Three technology shifts, three kinds of mess, and the question of who gets the bill
In July 1999, Hershey switched on a new order-management system six weeks before its biggest selling season. By the end of the quarter, sales were down 12 percent — not because the chocolate hadn't been made, but because the company could no longer reliably tell a truck where to take it. Hershey's own annual report booked $112 million of spend on the system and attributed the shortfall to "difficulties in order fulfillment... encountered since the start-up of a new integrated information system." [1]
In June 2026, the music streaming service Deezer reported that more than half the music uploaded to its platform each day — roughly 90,000 tracks — was fully AI-generated. Those tracks draw 1 to 3 percent of streams, and up to 85 percent of even that sliver was fraudulent. [2]
Both are messes. They are not the same kind of mess, and the difference is the entire story.
Every major technology shift produces slop. The interesting question was never how much. It's the cost structure underneath: what does it cost to make the mess, what does it cost to clean it, and — the part almost nobody prices in advance — who actually pays, and whether anyone writes it down. Underneath all three sits one variable: AI has driven the cost of producing plausible output toward zero while the cost of verifying it has barely moved. Everything else follows from that gap.
The mess that had an owner — ERP, 1990–2010
The ERP era ran on expensive mistakes. The average implementation in Panorama Consulting's 2008 survey of 1,322 organizations cost $8.5 million; 65 percent went over budget, 93 percent ran long, and 57 percent reported major operational disruption after go-live. Only about a fifth realized half the benefits they'd been promised. [3] In the manufacturing and distribution cohort of the same study, projects averaged 19.2 months, and more than two-thirds of the spend went to technical implementation rather than to the software itself. [4]
The failures were spectacular because they were load-bearing. FoxMeyer, the fourth-largest pharmaceutical distributor in the United States, replaced a mainframe that processed 420,000 orders a night with an SAP R/3 installation that managed 10,000. It lost $34 million of inventory and filed for bankruptcy in 1996. [5] Nike's supply chain project cost more than $100 million in lost sales [6]; Philip Knight's public comment — "I guess my immediate reaction is, 'this is what I get for our $400 million, huh?'" — is still the most honest sentence any CEO has said about enterprise software. [7] National Grid's post-go-live financial close stretched from four days to forty-three, stabilization ran roughly $30 million a month, and Wipro eventually paid $75 million to settle. [8]
Every one of those messes had an owner. A budget line. A signed statement of work. A named integrator. A general counsel who could calculate whether litigation was cheaper than absorption.
Concentrated pain creates a market. When one organization is bleeding $30 million a month, someone will sell them a tourniquet. So the ERP era grew a remediation economy on top of itself, and it was enormous. In 1998, ERP vendors booked $16.6 billion in software revenue; AMR Research put total ERP-related spend that year — services, hardware, databases and networking together — at roughly $70 billion. [9] Rescue became a product category. Panorama, Ultra, Pemeco, Lumenia and others still sell "ERP project recovery" as a named service line, a quarter-century later. [10]
The diagnosis was consistent, and it was never the software. Only 18 percent of projects in that manufacturing cohort deployed anything close to standard configuration; 35 percent customized heavily. [4] Thomas Davenport had already named the failure mode in Harvard Business Review in 1998: "An enterprise system, by its nature, imposes its own logic on a company's strategy, organization, and culture." [11] Twenty years later, after Lidl abandoned a reported €500 million project because SAP tracked inventory at retail price and Lidl insisted on purchase price, the German SAP user group restated the same finding in one line: "If a company wants to use the standard software, it has to adapt its own processes." [12]
The mess the market priced — dot-com, 1995–2003
The dot-com era inverted the economics. U.S. venture capital deployed $105 billion in 2000 alone — more than the four preceding years combined. [13] There were 370 technology IPOs in 1999, then 24 in 2001. [14] Webmergers counted 225 internet shutdowns in 2000, 537 in 2001, and at least 962 substantial companies through early 2003. [15]
But the cleanup mechanism here was price, not consulting. Global Crossing emerged from Chapter 11 with a total reorganization value of $407 million, against a peak market capitalization above $47 billion. ST Telemedia paid $250 million for 61.5 percent; creditors took the remaining 38.5 percent plus roughly $523 million in cash; existing shareholders were wiped out entirely. [16] marchFIRST, formed by a $5.7 billion all-stock merger closed the day after the Nasdaq's peak week, liquidated into $120 million of assets thirteen months later. [17] Razorfish, which billed $170 million in 1999, sold for $8.2 million. [18] Boo.com burned $135 million and sold its technology for $250,000. [19] The market wrote everything down to what it was worth and recycled the parts. Divine, Inc., which bought marchFIRST's remains, went bankrupt itself in 2003 and was auctioned for $54 million. [20]
And the losses landed, in the main, on people who had bought the risk on purpose. Shareholders and limited partners took the hit; that is what equity is for. Not entirely — employees lost jobs and worthless options, creditors took haircuts, suppliers and landlords ate the remainder, and telecom bondholders recovered barely twenty cents on the dollar. But the mechanism was legible, and every loss had somewhere to land.
The one part of the dot-com cleanup that was genuinely mispriced is the part worth remembering. When the accounting failures surfaced — the GAO counted 919 restatements between 1997 and mid-2002, with annual counts up about 145 percent — Congress responded with Sarbanes-Oxley. [21] The SEC estimated Section 404 would cost about $91,000 per company per year. [22] Financial Executives International measured the first-year reality for large filers at $4.36 million. [23] Roughly fiftyfold — though the two are not strictly comparable, since FEI's respondents averaged $5 billion in revenue while the SEC's figure was an all-filer average. Discount it heavily and it is still a permanent reminder that the cleanup is always priced by people looking at the last crisis.
The messes with no record — AI, 2023–
An ERP disaster required a board approval, a signed contract, a systems integrator on site, and eighteen months. The budget was the gate. It was a terrible gate — it let $500 million projects through — but it was a gate, and it meant every mess arrived attached to somebody's name.
Generative AI removed the gate without replacing it. The floor cost of producing a plausible artifact went to roughly zero, and the volume statistics follow exactly as you'd expect. NewsGuard tracked 49 AI-generated news sites in May 2023 and 3,749 by June 2026. [24] Graphite's three-detector study of 55,400 sampled articles estimated the AI-generated share of new web articles at 49.9 percent in Q1 2026 — an estimate rather than a census, since detectors are not ground truth and the line between human and AI writing blurs badly in hybrid workflows. [25] Deezer, which tags at ingest rather than inferring after the fact, went from roughly 30,000 AI tracks a day in September 2025 to about 90,000 in mid-2026. [26]
Volume is the obvious part, and it's the least interesting. What matters is that this era generates two different kinds of mess at once — and only one of them is new.
The familiar kind
The first kind looks exactly like ERP, and it is larger than the "AI pullback" coverage suggests. Spending on AI software — the closest published proxy for what enterprises actually put into AI programs — is still climbing steeply: Gartner puts the 2026 worldwide AI software market at $453 billion, up 60 percent year over year. [49]
A caution about the number you have probably seen instead. Gartner's headline "$2.59 trillion of AI spending in 2026" is not enterprise spending. More than half of it is infrastructure, which Gartner defines to include AI-processing semiconductors and devices; in the most recent vintage with a published breakdown, AI-capable smartphones were the single largest line in the table, ahead of both AI services and AI application software. Gartner's own analyst says spending to date "has primarily been driven by technology companies and hyperscalers," and that enterprises "have yet to really flex their spending potential." [49] It is a capital-expenditure-and-handsets number wearing an enterprise-budget costume, and it is quoted constantly as the latter.
But inside the budgets that are genuinely enterprise, individual programs are dying at ERP-era rates. A 2026 cost-governance survey found 40 percent of organizations had escalated AI cost overruns to the board, 33 percent had imposed emergency spending freezes, and a quarter had delayed or cancelled an AI initiative outright. [50] S&P Global put the share of AI proofs-of-concept abandoned before production at 46 percent [51] — though that last figure deserves less weight than it usually gets. A proof of concept is supposed to be cheap to kill; abandoning one is portfolio discipline, not failure. The cost-governance numbers are the troubling half, because an emergency spending freeze and a board escalation are not what running an experiment looks like. The pullback is real, but it is churn inside expanding budgets rather than a top-line cut — which is exactly how ERP behaved too.
The operational failures rhyme just as closely. Klarna announced in February 2024 that its AI assistant was doing the work of 700 full-time agents. [52] By May 2025 its CEO was telling Bloomberg that cost had been "a too predominant evaluation factor," that the result was "lower quality," and that the company was recruiting human agents again [53] — though it kept the AI, which still handled about two-thirds of inquiries, and the popular "Klarna fired the bots and rehired everyone" version is wrong. In March 2026 Amazon suffered four highest-severity outages in a single week; internal documents reviewed by Business Insider put the losses at roughly 6.4 million orders, and internal briefing material attributed a pattern of incidents to "Gen-AI assisted changes" with "high blast radius." [54] Amazon publicly disputed that framing, stating that no AI-written code was involved and that only one incident touched AI tooling at all — an engineer acting on advice an AI agent had inferred from an outdated internal wiki. [55]
These are owned messes. They have budgets, executives, named vendors and shareholders. Structurally, they are the ERP era running again on schedule.
What's missing from them
Except for one thing, and it is the thing that should worry a board.
In 1999 Hershey put its disaster in an SEC filing, in specific dollars, under its own name. Waste Management sued its vendor for more than $100 million and did it in open court. Wipro paid $75 million and everyone could read the number.
Search the public record now for a major corporation that has disclosed a charge, write-down, or impairment attributable to a failed AI program. They are remarkably hard to find — while a quarter of companies say they are cancelling AI initiatives over cost and two-thirds report overruns.
The tempting read is concealment. It isn't. It is accounting, and the accounting turns out to be the more interesting answer.
Under US GAAP, exploratory software work is expensed as incurred; capitalization begins only once management has committed funding and completion is probable. A failing pilot never clears that gate, so it never becomes an asset, so there is nothing to impair. The largest AI cost lines — model API calls, cloud inference, subscriptions — have no capitalization pathway at all. IAS 38 lands in the same place, expensing all research-phase spend. [57] And FASB's 2025 amendment tightens it further, explicitly denying capitalization to software carrying "novel, unique, or unproven functions or features" whose uncertainty has not been resolved through coding and testing. [58] That sentence describes generative AI almost by definition.
This is the real structural break with ERP, and it is sharper than "nobody admits it." ERP projects were capitalized. Hershey's own 1999 filing discloses $98.8 million of capitalized software and hardware against $13.2 million expensed — roughly nine dollars onto the balance sheet for every one through the income statement. [1] That is why the ERP era generated impairments, restatements, disclosures and lawsuits: its failures had somewhere to land. Today's AI failures dissolve into operating expense, indistinguishable afterwards from the ordinary cost of doing business.
The exception proves the rule. Zillow's home-price model failed in 2021 and produced a $304 million inventory write-down with $240–265 million more to follow. [59] That loss was legible only because Zillow had used the algorithm to buy houses; the write-down was against the houses. The model was never an asset. At a company not buying houses, the identical failure leaves no mark at all.
An unexpensed loss is a loss all the same. What it is not is a loss anyone can price, budget against, or learn from — and that is why the enterprise category, despite having clear owners, is not yet producing the remediation market ERP produced by this point in its cycle. Rescue is a hard sell to a company whose failure never surfaced as a number.
The genuinely new kind
Beneath the owned messes sits a second category with no owner at all. Nothing in either previous era looks like it — though, as we will see, it is not without precedent.
Consider curl. It is not a company. It is a free, open-source tool for moving data across networks, started in 1998 by a Swedish developer named Daniel Stenberg, and by his own count it sits inside more than twenty billion devices — phones, cars, televisions, medical equipment, and most of the servers the modern internet runs on. [56] Its security team is seven people, several of whom have about three hours a week to give it. [27]
In 2025, roughly 20 percent of the security reports curl received were AI slop and about 5 percent were real vulnerabilities. Each report, real or fake, "engages 3-4 persons... Perhaps for 30 minutes, sometimes up to an hour or three." [27] curl ended its bug bounty on January 31, 2026 — not because it stopped caring about security, but to remove the incentive drawing the flood. [28]
This is not one exhausted maintainer. HackerOne, the largest commercial bug bounty platform, suspended new submissions to its Internet Bug Bounty in March 2026, with valid-submission rates reportedly down from around 15 percent to under 5. [29] Submissions to the competing platform Bugcrowd more than quadrupled over three weeks that same month, and Nextcloud paused its program. [30] Jazzband, a collective maintaining widely-used Python packages, shut down altogether, its lead maintainer citing unsustainable AI-generated spam. [31]
Every one of those is infrastructure that commercial software sits on top of, maintained by people no company pays for the privilege.
The scientific record is worse, because there's no one to quit. A May 2026 Lancet audit checked 97.1 million references across 2.47 million papers in PubMed Central's Open Access subset and found the rate of papers carrying fabricated citations rising from 1 in 2,828 in 2023 to 1 in 277 in early 2026 — more than tenfold, with the steepest climb from mid-2024. [32]
The authors are careful about what that shows, and the article should be too. The timing coincides with the mainstreaming of AI writing tools; the study measured prevalence, not provenance. It did not test whether AI produced the fabrications, and fabricated citations long predate language models. What it does establish beyond argument is the response: at the time of the audit, 98.4 percent of the 2,810 affected papers had seen no publisher action of any kind.
That 98.4 percent is the number to sit with. In the ERP era, an unfixed mess meant someone's shipments stopped. In this era, an unfixed mess just stays there, quietly, in the corpus everyone else builds on.
Inside companies the same dynamic operates below the level of any budget. BetterUp Labs and Stanford's Social Media Lab surveyed 1,150 U.S. desk workers in September 2025 and found 40 percent had received "workslop" — AI-generated work product that looks finished and isn't — in the previous month, costing about two hours per instance and $186 per affected employee per month. For a 10,000-person organization: roughly $9 million a year. [33] Nobody will ever see that number on a P&L. It doesn't have a cost center. It shows up as a reviewer being vaguely slower than last year.
The ERP era's messes had owners, and the owners wrote them down. This era's owned messes leave no record — and beneath them runs a whole category that has no owner to keep one. That is the actual asymmetry, and it is worse than a volume problem, because volume problems get solved by capacity and this one doesn't.
The diagnosis hasn't changed
There's a temptation to treat this as a model-quality issue that will resolve itself with the next release. The evidence says otherwise, and it says it in a form that should feel familiar to anyone who lived through the first era.
The application security firm Veracode has been testing generated code against known vulnerability classes across more than 150 models. Syntax correctness now runs above 95 percent. Security pass rates sit at about 55 percent — "virtually identical to where they stood two years ago." For cross-site scripting — one of the oldest and best-understood web vulnerabilities there is — generated code passes 15 percent of the time. [34] On this dimension the models have become dramatically better at looking right without becoming better at being right. That is a claim about generated code and application security, not about model capability in general, which has moved a great deal elsewhere. But it is the dimension that matters here, because a security defect is precisely the kind of error that survives review by looking finished.
Stack Overflow, the question-and-answer site most working programmers live on, found in its 2025 survey that the top developer frustration, at 66 percent, was "AI solutions that are almost right, but not quite." [35] That is the best working definition of slop anyone has produced, and it's the same failure the ERP era diagnosed and refused to act on: the system will happily give you an output that satisfies the interface and violates the intent. Davenport's line still holds, with one word changed. The tool imposes its own logic. You either reconcile your process with it or you inherit its errors at scale.
There's a final, humbling detail. For twenty years the ERP industry ran on the Standish Group's 1994 finding that projects averaged 189 percent cost overruns. [36] Peer-reviewed work by Jørgensen and Moløkken-Østvold [37], and later by Eveleens and Verhoef in IEEE Software [38], took that number apart: the definition was never stated, the sample selected for failure stories, and comparable studies of the period put the real figure near 30 percent. The related folk statistic — "70 percent of ERP projects fail" — has no traceable source at all.
So the previous era generated its own low-quality, widely-repeated, decision-shaping content too. It just did it slowly, by hand. As the writer Francesco D'Isa put it, society has always eaten and re-eaten its own slop; AI seems to recycle mediocrity faster than ever before. [39] The slop isn't new. The metabolism is.
The cleanup industry is forming — around liability
It is forming, and you can see exactly where.
Gartner sizes the AI governance platform market at $492 million in 2026, reaching past $1 billion by 2030. [40] The UK's Department for Science, Innovation and Technology counted 524 companies in the AI assurance market contributing about £1.01 billion of value in 2024, projecting £18.8 billion by 2035. [41] Capital is following the tooling that tests and monitors AI output: Braintrust raised $80 million at an $800 million valuation in February 2026 [42]; LangChain raised $125 million at $1.25 billion in October 2025 on estimated ARR in the low tens of millions. [43] That is not a normal multiple. That is the market pricing the expectation that verifying AI output becomes a permanent line item.
Accenture booked $2.7 billion of advanced AI revenue in FY2025 against $5.9 billion of bookings [44] — though note carefully that this is implementation revenue, not remediation. No major firm has yet named a "fix our failed AI deployment" practice with revenue attached, the way ERP rescue became a product. Give it eighteen months.
But look at where the money is going, because it tells you the boundary. Every one of those markets sits on top of enterprise liability. A British Columbia tribunal ruled in 2024 that Air Canada owned what its chatbot said, rejecting the argument that the bot was "a separate legal entity." [45] Damien Charlotin's database of court decisions that explicitly found AI-fabricated citations stood at 1,962 cases as of late August 2026. [46] Deloitte Australia refunded A$97,000 on a A$440,000 government report containing more than a dozen fabricated references. [47] Where slop touches a contract, a regulator, or a courtroom, it acquires an owner — and the moment it has an owner, a market appears to serve it.
Where it doesn't, nobody is coming. There is no assurance vendor for curl's maintainers. Nothing in the £18.8 billion touches the 98 percent of contaminated papers that publishers haven't acted on. Wikipedia's response was volunteers: a cleanup project, a "Signs of AI Writing" guide, and a speedy-deletion criterion adopted in August 2025 — after the community spent two years failing to agree on anything broader and finally succeeded only by narrowing the rule to something objective enough to apply fast. [48]
That is the useful lesson from the first two eras, and it's the one nobody wants. ERP got a rescue industry because the pain was concentrated enough to fund one. The dot-com wreckage got cleared because the assets could be repriced and sold. Slop on the commons has neither property: cheap to produce, diffuse in its costs, worthless as salvage.
But this shape is not actually new, and claiming otherwise weakens the argument. Rao and Reiley put the externality ratio of email spam at roughly 100 to 1 — about $20 billion a year in costs to American firms and consumers, against some $200 million in worldwide spammer revenue. [60] Cheap to send, expensive to absorb, borne almost entirely by people who never sent anything. That is precisely the structure, twenty years early.
What matters is how it ended. Not with legislation — CAN-SPAM barely features in the causal story — but with absorption. A handful of providers large enough to amortize filtering across billions of users built the defenses and ate the cost. The externality never disappeared; it was relocated onto intermediaries who could carry it. Today's $5.5 billion email security market understates the true bill, because it counts only what is purchased and not what Google and Microsoft spend internally. [61]
So the question for this era is not whether the cleanup happens. It is who is large enough to absorb it. For enterprise AI, the answer is already forming: the governance and assurance market is exactly that absorption beginning. For the commons, there is no equivalent. The intermediaries are journals, package registries and volunteer maintainers, and 98.4 percent tells you how much absorption they are currently managing.
Which leaves the part that has to be built rather than waited for: a gate, put back deliberately where the budget used to sit by accident. Verification as a named, staffed, funded step rather than a virtue. Provenance as a requirement rather than a courtesy. And an honest institutional answer to the question the ERP era at least always knew how to answer: when this is wrong, whose name is on it?
The bill always comes. What changed is who gets it.
References
The ERP era
[1] Hershey Foods Corporation, 1999 Annual Report, Management's Discussion and Analysis. Q3 1999 net sales $1,066.7M vs $1,217.2M in Q3 1998 (−12.4%); $98.8M capitalized software and hardware plus $13.2M of expenses as of December 31, 1999. https://www.annualreports.com/HostedData/AnnualReportArchive/h/NYSE_HSY_1999.pdf Note: the $112M is disclosed project spend, not a loss figure — it is very widely misreported as a loss.
[3] Panorama Consulting Solutions, 2008 ERP Report, Part I (n = 1,322 organizations). https://cdn2.hubspot.net/hubfs/4439340/2008_ERP_Report_Part_I.pdf
[4] Panorama Consulting Solutions, 2008 ERP Report, Part IV (manufacturing and distribution cohort): average total cost $8,206,648; 19.2 months average duration; 68.2% technical implementation / 22.8% business implementation / 9.0% other; 18% plain vanilla, 35% heavy customization. https://cdn2.hubspot.net/hubfs/4439340/2008_ERP_Report_Part_IV.pdf
[5] Judy E. Scott, "The FoxMeyer Drugs' Bankruptcy: Was it a Failure of ERP?", University of Texas at Austin / AMCIS. https://zimmer.fresnostate.edu/\~sasanr/Teaching-Material/MIS/ERP/FoxMeyer.pdf
[6] Christopher Koch, "Nike Rebounds: How (and Why) Nike Recovered from Its Supply Chain Disaster," CIO, June 15, 2004. https://www.cio.com/article/264637/enterprise-resource-planning-nike-rebounds-how-nike-recovered-from-its-supply-chain-disaster.html
[7] Ann Sullivan, "Nike says i2 hurt its profits," Computerworld, March 5, 2001. https://www.computerworld.com/article/1445962/nike-says-i2-hurt-its-profits.html
[8] "Botched SAP implementation: National Grid and Wipro settle for $75m," The Register, August 6, 2018. https://www.theregister.com/2018/08/06/botched_sap_implementation_national_grid_wipro_settlement_75m/
[9] AMR Research, May 20, 1999: 1998 ERP vendor revenue $16.6 billion; 1998 ERP infrastructure sales $70 billion (inclusive of third-party services, hardware, databases and networking). https://www.itweb.co.za/article/amr-research-predicts-erp-market-will-reach-666-billion-by-2003/P3gQ2qGX1wEvnRD1
[10] Representative productized rescue practices: Panorama Consulting Group, "ERP Audit & Recovery" (https://www.panorama-consulting.com/services/project-recovery/); Ultra Consultants, "ERP Recovery" (https://ultraconsultants.com/erp-software-blog/erp-recovery/); Pemeco Consulting, "ERP Rescue" (https://pemeco.com/events/erp-rescue-a-proven-approach-to-recover-troubled-projects/); Lumenia Consulting, "ERP Project Recovery" (https://lumeniaconsulting.com/lumenia-services/erp-implementation/erp-project-recovery)
[11] Thomas H. Davenport, "Putting the Enterprise into the Enterprise System," Harvard Business Review 76(4), July–August 1998, pp. 121–131. https://hbr.org/1998/07/putting-the-enterprise-into-the-enterprise-system
[12] "Lidl cancels SAP introduction having sunk €500 million into it," Consultancy.uk, August 13, 2018. Jean-Claude Flury (DSAG) quoted. https://www.consultancy.uk/news/18243/lidl-cancels-sap-introduction-having-sunk-500-million-into-it Note: the €500M figure was reported by Handelsblatt and never confirmed by Lidl, which is privately held.
The dot-com era
[13] PricewaterhouseCoopers / NVCA MoneyTree Report (Thomson Reuters data), via the NVCA Yearbook: 2000 US venture investment $104,997.8M across 8,041 deals, against $55.9bn total for 1995–1998. https://leeds-faculty.colorado.edu/bhagat/NVCA_Yearbook_2014.pdf
[14] Jay R. Ritter, University of Florida, "Initial Public Offerings: Technology Stock IPOs." 370 tech IPOs in 1999; 24 in 2001. https://site.warrington.ufl.edu/ritter/files/IPOs-Tech.pdf
[15] Webmergers.com shutdown counts, via UPI ("Web closures more than double in 2001," December 27, 2001, https://www.upi.com/Archives/2001/12/27/Web-closures-more-than-double-in-2001/4841009429200/) and Out-Law/Pinsent Masons ("Nearly 5,000 internet companies have gone since 2000," March 13, 2003, https://www.pinsentmasons.com/out-law/news/nearly-5000-internet-companies-have-gone-since-2000)
[16] Global Crossing plan of reorganization, effective December 9, 2003. Total reorganization value $407 million, per the company's own filing (https://www.sec.gov/Archives/edgar/data/1061322/000119312503098911/dex991.htm); ST Telemedia $250 million for 61.5%, creditors 38.5% plus roughly $523 million in cash, prior equity cancelled with no consideration; post-emergence debt approximately $200 million against roughly $11 billion at end-2001. https://www.globalcustodian.com/global-crossing-emerges-from-chapter-11-with-singapore-technologies-telemedia-as-61-5-shareholder/
[17] Merger value: The Wall Street Journal, December 14, 1999 (https://www.wsj.com/articles/SB945094281606146732). Liquidation: "marchFIRST to Liquidate After Asset Sales Failed," The New York Times, May 1, 2001 (https://www.nytimes.com/2001/05/01/business/marchfirst-to-liquidate-after-asset-sales-failed.html)
[18] "SBI to acquire Razorfish for $8.2 million," Computerworld, 2003. https://www.computerworld.com/article/1332814/sbi-to-acquire-razorfish-for-8-2-million.html
[19] "Hitting The Wall At Boo," Newsweek, July 16, 2000. https://www.newsweek.com/hitting-wall-boo-161937
[20] Divine, Inc. — Chapter 11 filing February 25, 2003; assets auctioned April 2003 for approximately $54 million. https://en.wikipedia.org/wiki/Divine,\_Inc.
[21] U.S. General Accounting Office, GAO-03-138, Financial Statement Restatements: Trends, Market Impacts, Regulatory Responses, and Remaining Challenges, October 4, 2002. 919 restatements by 845 public companies, January 1, 1997 – June 30, 2002; annual count rose from 92 (1997) to 225 (2001). https://www.gao.gov/assets/gao-03-138.pdf
[22] SEC Release 33-8238, June 5, 2003: aggregate annual cost of the final Section 404 rule estimated at ~$1.24 billion, or ~$91,000 per company. Reported by GAO-03-933R. https://www.gao.gov/products/gao-03-933r
[23] Daniel L. Goelzer, PCAOB Board Member, "The Costs & Benefits of Sarbanes-Oxley Section 404," March 21, 2005, citing the Financial Executives International survey: $4.36 million average first-year Section 404 cost for companies averaging $5 billion in revenue. https://pcaobus.org/news-events/speeches/speech-detail/the-costs-benefits-of-sarbanes-oxley-section-404_126 Note: the FEI sample skews to large filers; the SEC figure was an all-filer average. Both are per company per year.
The AI era
[2] Deezer Newsroom, "AI music exceeds 50 percent of daily uploads," July 21, 2026. Also: up to 85% of streams of fully-AI tracks were fraudulent in 2025, while AI tracks account for 1–3% of all streams. https://newsroom-deezer.com/2026/07/ai-music-exceeds-50-percent-daily-uploads-deezer/
[24] NewsGuard: 49 AI-generated news sites, May 1, 2023 (https://www.newsguardtech.com/special-reports/newsbots-ai-generated-news-websites-proliferating/); 3,749 AI Content Farm sites across 16 languages, June 23, 2026 (https://www.newsguardtech.com/special-reports/ai-tracking-center/)
[25] Graphite, "AI now writes as many online articles as humans do," May 2026. 55,400 randomly sampled English-language Common Crawl articles classified by three independent detectors (Pangram, Copyleaks, GPTZero); Q1 2026: 49.9% primarily AI-generated. https://graphite.io/five-percent/research/ai-now-writes-as-many-online-articles-as-humans-do Note: this supersedes Graphite's single-detector October 2025 study, the source of the widely-repeated "AI overtook human writing in November 2024" claim.
[26] Deezer Newsroom, "AI-generated tracks represent 44% of new uploaded music," April 2026, which sets out the trajectory from ~30,000 tracks/day in September 2025. https://newsroom-deezer.com/2026/04/ai-generated-tracks-represent-44-of-new-uploaded-music/
[27] Daniel Stenberg, "Death by a thousand slops," July 14, 2025. https://daniel.haxx.se/blog/2025/07/14/death-by-a-thousand-slops/
[28] "Curl ending bug bounty program after flood of AI slop reports," BleepingComputer, January 22, 2026. https://www.bleepingcomputer.com/news/security/curl-ending-bug-bounty-program-after-flood-of-ai-slop-reports/
[29] "AI-Led Remediation Crisis Prompts HackerOne to Pause Bug Bounties," Dark Reading, April 8, 2026. Internet Bug Bounty submissions suspended effective March 27, 2026; John Morello (CTO, Minimus) quoted on valid submission rates falling from ~15% to below 5%. https://www.darkreading.com/application-security/ai-led-remediation-crisis-prompts-hackerone-pause-bug-bounties
[30] "Bug bounty platforms battle AI slop," Computing, May 18, 2026. https://www.computing.co.uk/news/2026/security/bug-bounty-platforms-battle-ai-slop
[31] Arjun Iyer, "The AI-Generated Code Crisis," The New Stack, April 9, 2026. Also reports CodeRabbit's finding of ~1.7× more issues in AI-co-authored pull requests across 470 open-source PRs. https://thenewstack.io/ai-generated-code-crisis/
[32] Maxim Topaz, Nir Roguin, P. Gupta et al., "Fabricated citations: an audit across 2.5 million biomedical papers," The Lancet, May 7, 2026. 125.6 million references extracted and 97.1 million verified across 2,471,758 papers from PubMed Central's Open Access subset, published January 1, 2023 – February 18, 2026; 4,046 fabricated references across 2,810 papers; rate of affected papers 1 in 2,828 (2023), 1 in 458 (2025), 1 in 277 (early 2026); 98.4% of affected papers had seen no publisher action at the time of the audit. https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(26)00603-3/fulltext — live dashboard: https://www.maxtopaz.com/citadel Note: the corpus is PubMed Central Open Access, not all of PubMed — Nature issued a correction on May 13, 2026 to that effect (https://www.nature.com/articles/d41586-026-00748-w), and much secondary coverage still carries the wrong scope. PMC OA over-represents open-access publishers, and the study found over a third of fabricated citations came from two of them, so the rate should not be generalized to all biomedical literature. The authors describe the increase as coinciding with the rise of AI writing tools; the study measures prevalence, not provenance, and makes no causal attribution. Retraction Watch's "4,406" is a transposition of 4,046.
[33] Kate Niederhoffer, Gabriella Rosen Kellerman, Angela Lee, Alex Liebscher, Kristina Rapuano and Jeffrey T. Hancock, "AI-Generated 'Workslop' Is Destroying Productivity," Harvard Business Review, September 22, 2025. Survey of 1,150 full-time U.S. desk workers by BetterUp Labs with the Stanford Social Media Lab. https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity — study page: https://www.betterup.com/workslop Note: BetterUp's page states 40%; the HBR article text renders it as 41%. Same study.
[56] Daniel Stenberg, "About Daniel," daniel.haxx.se: "he has created curl, one of the world's most installed software products, with over 20 billion installations." Stenberg founded curl in 1998 and has worked on it full-time since February 2019 via wolfSSL. https://daniel.haxx.se/about.html Note: the installation figure is the maintainer's own estimate, which is the only figure available — curl ships inside operating systems and embedded devices and is not centrally counted.
[57] Financial Accounting Standards Board, ASC 350-40, Intangibles — Goodwill and Other: Internal-Use Software: preliminary-project-stage costs are expensed as incurred; capitalization begins only when management authorizes and commits funding and it is probable the project will be completed. Training and data-conversion costs are excluded from capitalization at every stage. Capitalization is mandatory, not elective, when criteria are met. IFRS reaches an equivalent result via IAS 38, which expenses all research-phase expenditure. https://www.bu.edu/policies/files/2015/08/Accounting-for-Costs-of-Software.pdf · https://www.pkf-l.com/insights/capitalising-ai-tools-accounting-ias-38/
[58] FASB Accounting Standards Update 2025-06 (September 2025), replacing the stage framework with a probable-to-complete threshold explicitly not met where "the software being developed has novel, unique, or unproven functions or features, and the uncertainty related to those... has not been resolved through coding and testing" (ASC 350-40-25-12A). Expected to result in more software costs being expensed. Effective for annual periods beginning after December 15, 2027. https://dart.deloitte.com/USDART/home/publications/deloitte/heads-up/2025/fasb-asu-amends-software-costs-guidance
[59] Zillow Group Q3 2021 results, November 2, 2021: $304 million inventory write-down on Zillow Offers, with a further $240–265 million expected in Q4. Rich Barton: "we've determined the unpredictability in forecasting home prices far exceeds what we anticipated." https://www.sec.gov/Archives/edgar/data/1617640/000161764021000085/q32021991.htm Note: the charge was against home inventory, not against the model. The algorithm was never a capitalized asset — the loss became visible only because it had been used to buy physical assets.
[60] Justin M. Rao and David H. Reiley, "The Economics of Spam," Journal of Economic Perspectives 26(3), 2012, pp. 87–110: "American firms and consumers experience costs of almost $20 billion annually due to spam"; spammers and spam-advertised merchants "collect gross worldwide revenues on the order of $200 million per year"; "the 'externality ratio' of external costs to internal benefits for spam is around 100:1." https://www.aeaweb.org/articles?id=10.1257%2Fjep.26.3.87
[61] Fortune Business Insights, email security market: $5.46 billion in 2025, projected $6.06 billion in 2026 and $14.44 billion by 2034. https://www.fortunebusinessinsights.com/email-security-market-106607 Note: a commercial market-sizing estimate, and it counts only purchased email security — the largest share of actual defensive cost, what large providers spend filtering in-house, is disclosed nowhere.
[49] Gartner, "Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026," May 19, 2026. Table 1, "Worldwide AI Spending by Market, 2025-2027": AI Software $453,209M in 2026 against $282,9xx M in 2025; AI Infrastructure $1,431,509M of a $2,595,667M total. Gartner defines AI infrastructure as "AI-optimized IaaS, AI-optimized servers, AI network fabric, AI processing semiconductors and devices." John-David Lovelock, Distinguished VP Analyst: "Up to this point, AI spending has primarily been driven by technology companies and hyperscalers. Enterprises have yet to really flex their spending potential. That is coming and 2026 will be the inflection year." https://www.gartner.com/en/newsroom/press-releases/2026-05-19-gartner-forecasts-worldwide-ai-spending-to-grow-47-percent-in-2026 Note: the widely-quoted $2.59 trillion total is worldwide AI spending across all buyers, not enterprise spending. The device and semiconductor split inside the 2026 infrastructure line is not published; in the September 2025 edition of the same forecast — a $2.02 trillion base — hardware was 58% of the total and consumer-class devices alone were 27%, with GenAI smartphones the largest single line. A later Gartner revision (2Q26, July 24, 2026) reportedly moves the 2026 total to $2.67 trillion.
[50] Mavvrik with Benchmarkit, 2026 State of AI Cost Governance Report: a survey of 396 enterprise organizations across six sectors, fielded April–May 2026. 62% hit unexpected AI costs that materially altered business decisions; 40% required board-level escalation; 33% imposed emergency spending freezes; 25% delayed or cancelled an AI initiative. https://www.mavvrik.ai/blog/blog-ai-cost-governance-report-2026/ — coverage: https://www.cfodive.com/news/1-in-4-companies-delay-cancel-ai-projects-over-cost/827524/ Note: vendor-sponsored — Mavvrik sells AI cost-governance software. Respondent seniority, company-size bands and sampling frame are not disclosed.
[51] S&P Global Market Intelligence / 451 Research, reported by CIO Dive, March 14, 2025: organizations abandoned an average of 46% of AI proofs-of-concept before production; 1,000+ respondents across North America and Europe. https://www.ciodive.com/news/AI-project-fail-data-SPGlobal/742590/ Note: field dates not published in the coverage.
[52] Klarna press release, "Klarna AI assistant handles two-thirds of customer service chats in its first month," February 27, 2024: 2.3 million conversations in month one, "the equivalent work of 700 full-time agents," resolution time down from 11 minutes to 2. https://www.prnewswire.com/news-releases/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month-302072740.html
[53] Sebastian Siemiatkowski, Bloomberg interview, May 8, 2025, reported by CX Dive, May 9, 2025. https://www.customerexperiencedive.com/news/klarna-reinvests-human-talent-customer-service-AI-chatbot/747586/Note: Klarna did not remove the AI, which still handled roughly two-thirds of inquiries; this was a rebalancing toward a human escalation tier. Headcount had fallen from 5,527 (end 2022) to 3,422 (end 2024), largely through attrition and a hiring freeze, per Klarna's IPO filing.
[54] Amazon's March 2026 outages: four Sev-1 incidents in one week; internal briefing material describing incidents with "high blast radius" involving "Gen-AI assisted changes." The Register, March 10, 2026 (https://www.theregister.com/2026/03/10/amazon_ai_coding_outages); Fortune, March 12, 2026 (https://fortune.com/2026/03/12/amazon-retail-site-outages-ai-agent-inaccurate-advice/); CNBC, March 10, 2026 (https://www.cnbc.com/2026/03/10/amazon-plans-deep-dive-internal-meeting-address-ai-related-outages.html). Order-loss figures (~6.3 million on March 5, ~120,000 on March 2) originate with Business Insider, March 2026, reporting from internal documents: https://www.businessinsider.com/amazon-tightens-code-controls-after-outages-including-one-ai-2026-3 Note: verify the order figures against the Business Insider original before republication. The number is currently propagating largely through uncited secondary content.
[55] Amazon, "Amazon responds to inaccurate Financial Times report linking outages to AI," March 12, 2026: "The real cause was unrelated to AI and simply that our systems allowed an engineering team coding error to have broader impact than it should have." Amazon states none of the incidents involved AI-written code and only one involved AI tooling. https://www.aboutamazon.com/news/company-news/amazon-outage-ai-financial-times-correction Related and frequently conflated: the October 2025 AWS us-east-1 outage was not AI-related. AWS attributes it to a latent race condition in DynamoDB DNS management. https://aws.amazon.com/message/101925/
[34] Veracode, "Spring 2026 GenAI Code Security Update," March 24, 2026. 150+ models, 80 coding tasks: security pass rate ~55%, syntax correctness 95%+, XSS pass rate 15%, log injection 13%. https://www.veracode.com/blog/spring-2026-genai-code-security/
[35] Stack Overflow 2025 Developer Survey, AI section (n = 33,662 answering the AI questions). 66% cite "AI solutions that are almost right, but not quite" as their biggest frustration; 46% actively distrust AI accuracy against 33% who trust it. https://survey.stackoverflow.co/2025/ai
[36] The Standish Group, CHAOS Report, 1994. https://cs.franklin.edu/\~smithw/ITEC495_Resources/chaos%20report.pdf
[37] Magne Jørgensen and Kjetil Moløkken-Østvold, "How large are software cost overruns? A review of the 1994 CHAOS report," Information and Software Technology, 2006. Concludes that comparable surveys of the period put average cost overrun near 30%. https://web-backend.simula.no/sites/default/files/publications/Jorgensen.2006.4.pdf
[38] J. Laurenz Eveleens and Chris Verhoef, "The Rise and Fall of the Chaos Report Figures," IEEE Software 27(1), January/February 2010, pp. 30–36. https://www.cs.vu.nl/\~x/the_rise_and_fall_of_the_chaos_report_figures.pdf
[39] Francesco D'Isa, quoted in Brigitte Nerlich, "From sloppers to slopocalypse: the lexical productivity of AI slop," Making Science Public, University of Nottingham, January 2, 2026. https://makingsciencepublic.com/2026/01/02/from-sloppers-to-slopocalypse-the-lexical-productivity-of-ai-slop/
[40] Gartner press release, "Global AI Regulations Fuel Billion-Dollar Market for AI Governance Platforms," February 17, 2026. https://www.gartner.com/en/newsroom/press-releases/2026-02-17-gartner-global-ai-regulations-fuel-billion-dollar-market-for-ai-governance-platforms
[41] UK Department for Science, Innovation and Technology, "Trusted Third-Party AI Assurance Roadmap," September 2025: 524 companies, ~£1.01 billion of value in 2024, projected £18.8 billion by 2035. Summarised by Burges Salmon. https://www.burges-salmon.com/articles/102l5fo/trusted-third-party-ai-assurance-dsit-roadmap/
[42] "Braintrust lands $80M Series B funding round to become observability layer for AI," SiliconANGLE, February 17, 2026. https://siliconangle.com/2026/02/17/braintrust-lands-80m-series-b-funding-round-become-observability-layer-ai/
[43] "AI agent tooling provider LangChain raises $125M at $1.25B valuation," SiliconANGLE, October 20, 2025. https://siliconangle.com/2025/10/20/ai-agent-tooling-provider-langchain-raises-125m-1-25b-valuation/
[44] Accenture Q4 FY2025 earnings call transcript, September 25, 2025: advanced AI bookings $5.9 billion, advanced AI revenue $2.7 billion, against total FY2025 revenue of $69.7 billion. https://investor.accenture.com/\~/media/Files/A/accenture-v4/investors/earnings-reports/2025/accenture-fourth-quarter-fiscal-2025-conference-call-transcript-updated-10-8-2025.pdf
[45] Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia Civil Resolution Tribunal, decided February 19, 2024. https://www.mccarthy.ca/en/insights/blogs/techlex/moffatt-v-air-canada-misrepresentation-ai-chatbot
[46] Damien Charlotin, "AI Hallucination Cases" database. 1,962 cases as of August 26, 2026, across 50+ jurisdictions; includes only decisions where a court explicitly found reliance on AI-generated hallucinations. https://www.damiencharlotin.com/hallucinations/
[47] "Deloitte to refund Australian government for AI-hallucinated report," The Register, October 6, 2025; and CFO Dive, October 21, 2025. A$440,000 contract with the Department of Employment and Workplace Relations; more than a dozen fabricated references; A$97,000+ refunded. https://www.theregister.com/2025/10/06/deloitte_ai_report_australia/
[48] "Wikipedia editors adopt speedy deletion policy for AI slop articles," 404 Media, August 5, 2025 (https://www.404media.co/wikipedia-editors-adopt-speedy-deletion-policy-for-ai-slop-articles/); and Willemien Froneman, "Failed comprehensiveness, successful minimalism: Wikipedia's 3-year struggle to govern AI-generated content (2022–2025)," AI & SOCIETY 41(7), 2026, pp. 7045–7061 (https://link.springer.com/article/10.1007/s00146-026-03046-1)