Eighteen Errors: What an AI Research Tool Got Wrong About Semiconductor Forecasting

An AI research notebook summarized four sources faster than I read them. It also turned a 10% error into a 0.1% error, and the most persuasive explanation in its own podcast had no source behind it.

Predictive analytics, read through four sources

The Week 4 assignment asked for a focused exploration of one analytics domain inside the capstone client’s industry, built in Google’s NotebookLM, which now presents itself as Gemini Notebook. I chose predictive analytics and loaded four sources: a peer-reviewed case study in IEEE Access, the open journal of the Institute of Electrical and Electronics Engineers, on per-item model selection (Maheshwari et al., 2026), Deloitte’s 2026 semiconductor outlook (Kusters et al., 2026), the forecasting case from NXP Semiconductors and Amazon Web Services (AWS) (Boutane et al., 2023), and NVIDIA’s Form 10-Q.

[OBSERVED] The sources carry three lessons worth keeping. Choosing a model for each material lifted forecast accuracy from between 38% and 54% to between 70% and 85% in the plant studied (Maheshwari et al., 2026, p. 19108). Orders inflate under rationing, so a model trained on orders learns the inflation as demand (Lee et al., 1997, p. 551). And high-value AI chips carry roughly half of industry revenue on less than 0.2% of units (Kusters et al., 2026), so the forecast error that matters falls on a handful of products.

Eighteen errors, sorted by kind

Errors found

Unsupported (hallucinated)

Misattributed

Distorted

Where	What the tool said	What the source says
Video Overview	AutoPredictor’s WAPE was below 0.1%	Below 0.1, an error under 10%: a hundredfold understatement
Audio Overview	The model beat human forecasters in 2020 by digesting macroeconomic data while humans panicked	No source says this; the source assigns market indicators to NXP’s human experts
Draft note	A “custom 12-model ensemble framework”	One model selected per item; the paper gives both 12 and 13 models
Chat answer	A $105 billion “energy guarantee”	Credit support on a land, power, and shell buildout (Form 10-Q, p. 19)
Chat answer	A fourfold memory price rise applied to HBM	The rise applies to consumer DDR4 and DDR5 memory
Video Overview	The accuracy gain credited to NXP’s layered models	The gain belongs to the IEEE case alone

Table 1. Six of the eighteen documented errors. Observed. The full table, with locators for every row, appears in the reflective summary.

The mind map contained no errors, and the reason teaches something: I specified every branch and node in its prompt. Asked to summarize, the tool summarized well. Asked to explain, it supplied explanations, and the errors lived in those explanations.

THE FINDING WORTH CARRYING

The most persuasive explanation in the podcast had no source behind it: the interpretability risk the notebook describes, reproduced by the tool that summarized it.

The workaround that ended the correction loop

Regenerating a section to fix one error usually introduced another. The loop ended when I stopped asking the tool to write and started asking it to repeat verified text exactly, then saved that text as the note. Speed came from the tool; accuracy came from refusing to let it paraphrase a number.

What these numbers are not

The eighteen errors come from one notebook, built in October 2026, and the counts depend on my definitions: unsupported means no basis in any source, misattributed means a real figure assigned to the wrong method or actor, distorted means a sourced figure whose value, unit, or status changed. A different notebook would return a different list.

Provenance and marks

This post is part of the NVIDIA capstone for the M.S. Business Intelligence program at Full Sail University, from BIN660-O, Business Intelligence Case Studies, taught by Dr. Colton Tapoler. Epistemic marks appear before the claim they govern: [OBSERVED] read from a primary source and verified against it, [MODELED] produced by a stated method from stated inputs, [HYPOTHETICAL] a stated assumption or target with no measurement behind it, [UNVERIFIED] asserted somewhere but not resolved to a locator. AI disclosure: Anthropic’s Claude assisted with research, verification, arithmetic in code, and drafting; the author chose the analysis, checked every figure against its source, and answers for the content.

Non-affiliation notice. The author is not employed by, affiliated with, sponsored by or endorsed by NVIDIA Corporation, Microsoft Corporation, Google, Palantir Technologies, or any other company named here. Companies are named as subjects of an academic analysis prepared by an independent party from public disclosure only. No confidential information was used, and no representation is made on any company’s behalf.