<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:itunes="http://www.itunes.com/dtds/podcast-1.0.dtd" xmlns:googleplay="http://www.google.com/schemas/play-podcasts/1.0"><channel><title><![CDATA[Tauseef Ahmad Khan]]></title><description><![CDATA[Nutrition scientist at the University of Toronto. I study sugars, sweeteners and chronic disease, and write about how AI is changing science, learning and the university.]]></description><link>https://blog.tauseefkhan.ai</link><image><url>https://substackcdn.com/image/fetch/$s_!m7T9!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1962d911-9093-469b-832a-8e0be4605a36_835x835.jpeg</url><title>Tauseef Ahmad Khan</title><link>https://blog.tauseefkhan.ai</link></image><generator>Substack</generator><lastBuildDate>Sun, 22 Nov 2026 23:32:20 GMT</lastBuildDate><atom:link href="https://blog.tauseefkhan.ai/feed" rel="self" type="application/rss+xml"/><copyright><![CDATA[Tauseef Ahmad Khan]]></copyright><language><![CDATA[en]]></language><webMaster><![CDATA[tauseefahmadkhan@substack.com]]></webMaster><itunes:owner><itunes:email><![CDATA[tauseefahmadkhan@substack.com]]></itunes:email><itunes:name><![CDATA[Tauseef Ahmad Khan]]></itunes:name></itunes:owner><itunes:author><![CDATA[Tauseef Ahmad Khan]]></itunes:author><googleplay:owner><![CDATA[tauseefahmadkhan@substack.com]]></googleplay:owner><googleplay:email><![CDATA[tauseefahmadkhan@substack.com]]></googleplay:email><googleplay:author><![CDATA[Tauseef Ahmad Khan]]></googleplay:author><itunes:block><![CDATA[Yes]]></itunes:block><item><title><![CDATA[A Freight Train Through the Ivory Tower]]></title><description><![CDATA[AI now tutors students and drafts papers. Universities must decide what they can still certify.]]></description><link>https://blog.tauseefkhan.ai/p/a-freight-train-through-the-ivory</link><guid isPermaLink="false">https://blog.tauseefkhan.ai/p/a-freight-train-through-the-ivory</guid><dc:creator><![CDATA[Tauseef Ahmad Khan]]></dc:creator><pubDate>Mon, 12 Oct 2026 22:59:47 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!ohrc!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F372d251f-48ba-413d-affd-100017b0b61e_1456x816.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!ohrc!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F372d251f-48ba-413d-affd-100017b0b61e_1456x816.png 424w, https://substackcdn.com/image/fetch/$s_!ohrc!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F372d251f-48ba-413d-affd-100017b0b61e_1456x816.png 848w, https://substackcdn.com/image/fetch/$s_!ohrc!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F372d251f-48ba-413d-affd-100017b0b61e_1456x816.png 1272w, https://substackcdn.com/image/fetch/$s_!ohrc!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F372d251f-48ba-413d-affd-100017b0b61e_1456x816.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!ohrc!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F372d251f-48ba-413d-affd-100017b0b61e_1456x816.png" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/372d251f-48ba-413d-affd-100017b0b61e_1456x816.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:null,&quot;width&quot;:null,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;A freight train, driven by a human and an AI agent, entering an arch in an ivory university tower.&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="A freight train, driven by a human and an AI agent, entering an arch in an ivory university tower." title="A freight train, driven by a human and an AI agent, entering an arch in an ivory university tower." srcset="https://substackcdn.com/image/fetch/$s_!ohrc!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F372d251f-48ba-413d-affd-100017b0b61e_1456x816.png 424w, https://substackcdn.com/image/fetch/$s_!ohrc!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F372d251f-48ba-413d-affd-100017b0b61e_1456x816.png 848w, https://substackcdn.com/image/fetch/$s_!ohrc!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F372d251f-48ba-413d-affd-100017b0b61e_1456x816.png 1272w, https://substackcdn.com/image/fetch/$s_!ohrc!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F372d251f-48ba-413d-affd-100017b0b61e_1456x816.png 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><p><em>For this write-up, I found and read the sources myself, and outlined the piece; the draft was written by Claude Opus 5.5, and I revised it extensively.</em></p><p>Universities are where knowledge is made, verified and passed on by senior academics and their teams. They do this through two jobs: teaching the next generation, and deciding through peer review and publication which findings count. Now anyone can get a professor-level tutor at two in the morning, for free, and AI can draft a passable research paper overnight. So what is a university for? I am a nutritional epidemiologist, and fields like mine move at a glacial pace: paper by paper, presentation by presentation, conference by conference. AI is arriving like a freight train, and it will push aside anyone who does not hop on.</p><p>We seek knowledge to understand how the world works and to improve lives, our own, and those of others. Universities serve both ends and pass them on. They have relied on two kinds of evidence, and AI has weakened both: a student&#8217;s written work as proof of learning, and a scientist&#8217;s paper as proof of a finding. What remains for universities is the slow work of testing hypotheses against the world and training the people who do it.</p><h2>The teacher and the test</h2><p>For most of history, learning meant finding a learned person. Students in Baghdad, Damascus and C&#243;rdoba travelled for years to read a text with a named teacher, who then <a href="https://www.iranicaonline.org/articles/ejaza">wrote a licence</a> vouching for both the student and the copy they wrote. When <a href="https://yalebooks.yale.edu/book/9780300089554/paper-before-print/">cheap paper reached Baghdad</a> in the eighth century and books multiplied, trust still rested on who had taught you. Universities and journals later formalized a similar double promise: the degree vouches for the student, and peer review vouches for the work. AI strains both, starting with the degree.</p><p>An AI tutor can teach, if it is built to. In <a href="https://doi.org/10.1038/s41598-025-97652-6">a randomized trial in Harvard&#8217;s largest physics class</a>, a carefully designed tutor more than doubled learning gains over an active-learning class, on tests taken the same day. A careless one does harm. In <a href="https://doi.org/10.1073/pnas.2422633122">a trial with high-school mathematics students</a>, a standard chatbot raised practice scores by 48%, yet the students later scored 17% lower than controls on an exam taken without it. Their work improved while their learning fell.</p><p>That is the problem for the degree, because for centuries universities have judged learning by what students write. <a href="https://kevinbryanecon.com/teachwithai.html">Kevin Bryan, an economist at the University of Toronto</a>, puts it plainly: &#8220;the link between performance and student knowledge has been broken.&#8221; He now asks students for an early pr&#233;cis, shows them what a frontier model can already produce from it, and requires their final paper to be better. I expect the degree to move toward certifying what a graduate can do without the machine. Baghdad did something similar in 931: after a fatal medical error, it <a href="http://www.tertullian.org/fathers/ibn_abi_usaibia_02.htm">barred its physicians until each passed an examination</a> and received a certificate naming exactly what he could practise. The degree of the future will look more like that certificate than a transcript: a promise that a human has verified the ability.</p><h2>The gate moves</h2><p>Publishing, the second job, is strained harder. The tools that make papers cheap also make slop cheap. Submissions to <a href="https://blog.arxiv.org/2026/10/01/updated-rate-limit-policy/">arXiv</a>, the main preprint server for physics and AI, doubled in two years to 40,363 in September alone; on 1 October it limited each submitter to two a month. <a href="https://doi.org/10.1371/journal.pbio.3003152">Single-factor papers</a> mined from one US health and nutrition survey rose from about four a year to 190 in 2024, counted only to 9 October. People still have to judge what is true and what matters but with such immense load, how can they cope anymore! As usual, gatekeepers will resist any change to the status quo, but I doubt they will decide the outcome. The frontier of AI research already sits outside the universities, inside companies where <a href="https://casp.ac/reports/intelligence-explosion">AI writes most of the code</a> and much of that work can happen out of public view. A generation that can do research with AI will not wait for their permission.</p><p>That is why I expect journals, as they stand, to disappear, and better venues to replace them. Writing a paper is becoming routine work for machines, and papers are already <a href="https://doi.org/10.1038/s41586-026-11044-y">being turned into agents</a> that answer questions about the work and rerun its methods. The record will split in two: a few pages for people to read, and hundreds of pages of data, code and logs for agents, because <a href="https://arxiv.org/abs/2604.24658">a single narrative can no longer serve both readers</a>. <a href="https://doi.org/10.1073/pnas.2610214123">A peer-reviewed test</a> shows why the logs matter: an AI auditor checking the output of AI &#8220;scientist&#8221; systems did barely better than chance from the manuscript alone (54% accuracy) but reached 82% with the logs and code. The new venues are already being built. <a href="https://aixiv.science">aiXiv</a> is a preprint archive for research by human and AI scientists, with AI-assisted review; <a href="https://arxiv.org/abs/2503.18102">AgentRxiv</a>, a research prototype, lets AI agents post their findings and build on one another&#8217;s work. Prestige will move from a journal&#8217;s name to whether the results hold.</p><p>Agents will write and check, but a named person will still have to sign off. As machines perfect the technical links, <a href="https://www.linkedin.com/pulse/scaffolding-wisdom-human-architecture-judgment-discernment-aziz-gozic">Naveed Aziz argues</a>, human judgment becomes the link on which everything depends. The physicist Matthew Schwartz found the same when he let Claude loose on problems outside his field: its results were almost always technically correct, yet not all that interesting until an expert in that field steered the work (<a href="https://www.anthropic.com/research/claude-shaped-science">Schwartz, 2026</a>). That step, from a hypothesis to true knowledge, still needs scientists with a real grasp of the topic. It is also the slowest link. As papers multiply, those scientists will need agents of their own to check results and judge which ones matter. Research is becoming a joint human&#8211;agent effort at every stage: refining ideas, running analyses, writing the paper and reviewing it.</p><h2>Hypotheses and knowledge</h2><p>Machines can already take the first half of that step. They produce hypotheses faster than we can test them: <a href="https://ai.google/static/documents/AI-in-Science.pdf">in one survey of 637 scientists</a>, 41% said their backlog of untested hypotheses had grown, which the report links to AI. Most hypotheses begin as imaginative guesswork, as the immunologist <a href="https://doi.org/10.1371/journal.pbio.3003912">Peter Medawar argued</a>. A hypothesis becomes knowledge only by passing the test we set for it. A thousand years ago, <a href="https://www.harvardmagazine.com/2003/09/ibn-al-haytham-html">Ibn al-Haytham wrote</a> that the seeker after truth is not the one who trusts the writings of the ancients, &#8220;but rather the one who suspects his faith in them and questions what he gathers from them.&#8221; We, as scientists, should treat AI&#8217;s answers the same way.</p><p>In September, <a href="https://casp.ac/reports/intelligence-explosion">a group including Geoffrey Hinton and Yoshua Bengio</a> warned that AI may soon automate AI research itself, compressing years of progress into months or less. They also conceded that fields such as biology may have to rely more on slower, costlier feedback from the real world. The futurist <a href="https://www.noahpinion.blog/p/wheres-the-intelligence-explosion">Ramez Naam</a>, skeptical of such forecasts, draws the line more sharply: AI will become superhuman mainly where a machine can verify the answer, as in formal mathematics. On 6 October, <a href="https://www.scientificamerican.com/article/openai-unleashes-hundreds-more-math-results-upon-a-field-already-in-shock/">OpenAI released 372 results</a>, many checked by proof software, though mathematicians may need months to understand them. <a href="https://docs.google.com/document/d/e/2PACX-1vTSh-pyNP3Gi99WMmsinnLmE9V5CDI0HEm6WGbIPMNt3V5SGlClHF8-BetotKHOImrvQDSXmFdiOw8D/pub">Christian Szegedy</a> likens such systems to the Wright brothers&#8217; first plane: crude, yet already landing on rocks no climber had reached.</p><p>Nutritional epidemiology sits firmly on the slow side. A model can suggest that a sweetener might change appetite. Knowing whether it does takes an interventional trial or an observational study in real people. <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC2642865/">Ibn Sina</a> made the same demand of medicines a millennium ago: test them on the human body, because a drug might act differently in &#8220;the lion&#8217;s body or the horse&#8217;s body.&#8221; AI will multiply the hypotheses in nutritional science and similarly in clinical sciences; but trials in people will still set the pace.</p><h2>What survives</h2><p>Those trials need people who can run and judge them. AI can help train those people, as a well-built tutor does, but it also means fewer of them are needed. At US firms that adopted generative AI, <a href="https://doi.org/10.2139/ssrn.5425555">junior employment fell about 9%</a> relative to senior staff within six quarters. Even Schwartz <a href="https://www.anthropic.com/research/claude-shaped-science">admits he still does not know how to train graduate students</a>. Universities must keep some junior work for the young on purpose, or no one will be left who can check the machines.</p><p>So I expect the university to shrink toward what cannot be digitized: laboratories, clinics, cohorts, and mentors who model judgment. A machine can produce answers, but it does not ask what needs be tested in real world. Curiosity and the desire to improve lives are ours, and apprenticeship under human mentors is how the young will learn this craft. The old scholars trusted a text because of who taught it. We will trust a result because of who tested it, and universities must keep training those people.</p><p><em>Tauseef Khan is a Senior Research Associate in the Department of Nutritional Sciences at the University of Toronto. He studies dietary sugars, honey, sweeteners and chronic disease, and has recently been building AI workflows for systematic reviews and writing about how AI is changing science. The views here are his own.</em></p><div><hr></div><h2>Sources</h2><ul><li><p>Aziz N. The scaffolding of wisdom and the human architecture of judgment: discernment in the age of AI. LinkedIn, 25 Jan 2026.</p></li><li><p>Bastani H, Bastani O, Sungu A, et al.&nbsp;Generative AI without guardrails can harm learning: evidence from high school mathematics. <em>PNAS</em> 2025;122:e2422633122.</p></li><li><p>Bloom JM. <em>Paper Before Print: The History and Impact of Paper in the Islamic World.</em> Yale University Press, 2001.</p></li><li><p>Boboris K. Fair moderation, equitable access, and AI: arXiv&#8217;s updated rate limit policy. arXiv blog, 1 Oct 2026.</p></li><li><p>Bryan KA. Eight rules for teaching in AI world. kevinbryanecon.com, Aug 2026.</p></li><li><p>Chan A, Winter C, Barto A, et al.&nbsp;What if automating AI R&amp;D triggers an intelligence explosion? CASP Frontier AI Working Paper 2/2026.</p></li><li><p>Codreanu M, Imas A, Mateos-Garcia J, et al.&nbsp;<em>AI in Science: Early Insights.</em> Google, 2026.</p></li><li><p>Hosseini SM, Lichtinger G. Generative AI as seniority-biased technological change. SSRN working paper, 2025. doi:10.2139/ssrn.5425555.</p></li><li><p>Howlett J. OpenAI unleashes hundreds more math results upon a field already in shock. <em>Scientific American</em>, 6 Oct 2026.</p></li><li><p>Ibn Abi Usaybi&#703;a. <em>&#703;Uyun al-anba&#702; fi tabaqat al-atibba&#702;</em> (13th c.), quoting Thabit ibn Sinan on the 931 examination. English trans. L. Kopf.</p></li><li><p>Kestin G, Miller K, Klales A, et al.&nbsp;AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. <em>Sci Rep</em> 2025;15:17458.</p></li><li><p>Liu J, Pei J, Huang J, et al.&nbsp;The last human-written paper: agent-native research artifacts. arXiv:2604.24658, 2026.</p></li><li><p>Luo Z, Kasirzadeh A, Shah NB. The more you automate, the less you see: hidden pitfalls of autonomous AI scientists. <em>PNAS</em> 2026;123(41):e2610214123.</p></li><li><p>Miao J, Davis JR, Zhang Y, Pritchard JK, Zou J. Reimagining research papers as interactive and reliable AI agents. <em>Nature</em> 2026. doi:10.1038/s41586-026-11044-y.</p></li><li><p>Naam R. Where&#8217;s the &#8220;intelligence explosion&#8221;? <em>Noahpinion</em>, 27 Sep 2026.</p></li><li><p>Nasser M, Tibi A, Savage-Smith E. Ibn Sina&#8217;s Canon of Medicine: 11th century rules for assessing the effects of drugs. <em>J R Soc Med</em> 2009;102:78&#8211;80.</p></li><li><p>Sabra AI. Ibn al-Haytham. <em>Harvard Magazine</em>, Sep&#8211;Oct 2003 (on the <em>Aporias against Ptolemy</em>).</p></li><li><p>Schmidgall S, Moor M. AgentRxiv: towards collaborative autonomous research. arXiv:2503.18102, 2025.</p></li><li><p>Schwartz MD. Claude-shaped science. Anthropic research blog (guest post), 1 Oct 2026. The author was a visiting researcher at Anthropic during the project.</p></li><li><p>Stewart DJ. Ej&#257;za. <em>Encyclopaedia Iranica</em> VIII/3, 1998.</p></li><li><p>Suchak T, et al.&nbsp;Explosion of formulaic research articles, including inappropriate study designs and false discoveries, based on the NHANES US national health database. <em>PLoS Biol</em> 2025;23:e3003152.</p></li><li><p>Szegedy C. Quo vadis, mathematics? Is mathematics over, or just graduating? Essay, 3 Oct 2026.</p></li><li><p>Tregoning JS. Is the scientific paper still a fraud? <em>PLoS Biol</em> 2026;24(8):e3003912 (on Medawar, &#8220;Is the scientific paper a fraud?&#8221;, <em>The Listener</em>, 1963).</p></li><li><p>Zhang P, Hu X, Huang G, et al.&nbsp;aiXiv: a next-generation open access ecosystem for scientific discovery generated by AI scientists. arXiv:2508.15126, 2025 (platform: aixiv.science).</p></li></ul>]]></content:encoded></item><item><title><![CDATA[The agentic unslopper: how I stopped my AI grading its own homework]]></title><description><![CDATA[An AI that judges its own writing will usually like it. The fix is a loop that learns from every failure and leaves the final call to a human.]]></description><link>https://blog.tauseefkhan.ai/p/the-agentic-unslopper-how-i-stopped</link><guid isPermaLink="false">https://blog.tauseefkhan.ai/p/the-agentic-unslopper-how-i-stopped</guid><dc:creator><![CDATA[Tauseef Ahmad Khan]]></dc:creator><pubDate>Tue, 29 Sep 2026 18:06:17 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!vFeD!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2ea9c807-4df3-4223-a370-ea3764a46cac_1456x816.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!vFeD!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2ea9c807-4df3-4223-a370-ea3764a46cac_1456x816.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!vFeD!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2ea9c807-4df3-4223-a370-ea3764a46cac_1456x816.png 424w, https://substackcdn.com/image/fetch/$s_!vFeD!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2ea9c807-4df3-4223-a370-ea3764a46cac_1456x816.png 848w, https://substackcdn.com/image/fetch/$s_!vFeD!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2ea9c807-4df3-4223-a370-ea3764a46cac_1456x816.png 1272w, https://substackcdn.com/image/fetch/$s_!vFeD!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2ea9c807-4df3-4223-a370-ea3764a46cac_1456x816.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!vFeD!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2ea9c807-4df3-4223-a370-ea3764a46cac_1456x816.png" width="1456" height="816" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/2ea9c807-4df3-4223-a370-ea3764a46cac_1456x816.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:816,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:164074,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://tauseefahmadkhan.substack.com/i/218052344?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2ea9c807-4df3-4223-a370-ea3764a46cac_1456x816.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!vFeD!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2ea9c807-4df3-4223-a370-ea3764a46cac_1456x816.png 424w, https://substackcdn.com/image/fetch/$s_!vFeD!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2ea9c807-4df3-4223-a370-ea3764a46cac_1456x816.png 848w, https://substackcdn.com/image/fetch/$s_!vFeD!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2ea9c807-4df3-4223-a370-ea3764a46cac_1456x816.png 1272w, https://substackcdn.com/image/fetch/$s_!vFeD!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2ea9c807-4df3-4223-a370-ea3764a46cac_1456x816.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>I wrote AI slop again this morning. An agent drafted, I skimmed, and the text read smoothly enough to pass. Only on a second reading did I see what it was: tidy paragraphs that said less than they seemed to, with the same turns of phrase every few lines. It looked like slop, read like slop and smelled like slop. It belonged in the bin.</p><p>This keeps happening to researchers who use AI seriously, and this week the reason became clearer to me. A team at Meta (Jason Weston, Ilia Kulikov and Swarnadeep Saha) tested whether AI models can tell expert writing from model writing. They gave a model a published paper with one section removed, asked it to write the missing section, and then asked AI models, including the writers themselves, which version was better. The AI judges preferred the model&#8217;s section over the human original in 63.5% to 84.6% of comparisons. When the judges scored against rubrics that a model had written, they preferred the model every time. The work is a <a href="https://facebookresearch.github.io/RAM/blogs/unslop/">blog post</a> and has not yet been peer-reviewed.</p><p>So the most common quality check in agentic writing, asking the AI whether the draft is good, is biased towards yes. The AI is grading its own homework, and it likes its own handwriting.</p><h2>Two kinds of slop</h2><p>Prose slop is the kind of writing everyone notices. In one of my own writeups, an agent&#8217;s draft used the construction &#8220;X, not Y&#8221; nine times. Add a hedge after every claim, lists of three, signposting, and sections that retell the whole paper, and a reader stops trusting the text within a page. Readers can tell it was written by AI, and they switch off. Length is the trigger. Ask a model to improve a sentence or a paragraph and it usually does it very well. Ask it to write a whole essay or paper in one go, and the result is almost always slop.</p><p>The second kind of slop is more dangerous because it looks like care. A draft article of mine said that a study&#8217;s author &#8220;cautions against reading it as proof of harm&#8221;. It sounded balanced. The paper says the opposite: its findings &#8220;provide compelling evidence for the cognitive impacts of AI tool usage&#8221;. A fabricated caveat reads as diligence, so it survives self-review by the AI agent. My own work supplies other examples. In one citation, &#8220;considers&#8221; drifted to &#8220;accounts for&#8221;, turning a design judgment into a statistical adjustment the source never described or intended. A PDF converter turned 0.00247 into 00247. When reading an image, the vision model counted 13 marks on a figure in one run and 19 in the next.</p><p>Asking the model whether its work was good would have caught none of these. Each was caught by checking against something outside the model: the source text, a second extraction engine, a pixel count and my own eyes.</p><h2>A loop that learns</h2><p>I use Claude Code, Codex and several open models to help with my systematic reviews, meta-analyses and papers. Since the summer I have stopped hunting for the perfect prompt and built a loop instead. I call it the agentic unslopper. It has four parts.</p><p>First, every failure becomes a lesson with a tripwire. When I catch an agent&#8217;s mistake, the lesson goes into a shared memory, a folder of short notes that every agent reads, and it must name the check that would have caught it: a search pattern, an audit script or a gate in the workflow. The invented caveat produced such a rule: any sentence saying a source &#8220;cautions&#8221;, &#8220;notes&#8221; or &#8220;warns&#8221; must be found in the source before the draft is saved. The check matters more than the note. Agents forget written advice; a check runs every time.</p><p>Second, numbers never live in the model&#8217;s memory. Every value that is extracted, derived, calculated or estimated is written to disk the moment it exists, and later steps read it from the file. This is a hard rule in the instruction file that all my agents load. Data extraction for a meta-analysis runs through two AI engines plus a visual read of the page: one engine extracts, the second verifies, and a third model re-checks the tables. Any disagreement is resolved against the original PDF. I chose this set after comparing nine extraction engines on my own papers.</p><p>Third, my text is the default. When I edit a draft and an agent reviews my edits, it may flag only what can be verified: a changed number, a certainty claim the study design cannot carry, a citation or a quotation. It may not nudge my wording back towards its own. I learnt this the hard way: I once asked an agent to check and fix the citations in a manuscript, and it also &#8216;fixed&#8217; some of my prose, changing the meaning of what I wanted to convey. I made it a rule this week, after reading the Meta study. A second rule came with it: judge a draft against an expert example of the same kind, and list what is missing instead of giving a score.</p><p>Fourth, the loop feeds itself. Each workflow reads its own lessons at the start of every run, so a mistake made on Monday shapes Tuesday&#8217;s work. This is the recursive part, and the loop itself needs checking. A recent audit found 61 saved lessons about specific workflows, but only 5 in total had been wired into the workflow they concerned. The rest were orphan lessons that no workflow was reading. The agents were learning and then not using what they had learnt. A learning system decays quietly unless you audit the learning as well.</p><h2>What it cost, and what it bought</h2><p>The loop is slower than accepting the first draft. While preparing a perspective on AI in nutrition research for submission, three agents audited every cited claim in parallel, and each source was then checked against the original document. Of 31 cited claims, 24 were fully supported and 7 only partly. We made 13 corrections, most of them to wording that claimed a little more than the source said. None was dramatic; together they were the difference between a paper that is right and one that sounds right.</p><p>Unslopping can also overshoot. When I told an agent that a document I had written with AI&#8217;s help read like AI slop, it cut the text from 2,647 to 1,913 words and removed six of the nine &#8220;X, not Y&#8221; constructions. The result was disjointed and jarring. We restored the transitions and the useful signposts, and settled at 2,037 words. That lesson went into memory too: aim to cut 20&#8211;25%, keep the sentences that carry the argument from one point to the next, and stop there.</p><h2>Where the human stays</h2><p>In July I told my research group that writing is the main choke point in our research projects. AI can write methods and results in our standard laboratory style; the introduction and discussion are where we must put in our own effort. I still think so. A recent experiment by <a href="https://arxiv.org/abs/2607.27191">Kirgis and colleagues</a> reaches the same place from another direction: AI agents completed all the engineering for two research projects without human help, but the projects&#8217; original authors rejected both resulting papers.</p><p>So this is how I now write with AI. I ask the agent for an outline first. I review it, edit it and rewrite the points in my own words, then ask the agent to review it again. When typing is too slow, I switch on the microphone in Claude Code or Codex and talk through what I want to add, and the agent turns it into an outline. Only when I approve the outline do we write prose, one paragraph at a time, and I read and edit every sentence before moving on. I approve each paragraph separately in sequence. In my experience AI handles long-form writing badly: it loses the thread of the argument, and that is where slop creeps in. Reviewing and editing each line oneself is the key. The human mind drives the writing; the AI still remains just a tool that can write.</p><p>The unslopper does not take me out of the loop. My judgment is the gate: the agents stop there and ask for permission, and nothing moves forward until I decide. Because the workflows improve themselves, the output gets better with each round and my decisions get easier. The Meta team trained its model against rubrics built from expert writing. In my system the expert is me, and the rubric is every correction I have made. That is also why I cannot hand all my thinking to the agents. A judge who stops reading goes stale, and then the homework grades itself again.</p>]]></content:encoded></item></channel></rss>