Skip to content

Stockertim Dictionary

Artificial intelligence

Using AI well at work: how models work, prompting, workflows and agents, applied AI in finance, accounting, design, content, marketing and data, responsible AI and security (added 3 Nov 2026 with the AI at work courses). 235 terms, page 1 of 2.

This part of the Stockertim Dictionary explains 235 terms from artificial intelligence in plain English. A full entry gives the definition first, then how the term is used in practice and the free lessons that teach it.

Topics in this subject

  • Prompting and prompt engineering
  • Models and how they work
  • AI at work and productivity
  • Applied AI in business functions
  • Agents and automation
  • Evaluation and quality
  • Responsible AI and governance
  • AI safety and security
  • Abstractive summaryAn abstractive summary is a summary written in new words that paraphrases the source, reading smoothly but able to drift from what the source actually said.
  • Acceptable use policyAn acceptable use policy is an organisation's written rules on how staff may and may not use a technology, including which tools and which kinds of data are al…
  • Action itemAn action item is a task agreed in a meeting with a clear description, a named owner who accepted it and a due date, so progress can be followed up.
  • Agent specificationAn agent specification is a short written design for an AI agent setting its goal, inputs, allowed tools, step and spend caps, stop conditions, approvals, esca…
  • AI agentAn AI agent is an AI system that plans and carries out tasks of several steps by itself, choosing which permitted tools to use and when, in pursuit of a goal a…
  • AI coding assistantAn AI coding assistant is an AI tool that suggests, explains or writes computer code, from single lines to whole scripts, based on a description or existing co…
  • AI governanceAI governance is the set of policies, roles, registers and review processes an organisation uses to oversee how it buys, builds and uses AI.
  • AI impact assessmentAn AI impact assessment is a structured review, done before an AI use starts, of its purpose, the people it affects, the likely risks and the safeguards needed.
  • AI incidentAn AI incident is an event in which an AI system causes or nearly causes harm, a data breach, a security compromise or a serious error.
  • AI labelA visible notice telling the audience that a piece of content was generated or substantially changed using artificial intelligence.
  • AI supply chainThe AI supply chain is the set of providers, models, data sets, tools and services that an AI system depends on to work.
  • AI transparencyAI transparency means being open about when and how AI is used, what it does, what data it relies on and what its limits are.
  • AI use disclosureAn AI use disclosure is a statement telling readers or clients that AI tools helped produce a piece of work, briefly what they did and what people checked.
  • AI-assisted workAI-assisted work is work a person produces with help from an AI tool, such as a draft, summary or analysis, which the person checks, edits and takes responsibi…
  • Alert fatigueAlert fatigue is the loss of attention that follows when people receive too many alerts, most of them false, so genuine warnings are missed or rushed.
  • Analysis logA running record of the data, steps, code, prompts and decisions in an analysis, kept so that someone else can repeat the work and check it.
  • Anomaly detectionAnomaly detection is finding records that differ markedly from the normal pattern of the data, using rules, statistics or machine learning, so a person can loo…
  • Approval gateAn approval gate is a point in a workflow where an action waits until a named person approves it, placed before steps that are costly, visible to outsiders or…
  • Approved tools listAn approved tools list is a short list, kept by a business, of the AI and software tools staff may use for work because someone has checked their terms.
  • Artificial intelligenceThe field of building computer systems that perform tasks normally needing human judgement, such as recognising patterns, understanding language, making predic…
  • Attack surfaceThe attack surface is the full set of points where an attacker could try to get into, misuse or extract data from a system.
  • Automated biddingSoftware that sets the bids for ad placements automatically, auction by auction, to reach a goal the advertiser states, such as conversions within a budget.
  • Automatic captioningUsing speech recognition software to produce captions for video or audio, which a person then reviews and corrects before publishing.
  • Autonomy levelAn autonomy level describes how much an AI system may decide and do without a person approving each step, ranging from only making suggestions to acting entire…
  • Baseline measurementA baseline measurement records how a process performs before it is changed, such as time per item, error rate and cycle time, so the effect of the change can b…
  • Boundary valueAn input that sits exactly on a threshold in a rule, such as a sale of exactly 200 where a bonus starts at 200, the point at which formulas most often go wrong.
  • Brand suitabilityWhether the content an ad appears next to fits a particular brand's own standards and tone, beyond the basic safety limits most advertisers share.
  • Business intelligenceThe tools and practices that gather an organisation's data and turn it into reports and dashboards that people use to monitor performance and make decisions.
  • Callback verificationCallback verification is confirming a payment request or a change of bank details by phoning the person or company on a number you already hold, never a number…
  • Chain-of-thought promptingChain-of-thought prompting means asking a model to set out its intermediate steps before giving a final answer, which can help on multi-step problems and makes…
  • ChatbotA computer program that holds a conversation in text or speech, answering questions either from fixed scripts or by generating replies with a large language mo…
  • Churn predictionEstimating which current customers are likely to stop buying or cancel, usually with a model that learns from patterns seen before past cancellations.
  • Claim logA list of every factual claim in a piece of content, with the source used to check it, who checked it and whether it was confirmed, corrected or removed.
  • Claim substantiationHaving evidence, such as test results, data or a fair comparison, that proves a marketing claim is true before the claim is published.
  • Close checklistA close checklist is the list of tasks, owners, deadlines and review steps an accounting team works through to close the books at the end of each period.
  • ClusteringAn unsupervised learning method that groups items so that those in the same group are more alike than those in other groups, for example shares whose returns t…
  • Code reviewCode review is a structured read of code by a person before it is used, to find errors, security risks and unclear parts, and to confirm it does what was inten…
  • Compounding errorCompounding error is the way small error rates at each step of a task multiply together, so that a task of many steps fails far more often than any single step…
  • ComputeCompute is the processing power, usually from specialised chips in data centres, used to train and run AI models, and a large driver of their energy use and co…
  • Concept directionA distinct creative route, with its own central idea and look, that a designer explores and presents before one is chosen for development.
  • Conditional sumAdding up only the values that meet one or more stated conditions, for example the amounts of invoices that are both paid and in one region.
  • Confidence scoreA confidence score is a number a model attaches to a result to show how sure it is, often used to send low-confidence items to a person for review.
  • Confidential informationConfidential information is information an organisation or person is entitled to keep private, such as client data, staff records, contracts or results not yet…
  • ConnectorA connector is a ready-made link between an AI application or workflow tool and another system, handling the connection and sign-in so that system's data and a…
  • Content atomisationBreaking one substantial piece of content, such as a webinar or report, into many smaller pieces adapted for different formats and channels.
  • Content credentialsMetadata attached to an image or file that records how it was made and edited, including whether and how AI tools were used.
  • Content lifecycleThe stages a piece of content passes through, from the first idea and research to making, editing, publishing, distribution and review of how it performed.
  • Context windowThe context window is the maximum amount of text, measured in tokens, that a language model can take into account at once, including instructions, pasted mater…
  • Contrast ratioA measure of the difference in brightness between text and its background, used in published accessibility guidelines to judge whether text is readable.
  • Control totalA total calculated independently from the source data and compared with the processed or summarised data to confirm that nothing was lost or counted twice.
  • Copy variantAn alternative version of an ad, headline or message, written so that it can be tested against other versions to see which works better.
  • Cost per runCost per run is what one complete run of an automation costs, adding model usage for text sent and generated, per-action fees and an allowance for retries.
  • Cross-validationTesting a model several times, each time holding out a different part of the data for evaluation, to estimate how well it performs on data it has not seen.
  • Data clean roomA secure environment where two organisations can analyse their combined data without either seeing the other's raw customer records.
  • Data dictionaryA document that lists each field in a data set with its meaning, its type and the values it is allowed to hold, so everyone reads the data the same way.
  • Data leakageWhen information a model would not really have at prediction time, from the future or from the answer itself, slips into its training data or features and make…
  • Data loss preventionData loss prevention is the set of tools and rules that detect and block sensitive data leaving an organisation through email, uploads, prompts or other channe…
  • Data minimisationData minimisation is the principle of collecting, using and sharing only the personal or confidential data actually needed for a specific purpose, and no more.
  • Data poisoningData poisoning is the corruption of the data a model learns from, so that the model behaves wrongly, either generally or in specific situations.
  • Data storytellingExplaining what data shows through a clear message, supporting charts and enough context for the audience to understand the finding and act on it.
  • Data validation ruleA rule that checks each value meets a condition, such as falling within a date range or appearing on an allowed list, so that bad entries are caught.
  • Decision logA decision log is a record of a decision that notes the question, the evidence and tools used, the checks made, who decided and why.
  • Decision pointA decision point is a step in a process where the path divides according to a condition or a judgement, such as whether an invoice matches its purchase order.
  • DeduplicationFinding and removing repeated records in a data set, after first deciding which fields must match for two rows to count as the same record.
  • Defence in depthDefence in depth is the practice of using several independent layers of protection, so that one failure does not expose everything.
  • DelimiterA delimiter is a marker, such as triple quotes or a heading, that separates your instructions from pasted material in a prompt so the model can tell them apart.
  • Design briefA written statement of a design job that sets out the goal, audience, requirements, constraints, budget and deadline agreed with the client.
  • DeskillingDeskilling is the gradual loss of skill that can follow when people hand a task to a tool and stop practising it, leaving them less able to check or step in.
  • Diffusion modelA type of generative model that creates an image by starting from random noise and removing that noise step by step, guided by a written prompt.
  • Document extractionDocument extraction is pulling specific facts or figures out of a document, such as revenue or a reporting date, into a structured form like a table.
  • Dynamic contentParts of an email, web page or ad that change automatically for each viewer, such as a greeting, a product block or a local address.
  • Editorial sign-offThe final approval of a piece of content by a named, responsible person before it is published, recorded so the team knows who approved it.
  • EmbeddingAn embedding is a long list of numbers that represents the meaning of a piece of text, so that texts with similar meanings end up with similar lists even when…
  • Error pathAn error path is the branch a workflow follows when a step fails, typically retrying, moving the item to a queue for people and alerting a named owner, so fail…
  • Evaluation rubricAn evaluation rubric is a written set of criteria, with a scoring rule for each, used to judge AI outputs consistently across many cases and reviewers.
  • Exception queueAn exception queue is the list of items an automated process could not handle confidently or that broke a rule, held for a person to review, correct and approv…
  • Exception rateThe exception rate is the share of items in a process that cannot follow the normal path and need extra handling, expressed as a percentage of all items handle…
  • Excessive agencyExcessive agency is giving an AI system more permissions, tools or autonomy than its task needs, which widens the harm a mistake or a misleading input could ca…
  • Expected outputAn expected output is the correct result for a test case, written down before the system runs, against which the actual output is compared and scored.
  • Exploratory data analysisA first look at a data set to understand its shape, quality and patterns, such as gaps, odd values and spread, before any formal analysis begins.
  • Exponential backoffExponential backoff is a retry pattern in which the wait between attempts multiplies each time, for example 10 seconds, then 30, then 90, giving a struggling s…
  • Extractive summaryAn extractive summary is a summary built from sentences or phrases taken directly from the source, so every line can be traced back exactly.
  • Fabricated citationA fabricated citation is a reference produced by a model that looks genuine but points to a source that does not exist, or to a real source that does not say w…
  • Fact-checkingVerifying the factual claims, numbers, quotes, names and dates in a piece of content against reliable sources before it is published.
  • Fairness metricA fairness metric is a number that compares how a model or process treats different groups, such as the ratio of their approval or selection rates.
  • Fake reviewA review that does not reflect a genuine customer's real experience, including reviews written by staff, paid for, swapped or generated with AI.
  • False negativeA miss: a case that really is positive, such as a fraudulent payment or a real attack, that a test or detection system wrongly classes as negative.
  • False positiveA false alarm: an innocent case that a test or detection system wrongly flags, such as a genuine payment blocked as fraud.
  • FeatureIn machine learning, one measurable input a model uses to make a prediction, such as a share's past one-month return or a company's debt-to-equity ratio.
  • Feature engineeringTurning raw data into useful model inputs, for example converting a price series into returns, volatility or ratios, and aligning each value with the time it b…
  • Few-shot promptingFew-shot prompting means including a small number of worked examples in a prompt so the model can follow the pattern, labels and format they show.
  • Field mappingField mapping is the workflow setting that says which value from one step or system fills which field in the next, including any change of format such as dates…
  • Formula change logA simple record of each change made to a workbook's formulas, saying what changed, where, why, whether AI drafted it and who checked it.
  • Formula test sheetA small, separate sheet of made-up rows, each with its correct result worked out by hand, used to try a new or AI-drafted formula before it touches real data.
  • Full-population testingFull-population testing is testing every item in a data set against chosen criteria, rather than a sample, so no item meeting those criteria is missed.
  • Function callingFunction calling is a structured way for a model to request a specific software function by name with named inputs, which the surrounding software checks and r…
  • Fuzzy matchingFuzzy matching is matching records that are similar but not identical, such as a slightly different payee name, reference or amount, usually producing proposed…
  • Generative AIArtificial intelligence that creates new content, such as text, images, audio, video or computer code, in response to a prompt, based on patterns learned from…
  • Go/no-go checklistA go/no-go checklist is a short list of yes-or-no questions answered before an automation is built or launched, where certain answers stop the project or chang…
  • GroundingGrounding means tying a model's answer to supplied sources so that each claim can be traced to a specific passage and checked by a person.
  • GuardrailA guardrail is a rule or check that limits what an AI system may say or do, such as blocking a topic, capping a payment or requiring approval before an action.
  • Human checkpointA human checkpoint is a fixed point in a workflow where a named person must review and approve the work before it moves to the next step or leaves the team.
  • Human handoverHuman handover is passing a customer conversation from an automated assistant to a person, together with the full conversation history, whenever the assistant…
  • Human on the loopHuman on the loop is an arrangement where people supervise an AI system as it runs and intervene when needed, rather than approving every action first.
  • Indirect prompt injectionIndirect prompt injection is an attack in which hidden instructions are planted in content an AI system reads, such as a web page, document or email.
  • Inference (privacy)In privacy, inference means deriving new and possibly sensitive information about a person, such as health or beliefs, from other data about them.
  • InpaintingRegenerating a selected part of an image, such as an object or a patch of background, while keeping the rest of the picture unchanged.
  • Insecure output handlingInsecure output handling is passing AI output into other systems, such as browsers, databases or scripts, without checking it, so that it can cause harm.
  • Invented functionA function name that an AI assistant writes into a formula although the spreadsheet program has no such function, which usually shows up as a #NAME? error.
  • Irreversible actionAn irreversible action is one that cannot be fully undone once taken, such as sending money, deleting records or messaging customers, and so needs a person's a…
  • Iterative promptingIterative prompting means improving an output over several rounds of targeted follow-up instructions, each aimed at one specific fault in the previous draft.
  • JailbreakA jailbreak is an attempt to get an AI system to ignore its safety rules or usage limits, usually through framing, persuasion or disguised requests.
  • Journal entry testingJournal entry testing is audit analysis of accounting journal entries to identify unusual or higher-risk postings, such as entries at odd times, by unexpected…
  • Knowledge cut-offThe date after which an AI model has no information from its training, so it may not know about recent events, prices, rules or products.
  • Known-answer testA check that runs code on a small data set whose correct result has been worked out by hand, so any difference between the two exposes a mistake.
  • LabelIn supervised machine learning, the known answer each training example carries and the model learns to predict, such as whether a loan was repaid or next month…
  • Large language modelAn AI model trained on very large amounts of text to predict the next word, which lets it answer questions, summarise documents, translate and write in a human…
  • Lateral readingLateral reading is checking a claim by leaving the source and looking at what other independent sources say about the claim and about the source itself.
  • Lead scoringRanking potential customers by how likely they are to buy, by adding points for signals using rules or by using a model trained on past leads.
  • Line editEditing that works through a text sentence by sentence to improve clarity, rhythm, word choice and tone, without changing the overall structure.
  • Live samplingLive sampling is the regular checking of a small random selection of a running system's real outputs against their source documents, to catch errors the test s…
  • Local search listingA local search listing is a business's entry in a map or local search service, giving its name, address, phone number, hours and description.
  • Lookup functionA spreadsheet function that finds a value in a table, such as a product code, and returns related information from the same row, such as its price.
  • Machine learningA branch of artificial intelligence in which software learns patterns from example data and uses them to make predictions or decisions, rather than following r…
  • Manual fallbackA manual fallback is the documented way people carry on a process by hand when its automation is paused, broken or rolled back, so work continues without the s…
  • Mass-produced contentLarge numbers of pages produced quickly and mainly to attract search traffic, with little original value for the people who read them.
  • Match rateMatch rate is the share of items an automated reconciliation paired successfully, found by dividing the matched items by all items.
  • MCP serverAn MCP server is a program that offers tools, resources or prompt templates to AI applications using the Model Context Protocol, usually wrapping one system su…
  • Model Context ProtocolThe Model Context Protocol is an open standard that defines how AI applications connect to outside tools and data through servers offering tools, resources and…
  • Model validationModel validation is independent testing that a model works as intended and is fit for its use, carried out before the model is used and repeated during use.
  • Multi-armed banditAn adaptive testing method that shifts more traffic towards the better-performing versions as results come in, while still giving weaker versions a small share.
  • Natural language processingThe branch of artificial intelligence that lets computers read, interpret and produce human language, in text or speech, for tasks such as translation, summari…
  • Natural language queryA question put to a data tool in everyday words rather than code, which the tool translates into a query and answers from the data it holds.
  • Negative promptA list of things an image generator should avoid putting in the image, such as blur, extra text or a cluttered background.
  • Neural networkA machine learning model made of layers of connected mathematical units that adjust their connection strengths during training, loosely inspired by neurons in…
  • News triageNews triage is sorting incoming news by relevance and urgency so that people read the stories that matter most first and less important items wait.
  • No-code automationNo-code automation is building automated workflows with visual tools, choosing triggers, conditions and actions from menus and connecting them, instead of writ…
  • Non-human identityA non-human identity is an account, key or credential used by software, such as an AI agent or automated script, rather than by a person.
  • ObservabilityObservability is being able to see what an automated system did and why, from its logs, traces and metrics, so that problems can be found, explained and fixed.
  • Off-by-one errorA common coding mistake in which a loop, range, index or count starts or stops one step too early or too late, so an item is skipped or counted twice.
  • Optical character recognitionOptical character recognition is technology that reads images of printed or handwritten text, such as scanned invoices or receipts, and turns them into text a…
  • OrchestrationOrchestration is coordinating several models, tools and steps so they run in the right order, pass information between them, handle failures and keep a record…
  • OriginalityThe quality of work that adds new ideas, experience, evidence or expression of its own, rather than restating or rearranging what already exists.
  • Outline-first draftingA way of working where the structure and key points of a piece are agreed before any full text is written, so drafting follows a clear plan.
  • OutpaintingExtending an image beyond its original edges with generated content that continues the scene, for example to fit a wider format.
  • Output constraintAn output constraint is a limit set in a prompt on the length, scope, structure or content of an answer, such as a word range or a list of topics to leave out.
  • Output verificationOutput verification is checking an AI output against sources, independent calculations and common sense before relying on it, with effort matched to the stakes.
  • Over-relianceOver-reliance is depending on a tool or source more than its proven reliability justifies, for example accepting its outputs without the checks its error rate…
  • Package hallucinationPackage hallucination is an AI tool suggesting a software package that does not exist, a name that someone may later register with harmful code.
  • Pairwise comparisonPairwise comparison judges two versions of a system by running both on the same cases under the same settings and comparing their results case by case, not onl…
  • Pass rateA pass rate is the share of items, test cases or people that meet the standard of a test or check, worked out as the number passed divided by the number tested.
  • People-first contentContent created mainly to help a real reader with a real need, rather than mainly to rank in search results or attract clicks.
  • Permission scopeA permission scope is a named slice of access granted to an application or connector, such as reading orders or creating draft bills, so it receives only the r…
  • PersonalisationTailoring marketing content or offers to a person or a group of customers based on data about them, such as what they bought or said they prefer.
  • Polling triggerA polling trigger starts a workflow by checking a source on a fixed schedule, such as every five minutes, and launching one run for each new item found since t…
  • Predictive analyticsUsing historical data with statistical or machine-learning models to estimate future outcomes, such as which customers are likely to buy, leave or respond to a…
  • Predictive modelA predictive model is a model trained on past data to estimate a specific outcome, such as the chance that a loan defaults or a payment is fraudulent.
  • Process miningProcess mining is analysing the event logs of business systems to reconstruct how a process really runs, case by case, before deciding how to improve or automa…
  • PromptA prompt is the set of instructions and material you give an AI model in order to get a response, from a short question to a long brief with pasted documents.
  • Prompt chainingPrompt chaining means splitting a job into a planned sequence of prompts, where the output of each step becomes the input to the next.
  • Prompt engineeringPrompt engineering is the practice of designing, testing and improving prompts so that an AI model's outputs are reliably useful for a particular task.
  • Prompt injectionPrompt injection is text crafted to override or hijack the instructions an AI system was given, either typed directly or hidden in content the system reads.
  • Prompt libraryA prompt library is a shared, maintained collection of tested prompt templates, each with a name, version, owner, test results and notes on when to use it.
  • Prompt templateA prompt template is a reusable prompt with marked fields to fill in for each new case, so the tested wording stays the same while the details change.
  • Prompt test setA prompt test set is a fixed collection of inputs, each with an expected result, used to compare prompt versions fairly on exactly the same cases.
  • Proposal templateA proposal template is a reusable outline for quotes or proposals with fixed sections, such as the problem, the work, the timeline and the price, filled in afr…
  • ReadabilityReadability is how easily a reader can understand a text, shaped by sentence length, choice of words, structure and layout rather than by the topic alone.
  • Recalculation checkA recalculation check is independently recomputing a figure produced by a person or a tool, from the original inputs, to confirm it is right before use.
  • Receipt captureReceipt capture is photographing, scanning or forwarding receipts so that software can read details such as the date, supplier and amount and record them in th…
  • Record to reportRecord to report is the end-to-end accounting process from recording transactions, through reconciling and closing the books, to producing financial statements…
  • Red lineA red line is a use of AI that an organisation decides it will not allow, whatever the claimed benefit, because the harm is judged too great.
  • Red teamingRed teaming is the practice of deliberately testing a system by acting as an attacker or misuser, to find weaknesses before others do.
  • Reference imageAn image supplied to a generator to guide the composition, colour, pose or style of what it produces, alongside the written prompt.
  • Regression testingRegression testing means re-running earlier tests after any change to a system, such as a new prompt, model or document, to confirm that nothing that used to w…
  • Responsible AIResponsible AI means designing, buying and using AI in ways that are fair, safe, transparent and accountable, with clear human responsibility for the results.
  • Retrieval-augmented generationRetrieval-augmented generation is a system design in which relevant documents are first searched for and fetched, then given to the model with the question so…
  • Review controlA review control is a control in which a person checks work, such as entries coded by a tool or a draft reconciliation, before it is accepted, and records the…
  • Review thresholdA review threshold is a rule set in advance for how closely AI output must be checked before use, rising with the stakes from a quick read for internal notes t…
  • Robotic process automationSoftware bots that copy the clicks and keystrokes a person makes in computer systems, so that repetitive, rule-based office tasks run automatically without cha…
  • Role promptingRole prompting means telling a model to answer from a stated perspective, such as an editor or a tutor, which shapes vocabulary and focus but adds no new knowl…
  • RollbackA rollback returns a system to its last known good version, such as the previous prompt, settings or workflow, when a change causes errors, so service recovers…
  • RoutingRouting means sending each incoming item to the step, model or person most suited to handle it, usually based on its type, difficulty or how confident a classi…
  • RunbookA runbook is a short operating guide for a live system that tells the person on duty how to check it, what each alert means and the exact steps to pause, fix o…
  • Seed (generative AI)The number that sets the starting noise of a generation, so the same seed, prompt and settings can reproduce a result in many tools.
  • Sensitive data exposureSensitive data exposure is sensitive information escaping through prompts, uploads, connected tools, logs or AI outputs to people or systems that should not ha…
  • Sentiment analysisUsing software to score text as positive, negative or neutral in tone; in finance it is applied to news, filings, social media and earnings-call transcripts, a…
  • SHAP valueA way to explain one prediction by sharing out the gap between that prediction and the model's average output among the inputs, giving each a positive or negat…
  • Silent errorA mistake that lets code or a formula run to the end without any warning yet produce a wrong result, such as numbers stored as text being joined instead of add…
  • Simpson's paradoxA pattern that appears within each of several groups but reverses or disappears when the groups are combined, usually because the groups differ in size or mix.
  • Speech-to-textSpeech-to-text is technology that converts spoken audio into written text, such as a meeting transcript, and can mishear names, numbers and who was speaking.
  • Spend capA spend cap is a hard limit on how much an automated system may spend on model use or paid actions per run, day or month, enforced outside the model so it hold…
  • Spot checkA spot check is a quick manual check of a few items, chosen deliberately, to catch errors in a larger output such as a spreadsheet column or a long list.
  • Spurious precisionSpurious precision is giving a number more exactness than the evidence supports, such as a forecast stated to the last unit when the real uncertainty is wide.
  • Stop conditionA stop condition is a rule that makes an agent end a run and hand over, for example when it reaches its step cap, finds missing data or meets a case outside it…
  • Structured outputStructured output is model output in a fixed, machine-readable shape, such as a table or a set of named fields, so another tool or a person can use it directly.
  • Style referenceAn approved example image or written set of style rules used to keep a series of generated images in one consistent look.
  • Style sheetA list of a publication's choices on spelling, preferred terms, numbers, dates and format, kept so that every piece is consistent.
  • Supervised learningMachine learning from examples that come with known answers, so the model learns to predict a category or a number for new cases.
  • SycophancySycophancy is the tendency of an AI model to agree with, praise or flatter the user rather than stay accurate, especially when the user signals a preferred ans…
  • Synthetic dataArtificial data generated to resemble real data in structure and statistics without describing real people, often used to test formulas, queries and tools safe…

Other subjects

Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.

English