Literacy Foundation

Start Where You Already Are

• Researcher use of AI tools rose from 57% (2024) to 84% (2025). ^[16]
• Use specifically for research and publication tasks: 45% → 62%. ^[16]
• 80% reach for a general chatbot; only 25% use an AI research assistant. ^[16]
• A 20-country survey of publishing academics finds use is mainly for writing, not analysis. ^[15]
• The top concern is always the same: inaccuracy — 74% call it a major barrier (2024). ^{[15] [16] [17]}

Adoption figures: Wiley ExplanAItions survey, fielded Aug 2025 — still the latest wave as of Aug 2026; publisher self-reported data ^[16]. Barrier figure: Ithaka S+R, self-reported survey data ^[17]. Peer-reviewed corroboration: ^[15].

Literacy Foundation

It Predicts the Next Token. That Is the Whole Trick.

• Training compresses an enormous body of text into statistical patterns.
• No facts are stored in a database — there is nothing to look up.
• Output is produced one token at a time, each conditioned on everything before it.
• Fluency and truth are separate properties; the model optimises the first.
• So a real citation and a fabricated one come out of the identical process. ^[4]

Literacy Foundation

Tokens: What the Model Actually Reads

• A token is a chunk of text — a word, part of a word, a space or a mark. ^[1]
• English rule of thumb: 1 token ≈ 4 characters ≈ ¾ of a word; 100 tokens ≈ 75 words. ^[1]
• Everything is counted, priced and limited in tokens — never in pages or papers.
• Tokenizers differ: the same text can produce ~30% more tokens on a newer model. ^[2]
• Practical translation: "my 40-page PDF" means "roughly 20,000 tokens".

Literacy Foundation

The Context Window: The Model's Working Memory

• The window holds everything the model can see at once: prompt, files, its own output. ^[1]
• Frontier models in August 2026 advertise 1,000,000-token windows. ^{[2] [3]}
• Extended thinking — now on by default — spends part of that window on hidden "thinking" tokens. ^{[2] [10]}
• Nothing outside the window exists — unless the product deliberately re-injects it (saved memory, project files).
• A separate limit: reliable knowledge can end months before today's date. ^{[2] [3]}

Literacy Foundation

A Big Window Is Not Perfect Recall

• Accuracy depends on where in the context the relevant passage sits. ^[9]
• Highest at the beginning and the end; it degrades significantly in the middle. ^[9]
• Measured on the 2023-era long-context cohort ^[9] — no current-generation replication exists, so assume it still applies.
• So: put what you most need used at the top or the bottom of your prompt.
• And: five well-chosen papers beat fifty dumped in "because it fits".

Literacy Foundation

Why Hallucination Is Structural

• Hallucinations "originate simply as errors in binary classification". ^[4]
• Where true and false statements cannot be told apart, errors "arise through natural statistical pressures". ^[4]
• The rarer the fact, the worse it gets: citation accuracy falls 27.7 points from popular to recent papers. ^[7]
• Vendors now document the tradeoff themselves — Anthropic's Opus 5 card: "accuracy is 11% higher than Opus 4.8, but its rate of hallucinations is also 6% higher". ^[27]
• Confident, well-formed and wrong is therefore the expected failure mode — not an anomaly.

Literacy Foundation

And Why It Persists: We Grade It In

• Most benchmarks score accuracy only, so "I don't know" scores the same as a wrong guess. ^[4]
• Models are "optimized to be good test-takers" — guessing when uncertain raises the score. ^[4]
• The tradeoff is live in 2026: Mythos 5 scored 82% on missing-reference honesty, behind its predecessors' 91–94% — the newest models are tuned to answer more, not abstain more (vendor's own card). ^[28]
• What you can do today: give the model explicit permission to abstain, in your prompt. ^[4]

Literacy Foundation

Hallucination, Measured: The Citation Case

• The 2023 baseline (GPT-3.5/GPT-4 era): of 636 citations in generated literature reviews, 55% were fabricated (GPT-3.5), 18% (GPT-4). ^[5]
• Even among the real ones, 43% and 24% respectively carried substantive errors. ^[5]
• The current generation (2026): with web search on, frontier models still produced only 50.9% fully correct BibTeX entries. ^[7]
• Deep-research agents: 3–13% of citation URLs hallucinated, 5–18% non-resolving. ^[8]
• An audit of 111M references estimates ~146,932 hallucinated citations in the 2025 literature. ^[6]

Literacy Foundation

Extended Thinking: No Longer a Separate Model Class

• The standard-vs-reasoning split has collapsed: current models spend hidden "thinking" tokens by default. ^{[2] [10]}
• Claude Fable 5: "Adaptive thinking is always on" — it cannot be switched off. ^[2]
• Effort is a dial, not a model choice: ChatGPT gives Plus/Pro users a slider for "how much effort"; Gemini exposes minimal→high. ^{[29] [3]}
• The ancestor is chain-of-thought prompting (2022, PaLM era), which lifted multi-step accuracy sharply. ^[12]
• The cost is unchanged: thinking tokens are billed as output and eat your context window. ^[10]

Literacy Foundation

What Extended Thinking Still Does Not Fix

• The statistical origin of hallucination is untouched — a careful wrong answer is still wrong. ^{[4] [27]}
• You cannot audit it: the raw chain of thought is hidden or summarised ^[2], and models "often deny intending to use" hints they demonstrably used. ^[11]
• Longer thinking is not grounding: nothing was retrieved unless a tool actually retrieved it.
• Raise the effort dial for hard multi-step work; verify facts exactly as before.

Series Artefact

Using the Map: Three Rules

• Rule 1 — Could you catch the error? If you could not spot it, the task is not in the Safe zone.
• Rule 2 — Rarity moves tasks rightward. The more specific and recent the fact, the worse the accuracy. ^[7]
• Rule 3 — Grounding changes the zone, not the obligation. Retrieval cuts fabrication but does not remove it. ^{[7] [8]}

Series Artefact

Tier 3: Deep-Research Agents

• They plan, search, read, iterate, and return a cited report over several minutes. ^[23]
• Some let you review and edit the research plan before it runs. ^[23]
• They produce substantially more citations per query than a search-augmented chatbot. ^[8]
• And they hallucinate citation URLs at a higher rate — 10.7% pooled, vs 4.8% for search-augmented models. ^[8]
• Treat the report as a well-organised lead list, never as a verified synthesis.

Systematic Prompting

CRIT: A Prompt Structure That Travels

• C — Context: your field, your question, and the material you are supplying.
• R — Role & register: who the output is for, and how it should read.
• I — Instructions & constraints: format, length, scope — and what to do when unsure.
• T — Task: one specific request, plus how you will check the result.
• Then iterate: a prompt is a first draft, not an incantation.

Systematic Prompting

What the Evidence Says About Prompting

• Prompting is a documented literature: 58 text-prompting techniques catalogued. ^[13]
• Step-by-step ("chain-of-thought") prompting produced large gains in 2022 — today's models do this by default, so save the instruction. ^{[12] [2]}
• But role prompting is oversold: 162 personas, 2,410 questions, no accuracy gain (2024 study). ^[14]
• So use roles to set register; use context, constraints and retrieval to buy correctness.
• Highest-value single instruction: permission to say "I don't know". ^[4]

Systematic Prompting

Evaluate Every Output: The Four Checks

• 1. Grounding — where did this come from? If nothing was retrieved, nothing was checked.
• 2. Citations — does every reference resolve to a real record you can open? ^{[5] [7]}
• 3. Claim-to-source — does that source actually say what the output claims it says? ^[6]
• 4. Omission — what was left out? Summaries lose the caveats first. ^[18]

Live Demo & Takeaways

What We Do Next — and What You Take Home

• Demo A: one research question, asked naively and then with CRIT, in two tools side by side.
• Demo B: we set a citation trap live, then run the verification checks — whatever the model does. ^{[5] [7]}
• You leave with: the CRIT template, the capability/failure map, the tool taxonomy, a hallucination checklist.
• Next session: literature discovery and synthesis — Tier 2 against Tier 3, head to head.