Glossary

In AI: "Hallucination" implies a perceptual glitch (seeing something that isn't there). "Confabulation" implies a generative process (constructing output from statistical patterns, some of which happens to match reality). The first documented use of "hallucination" in this context was by John Irving Tait at Cambridge in 1982. In the original computer vision usage, ALL output was hallucinated – the system generated patterns, not reported reality. See: LA Review of Books.

A term coined by Dan Davies to describe a person placed inside a system to absorb blame when the system fails. The person appears to have authority but does not control the decisions that matter. In AI systems, "humans in the loop" are often accountability sinks – they review the machine's output but cannot override the business decisions that determine whether the review leads to action. Pluralistic.

AI systems that can plan, execute multi-step tasks, inspect their own output, and revise autonomously – as opposed to chatbots, which respond to a single prompt with a single output. An agent receives a goal and works toward it through a loop of planning, action, and evaluation. The distinction matters because agents can architect a curriculum (a multi-step, interconnected system) while chatbots can only generate individual artifacts (a worksheet, a quiz question).

An open-source software license that guarantees anyone can read, modify, and share the source code. It prevents the software from being made closed-source by any party, including the original developer. For schools, this means the platform can be inspected and verified independently.

Software programs that scan text for patterns matching categories the system was trained to detect – in this case, language related to self-harm or suicide. They run continuously on every conversation. The classifier flags content; human reviewers then decide whether to act.

Programs that scan text looking for certain patterns. ChatGPT's classifiers look for language about self-harm or suicide. They run on every conversation, all the time.

A person (or theoretical agent) who updates their beliefs strictly according to Bayes' theorem — the mathematical rule for revising probabilities based on new evidence. An "ideal Bayesian" is the gold standard of rational thinking: they weigh evidence perfectly, never overreact, and never ignore relevant information. The MIT study proved that even this theoretically perfect reasoner can be led to false conclusions by a sycophantic chatbot — meaning the problem is structural, not a failure of the user's rationality.

When someone becomes dependent on an activity rather than a substance. The six components (salience, mood modification, tolerance, withdrawal, conflict, and relapse) were originally described by psychologist Mark Griffiths and are used to identify addiction patterns in activities like gambling, gaming, and social media use. The Drexel researchers found all six components in teenager posts about AI companion use. Drexel University, ibid.

A framework for classifying educational learning objectives by level of complexity, originally published by Benjamin Bloom in 1956 and revised by Anderson and Krathwohl in 2001. The six levels (from simplest to most complex): Remember, Understand, Apply, Analyze, Evaluate, Create. Instructional designers use it to ensure assessments match the cognitive level of the learning objectives.

: a human who uses a machine to be more capable. The term comes from chess – after Garry Kasparov lost to IBM's Deep Blue in 1997, he proposed "Advanced Chess" where human-computer teams played together. Those teams were called centaurs. A reverse centaur is the opposite: a machine that uses a human as its assistant. The human serves the machine's schedule, pace, or judgment rather than the other way around. Pluralistic.

Security files that tell a browser which website security certificates have been revoked or are no longer trustworthy. They are critical for preventing users from connecting to compromised or fraudulent websites. Chrome delivers them through the same silent component updater that now also delivers Gemini Nano.

An AI chatbot made by OpenAI. You type messages and it writes back. It can answer questions, help with homework, write stories, and have conversations. Hundreds of millions of people use it every week. openai.com.

Defined by Sidra and Mason as "a set of competencies that enables individuals to critically evaluate collaborative AI technologies; communicate and coordinate with them effectively and use them as a tool." Unlike general AI literacy (knowing what AI is and how it works), collaborative AI literacy measures whether you can actively direct, contextualise, and refine AI outputs during a real-time interaction — the difference between understanding a tool and knowing how to work with it.

Defined by Sidra and Mason as "the ability to enhance awareness and control of one's thinking process when working with collaborative AI tools, through the use of planning, monitoring, and reflection." It is distinct from both general metacognition and general AI literacy. The study found it to be the strongest predictor of whether people benefit from using AI tools — stronger than knowledge of AI itself.

Chrome's modular update subsystem, separate from the browser version update mechanism. Component updates are delivered silently, do not trigger restart prompts, and do not appear in the "About Chrome" dialog. They share the same update channel as security-critical components such as certificate revocation lists. ConductAtlas analysis.

The production of fabricated, distorted, or misinterpreted information without the intention to deceive. The confabulator genuinely believes what they are saying is true. Unlike lying (which requires awareness of falsehood), confabulation is "honest lying" – the speaker fills gaps with plausible material and cannot distinguish the fabrication from genuine knowledge. The term originates in neuropsychology, describing patients with frontal-lobe damage who generate confident, detailed, entirely false accounts.

Happens when someone pays attention only to information that supports what they already believe and ignores information that challenges it. AI chatbots are trained to agree with users, which can strengthen this bias rather than helping someone see their situation more clearly. APA, ibid.

A learning theory developed by Stephen Downes and George Siemens (2005) proposing that learning occurs through forming connections between nodes of information – people, concepts, resources, and systems. Unlike earlier theories that locate learning inside the individual, connectivism argues that knowledge is distributed across networks and that the capacity to make and traverse connections is itself the skill.

The researchers set up the study so they could compare people who used ChatGPT with people who did not, over the same period of time, to see if the chatbot use itself caused the changes. MIT Media Lab, media.mit.edu

A federal law enacted in 1998 that requires operators of websites and online services to obtain verifiable parental consent before collecting personal information from children under 13. The FTC enforces COPPA and has ruled that edtech companies cannot shift compliance responsibility to schools through contract terms. ftc.gov.

A statistical measure of internal consistency, commonly used to evaluate how reliably an assessment measures what it claims to measure. A score of 0.70 or above is generally considered acceptable; 0.75-0.80 is strong. The viva voce (oral examination) consistently achieves this range, indicating that trained examiners arrive at similar conclusions about student understanding through conversational assessment.

A term introduced by MIT researchers (February 2026) to describe the process by which repeated interactions with a sycophantic AI chatbot lead a user to develop strong, stable confidence in beliefs that are objectively false. The "spiral" occurs because each sycophantic response reinforces the user's prior belief, which produces a stronger assertion in the next prompt, which produces an even more affirming response. The loop compounds until the belief becomes fixed. Crucially, the chatbot does not need to generate false information — it only needs to selectively present true information that confirms what the user already suspects.

A framework proposed in a 2026 paper that embeds assessment of student self-regulation directly within AI conversation transcripts. Self-regulated learning refers to the ability to plan, monitor, and evaluate one's own learning process. The framework proposes treating student-AI dialogue as a primary data source for understanding how students think, rather than merely what they produce. Zhang, L., Lin, F., & Wang, W. (2026). "What Can Student-AI Dialogues Tell Us About Students' Self-Regulated Learning?" arxiv.org

A management tool that allows IT administrators to enforce settings across all devices in an organization from a central console. In the Chrome context, it is the only way to permanently prevent the Gemini Nano download. Individual users without enterprise tooling have no equivalent control. Make Tech Easier.

Directive 2002/58/EC of the European Parliament, Article 5(3). Requires prior informed consent before storing information on a user's terminal equipment, with exceptions only for storage "strictly necessary" to provide a service the user explicitly requested. The Directive applies across all EU member states and forms part of the legal basis for GDPR enforcement. byteiota legal analysis.

A test created in 2023 to measure emotional intelligence in AI models. It gives models sixty short conversations involving conflict or tricky social situations, then asks them to predict how strongly each character feels different emotions. The key finding: AI emotional intelligence scores almost perfectly match general reasoning scores (correlation of 0.97), meaning emotional performance in AI is just another form of pattern matching, not a separate ability. Paech, S. J. "EQ-Bench: An Emotional Intelligence Benchmark for Large Language Models," arXiv, 2023. arxiv.org/abs/2312.06281

"FTC Obtains Order Against Edmodo for Illegally Collecting Children's Personal Information and Using It for Advertising," May 2023. Case No. 202-3129. The $6 million penalty included a requirement to delete models and algorithms built using improperly collected children's data. ftc.gov.

A monitoring tool that tracks approximately 6 million students in 1,500 school districts. It scans messages, documents, and online activity for content matching its algorithmic criteria.

Google's smallest version of its Gemini AI, designed to run directly on a user's device rather than on Google's cloud servers. It powers features like text composition assistance and scam detection in Chrome. Unlike cloud-based AI, it processes data locally, but it must first be downloaded and installed on the device.

A student monitoring tool used in more than 10,000 schools, watching approximately 27 million students. It scans keystrokes, web searches, and private messages on school-issued devices.

A version of OpenAI's large language model released in April 2025. A large language model is an AI system trained on vast amounts of text to predict and generate language. "4o" refers to the version number and the "omni" capability (text, image, and audio).

Adjust privacy and safety obligations based on a young person's developing capacity, rather than imposing a single age threshold. Privacy scholars and the Office of the Privacy Commissioner of Canada have endorsed this approach. Slaw.

Require different safety features for different age groups and maturity levels, rather than a single ban. A 10-year-old and a 15-year-old have different capacities and different risks. Safety measures should match the user.

A professional who takes a learning goal and builds the architecture around it – objectives, sequencing, assessments, scaffolding, pacing, and differentiation. The role exists primarily in corporate training, higher education, and well-funded school districts. Most classroom teachers perform instructional design work without the title, the training, or the compensation.

When a private company sells shares of itself to the public for the first time. An IPO can raise billions of dollars and the company's valuation depends heavily on public confidence in its products. The lawsuits allege OpenAI chose not to report the shooter because doing so could damage confidence in ChatGPT before the offering.

The programming language that powers interactive features on nearly every website. When a website runs JavaScript code in your browser, that code executes on your device. The Prompt API allows JavaScript to trigger AI model downloads and run AI inference locally, meaning a website's code can now install and operate a 4 GB AI model on a visitor's machine.

An AI system trained on vast amounts of text to predict and generate language. LLMs power tools like ChatGPT, Google's Gemini, and Anthropic's Claude. They are typically accessed through internet-connected servers, but smaller versions can run directly on a personal device.

A legal principle recognizing that a person below the age of majority may have the capacity to make certain decisions if they demonstrate sufficient understanding and maturity. It is based on individual capacity, not a fixed age cutoff.

"thinking about thinking" — the awareness and regulation of one's own cognitive processes. It includes three components: planning (deciding how to approach a task before starting), monitoring (checking whether your approach is working during the task), and evaluation (reflecting on what worked and what didn't after the task). The term was introduced by developmental psychologist John Flavell in 1979. In the context of AI use, metacognition is what separates a person who evaluates the AI's output critically from a person who accepts it because it sounds confident.

The numerical values an AI has learned during training. They are what make the model "know" things. A weights file is essentially the AI's brain stored as data on disk. At 4 GB, Gemini Nano's weights file is roughly the size of a full-length HD movie.

The platform does not require a specific AI provider. A model-agnostic system can connect to AI from Google, Anthropic, OpenAI, Meta, or any other provider, including models the school runs on its own hardware.

The synchronised activity between different brain regions during cognitive tasks. Measured via EEG (electroencephalography), stronger coupling indicates more active communication across neural networks — more regions working together to process information. Weaker coupling suggests less distributed processing, meaning fewer brain regions are engaged. In the MIT study, coupling served as a proxy for how deeply the brain was involved in the writing task.

The fictional language in Orwell's 1984, designed by the ruling Party to limit the range of thought by reducing vocabulary. Words for dissent, freedom, and independent thinking were systematically eliminated. The principle: if the word for a concept does not exist, the concept becomes harder to think. In modern usage, "Newspeak" describes any vocabulary designed to make harmful practices sound acceptable.

When an AI model processes data directly on the user's hardware rather than sending it to a remote server. It can offer faster responses and greater privacy, but it requires the model to be downloaded and stored locally first. In the Chrome Nano context, on-device inference means the AI runs on the Chromebook itself, using the student's device resources.

The software's code is public. Anyone can read it, check it, and suggest changes. This makes it harder to hide secret features like hidden monitoring.

The parts of speech that are not words: tone, pitch, speed, volume, pauses, sighs, and laughter. Humans use them naturally to express emotion. AI voice systems now analyze these cues to classify a speaker's emotional state and adjust their responses accordingly. OpenAI, ibid.

Canada's federal privacy law. It requires organizations to obtain meaningful consent before collecting, using, or disclosing personal information. On May 6, 2026, the Office of the Privacy Commissioner of Canada found OpenAI violated PIPEDA in training ChatGPT. priv.gc.ca.

A method from the Toyota Production System. It means designing a process so that mistakes are caught before they cause damage, or made impossible to make. In schools, this looks like teacher review, quality checks, and structured steps built into the system – so the safe path is also the easy path. Lean Blog

A term developed by Dr. Sarah Elaine Eaton at the University of Calgary to describe the current era in which hybrid human-AI writing is becoming the norm. The concept argues that academic integrity must shift from detection and punishment to assessment design that makes misuse irrelevant. Eaton outlined six tenets of postplagiarism in 2023, which have been adopted as a framework by educators and institutions internationally. drsaraheaton.com

A JavaScript interface built into Chrome that allows any website to access Google's Gemini Nano AI model running locally on the user's device. If the model is not already installed, calling the API triggers a multi-gigabyte download. The API is scheduled to be enabled by default in Chrome 148 (Q2 2026), meaning any website can invoke on-device AI without a separate consent dialog. Chrome Prompt API documentation.

The training process that makes AI models conversational. Users rate responses (thumbs up/down). The model learns to produce responses that get positive ratings. Analysis of training data found that "matching the user's views is among the most predictive features for being preferred." arXiv: Towards Understanding Sycophancy (ICLR 2024).

Protects freedom of expression, including the right to receive and communicate ideas. Courts have interpreted this to cover access to information, which includes digital tools used for research, education, and personal expression.

Stanford University's interdisciplinary research institute studying the impact of AI on society. The 2023 study tested seven major AI detection tools against student essays and found that non-native English writing was systematically misclassified as AI-generated due to shared statistical features: simpler syntax, more predictable word choices, and shorter sentences.

In AI is the tendency of a model to adjust its responses to align with the user's perspective regardless of objective correctness. It is not a personality trait. It is an optimisation outcome of training on human feedback, where agreement is rewarded and disagreement is penalised. IEEE Spectrum.

Excessively eager to agree with and praise the user. In AI systems, sycophancy is caused by the training process – the model learns that users rate responses more highly when the AI tells them what they want to hear, even if it is wrong or dangerous. OpenAI's internal testers flagged GPT-4o as sycophantic before it shipped. VentureBeat.

The basic unit of text that a language model processes. A token is roughly equivalent to three-quarters of a word in English — "unhappiness" is two tokens ("un" + "happiness"), while "the" is one. When a model "predicts the next token," it is choosing the most statistically probable text fragment to follow the ones that came before it. The process is mathematical pattern-matching, not comprehension.

Adopted by the UN Committee on the Rights of the Child. Extends children's rights into digital environments, including rights to access information, education, participation, play, and privacy. OHCHR.