← series index
# Session 1: AI Foundations for Researchers
*AI for Researchers — attendee handout · Landscape as of August 2026*
---
## At a glance
| | |
|---|---|
| **Session** | 1 of 7 — AI Foundations for Researchers |
| **You'll learn** | How large language models actually work, why hallucination is a structural property rather than a bug, and how to judge which research tasks are safe to hand to AI. |
| **You'll practice** | Turning a weak prompt into a structured one, and verifying an AI-generated citation before you trust it. |
| **Prerequisite** | None. No coding required. |
| **Tools referenced** | ChatGPT (GPT-5.6), Claude (Fable 5 / Opus 5), Gemini (3.1 Pro / 3.7 Flash), Kimi (K3); Elicit, Consensus, Semantic Scholar, SCiNiTO, Scite, ResearchRabbit, Gemini Notebook (formerly NotebookLM); ChatGPT deep research, Gemini Deep Research — illustrative, no endorsement implied. |
---
## The one-paragraph mental model
A large language model is trained to predict the next **token** — a chunk of text roughly ¾ of a word long [1]. Nothing is stored as a retrievable fact; the training text is compressed into statistical patterns, and the answer is generated one token at a time. Current models also spend hidden "thinking" tokens before answering — extended thinking is on by default now, not a separate model class [2][10]. Everything the model can see at once lives in its **context window** — your prompt, your attachments and its own output — and nothing outside that window exists for it unless the product deliberately re-injects it (saved memory, project files) [1][2]. Because a false statement and a true one are produced by the identical process, hallucinations "originate simply as errors in binary classification" and "arise through natural statistical pressures" [4]. They persist because guessing outscores an honest "I don't know" — the newest models are tuned to answer more, not abstain more, and their vendors document the tradeoff in their own model cards [4][27][28]. **Fluent, well-formed and wrong is the expected failure mode — not a glitch.**
---
## Prompt templates
Use these as starting points; adapt the bracketed parts to your own material.
**Template 1 — the CRIT skeleton (use this for anything that matters)**
```
Context: I am a researcher in [field] working on [specific question].
I have attached [what you are supplying: papers, dataset, draft].
Role and register: write for [audience]. [Tone: neutral academic prose /
plain language for a lay reader / etc.]
Instructions and constraints:
- Use ONLY the material I have attached. Do not add outside knowledge.
- Every numeric or factual claim must be followed by (source, p. X).
- If something I ask for is not in the attached material, write
"not reported in the attached sources" — do not estimate it.
- Where sources disagree, say so explicitly rather than averaging.
- [Length / format constraint, e.g. 250 words maximum; markdown table.]
Task: [one specific request].
Then list separately the two biggest limitations of this answer.
```
*When to use:* any task whose output you will rely on or show to someone else.
*Watch out for:* the "Role" line is the weakest part. A controlled study across 162 personas and 2,410 factual questions found that adding a persona **does not improve accuracy** [14]. Use the role to set tone and audience; use context, constraints and retrieval to buy correctness.
**Template 2 — the abstention clause (paste into any prompt)**
```
If you are not certain a fact, figure, quotation or reference is correct,
say "I am not certain" and leave it out. An incomplete answer is more
useful to me than a complete one containing an invented item.
```
*When to use:* every time you ask for facts, numbers, quotations or references.
*Watch out for:* it reduces the failure rate; it does not eliminate it [4] — Anthropic's own card shows its newest model fabricating a response more often than its predecessors when a referenced source is missing (82% vs 91–94% honest handling) [28]. Still verify.
**Template 3 — the source-bounded reading prompt**
```
Answer using only the document(s) I have attached. For every claim, quote
the sentence you took it from and give its page or section. If the answer
is not in the document, say so. Do not use general knowledge.
```
*When to use:* asking questions about a paper, protocol or dataset you have uploaded.
*Watch out for:* long documents are not read evenly — accuracy is highest at the beginning and end of a context window and degrades significantly in the middle, even in long-context models (a 2023-era measurement with no current-generation replication; assume it still applies) [9]. Put the critical passage near the top or bottom, and prefer five well-chosen papers to fifty.
---
## The Capability/Failure Map for Research Tasks
The series' standing reference. "Safe" means safe to **draft** with — it never means safe to skip reading.
| Zone | Typical research tasks | Why it lands here | Your obligation |
|---|---|---|---|
| **Safe** <br>*draft with it, still read it* | Polishing, translating and restructuring text you wrote; brainstorming search terms; explaining a concept you can check against a textbook | You supply the ground truth, so errors are visible to you | Read every line. Disclose AI writing assistance where the journal requires it [25] |
| **Unreliable** <br>*needs checking, every time* | Summarising a paper you have; first-pass title/abstract screening; drafting cross-paper extraction tables; generating analysis code | Evidence is mixed and task-dependent; across 37 studies of LLMs in systematic reviews (2025 scoping review), fully validated applications remain rare [18] | Spot-check against the source. Never let AI be the only pass [18] |
| **Dangerous** <br>*do not delegate* | Asking for references from memory; treating an AI synthesis as evidence; uploading confidential manuscripts, grant applications or unpublished data | Fabrication is structural, not an accident — vendors document the accuracy-vs-hallucination tradeoff in their own model cards [5][7][27][28]; funders prohibit AI in peer review [26] | Retrieve citations from a database instead. Keep confidential material out of these tools [26] |
**Three rules for using the map**
1. **Could you catch the error?** If you could not spot it, the task is not in the Safe zone.
2. **Rarity moves tasks rightward.** Citation accuracy fell 27.7 percentage points from popular to recent papers in a 2026 benchmark [7].
3. **Grounding changes the zone, not the obligation.** Retrieval cuts fabrication; it does not remove it [7][8].
---
## Tool comparison — the 2026 tool landscape taxonomy
| Tier | Best for | Access / cost | Key limitation |
|---|---|---|---|
| **Tier 1 — General chatbots** <br>ChatGPT (GPT-5.6), Claude (Fable 5 / Opus 5), Gemini (3.1 Pro / 3.7 Flash), Kimi (K3) | Language work: drafting, rewriting, translating, explaining, brainstorming | Free tiers plus paid plans; often available through institutional licences | Leans on memory even when search is available [7]. The worst possible source of citations [5][7] |
| **Tier 2 — Research-specific tools** <br>Elicit, Consensus, Semantic Scholar, SCiNiTO, Scite, ResearchRabbit | Finding and comparing real papers; structured extraction across many papers | Mostly freemium; several offer institutional subscriptions | Bounded by corpus coverage; paywalled full text is often unavailable to the tool [19][20][21] |
| **Tier 2b — Source-grounded notebooks** <br>Gemini Notebook (was NotebookLM, renamed July 2026) | Q&A restricted to documents you upload yourself | Free tier; enterprise/education tiers | Grounded in *your* corpus only — it cannot tell you what your corpus is missing [22] |
| **Tier 3 — Deep-research agents** <br>ChatGPT deep research, Gemini Deep Research | A fast, structured first map of an unfamiliar area; lead generation | Bundled with paid chatbot plans; API access | Produces more citations per query *and* hallucinates citation URLs at a higher rate than search-augmented chatbots — 3–13% fabricated, 5–18% non-resolving [8] |
*Landscape as of August 2026. One caveat: Tier-2 tools such as Consensus and Scite now ship MCP servers, so a Tier-1 chatbot can call them directly [30][31] — read the tiers as what you are reaching for, not walls. Tool capabilities and pricing change quickly — verify before relying on this table. Examples are illustrative, not endorsements.*
---
## Checklist — "spot a hallucination" before you rely on AI output
Run these in order. Stop at the first failure.
- [ ] **Grounding** — Was anything actually retrieved? If the tool did not search a real corpus, no claim in the output has been checked against anything.
- [ ] **The DOI test** — Paste every DOI into `https://doi.org/`. "DOI Not Found" is the fastest single indicator of a fabricated reference.
- [ ] **The quoted-title test** — Search the exact title in quotation marks in Google Scholar or a Tier-2 tool. Zero results, or a *different* paper by that author, means fabrication.
- [ ] **Real person, invented paper** — Fabricated references usually name real researchers who genuinely work in the field. Confirming the *author* exists proves nothing; confirm the *paper* exists.
- [ ] **Claim-to-source** — Open the source and confirm it actually says what the output says it says. A real citation attached to a claim it does not support is the failure mode that survives every automated check.
- [ ] **Volume, year and pages** — Check them against the journal's own archive. Fabrications frequently attach plausible metadata to a volume that does not exist for that year.
- [ ] **Omission** — Ask what was left out. Summaries lose caveats, limitations and sample restrictions first [18].
- [ ] **Disclosure** — If any of this text reaches a manuscript, check the journal's requirement: ICMJE-aligned journals require authors to disclose the use of AI-assisted technologies at submission, and to report AI writing assistance in the acknowledgements [25]. *(Session 4 covers this in depth.)*
- [ ] **Confidentiality** — Never paste a manuscript under review, a grant application, or unpublished participant data into a general AI tool. NIH prohibits generative AI in its peer-review process, and uploading application content violates its confidentiality requirements [26].
---
## Further reading
- Kalai, Nachum, Vempala & Zhang (2025). *Why Language Models Hallucinate*. arXiv:2509.04664 — https://arxiv.org/abs/2509.04664 — the clearest available account of why hallucination is structural rather than a fixable bug; readable without a machine-learning background.
- Walters & Wilder (2023). *Fabrication and errors in the bibliographic citations generated by ChatGPT*. Scientific Reports 13, 14045 — https://doi.org/10.1038/s41598-023-41032-5 — the study that put hard numbers on fabricated references; short, open access, and the single best paper to hand a sceptical colleague.
- Zhao et al. (2026). *LLM hallucinations in the wild: Large-scale evidence from non-existent citations*. arXiv:2605.07723 — https://arxiv.org/abs/2605.07723 — an audit of 111 million references showing hallucinated citations are already embedded in the published record.
- Liu et al. (2024). *Lost in the Middle: How Language Models Use Long Contexts*. TACL 12 — https://aclanthology.org/2024.tacl-1.9/ — why a million-token context window does not mean the model read your document evenly; measured on the 2023-era cohort, and still the reason to place key passages deliberately.
- Lieberum et al. (2025). *Large language models for conducting systematic reviews: on the rise, but not yet ready for use — a scoping review*. J Clin Epidemiol 181, 111746 — https://www.jclinepi.com/article/S0895-4356(25)00079-4/fulltext — a sober, methodologically careful survey of what LLMs can and cannot yet do in evidence synthesis.
- Schulhoff et al. (2025). *The Prompt Report: A Systematic Survey of Prompting Techniques*. arXiv:2406.06608 — https://arxiv.org/abs/2406.06608 — the reference work if you want to go beyond CRIT; 58 catalogued text-prompting techniques.
*Full annotated source list for this session (31 entries, with the exact claim each one grounds): see `sources.md`.*
---
*AI for Researchers · Session 1: AI Foundations for Researchers · Landscape as of August 2026*