Clear language for consequential choices. Search nineteen terms, compare the distinctions, and follow each question into fiction.
19 terms
No matching term. Try a broader word, or clear the search to browse all nineteen entries.
Authentication
Authentication checks a particular claim about identity, source, or integrity. The scope depends on the system and the check. NIST’s cybersecurity glossary includes several context-dependent meanings, which is a reason to ask what has been authenticated rather than treating the word as a complete verdict. NIST CSRC glossary
An authenticated account does not make every message sent from it accurate. A signature associated with a document does not, by itself, settle every factual assertion in the document. The useful result is narrower: a stated claim has been checked by a stated process, under particular assumptions.
Imagine a workplace recording whose source and integrity have been checked. The check may support confidence in the file while leaving the person’s intention unresolved. Asking about that intention does not invalidate the check; it identifies another question.
Read with “Can authentic evidence lead to a false conclusion?” Related terms: Content provenance; Inference. Fiction: The Evidence Was Perfect.
Automation bias
Automation bias is excessive deference to an automated system. A person may give its output more weight than the available reasons justify, including when there is information that deserves further examination. NIST discusses this risk in its Generative AI Profile. NIST AI 600-1, section 2.7
It is not a synonym for trusting software. Trust can be well supported. The question is whether the person’s reliance fits the task, evidence, and consequences, and whether discrepancies can still receive attention.
Imagine a reviewer noticing an incorrect date in a recommendation but approving it because the system has usually been dependable. The prior performance matters; the current date still needs checking. A process that permits investigation and correction is more useful than a rule requiring either automatic agreement or automatic suspicion.
Read with “When does helpful AI become someone else’s decision?” Related terms: Decision support; Meaningful human oversight. Fiction: The Evidence Was Perfect.
Consent
Consent is permission for a particular arrangement. In these guides, the ethical questions concern what a person understands, which use they agree to, who will be affected, and how the arrangement can be reconsidered. This is not a universal legal definition or a statement that consent supplies every permission a service needs.
A person might agree to preserve an interview while declining to have new messages generated in their voice. A recipient might welcome access to recordings but decline incoming conversations. Those choices should be described separately before they are combined into one setting.
The digital-afterlife research by Hollanek and Nowaczyk-Basińska emphasizes the interests of both represented people and those expected to interact with simulations. Our practical examples apply that distinction without claiming that a discussion sheet resolves legal rights. Research paper
Read with “Who gets to speak for someone who has died?” Related terms: Digital legacy; Griefbot and deadbot. Fiction: Your Mother Would Have Wanted This.
Content Credentials and C2PA
Content Credentials are a way to express signed provenance information associated with digital content. C2PA, the Coalition for Content Provenance and Authenticity, develops the relevant open technical standard. A credential can include assertions about a file’s creation or editing history; what is available depends on the particular record. C2PA FAQ
A reader should ask what the credential says and what its validation establishes. An interface’s badge is not a substitute for those questions. Nor should a missing credential be treated as a complete explanation of how a file was made.
Imagine a photo with an available history showing that it was cropped. That information can help a reader assess its use. It does not reveal everything outside the crop or decide whether the accompanying caption is fair. The credential supplies evidence about the media; interpreting a claim still requires reasoning.
Read with “Can authentic evidence lead to a false conclusion?” Related terms: Content provenance; Authentication. Fiction: The Evidence Was Perfect.
Content provenance
Content provenance is information about the origin and history of a digital item. It may help a reader understand where material came from and what happened to it. The available information can be incomplete, and the method used to record or verify it matters. NIST’s digital-content-transparency report examines provenance alongside other approaches. NIST AI 100-4
Provenance answers questions about a record’s history. The interpretation attached to that record may require additional evidence. A documented sequence of edits, for example, does not determine whether the final image has been placed beside an accurate account of an event.
Imagine finding the original version of a cropped workplace photo. That might resolve what the crop omitted. It would not necessarily establish what happened earlier, who authorized the activity, or which policy applied. Each additional question should be connected to the evidence capable of answering it.
Read with “Can authentic evidence lead to a false conclusion?” Related terms: Content Credentials and C2PA; Inference. Fiction: The Evidence Was Perfect.
Decision support
Decision support supplies information or recommendations to someone making a choice. It differs from automatically executing that choice, though a single service may include both functions. NIST’s AI Risk Management Framework describes arrangements in which systems act, defer decisions to people, or provide an additional opinion. NIST AI RMF 1.0, Appendix C
The important practical question is what the arrangement permits. A suggested appointment, a provisional reservation, and a changed calendar entry have different consequences even if all appear under the same “assistant” label.
Imagine a tool forecasting the costs of opening a small shop. It can help someone see a risk without deciding how much that person values the attempt. If the plan relies on a relative’s unpaid work, the relative also has a decision to make. Better information does not eliminate the need to identify whose choices are involved.
Read with “When does helpful AI become someone else’s decision?” Related terms: Human autonomy; Meaningful human oversight. Fiction: The Right to Be Wrong.
Digital legacy
Digital legacy is used here for the digital material and relationships to services that a person leaves behind: recordings, correspondence, photographs, accounts, and other traces of a life. It is broader than a conversational recreation. An archive can remain valuable without being used to simulate its subject.
Possessing a file does not answer every ethical question about using it. A letter has a writer and a recipient; a family recording may include people with different expectations about its audience. The proposed use needs to be described before the relevant interests can be considered.
Imagine inheriting recordings of an uncle telling stories. Keeping them, letting relatives search them, and generating new advice in his voice are separate proposals. Calling all three “preserving his legacy” would leave the consequential differences unexplained. The digital-afterlife literature provides a framework for examining the people involved. Hollanek and Nowaczyk-Basińska
Read with “Who gets to speak for someone who has died?” Related terms: Consent; Griefbot and deadbot. Fiction: The Last Person Who Remembers.
Griefbot and deadbot
These terms describe AI-enabled simulations associated with deceased people. Terminology varies, and a name does not establish that a service has a therapeutic purpose, works in a particular way, or preserves the person it represents. The digital-afterlife research cited here uses “deadbot” for an AI-enabled representation of a deceased individual. Hollanek and Nowaczyk-Basińska
A simulation may be meaningful to a recipient while still producing statements the represented person never made. Distinguishing generated replies from quotations or recordings helps make that relationship understandable.
Imagine receiving birthday wishes in a late parent’s voice. The first questions are practical: who arranged the message, which material informed it, and did the recipient agree to receive it? The familiarity of the voice cannot answer those questions. Nor does an individual’s appreciation establish how another person would experience the same arrangement.
Read with “Who gets to speak for someone who has died?” Related terms: Digital legacy; Consent; Retrieval-augmented generation. Fiction: The Last Person Who Remembers.
Human autonomy
Human autonomy concerns people’s ability to direct their lives and choices. Philosophical accounts differ; this guide uses the term to examine whether assistance supports a person’s purposes and leaves them able to understand and shape consequential decisions. Prunkl’s work argues for clarity about the concept when discussing AI. Human autonomy in the age of artificial intelligence
This differs from a software system’s operational autonomy: its ability to perform tasks without repeated intervention. Increasing the latter does not tell us, by itself, what happens to the former.
Imagine delegating routine travel bookings within a budget. The arrangement may free you to spend time on something you value. If it begins excluding journeys for reasons you cannot inspect or change, a different question arises. The useful boundary includes who sets the goal, which constraints apply, and how the arrangement can be revised.
Read with “When does helpful AI become someone else’s decision?” Related terms: Consent; Decision support. Fiction: The Right to Be Wrong.
Inference
An inference is a step in reasoning from information to a conclusion. In this glossary the word has that ordinary reasoning sense. It should not be confused with “model inference,” the technical use of a trained model to produce outputs. A glossary should make clear which meaning a page needs.
Suppose a recording shows someone carrying a box from a cupboard. The observation may support the conclusion that a box left the room. A conclusion about theft requires additional premises concerning contents, ownership, and permission. Identifying those premises makes the claim easier to examine.
The example does not imply that intention can never be established or that all explanations are equally plausible. It shows why a report should distinguish what a source directly supports from the reasoning added to it. The technical provenance context is discussed separately in C2PA’s explainer. C2PA on provenance and truth
Read with “Can authentic evidence lead to a false conclusion?” Related terms: Authentication; Content provenance. Fiction: The Evidence Was Perfect.
Meaningful human oversight
This guide uses meaningful human oversight to describe an arrangement in which a person can understand enough about a system’s consequential use to investigate concerns and affect what happens. “Meaningful” identifies a practical requirement, not a universal certification. The definition is an editorial synthesis informed by NIST’s guidance on human roles and oversight. NIST AI RMF Playbook
A named reviewer is only part of the arrangement. Ask what they can inspect, how much time they have, which decisions they can change, and where an unresolved concern goes. Those questions also help distinguish a reviewer’s responsibility from that of the people who set the task or deployed the system.
Imagine a care assessment that a case officer may approve but cannot amend. If the only alternative is to stop the entire service, the oversight process has a consequential limitation worth making explicit. Adding another approval box would not explain how an individual error gets corrected.
Read with “When does helpful AI become someone else’s decision?” Related terms: Decision support; Automation bias. Fiction: The Price of Being Useful.
Retrieval-augmented generation
Retrieval-augmented generation, often shortened to RAG, combines a generative model with retrieved material that can inform its response. Lewis and colleagues’ 2020 paper describes an approach joining model parameters with an external retrieval mechanism for language tasks. Foundational paper
Retrieving a passage and generating an answer from it are different operations. A quotation can be checked against its source. A generated answer may combine, interpret, or go beyond retrieved passages, so the presence of source material does not make every sentence a quotation.
Imagine a family service retrieving a letter about a birthday and then answering a question addressed to its deceased writer. The letter may be authentic. The answer is still a new output unless it reproduces identifiable words from the record. Showing the passage and distinguishing any generated addition would help the recipient understand what they are receiving.
Read with “Who gets to speak for someone who has died?” Related terms: Digital legacy; Griefbot and deadbot. Fiction: The Last Person Who Remembers.
AI memory
“It remembers me” can describe several different things. A system might have the current conversation in view, load a saved preference, retrieve an old document, or produce an answer influenced by training. These are different places to look when asking where information is kept and what a deletion control changes.
Machine unlearning
Machine unlearning concerns removing or reducing the influence of specified training data on a trained model. It addresses a different problem from deleting a document in a folder or refusing to display a certain answer.
Confidence calibration
Confidence calibration concerns whether probability estimates match observed frequencies across an appropriate set of cases. Roughly, among many comparable predictions assigned a 90 percent probability of being correct, a well-calibrated system should be correct about 90 percent of the time.
Proxy measures
A proxy measure is an observable indicator used to stand in for something harder to measure. Completed exercises may stand in for learning; completed visits may stand in for care. The indicator can be useful without capturing the whole goal.
AI sycophancy
AI sycophancy is a tendency for a response to align with a user's expressed beliefs or preferences at the expense of a more warranted answer. It can appear as flattering assessment, an unwarranted change of position, or agreement with a one-sided account.
Prompt injection
Prompt injection is an attempt to redirect an AI system through instructions placed in material it encounters. For an assistant reading a web page, email or document, the crucial boundary is between content to examine and instructions authorized to govern its behavior.
Algorithmic monoculture
Algorithmic monoculture describes many decision-makers relying on the same or closely shared algorithmic approach. Different organizations can appear to offer separate judgments while drawing on a common ranking or set of assumptions.