Retrospective start-here guides / 100 guides
Launches, documentation, regulator guidance, incidents, studies and the older precedents behind everyday AI use, each read from the task outward: what the tool does, what to check, and where it fails.
Historical event dates and source dates are separate from the preparation date of this local edition. Every guide is a retrospective draft prepared 16 September 2026; none was published on its historical date.
100 guides
Recovered ELIZA source code and the restoration project's own account show why a scripted reply can feel understood without any comprehension.
A self-mocking 2001 campaign shows how Microsoft itself described the failure of its default, on-screen Office Assistant.
Google's current documentation states how autocomplete predictions are generated and moderated, and what they are not meant to be.
The 2011 iPhone 4S announcement named Siri's capabilities and its beta status the same day, a pattern worth checking in any assistant today.
OpenAI's 2020 paper and API post explain the few-shot mechanism behind the advice to put examples inside a prompt.
GitHub's 2021 and 2022 posts show why programmers, who can run and check code instantly, adopted generation first.
OpenAI's launch post shows the model behind ChatGPT predated the product and states its own reliability limitations.
Microsoft's own first-week post names long conversations, not a single bad reply, as the cause of erratic Bing chat behavior.
OpenAI's GPT-4 materials pair a bar-exam score with an explicit warning about hallucination and overreliance.
Meta's 2023 releases and license show what running a model locally changes about cost and privacy, and the rules that still apply.
Vendor documentation defines the token as the unit models read, price and limit, and a token is not a word.
Vendor documentation calls the context window a model's working memory, and describes what happens as a chat fills it.
API documentation describes temperature and sampling settings as controlled randomness, not a bug in the model.
A September 2025 research paper and vendor help pages define hallucination and explain why guessing is often rewarded.
Vendor documentation and published system-prompt text show whose instructions shape a chat before a user types.
Vendor documentation explains retrieval as grounding answers in a document's own text, with stated limits.
Help pages describe when assistants search the live web and warn that citations can be wrong or outdated.
A February 2024 rollout and later expansions show what memory keeps and how a user can see or clear it.
OpenAI's September 2024 o1 launch and current vendor docs show what extra thinking time buys and costs.
Vendor deprecation pages set retirement dates for models and show what breaks when one arrives.
The guide's current order puts message roles and structure ahead of phrasing tricks, and treats reasoning models differently.
The vendor's current reference lists clarity, context and examples before role-play, and names which advice is model-specific.
The API guide's role-constraints-context-task template differs from Workspace's persona-task-context-format sentence.
The 2022 paper tested three large models on math and reasoning tasks and tied the gain to model scale, not wording alone.
The 2020 GPT-3 paper named this in-context learning, and current vendor guides warn that too-similar examples can backfire.
The 2023 paper found a U-shaped accuracy curve by position, a limit worth checking before trusting a long prompt.
Two studies found prompt formatting and phrasing shifted accuracy sharply and inconsistently across models.
A 2023 study of 162 personas found no general accuracy gain, while a vendor guide frames roles as a tone setting.
OpenAI's help page names the failure as hallucinating, and a 2023 case study found every requested citation fabricated.
Comparing three vendors' current libraries shows a template is a starting structure to adapt, not a finished answer to copy.
Vendor file-upload pages describe what a pasted document becomes, and the checks that catch a draft's invented detail.
Two 2023 evaluations rate AI summaries highly overall, but neither measured whether every number and qualifier survived.
A 2023 GPT evaluation and Google's own 2024 language expansion both show translation quality is uneven, not uniformly good.
Vendor documentation says to review generated code and formulas before relying on them, and describes each tool's sandbox.
A 2021 study found vulnerabilities in about 40% of Copilot's suggestions, and GitHub's own survey reports faster task completion.
OpenAI's GPT-4V system card documents hallucinations testers found, and Anthropic lists where its own vision tool is approximate.
OpenAI's 2023 launch, its current voice FAQ, and the GPT-4o system card each describe a different layer of what voice changes.
A travel-planning benchmark found GPT-4 solved 0.6% of realistic requests, and OpenAI describes how its search and links work.
Usage policies classify medical guidance as high-risk, and a JAMA study of one forum found preference without testing accuracy.
A federal judge's 2023 sanctions order documents fabricated case citations from ChatGPT, and usage policies address licensed advice.
OpenAI, Anthropic and Google document different defaults and different toggles for whether a conversation trains their model.
Help pages from OpenAI, Google and Anthropic each define a different reach for temporary chats, history settings and deletion.
The Garante's March 2023 order and its April follow-up record why it paused ChatGPT and what OpenAI changed to resume service.
OpenAI's own incident report names the redis-py bug that exposed chat titles and some billing details in March 2023.
ICO guidance on AI and data protection sets a lawful-basis test and a distinct meaning for accuracy that a chatbot's fluency can hide.
The EDPB ChatGPT taskforce's 2024 report gives preliminary views on lawful basis, fairness, accuracy and data-subject rights.
Two 2023 FTC business-blog posts set out what an AI marketing claim must prove and what a chatbot-enabled deception still violates.
OpenAI, Anthropic and Google document contractual no-training defaults for paid business tiers that differ from consumer settings.
Help pages and a 2025 removal show that a shared AI chat link is public by default and was briefly made searchable.
An FTC alert says short audio clips are enough to clone a voice, and recommends verifying by calling the person back on a known number.
An FCC ruling treats AI-generated voices in robocalls as 'artificial' under a 1991 law, opening state enforcement against voice-cloning scams.
A security classification and vendor guidance describe how hidden text in a document or webpage can redirect an assistant that reads it.
Two NCSC blog posts from August 2023 explain prompt injection and data poisoning in plain terms and what organisations can do about them.
A joint CISA-NCSC guidance document sets secure-by-design practices for building AI systems, agreed globally for the first time, the agencies state.
Content Credentials and SynthID document how an image was made, but the standards say that absence of a credential proves nothing.
OpenAI, Anthropic and Google state different minimum ages and parental-control approaches in their current terms and help pages.
Google recommends 2-Step Verification for any account; Claude's own help pages describe logging in without a password at all.
Anthropic's and Google's help pages describe separate paths for flagging harmful output and for reviewing account or policy decisions.
A pre-registered study of 758 BCG consultants found large gains on tasks inside AI's frontier and worse accuracy on one task outside it.
An NBER study of 5,179 support agents found a generative AI assistant raised productivity most for less experienced workers.
A randomized experiment on 444 professionals found ChatGPT cut writing time and raised quality most for weaker writers.
A randomized controlled trial of 95 freelance programmers found GitHub Copilot cut time to build one HTTP server by 55.8%.
METR timed 16 experienced open-source developers and found AI tools increased task time by 19% even as developers believed they were faster.
Pew's American Trends Panel surveys found 34% of US adults had used ChatGPT by 2025, about double the 2023 share, with use split by age and purpose.
The Anthropic Economic Index analysed about a million Claude.ai conversations by occupational task, and Anthropic names its own method's blind spots.
An OpenAI and Harvard analysis of 1.5 million ChatGPT messages found about 70% of consumer use is non-work, mostly practical guidance.
Two controlled experiments found people over-rely on confident AI explanations and misjudge when to trust their own answer instead.
Pew and a UK government tracker found most people cannot reliably identify AI in everyday products or say how their data is used.
Vendors' own pricing pages, read on the retrieval date, show what a subscription actually adds beyond a higher message limit.
Microsoft and Google's own documentation shows what a workplace assistant can read from an account and which setting controls it.
OpenAI's GPTs and Projects and Anthropic's Projects save instructions and files for reuse, but keep each one's memory separate.
Ollama, LM Studio and llama.cpp's own documentation states the hardware a local model needs and what running it locally buys.
Google's own announcement and help page describe how AI Overviews are generated and state plainly that they can be wrong.
Perplexity's own documentation describes how it attaches numbered citations and states plainly that its answers can be wrong.
Google's NotebookLM launch post and current help pages describe a tool restricted to a user's own uploaded sources.
Microsoft Edge and Chrome document what their built-in AI sidebars can see across open tabs and how to limit it.
Developer pricing pages price the same models per million tokens, a different unit from a subscription's flat monthly fee.
Vendor status pages, read on the retrieval date, show recent incidents and make the case for a fallback plan.
NPR, Harvard and the UK government each publish a written AI-use policy, and comparing them shows what a policy actually has to cover.
The UK's January 2024 generative AI framework set ten principles for civil servants and was withdrawn thirteen months later.
OMB's March 2024 memo set minimum practices and a public opt-out for risky federal AI, before a 2025 memo rescinded and replaced it.
Nature, ICMJE and Amazon KDP each set their own disclosure line for AI-assisted writing, and the lines do not match.
JCQ and Oxford both treat undeclared AI-generated work as misconduct, with the default set to not permitted.
OpenAI withdrew its text classifier for low accuracy, and a Stanford study found detectors penalise non-native English writers most.
Two 2025 randomised trials found generative AI can help or hurt learning depending on whether it is built to tutor.
UNESCO's competency framework and the EU AI Act's Article 4 both set out what understanding AI is supposed to look like.
Article 50 of the EU AI Act, in force from August 2026, requires disclosure that content or a conversation came from AI.
OpenAI and Anthropic assign their rights in AI output to the user, but the US Copyright Office says a prompt alone is not authorship.
A routine built from vendor guidance and a fact-checking study for testing an assistant's answer before acting on it.
Grounded, reasoning and local assistants are built for different jobs, as their own documentation describes.
Regulator guidance and vendor data pages describe why other people's data and secrets need a reason before pasting.
Three vendors' own prompting guides describe the same shape: goal, input, constraints, format and a check.
Memory, activity history and model-training controls, as three vendors' own help pages describe them.
What three vendors' own model documentation states about limits, cutoffs and safety testing, for a non-expert reader.
Usage policies and an over-reliance study mark where the documentation itself says to stop relying on an assistant.
Editorial policy, a share-link FAQ and a government playbook describe what counts as a usable record of AI use.
A plain, checkable first explanation for a new user, built from three vendors' own explainer pages.
Release notes and settings pages from three vendors show why models change and configurations do not stay put.
The CNIL's 2024 guidance on generative AI tells organisations and staff what counts as safe to type into a consumer AI tool.
NIST's voluntary framework organizes AI risk management into four functions and was built with input from over 240 organizations, NIST states.
Try another category or clear your search.
