← series index

# Sources — Session 1: AI Foundations for Researchers

*AI for Researchers · dated, annotated reference list · compiled 2026-07-27, revised 2026-09-02 · Landscape as of August 2026*

Entry format (one line per source, numbering matches the `[n]` footnote markers used in `slides.html` and `handout.md`):

`- [n] Author/Org (Year). *Title*. URL — accessed YYYY-MM-DD. [type: peer-reviewed | policy | primary-doc | news] — one-line note on what claim(s) it grounds.`

Allowed `type` values:
- `peer-reviewed` — peer-reviewed paper or arXiv/preprint literature
- `policy` — official publisher/funder/institutional policy page
- `primary-doc` — primary tool/vendor documentation (not a blog or listicle)
- `news` — recent news/benchmark article; informs tool discovery only, never the sole ground for a factual claim

---

## References

- [1] OpenAI (2026). *What are tokens and how to count them?* https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them — accessed 2026-07-27. [type: primary-doc] — Grounds the tokenization slide: "Tokens are the building blocks of text that OpenAI models process"; rules of thumb 1 token ≈ 4 characters, 1 token ≈ ¾ of a word, 100 tokens ≈ 75 words; input / output / cached / reasoning token categories.
- [2] Anthropic (2026). *Models overview & Fable 5 model docs — Claude Platform Docs*. https://platform.claude.com/docs/en/models/overview — accessed 2026-09-02 (first captured 2026-07-27). [type: primary-doc] — Grounds the context-window and knowledge-cutoff figures for the Claude family as of September 2026 (Claude Fable 5.1 / Opus 5 / Sonnet 5 = 1M-token context, 128k max output, cutoffs Jun 2026 / May 2026 / Jan 2026; Haiku 4.5 = 200k, cutoff Feb 2025); also grounds the extended-thinking-by-default claims — "Adaptive thinking is always on" and "The raw chain of thought is never returned" (thinking blocks come back summarised or omitted) on the Fable 5 model docs. The ~30% tokenizer-change note is retained from the 2026-07-27 capture of the models overview.
- [3] Google (2026). *Gemini 3 Developer Guide — Gemini API*. https://ai.google.dev/gemini-api/docs/gemini-3 — accessed 2026-07-27. [type: primary-doc] — Grounds Gemini context/knowledge-cutoff figures (gemini-3.1-pro-preview, gemini-3.1-flash-lite: 1M input / 64k output, knowledge cutoff Jan 2025) and the `thinking_level` control (minimal / low / medium / high), described as "relative allowances for thinking rather than strict token guarantees". (Gemini 3.1 Pro remained labelled "Preview" as of 2026-09-02.)
- [4] Kalai, A. T., Nachum, O., Vempala, S. S., & Zhang, E. (2025). *Why Language Models Hallucinate*. arXiv:2509.04664. https://arxiv.org/abs/2509.04664 — accessed 2026-07-27. [type: peer-reviewed] — The load-bearing source for "hallucination is structural": hallucinations "originate simply as errors in binary classification" and "arise through natural statistical pressures"; they persist because "language models are optimized to be good test-takers, and guessing when uncertain improves test performance". Also grounds the mitigation of instructing/rewarding abstention.
- [5] Walters, W. H., & Wilder, E. I. (2023). *Fabrication and errors in the bibliographic citations generated by ChatGPT*. Scientific Reports 13, 14045. https://www.nature.com/articles/s41598-023-41032-5 — accessed 2026-07-27. [type: peer-reviewed] — Grounds the headline citation-fabrication figures: across 636 citations in 84 generated literature reviews, "55% of the GPT-3.5 citations but just 18% of the GPT-4 citations are fabricated", and "43% of the real (non-fabricated) GPT-3.5 citations but just 24% of the real GPT-4 citations include substantive citation errors". A 2023 study of the GPT-3.5/GPT-4 generation — kept deliberately as the labelled historical baseline; the current-generation companion number is [7].
- [6] Zhao, Z., Wang, Y., Stuart, T., de Vaan, M., Ginsparg, P., & Yin, Y. (2026). *LLM hallucinations in the wild: Large-scale evidence from non-existent citations*. arXiv:2605.07723. https://arxiv.org/abs/2605.07723 — accessed 2026-09-02. [type: peer-reviewed] — Grounds the "it is already in the record" claim: an audit of 111 million references across 2.5 million papers (arXiv, bioRxiv, SSRN, PubMed Central) finding "a sharp rise in non-existent references following widespread LLM adoption, with a conservative estimate of 146,932 hallucinated citations in 2025 alone", concentrated in small and early-career author teams, with existing moderation capturing "only a fraction".
- [7] Rao, D., & Callison-Burch, C. (2026). *BibTeX Citation Errors in Scientific Publishing Agents: Evaluation and Mitigation*. COLM 2026; arXiv:2604.03159v2 (2026-08-08). https://arxiv.org/abs/2604.03159 — accessed 2026-09-02. [type: peer-reviewed] — Grounds "web search does not fix citations", now on the current generation: a 931-paper benchmark across three search-enabled frontier models (GPT-5, Claude Sonnet 4.6, Gemini 3 Flash) gives 83.6% field-level accuracy but "only 50.9% of entries are fully correct", with accuracy dropping 27.7 percentage points from popular to recent papers, "revealing heavy reliance on parametric memory even when search is available"; deterministic retrieval from authoritative records raises fully correct entries to 78.3%.
- [8] Rao, D., Wong, E. W. M., & Callison-Burch, C. (2026). *Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents*. arXiv:2604.03173. https://arxiv.org/abs/2604.03173 — accessed 2026-09-02. [type: peer-reviewed] — Grounds the deep-research-agent caveat: across 10 models/agents (53,090 URLs on DRBench; 168,021 on ExpertQA) "3–13% of citation URLs are hallucinated … while 5–18% are non-resolving overall", and "deep research agents generate substantially more citations per query than search-augmented LLMs but hallucinate URLs at higher rates" — "Pooling across the two deep research agents, the hallucination rate is 10.7% [10.2, 11.2] versus 4.8% [4.3, 5.2] for the eight search-augmented models"; non-resolving rates vary by field from 5.4% to 11.4%.
- [9] Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2024). *Lost in the Middle: How Language Models Use Long Contexts*. Transactions of the Association for Computational Linguistics 12, 157–173. https://aclanthology.org/2024.tacl-1.9/ — accessed 2026-07-27. [type: peer-reviewed] — Grounds "a big context window is not perfect recall": performance "is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models". Measured on the 2023-era long-context cohort; no current-generation replication located as of 2026-09-02 — taught as a labelled baseline whose advice (deliberate context placement) still stands.
- [10] OpenAI (2026). *Reasoning models — OpenAI API guide*. https://developers.openai.com/api/docs/guides/reasoning — accessed 2026-07-27. [type: primary-doc] — Grounds what a reasoning model is: models "introduce reasoning tokens in addition to input and output tokens … to 'think,' breaking down the prompt and considering multiple approaches"; those tokens "occupy space in the model's context window and are billed as output tokens"; reasoning effort is selectable (none / low / medium / high / xhigh / max) and the models "work especially well for complex problem solving, coding, scientific reasoning, and multi-step agentic workflows".
- [11] Walden, W., & Wanner, M. (2026). *Reasoning Models Will Sometimes Lie About Their Reasoning*. arXiv:2601.07663. https://arxiv.org/abs/2601.07663 — accessed 2026-07-27. [type: peer-reviewed] — Grounds "the visible reasoning trace is not an audit trail": models "may acknowledge the presence of hints, [but] they will often deny intending to use them — even when permitted to use hints and even when it can be demonstrated that they are using them", raising "broader challenges for CoT monitoring and interpretability".
- [12] Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022). *Chain-of-Thought Prompting Elicits Reasoning in Large Language Models*. NeurIPS 2022; arXiv:2201.11903. https://arxiv.org/abs/2201.11903 — accessed 2026-07-27. [type: peer-reviewed] — Grounds the origin of step-by-step reasoning as a prompting technique: "prompting a 540B-parameter language model with just eight chain of thought exemplars achieves state of the art accuracy on the GSM8K benchmark of math word problems".
- [13] Schulhoff, S., Ilie, M., Balepur, N., et al. (2025). *The Prompt Report: A Systematic Survey of Prompting Techniques*. arXiv:2406.06608 (v6, Feb 2025). https://arxiv.org/abs/2406.06608 — accessed 2026-07-27. [type: peer-reviewed] — Grounds "prompting is a documented literature, not folklore": a PRISMA-based survey presenting "a detailed vocabulary of 33 vocabulary terms, a taxonomy of 58 LLM prompting techniques, and 40 techniques for other modalities".
- [14] Zheng, M., Pei, J., Logeswaran, L., Lee, M., & Jurgens, D. (2024). *When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models*. Findings of ACL: EMNLP 2024, 15126–15154. https://aclanthology.org/2024.findings-emnlp.888/ — accessed 2026-07-27. [type: peer-reviewed] — Grounds the honest caveat on role prompting: across 162 roles, 4 LLM families and 2,410 factual questions, "adding personas in system prompts does not improve model performance … compared to the control setting where no persona is added", and automatically finding a good persona performs "no better than random selection".
- [15] Mohammadi, E., Thelwall, M., Cai, Y., Collier, T., Tahamtan, I., & Eftekhar, A. (2026). *Is generative AI reshaping academic practices worldwide? A survey of adoption, benefits, and concerns*. Information Processing & Management 63(1), 104350. https://doi.org/10.1016/j.ipm.2025.104350 — accessed 2026-07-27. [type: peer-reviewed] — Peer-reviewed grounding for the adoption picture: a 20-country survey of publishing academics showing widespread adoption with large disciplinary/national/gender differences ("females were 10% less likely to use Gen AI frequently … for research"), tools used "mainly … for academic writing rather than data analysis", and top concerns of "inaccuracy, plagiarism, discouraging critical thinking, a lack of transparency and explainability, intellectual property rights violations, and data privacy risks".
- [16] Wiley (2025). *AI Adoption Jumps to 84% Among Researchers as Expectations Undergo Significant "Reality Check"* (ExplanAItions study, wave 2). https://newsroom.wiley.com/press-releases/press-release-details/2025/AI-Adoption-Jumps-to-84-Among-Researchers-as-Expectations-Undergo-Significant-Reality-Check/default.aspx — accessed 2026-07-27. [type: news] — Publisher's own report of its 2,430-researcher survey (fielded August 2025): AI-tool usage "surged from 57% in 2024 to 84% in 2025"; use for research and publication tasks "grew significantly to 62% from 45%"; concerns about inaccuracies and hallucinations rose "from 51% to 64%"; "80% using mainstream tools like ChatGPT compared to just 25% using AI research assistants". **Review ruling (Ehsan checkpoint, 2026-07-27): publisher first-party surveys are citable when quoted accurately AND labelled on the slide as self-reported survey data.** Slide 6 therefore carries a visible provenance line naming this as publisher self-reported data. Retains type `news` (the template vocabulary has no `survey`/`report` value) and is still corroborated by the peer-reviewed [15]. Re-checked 2026-09-02 (web search of wiley.com study pages and newsroom): the Aug-2025 wave remains the latest published ExplanAItions wave — the slide labels the figures with their survey years accordingly.
- [17] Ruediger, D., McCracken, C., & Skinner, M. (2024). *Adoption of Generative AI by Academic Biomedical Researchers*. Ithaka S+R. https://sr.ithaka.org/publications/adoption-of-generative-ai-by-academic-biomedical-researchers/ — accessed 2026-07-27. [type: news] — Survey report (2,459 complete responses, fielded Feb–Mar 2024) grounding the "interest exceeds confident use" pattern: 63% had experimented with generative AI but only 7% used it regularly for research; 74% cited insufficient accuracy as a moderate-to-large barrier. **Same review ruling as [16] applies:** first-party survey data is citable when quoted accurately and labelled on the slide as self-reported; slide 6's provenance line names Ithaka S+R as the self-reporting body. Retains type `news`; corroborated by the peer-reviewed [15].
- [18] Lieberum, J.-L., Toews, M., Metzendorf, M.-I., Heilmeyer, F., Siemens, W., Haverkamp, C., Böhringer, D., Meerpohl, J. J., & Eisele-Metzger, A. (2025). *Large language models for conducting systematic reviews: on the rise, but not yet ready for use — a scoping review*. Journal of Clinical Epidemiology 181, 111746. https://www.jclinepi.com/article/S0895-4356(25)00079-4/fulltext — accessed 2026-07-27. [type: peer-reviewed] — Grounds the "needs checking" zone of the capability/failure map: 37 included articles covering 10 of 13 defined systematic-review steps, most often literature search (41%), study selection (38%) and data extraction (30%); "about half of the studies found LLMs promising, a quarter were neutral, and one-fifth found them not promising"; conclusion that "LLMs should only be used with caution and under human supervision".
- [19] Consensus (2025). *Welcome to Consensus*. https://consensus.app/home/blog/welcome-to-consensus/ — accessed 2026-07-27. [type: primary-doc] — Primary description of a Tier-2 research-specific tool: a corpus of "over 200 million scientific documents" aggregated from Semantic Scholar, OpenAlex and its own crawl; hybrid semantic + BM25 retrieval, quality re-ranking, and the stated design rule "We only use AI after we search the scientific literature."
- [20] Elicit (2026). *Elicit: AI for scientific research*. https://elicit.com/ — accessed 2026-07-27. [type: primary-doc] — Primary description of a Tier-2 structured-extraction tool: search, summarize and extract data across "over 125 million papers"; report generation "inspired by systematic reviews"; automation of screening and data extraction with claimed workload savings. Vendor claims are reported as vendor claims on slides, never as measured effect sizes.
- [21] SCiNiTO (2026). *What is SCiNiTO?* https://www.scinito.ai/help/getting-started/what-is-scinito — accessed 2026-07-27. [type: primary-doc] — Primary description of another Tier-2 option, included for landscape completeness: "an AI research assistant built on top of OpenAlex" bundling search across 500M+ works, source-citing AI chat, PDF analysis, a journal recommender, a manuscript reviewer agent and shared research Spaces.
- [22] Google (2026). *Gemini Notebook (formerly NotebookLM)*. https://notebooklm.google/ — accessed 2026-07-27. [type: primary-doc] — Primary documentation for the source-grounded notebook category, and for the July 2026 rename: "NotebookLM is now Gemini Notebook as of July 2026"; responses "will always be grounded in the information you provide", which the vendor describes as "reducing the likelihood of AI errors and hallucinations". Rename (2026-07-16) and free-tier limits re-confirmed 2026-09-02 via Google's blog and support pages.
- [23] Google (2026). *Gemini Deep Research Agent — Gemini API*. https://ai.google.dev/gemini-api/docs/interactions/deep-research — accessed 2026-07-27. [type: primary-doc] — Primary documentation for the Tier-3 deep-research-agent category: an agent that "autonomously plans, executes, and synthesizes multi-step research tasks" producing "detailed, cited reports"; tasks "involve iterative searching and reading and can take several minutes to complete"; collaborative planning lets the user review and edit the research plan before execution.
- [24] OpenAI (2025). *Introducing deep research*. https://openai.com/index/introducing-deep-research/ — accessed 2026-09-02. [type: news] — Vendor announcement used only for category discovery and for the vendor's own description of the workflow (independent multi-step browsing over many sources, producing a cited report in tens of minutes). Not used as grounds for any accuracy claim; accuracy claims about this category come from [8].
- [25] ICMJE (2025). *Recommendations — Defining the Role of Authors and Contributors, §4 Artificial Intelligence (AI)-Assisted Technology*. https://www.icmje.org/recommendations/browse/roles-and-responsibilities/defining-the-role-of-authors-and-contributors.html — accessed 2026-07-27. [type: policy] — Grounds the disclosure claim on the capability/failure map: "At submission, the journal should require authors to disclose whether they used artificial intelligence (AI)-assisted technologies (such as Large Language Models [LLMs], chatbots, or image creators) in the production of submitted work", and "Use of AI for writing assistance should be reported in the acknowledgment section." (Policy is covered in depth in Session 4.)
- [26] NIH (2023). *The Use of Generative Artificial Intelligence Technologies is Prohibited for the NIH Peer Review Process* (NOT-OD-23-149). https://grants.nih.gov/grants/guide/notice-files/NOT-OD-23-149.html — accessed 2026-07-27. [type: policy] — Grounds the "dangerous" zone: NIH "prohibits NIH scientific peer reviewers from using natural language processors, large language models, or other generative Artificial Intelligence (AI) technologies for analyzing and formulating peer review critiques"; uploading application content to online generative AI tools "violates the NIH peer review confidentiality and integrity requirements".
- [27] Anthropic (2026). *System Card: Claude Opus 5* (2026-07-24). https://www-cdn.anthropic.com/b514064af1408018e64b1ad24e7d5e75850b4ffd/Claude%20Opus%205%20System%20Card.pdf — accessed 2026-09-02. [type: primary-doc] — Vendor system card (labelled as such on slides) grounding "the tradeoff is structural and vendors document it": on the AA-Omniscience factuality benchmark, "accuracy is 11% higher than Opus 4.8, but its rate of hallucinations is also 6% higher"; the executive summary adds "The model hallucinates factual claims slightly more than Opus 4.8, despite being more accurate overall", and observed agentic failure modes include "Fabricating execution output, file contents, or citations for work it had not actually done".
- [28] Anthropic (2026). *System Card: Claude Fable 5 & Claude Mythos 5* (2026-06-09). https://anthropic.com/claude-fable-5-mythos-5-system-card — accessed 2026-09-02. [type: primary-doc] — Vendor system card (labelled as such on slides) grounding the guess-more-vs-know-more tradeoff: on the 500-prompt missing-reference honesty set, "Mythos 5 scored 82%, behind Claude Mythos Preview (94%), Claude Opus 4.7 (92%), and Claude Opus 4.8 (91%)" — the model "is more inclined to attempt an answer than to decline"; red-teamers list "hallucinated citations and data" among "the consistent failure modes".
- [29] OpenAI (2026). *GPT-5.6 August update — Deployment Safety Hub*. https://deploymentsafety.openai.com/gpt-5-6-august-update — accessed 2026-09-02. [type: primary-doc] — Grounds the effort-dial claim for ChatGPT: with the 2026-08-06 update, Plus and Pro users get updated GPT-5.6 Sol plus "a slider that lets them choose how much effort ChatGPT uses for a response"; Free and Go users get a new default model for everyday chats.
- [30] Consensus (2026). *Consensus MCP server — developer docs*. https://docs.consensus.app/docs/mcp — accessed 2026-09-02. [type: primary-doc] — Grounds the taxonomy caveat: the Consensus MCP server lets ChatGPT, Claude, and other MCP clients "search over 200 million peer-reviewed academic research papers directly from the conversation" — i.e., a Tier-1 chatbot calling a Tier-2 tool directly.
- [31] Research Solutions (2026). *Research Solutions Launches Scite MCP, Connecting ChatGPT, Claude, & Other AI Tools To Scientific Literature* (2026-02-26). https://researchsolutions.investorroom.com/2026-02-26-Research-Solutions-Launches-Scite-MCP,-Connecting-ChatGPT,-Claude,-Other-AI-Tools-To-Scientific-Literature — accessed 2026-09-02. [type: primary-doc] — Second ground for the taxonomy caveat: Scite MCP works with "ChatGPT, Claude, Copilot, Cursor, Claude Code, and any tool that supports the Model Context Protocol", exposing "over 250 million scientific articles" with citation contexts.

<!--
Counts for this session (checked 2026-07-27; revised 2026-09-02):
  total entries          = 31   (target >= 15)
  peer-reviewed/preprint = 12   ([4] [5] [6] [7] [8] [9] [11] [12] [13] [14] [15] [18]; target >= 5)
  policy                 = 2    ([25] [26])
  primary-doc            = 14   ([1] [2] [3] [10] [19] [20] [21] [22] [23] [27] [28] [29] [30] [31])
  news                   = 3    ([16] [17] [24]) — corroboration / tool discovery only, never sole grounds
-->

### Dropped during research (and why)

- *"LLM Hallucination Rate Up to 82%: 40+ Stats (2026)"* (sqmagazine.co.uk) and *"Latest AI Hallucination Rates & Benchmarks … July 2026"* (suprmind.ai) — SEO statistics roundups with no traceable methodology. Dropped; the equivalent claims are grounded in [5], [7] and [8] instead.
- *"Citation Hallucination Rates in LLM-Generated Research: A 2025 Benchmark Across 10 Models"* (hub.stabilarity.com) — aggregator blog restating figures from primary studies without linking to them. Dropped in favour of [7].
- Chroma's "context rot" technical report and the many secondary write-ups of it — a vendor technical report rather than peer-reviewed work; the same teaching point is made from the peer-reviewed [9] instead.
- Blog comparisons of Elicit / Consensus / Semantic Scholar (paperguide.ai, thedrive.ai, agentledgrowth.com and similar) — affiliate-style listicles. Dropped; the taxonomy slide uses each tool's own documentation ([19], [20], [21], [22]).
- The claim "~65% of enterprise AI failures in 2025 were attributed to context drift" — traced only to secondary marketing pages with no underlying study. Dropped entirely; no slide makes this claim.
- Corpus size figures for research tools quoted in secondary comparisons (e.g. "Elicit indexes 138M papers") conflicted with the vendors' own current pages. Only the vendors' own current figures are used, and they are labelled as vendor-stated.

---

*AI for Researchers · Session 1: AI Foundations for Researchers · Landscape as of August 2026*