A lot of AI vocabulary is used to mean more than it does. “Open source” often means only that a file can be downloaded, “reasoning” means extra hidden text, and “memory” is a product feature, not a mind. Each entry below gives the common wrong use, what the term means according to the people who define or build it, and why the gap matters. For plain definitions, see the AI glossary and the prompting terms glossary. Wrong beliefs about AI in general are on common AI misconceptions, and prompting folklore is on prompting myths.
- Open weights are not open source. The standards body’s definition also asks for the training code and detailed information about the training data.
- A reasoning model writes hidden working text before it answers. You pay for that text, and it may not show how the answer was really reached.
- A token is not a word, and a context window is not memory. Memory is a separate product feature that saves details between chats.
- The phrase trained on your data can mean four different things. Only training and fine-tuning change the model itself.
- AGI and agent have no single agreed meaning. Ask which one the speaker is using.
- A provenance label says where a file came from, not whether it is true. A missing label proves nothing.
“Open” means less than it sounds
Wrong: if you can download it, it is open source
The Open Source Initiative, which keeps the definition of open source software, published an Open Source AI Definition. It asks for four freedoms, starting with the freedom to “Use the system for any purpose and without having to ask for permission.” It also says what has to be released so you can actually change a system. Weights alone are not enough. The list includes the training code, described as “The complete source code used to train and run the system”, and data information: “Sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system.”Open Source Initiative, The Open Source AI Definition – 1.0, read at source 22 Sep 2026: “Sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system.”
Most downloadable models are better described as open weights: you can run and adapt them, but you cannot rebuild them. Many also come under licences that limit how you use them, which fails the “any purpose” test. The trade-offs are covered on open or closed, and the licences themselves are compared on AI labs outside the US.
When a model is called open, read the licence and check whether the training code and data information were published. The word alone tells you neither.
What the model is actually doing
Wrong: a reasoning model thinks like a person
Vendors describe something narrower. OpenAI’s guide says “Reasoning models use internal reasoning tokens before producing a response.” That hidden text costs money: “Reasoning tokens are not visible as answer text, but they count toward output usage and are billed as output tokens.”OpenAI, Reasoning models and Understanding and counting tokens, both read at source 22 Sep 2026: “Reasoning models use internal reasoning tokens before producing a response.”
The working a model shows is not necessarily how it got its answer. In a 2025 test, researchers slipped models a hint and checked whether their written reasoning admitted using it. Anthropic reported that “Claude 3.7 Sonnet mentioned the hint 25% of the time, and DeepSeek R1 mentioned it 39% of the time.” Its post puts the general point plainly: “There’s no specific reason why the reported Chain-of-Thought must accurately reflect the true reasoning process”.Anthropic, Reasoning models don’t always say what they think, 3 Apr 2025, read at source 22 Sep 2026: “Claude 3.7 Sonnet mentioned the hint 25% of the time, and DeepSeek R1 mentioned it 39% of the time.”
Wrong: a hallucination means the AI is broken, or lying
The US standards body NIST uses the word confabulation, defined as “The production of confidently stated but erroneous or false content”. It lists this among the risks “unique to or exacerbated by the development and use of GAI”, not as a rare fault.NIST, AI 600-1, Generative Artificial Intelligence Profile, July 2024, read at source 22 Sep 2026: “The production of confidently stated but erroneous or false content”. A 2025 paper gives one reason it continues: “language models hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty”.Kalai, Nachum, Vempala & Zhang, Why Language Models Hallucinate, arXiv, 4 Sep 2025, read at source 22 Sep 2026. Lying would need an intention. Guessing is a better picture, and it means a confident tone tells you nothing. More on why AI makes things up.
Wrong: a token is a word
OpenAI: “A token can represent a character, part of a word, a whole word, or punctuation.” It gives a rough English estimate, “1 token is approximately three-quarters of a word”, and warns that “A token count is not the same as a word count.”OpenAI, Understanding and counting tokens, read at source 22 Sep 2026: “1 token is approximately three-quarters of a word.” Anthropic’s figure for its own models is different: “For Claude, a token approximately represents 3.5 English characters, though the exact number can vary depending on the language used.”Anthropic, Glossary, read at source 22 Sep 2026. Prices and limits are counted in tokens, so the same text can cost different amounts on different models, and more in some languages than in English.
These words describe how the machine works, not human qualities. Budget for hidden reasoning tokens, and check facts however confident the answer sounds.
What the model knows about you
Wrong: it follows our conversation, so it has memory
The context window is the text a model can see during one request. Anthropic defines it as “all the text a language model can reference when generating a response, including the response itself”. More is not always better: “As token count grows, accuracy and recall degrade, a phenomenon known as context rot.”Anthropic, Context windows, read at source 22 Sep 2026.
Memory is a separate product feature. OpenAI: “When Memory is enabled, ChatGPT can remember relevant preferences and details from your chats and other available sources.” It is selective: “Memory does not retain every detail from every conversation.”OpenAI, Memory FAQ, read at source 22 Sep 2026: “Memory does not retain every detail from every conversation.” A long chat is not memory, and a remembered detail is not the whole conversation.
Wrong: any tool that uses my files was trained on my data
This phrase covers four different things. Training builds the model in the first place. Fine-tuning, in Anthropic’s words, “is the process of further training a pretrained language model using additional data.” Retrieval (RAG) changes nothing in the model: “In RAG, a language model is augmented with an external knowledge base or a set of documents that is passed into the context window.”Anthropic, Glossary, read at source 22 Sep 2026: “Fine-tuning is the process of further training a pretrained language model using additional data.” Memory is the product feature above. OpenAI handles whether your chats are used for training as its own setting: “For personal ChatGPT accounts, OpenAI may use chats and remembered information to improve models when Improve the model for everyone is on.”OpenAI, Memory FAQ, read at source 22 Sep 2026.
Why it matters: a tool that knows your documents through retrieval has not absorbed them. Remove a document and it is gone from the next answer. Data used in training cannot be pulled back out that way. What happens to what you type is covered on what happens to what you type.Reasoning, September 2026 — follows from the Anthropic glossary: retrieved documents are passed into the context window at the time of the query, while fine-tuning is further training of the model.
When a vendor says a tool uses your data, ask which of the four it means: training, fine-tuning, retrieval or memory. Only the first two change the model.
Big labels with no fixed meaning
Wrong: AGI is a milestone everyone agrees on
It is not. OpenAI’s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work”, which is an economic test.OpenAI, OpenAI Charter, read at source 22 Sep 2026. A 2023 research paper proposes levels instead of a finish line. To build them, its authors say, “we analyze existing definitions of AGI, and distill six principles that a useful ontology for AGI should satisfy.”Morris et al., Levels of AGI, arXiv, v5 revised 24 Sep 2025, read at source 22 Sep 2026. The four main meanings are set out on what AGI means. A claim that AGI “is here” means little until you know which meaning is being used.
Wrong: any chatbot with a tool is an agent
Anthropic separates two things people call agents. “Workflows are systems where LLMs and tools are orchestrated through predefined code paths.” Agents are “systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.” Its advice for builders is restrained: “This might mean not building agentic systems at all.”Anthropic, Building effective agents, 19 Dec 2024, read at source 22 Sep 2026: “Workflows are systems where LLMs and tools are orchestrated through predefined code paths.” The difference matters for risk. A fixed workflow does what its code says. An agent decides its own next step, so it needs tighter permissions and more checking; see browsing agents.
For AGI and agent, ask what the speaker means and what would prove it. If they cannot say, it is a label, not a claim you can check.
Settings, sizes and labels
Wrong: temperature zero gives the same answer every time
Anthropic says otherwise: “Even with temperature set to 0, the results will not be fully deterministic and identical inputs may produce different outputs across API calls.”Anthropic, Glossary, read at source 22 Sep 2026. The setting is also not available everywhere. Google’s release notes for 21 July 2026 say: “The sampling parameters temperature, top_p and top_k are now deprecated.”Google, Gemini API release notes, entry dated 21 July 2026, read in a browser 22 Sep 2026 (the page refuses automated fetches): “Deprecated parameters: The sampling parameters temperature, top_p and top_k are now deprecated.” The details are on Gemini temperature deprecated.
Wrong: more parameters means a better model
Parameter count is size, not quality. In 2022 researchers trained a model called Chinchilla with “the same compute budget as Gopher but with 70B parameters” and more training data, and reported that “Chinchilla uniformly and significantly outperforms Gopher (280B)”, along with other, larger models.Hoffmann et al., Training Compute-Optimal Large Language Models, arXiv, 29 Mar 2022, read at source 22 Sep 2026: “Chinchilla uniformly and significantly outperforms Gopher (280B), GPT-3 (175B), Jurassic-1 (178B), and Megatron-Turing NLG (530B) on a large range of downstream evaluation tasks.” Those models are old now. The point that lasts is that training data and methods count as much as size, so compare models on your own tasks, not on their size.
Wrong: a Content Credentials label proves it is real
The C2PA standard records where a file came from and how it was edited. Its own explainer is direct: “Content Credentials do not provide value judgments” about whether the provenance data is true. They show only “whether the provenance information is well-formed and free from tampering, valid and trusted”. They are also fragile: “the Content Credential may be routinely removed or corrupted by legacy or non-Content Credential capable platforms during distribution.”C2PA Technical Working Group, C2PA Content Credentials Explained, September 2025, read at source 22 Sep 2026. A label tells you who signed a claim about the file. A missing label tells you nothing, because platforms often strip it. See how AI content gets labelled.
Settings, sizes and labels feel exact but are not. Test the setting, compare models on your own work, and treat a provenance label as a record of who signed, not a verdict on truth.
What this page could not verify
- How common each misuse is. No survey measuring how often people misuse these terms was found. The misuses listed are ones that recur in product copy and coverage, chosen by judgement.
- How memory is built inside each product. OpenAI’s help page describes what Memory does, not how it is implemented. Other vendors’ memory features were not checked for this page.
- Whether any current downloadable model meets the OSI definition. This page did not check individual models against the definition.
- The Gemini release notes could only be read in a browser, so that quote was checked by eye and saved from the page text, not fetched automatically.
Open Source Initiative, The Open Source AI Definition – 1.0 · OpenAI, Reasoning models · OpenAI, Understanding and counting tokens · OpenAI, Memory FAQ · OpenAI, OpenAI Charter · Anthropic, Reasoning models don’t always say what they think (2025) · Anthropic, Glossary · Anthropic, Context windows · Anthropic, Building effective agents (2024) · NIST, AI 600-1 (2024) · Kalai et al., Why Language Models Hallucinate (2025) · Morris et al., Levels of AGI (2023, rev. 2025) · Hoffmann et al., Training Compute-Optimal Large Language Models (2022) · Google, Gemini API release notes · C2PA, C2PA Content Credentials Explained (2025). All read at source 22 Sep 2026; the Google page in a browser.