs less attention than an analysis that will inform a major company expenditure. Reviewing every proposal in equal detail can unnecessarily slow down the work. Giving every document a cursory sign-off, on the other hand, increases the risk of poor decisions.<\/p> The OpenAI example shows why checking results matters: finding one error also called into question papers that relied on the same reasoning. But it doesn’t tell us how quickly the entire collection can be verified. Companies need to address a similar issue when planning their team’s work: who will assess new proposals, to what extent and when.<\/p> A company that produces twice as many analyses with AI should therefore examine the outcome. Is it making sound decisions faster? Is it completing projects more efficiently? If all that’s grown is the pile of documents awaiting approval, it’s hard to call that a productivity gain for the team as a whole.<\/p>","protected":false},"excerpt":{"rendered":" AI allows us to prepare analyses, propose solutions and write research papers faster. But someone still has to assess their accuracy and usefulness. Can the people responsible for reviewing this work keep up with the growing volume of results? And where does that leave the benefits of working faster?<\/p>\n","protected":false},"author":465,"featured_media":20524,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"rank_math_lock_modified_date":false,"footnotes":""},"categories":[888,832],"tags":[],"popular":[],"difficulty-level":[],"ppma_author":[892],"class_list":["post-20574","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-business-2","category-editors-picks"],"acf":[],"authors":[{"term_id":892,"user_id":465,"is_guest":0,"slug":"kmironczuk","display_name":"Krzysztof Miro\u0144czuk","avatar_url":{"url":"https:\/\/haimagazine.com\/wp-content\/uploads\/2025\/10\/awatar-2.png","url2x":"https:\/\/haimagazine.com\/wp-content\/uploads\/2025\/10\/awatar-2.png"},"first_name":"Krzysztof","last_name":"Miro\u0144czuk","user_url":"","job_title":"","description":"Od lat zajmuj\u0119 si\u0119 nowymi technologiami w biznesie, edukacji i codziennym \u017cyciu. W centrum mojej uwagi pozostaje cz\u0142owiek \u2013 i to, by technologia wyr\u00f3wnywa\u0142a szanse, zamiast tworzy\u0107 bariery."}],"_links":{"self":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts\/20574","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/users\/465"}],"replies":[{"embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/comments?post=20574"}],"version-history":[{"count":1,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts\/20574\/revisions"}],"predecessor-version":[{"id":20575,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts\/20574\/revisions\/20575"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/media\/20524"}],"wp:attachment":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/media?parent=20574"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/categories?post=20574"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/tags?post=20574"},{"taxonomy":"popular","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/popular?post=20574"},{"taxonomy":"difficulty-level","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/difficulty-level?post=20574"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/ppma_author?post=20574"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}Ideas are piling up<\/strong><\/h4>
Checking is work too<\/strong><\/h4>
Who’s going to do all this?<\/strong><\/h4>
Organizing the review process<\/strong><\/h4>