Human knowledge practice questions.

The AlphaFold Protein Structure Database, released by DeepMind with EMBL-EBI from 2021 and extended to over 200 million entries in 2022, supplies a predicted structure for almost every protein known to science. A structural biologist argues that it cannot replace laboratory determination. Which reason best supports that argument?

• A prediction is a hypothesis until an experiment tests it
• The database may be searched only by laboratories holding a paid subscription to it
• The predictions cover only those proteins whose structures had already been solved
• The files cannot be opened by the visualisation software that laboratories already use

Answer: A prediction is a hypothesis until an experiment tests it

AlphaFold returns a confidence-scored model, so it moves where experimental effort is spent rather than removing the need for it, and treating a prediction as a finding is the standard error. Access is free and coverage is wide, which is why the status of the claim, not the size of the database, is what a scientist disputes.

Campaigns such as WikiProject Women in Red, started in 2015, report that fewer than one in five biographies on English Wikipedia are about women. Which feature of Wikipedia's own rules best explains how this gap is reproduced?

• Only editors who hold a recognised academic qualification may write biographies
• The site ranks biographies of women lower in its internal search results
• Notability must be shown from existing published coverage, which is itself uneven
• The site's licence prevents biographies from being translated into another language edition

Answer: Notability must be shown from existing published coverage, which is itself uneven

Wikipedia deliberately follows published sources instead of leading them, so an existing gap in press and academic coverage is inherited rather than created by the encyclopaedia. The implication is that a reference work can be open to every editor and still reproduce a selection made elsewhere.

In June 2023 a court in New York sanctioned lawyers who had filed a brief citing decisions that did not exist; the citations had been produced by a chatbot and were never checked. Which conclusion about knowledge produced by a large language model does the case support?

• Text is generated for plausibility, so verification is a separate step
• The model had been trained on a database of court records that was out of date
• The model reproduced passages word for word from decisions it had memorised
• The model's answer was correct until citation software reformatted the references

Answer: Text is generated for plausibility, so verification is a separate step

A language model is optimised to produce text that reads like its training corpus, not to confirm that a referenced item exists, so fluent output and verified output are different things. The examinable point is where responsibility moved: the tool shifted the work from drafting to checking, and the checking was skipped.

A search engine answers a health query with a single extracted box at the top of the results page instead of a list of links. Which implication for human knowledge does this raise most directly?

• The pages from which the text was taken are removed from the search index
• The extracted answer cannot be corrected once it has been generated for a query
• The engine must license each publisher's text before any link may be displayed
• Disagreement between sources is hidden behind one settled-looking answer

Answer: Disagreement between sources is hidden behind one settled-looking answer

The box does not change what the sources say; it changes how many of them a reader ever sees, which moves editorial power to whoever decides what is extracted. The impact is a faster answer, the implication is that the plurality on which verification depends becomes invisible.

The BBC Domesday Project of 1986 stored community surveys on videodiscs readable only by one model of microcomputer with a matching disc player; by the mid-1990s working readers were scarce, and a rescue project recovered the content in 2002. Which risk to the knowledge record does this illustrate?

• Magnetic media lose the charge that holds the data within about ten years
• Hardware and formats can vanish while the medium is still readable
• Copyright law can forbid an archive from keeping material it has collected
• Encryption keys are lost when the team that created an archive disperses

Answer: Hardware and formats can vanish while the medium is still readable

The discs had not decayed; the machines and the software that understood them had gone, which is why preservation is a continuing migration rather than a single act of storage. Digitising material therefore creates a maintenance obligation instead of removing one.

Researchers showed in a 2024 Nature paper that when language models are trained largely on text produced by earlier models, output quality degrades over generations - so-called model collapse. As AI-written text spreads across the web, the finding points to the implication that

• search engines will remove AI-written pages from their indexes
• human writing will always remain identifiable by automated detection tools
• a web that feeds on its own output erodes as a record of human knowledge
• future language models will need far less training data than current ones

Answer: a web that feeds on its own output erodes as a record of human knowledge

Knowledge institutions depend on provenance: if new text is generated from old generated text, errors persist and diversity narrows with no external check against the world. The distractors promise remedies — reliable detectors, wholesale index purges — that the finding does not support and that would not restore provenance in any case.