WEBVTT Kind: captions Language: en NOTE Module 6: Seminar narrated walkthrough. Generated from the narration script and per-segment audio durations (0.7 s gap per slide). s01-001 00:00:00.000 --> 00:00:03.462 Welcome to Module 6 of Claude Code for Accounting Research. s01-002 00:00:03.512 --> 00:00:06.533 This one is a seminar on AI, the production of knowledge, s01-003 00:00:06.583 --> 00:00:09.096 and the academic research market. s01-004 00:00:09.146 --> 00:00:10.563 There's no lab and nothing to install, s01-005 00:00:10.613 --> 00:00:13.230 so this video walks through the framing slides. s02-006 00:00:13.980 --> 00:00:15.973 So start with what you built. s02-007 00:00:16.023 --> 00:00:20.645 In Module 2 you built an R&D intensity panel from the SEC's Financial s02-008 00:00:20.695 --> 00:00:23.502 Statement Data Sets, in an eighty-minute lab. s02-009 00:00:23.552 --> 00:00:27.435 Around 2019, that was about a week of an RA's time. s02-010 00:00:27.485 --> 00:00:31.736 Module 3, the EDGAR risk-factor pipeline and a reusable skill, s02-011 00:00:31.786 --> 00:00:35.924 was also an eighty-minute lab, against two to three weeks for an RA. s02-012 00:00:35.974 --> 00:00:39.451 Module 4, a filtered, verified Compustat pull, took the same eighty minutes, s02-013 00:00:39.501 --> 00:00:42.714 against several days. s02-014 00:00:42.764 --> 00:00:46.313 Module 5, a hardened repo with a paper trail and git, took about three hours, s02-015 00:00:46.363 --> 00:00:49.762 against days of audit work. s02-016 00:00:49.812 --> 00:00:52.616 That speed-up raises a question that isn't technical, s02-017 00:00:52.666 --> 00:00:54.810 and this module is about that question. s03-018 00:00:55.560 --> 00:01:00.662 In Module 1 you saw the research pipeline: idea, data, measure, analysis, paper. s03-019 00:01:00.712 --> 00:01:02.181 Here it's a harder map. s03-020 00:01:02.231 --> 00:01:07.295 Each stage is plotted by how automatable it is and how much of the paper's value s03-021 00:01:07.345 --> 00:01:08.003 it carries. s03-022 00:01:08.053 --> 00:01:11.839 Most of this course worked in the upper left, s03-023 00:01:11.889 --> 00:01:16.016 execution work an agent does faster than you and about as well. s03-024 00:01:16.066 --> 00:01:19.338 The upper right is identification, contribution, and interpretation, s03-025 00:01:19.388 --> 00:01:22.467 where a paper succeeds or fails. s03-026 00:01:22.517 --> 00:01:23.395 The line between the two moves. s03-027 00:01:23.445 --> 00:01:27.553 Measurement construction, which you worked on in Module 3, s03-028 00:01:27.603 --> 00:01:30.520 has been moving toward the automatable side. s03-029 00:01:30.570 --> 00:01:33.956 So, on your own, place your dissertation's stages on this map, s03-030 00:01:34.006 --> 00:01:36.119 and don't discuss it yet. s03-031 00:01:36.169 --> 00:01:39.430 Block 1 opens with two volunteers defending their placement. s04-032 00:01:40.180 --> 00:01:42.153 A paper is an artifact. s04-033 00:01:42.203 --> 00:01:45.758 Once it exists, it's transferable and cheap to copy. s04-034 00:01:45.808 --> 00:01:50.086 Your understanding is capital, and you only build it by doing the work. s04-035 00:01:50.136 --> 00:01:52.697 For most of the profession's history the two came bundled, s04-036 00:01:52.747 --> 00:01:55.995 so the paper was good evidence about the researcher. s04-037 00:01:56.045 --> 00:01:58.552 Agentic AI unbundles them. s04-038 00:01:58.602 --> 00:01:59.658 The evidence here isn't reassuring. s04-039 00:01:59.708 --> 00:02:04.463 Macnamara and coauthors argue that AI takes over the sub-tasks where expertise s04-040 00:02:04.513 --> 00:02:09.317 gets built, and that skills can decay without the person noticing. s04-041 00:02:09.367 --> 00:02:13.396 Gerlich's survey evidence finds that higher educational attainment weakens that s04-042 00:02:13.446 --> 00:02:15.273 effect but doesn't eliminate it. s04-043 00:02:15.323 --> 00:02:18.202 The personal audit is the test from Module 5: s04-044 00:02:18.252 --> 00:02:20.850 can you discuss the finding without the LLM? s05-045 00:02:21.600 --> 00:02:25.407 Block 2 is about provenance: who or what produced a result, s05-046 00:02:25.457 --> 00:02:27.883 and who's accountable for it. s05-047 00:02:27.933 --> 00:02:32.245 The quote on the slide is from a 1979 IBM training manual... s05-048 00:02:32.295 --> 00:02:33.904 a computer can never be held accountable. s05-049 00:02:33.954 --> 00:02:36.519 In 2023, Nature's editors set two rules. s05-050 00:02:36.569 --> 00:02:38.627 No language model gets credited as an author, s05-051 00:02:38.677 --> 00:02:43.644 because authorship carries accountability a model can't take on. s05-052 00:02:43.694 --> 00:02:46.782 And any use of these tools has to be documented. s05-053 00:02:46.832 --> 00:02:52.479 The second rule is where the profession is putting its practical effort now: s05-054 00:02:52.529 --> 00:02:56.096 disclosure policies, audit trails, and verification. s05-055 00:02:56.146 --> 00:02:59.790 You built that kind of documentation in Module 5, with the paper trail. s06-056 00:03:00.540 --> 00:03:04.424 This slide is a case study, and it's labeled that way on purpose. s06-057 00:03:04.474 --> 00:03:07.409 We don't cite this paper as evidence for anything. s06-058 00:03:07.459 --> 00:03:11.878 In 2024 a working paper claimed that a randomized rollout of an AI s06-059 00:03:11.928 --> 00:03:16.570 materials-discovery tool raised researcher output by 44%. s06-060 00:03:16.620 --> 00:03:18.673 It circulated to acclaim from prominent economists, s06-061 00:03:18.723 --> 00:03:22.116 and then it was found to be fabricated. s06-062 00:03:22.166 --> 00:03:25.877 The paper was withdrawn, and the author was severed from MIT. s06-063 00:03:25.927 --> 00:03:30.965 So the field's provenance-checking machinery failed before AI was writing papers. s06-064 00:03:31.015 --> 00:03:34.718 The open question for Block 2 is whether AI makes that machinery stronger, s06-065 00:03:34.768 --> 00:03:39.135 through audit trails and replication at near-zero cost, or weaker, s06-066 00:03:39.185 --> 00:03:41.450 through volume and plausibility at scale. s07-067 00:03:42.200 --> 00:03:44.374 Block 3 is about the market. s07-068 00:03:44.424 --> 00:03:48.672 Placement, tenure, prestige, and salary all clear through publications, s07-069 00:03:48.722 --> 00:03:53.577 so papers are the currency, and they've always been expensive to produce. s07-070 00:03:53.627 --> 00:03:56.589 The pre-AI baseline was already strained. s07-071 00:03:56.639 --> 00:03:58.817 Bloom, Jones, Van Reenen and Webb, 2020, s07-072 00:03:58.867 --> 00:04:02.923 document that ideas are getting harder to find: s07-073 00:04:02.973 --> 00:04:07.176 research productivity has fallen even as research effort has grown. s07-074 00:04:07.226 --> 00:04:08.369 Lusher, Yang and Carrell, 2023, s07-075 00:04:08.419 --> 00:04:13.112 find that when more working papers come out in the same week, each one does worse, s07-076 00:04:13.162 --> 00:04:16.341 and not just the weak ones. s07-077 00:04:16.391 --> 00:04:18.207 The figure follows that logic. s07-078 00:04:18.257 --> 00:04:21.145 Production cost falls, paper volume goes up, s07-079 00:04:21.195 --> 00:04:24.462 and referee time and reader attention don't scale with it. s07-080 00:04:24.512 --> 00:04:27.995 So the market reprices what attention still depends on: taste in questions, s07-081 00:04:28.045 --> 00:04:32.150 data access, credibility, and accountability. s08-082 00:04:32.900 --> 00:04:35.603 Here's how the discussion runs in the live session. s08-083 00:04:35.653 --> 00:04:40.087 Three blocks of about thirty minutes each, with a five-minute stretch between them. s08-084 00:04:40.137 --> 00:04:44.937 Each block starts with two-minute position statements from the debaters assigned s08-085 00:04:44.987 --> 00:04:48.554 in Module 5, then open discussion, then I synthesize. s08-086 00:04:48.604 --> 00:04:52.461 Several of you will be defending positions you don't actually hold, s08-087 00:04:52.511 --> 00:04:53.974 and that's part of the exercise. s08-088 00:04:54.024 --> 00:04:57.784 The ground rule is no catastrophizing and no dismissing. s08-089 00:04:57.834 --> 00:05:01.394 If you raise a provocation, pair it with its counterargument. s08-090 00:05:01.444 --> 00:05:03.970 And the memo is due one week after the seminar. s09-091 00:05:04.720 --> 00:05:10.112 Block 1 asks: if AI produces the artifact, what happens to the learning? s09-092 00:05:10.162 --> 00:05:14.435 We start from the placement you just did, with two volunteers defending theirs. s09-093 00:05:14.485 --> 00:05:19.222 Then the question is which pipeline stages build intuition worth protecting. s09-094 00:05:19.272 --> 00:05:21.386 The accounting parallel is AI-assisted audit judgment. s09-095 00:05:21.436 --> 00:05:22.850 There's also the institutional version. s09-096 00:05:22.900 --> 00:05:29.006 If researchers no longer have to build the capital to produce the artifact, s09-097 00:05:29.056 --> 00:05:31.023 does the profession keep building it on purpose, s09-098 00:05:31.073 --> 00:05:35.235 the way we still teach proofs to students who'll never prove anything new? s09-099 00:05:35.285 --> 00:05:39.708 The block closes on this question: is PhD training a production process, s09-100 00:05:39.758 --> 00:05:41.790 or a transformation of the student? s10-101 00:05:42.540 --> 00:05:47.166 Block 2 asks whether it matters who or what produced a result. s10-102 00:05:47.216 --> 00:05:51.068 Take two identical extraction results from the Module 3 pipeline. s10-103 00:05:51.118 --> 00:05:53.962 An RA hand-coded one over three weeks. s10-104 00:05:54.012 --> 00:05:57.233 The other came from a Claude session with a full audit trail: cached inputs, s10-105 00:05:57.283 --> 00:06:02.053 a quality report, committed code, and a stratified human audit sample. s10-106 00:06:02.103 --> 00:06:03.964 Are they equivalent? s10-107 00:06:04.014 --> 00:06:07.481 A spotless audit trail can still deliver a wrong estimate. s10-108 00:06:07.531 --> 00:06:08.595 Module 5 had that example, s10-109 00:06:08.645 --> 00:06:14.129 where DECISIONS.md claimed a robustness check the commit history didn't contain. s10-110 00:06:14.179 --> 00:06:18.660 A ten-minute segment then asks what's being verified when a second model reviews s10-111 00:06:18.710 --> 00:06:20.002 the first model's code, s10-112 00:06:20.052 --> 00:06:24.331 since correlated training plausibly produces correlated blind spots. s10-113 00:06:24.381 --> 00:06:28.410 The block closes on the retraction question: stronger or weaker? s11-114 00:06:29.160 --> 00:06:31.862 Block 3 is returns to research. s11-115 00:06:31.912 --> 00:06:35.440 You built the Module 4 pipeline in an eighty-minute lab. s11-116 00:06:35.490 --> 00:06:37.365 Now suppose every accounting PhD student can. s11-117 00:06:37.415 --> 00:06:42.311 What does the JAR submission queue look like in five years? Who referees it? s11-118 00:06:42.361 --> 00:06:45.244 What happens to the signal value of a publication? s11-119 00:06:45.294 --> 00:06:48.746 The candidates for what stays scarce are question taste, s11-120 00:06:48.796 --> 00:06:53.311 data access and institutional knowledge, credibility, and accountability. s11-121 00:06:53.361 --> 00:06:54.271 Then there's homogenization. s11-122 00:06:54.321 --> 00:06:58.752 If everyone's referee reports pass through the same models, s11-123 00:06:58.802 --> 00:07:01.519 does voice still signal anything? s11-124 00:07:01.569 --> 00:07:05.793 The block closes with each of you naming one input you think stays scarce, s11-125 00:07:05.843 --> 00:07:08.070 and that answer is where your memo starts. s12-126 00:07:08.820 --> 00:07:13.103 The memo is titled What I will still be paid for in 2036. s12-127 00:07:13.153 --> 00:07:15.946 It's one page, due a week after the seminar. s12-128 00:07:15.996 --> 00:07:19.759 Commit to specific parts of the research production function that you think stay s12-129 00:07:19.809 --> 00:07:20.678 human. s12-130 00:07:20.728 --> 00:07:24.981 Defend that with at least one assigned reading and one experience from the build s12-131 00:07:25.031 --> 00:07:25.467 modules. s12-132 00:07:25.517 --> 00:07:29.964 And name one concrete investment you'll make in your remaining PhD years. s12-133 00:07:30.014 --> 00:07:33.695 Grading is credit or no credit on defended specificity. s12-134 00:07:33.745 --> 00:07:38.312 An optimist and a pessimist can both earn credit, and a vague memo can't. s12-135 00:07:38.362 --> 00:07:42.224 AI help is fine everywhere in this course except the argument of this memo, s12-136 00:07:42.274 --> 00:07:46.611 which has to be your own, and you disclose what you used in one line. s12-137 00:07:46.661 --> 00:07:47.810 See you at the seminar.