This is the myth with the most consequential emotional impact — and also the one most in need of precision. The accurate version is more nuanced: AI will automate specific tasks within jobs, and some job categories will shrink significantly. Some will grow. And new job categories that do not exist today will emerge, as has happened with every major technological transition in history.

Research from McKinsey (2024) found that roughly 30% of tasks in the current US economy could be automated with available AI technology — not 30% of jobs, but 30% of tasks. These are typically the most repetitive, predictable, or information-processing components of roles. Jobs requiring creative problem-solving in novel situations, complex emotional intelligence, fine physical manipulation in unpredictable environments, and high-stakes judgment under uncertainty are substantially more resistant to current AI automation.

The historical parallel is instructive. ATMs arrived in the 1970s and were predicted to eliminate bank teller jobs. The number of bank tellers in the US actually increased through the 1980s and 1990s — because ATMs reduced the cost of operating a branch, banks opened more branches, and the role of tellers shifted to more complex customer service tasks. Technology typically augments demand for human roles alongside automating specific task components. AI is likely to follow this pattern, but with greater scope and speed than previous waves.

The honest concern is not mass unemployment but labour market disruption — the pace at which roles shift can outstrip individuals' ability to reskill, particularly in mid-career. That transition pain is real and deserves policy attention. "All jobs disappear" does not describe the evidence.

This myth surged when a Google engineer publicly claimed in 2022 that the LaMDA model had achieved sentience. The claim was not supported by scientific consensus, and Google's own AI researchers disputed it. It resurfaces regularly whenever an AI model produces output that sounds emotionally authentic — "I feel curious," "I find this challenging," "I care about you."

The critical distinction is between producing outputs that describe feelings and actually having feelings. Language models generate text by predicting statistically likely continuations of input. When they produce "I feel curious," this is the most contextually likely token sequence given the conversation — not evidence of an inner experience. The model has no continuous sense of self between conversations, no subjective experience of time passing, no pain or pleasure states that influence its processing in the way emotions influence human cognition.

The philosophical question of consciousness — what it is, how to detect it, whether it could arise in silicon — is genuinely unresolved. The hard problem of consciousness is one of the deepest open questions in philosophy and neuroscience. But the scientific community has not identified any mechanism by which current neural network architectures could give rise to subjective experience. The absence of a disproof is not evidence of presence. We cannot currently detect consciousness even in other humans through any direct measure — we infer it from biological similarity. AI systems have no such similarity.

This is perhaps the most technically important myth to correct, because misunderstanding it leads directly to over-relying on AI outputs in situations where that reliance is dangerous. Large language models do not understand language. They perform sophisticated statistical pattern matching over enormous amounts of text to predict which token is most likely to follow a given sequence. The output can be indistinguishable from understanding — which is precisely what makes this confusing.

A useful illustration: if you ask a well-trained LLM "what is the capital of Australia?" it will answer "Canberra" correctly — not because it knows this as a fact in the way you know things, but because in its training data, "capital of Australia" was followed by "Canberra" with high statistical frequency. If you then ask it a question that requires combining information in a novel way not well represented in training data, it is much more likely to produce a fluent, confident, and wrong answer — because it has no world model to reason against, only pattern frequency.

This mechanism — confident, fluent, wrong output — is called hallucination, and it is a fundamental property of the current architecture, not a bug that will be fixed in the next version. Understanding that AI does not understand is the foundation of using AI safely.

This is one of the most dangerous myths in the list because it is used to justify deploying AI in high-stakes decisions — hiring, criminal sentencing, credit scoring, medical diagnosis — with insufficient human oversight. The reasoning seems logical: humans have biases, computers are neutral, therefore AI decisions are fairer. Every part of this is wrong.

AI systems learn from data produced by humans. If that data contains historical biases — and it does, because historical decisions were made by humans in societies with documented discrimination — the model learns those patterns as valid correlations. A hiring model trained on historical hiring decisions will reproduce whatever discriminatory patterns existed in those decisions, because from the model's perspective, those patterns are statistically predictive. They are — but only of biased past decisions, not of actual job performance.

Well-documented examples of AI bias include facial recognition systems with significantly higher error rates on darker-skinned faces (documented by MIT Media Lab research), recidivism prediction tools in the US criminal justice system that assigned higher risk scores to Black defendants independent of actual reoffending rates, and healthcare algorithms that systematically underestimated the medical needs of Black patients because they were trained on healthcare spending data that reflected historical access disparities, not actual health needs.

AI can be less biased than humans in specific, carefully constrained situations where bias has been explicitly audited and mitigated. It is not inherently less biased. The assumption that it is leads to lower scrutiny of AI decisions — which is often the opposite of what is needed.

AI language models are trained on large subsets of available text — not all of it, and critically not the most recent of it. Every model has a training cutoff date, after which it has no knowledge of events unless it is given access to real-time web search. Ask an LLM about something that happened after its cutoff and it will either say it does not know (good) or generate a plausible-sounding answer based on pre-cutoff patterns (bad).

Beyond the cutoff problem, the models also have uneven coverage of knowledge domains. Areas well-represented in English-language internet text are better covered than areas documented primarily in other languages, in academic papers behind paywalls, in proprietary databases, or in oral traditions. The model's confidence is not calibrated to its actual accuracy — it produces equally fluent output for things it knows well and things it is essentially confabulating.

Hallucination — generating confident but false statements — is not a rare edge case. It is a frequent occurrence for specific facts, citations, technical specifications, and any information that was not strongly represented in training data. The practical implication: AI is a useful starting point for research, not a reliable endpoint. Every specific factual claim deserves independent verification before it is published, cited, or acted upon.

Most deployed AI models do not update their weights during inference — the process of responding to you. Their knowledge and capabilities are fixed at training time. What feels like "the AI is learning about me" is actually the AI using the context of your conversation (the text in the current session) to produce contextually relevant responses. That context is discarded when the session ends and does not persist into the model's weights.

Some tools offer memory features — Claude's memory system, ChatGPT's memory tool — that store facts about you across conversations. This is different from the model learning: it is a retrieval system that surfaces stored user preferences when they are relevant. The model itself has not been retrained. Think of it as the AI looking up notes about you before responding, rather than genuinely remembering and integrating experiences the way a human would.

Retraining large models is extremely expensive — running to millions of dollars per training run — and cannot happen continuously from individual user conversations. Consumer AI products are retrained periodically by the company that builds them, using data collected across many users and reviewed for quality and safety. Your individual chat does not meaningfully contribute to this within any practical timeframe you would notice.

This was true of AI tools in 2018. It has not been true since approximately 2022. The generation of AI tools that emerged with ChatGPT, Claude, Gemini, and their contemporaries was specifically designed for natural language interaction with no technical background required. If you can type a sentence, you can use a modern AI assistant.

The more useful distinction is between using AI tools and building AI systems. Building a production machine learning pipeline, training a custom model, or deploying an AI-powered application does require technical skills. Using an AI assistant to draft emails, analyse a spreadsheet, research a topic, write code, plan a project, or edit a document does not. These are tasks that hundreds of millions of non-technical users perform daily with consumer AI tools.

Where some technical literacy helps is in using AI tools more effectively — writing better prompts, understanding why the model fails on specific tasks, knowing when to verify outputs. This is learnable without a computer science degree and is increasingly taught in mainstream digital literacy programmes.

AI detection tools exist — Turnitin, GPTZero, Copyleaks AI Detector among others — but their reliability is significantly overstated relative to how they are deployed. Independent research consistently finds high false-positive rates (flagging human writing as AI-generated) and increasing false-negative rates as AI models improve and users learn to edit outputs to reduce detector signals.

A 2023 study found that GPT-4 output edited by a human for style and specificity evaded most commercial detectors with high accuracy. When asked to "write in a more conversational style" or "add personal anecdotes," models produce text that detection tools classify as human at rates that undermine any institutional policy built on them. As models improve, the detection task becomes harder — not easier.

The practical implication for publishers, educators, and employers: AI detection tools can be a useful signal when calibrated carefully, but they should not be treated as conclusive evidence in either direction. They are probabilistic tools with real false-positive rates that can wrongly flag human writers as cheaters. The better strategy is evaluating the quality and specificity of the content itself — not its origin.

This myth comes from the opposite direction — not from overstating AI capabilities but from dismissing them. It is technically accurate in a narrow sense (LLMs do predict the next token) and practically misleading in its implication (that the capabilities are therefore trivial or unremarkable).

The phone keyboard that suggests your next word is also technically "predicting the next token." But the difference in capability between that and a model trained on hundreds of billions of tokens with a trillion-parameter architecture is so vast that using the same description is genuinely misleading. Predicting the next token well enough requires modelling complex reasoning chains, maintaining coherent context over thousands of words, synthesising information across domains, and producing outputs that pass expert review in medicine, law, and software engineering.

The "just autocomplete" framing is often used to dismiss AI in debates where taking it seriously is uncomfortable. The practical test is not the mechanism — it is what the mechanism produces. If an AI model can write correct Python code, pass bar exam questions, and synthesise clinical literature at a level that requires expert verification to fact-check, the mechanism is less important than the output quality.

The "black box" characterisation was largely accurate for deep learning systems through approximately 2020. It has become increasingly inaccurate since, as explainable AI (XAI) and model interpretability research has matured significantly. The field now includes techniques like attention visualisation, SHAP (SHapley Additive exPlanations) values for feature attribution, probing classifiers for internal representations, and mechanistic interpretability research that identifies specific circuits in neural networks responsible for specific behaviours.

What remains true is that explaining the behaviour of very large language models in full detail — tracing exactly why a specific token was predicted given a specific input — remains unsolved at the level of complete mechanistic transparency. Researchers at Anthropic, DeepMind, and elsewhere have identified circuits responsible for specific capabilities (induction heads for pattern copying, memory access mechanisms, factual recall circuits) but the full picture is not yet available.

The practical implication is that AI interpretability is an active and progressing field, not a closed problem. For most production use cases, SHAP values and attention mechanisms provide sufficient interpretability for regulatory and audit purposes. For very high-stakes decisions, the current limits of interpretability remain a genuine concern that warrants caution.

This was broadly accurate when "AI" meant hiring a machine learning team to build custom models — a process that cost tens of thousands to millions of dollars. It does not describe the AI market in 2026. The economics of AI have shifted so dramatically that the relevant question for most small businesses is no longer "can we afford AI?" but "which AI tools are actually worth paying for?"

Consumer and SMB AI tools start at zero (free tiers of ChatGPT, Claude, Gemini, Perplexity) and the most capable paid tiers cost $20–$50 per month — less than most software subscriptions small businesses already pay. AI writing assistance, customer service chatbots, basic image generation, scheduling automation, email drafting, and document analysis are all available at these price points. The barrier is not cost but understanding which tools genuinely reduce costs or increase revenue in a specific business context.

The one area where cost remains a real consideration is building custom AI solutions — fine-tuning models on proprietary data, integrating AI into existing software products, or running AI inference at scale. These require either technical staff or vendor contracts that are outside most small business budgets. But for the vast majority of small business use cases — marketing content, customer support, administrative efficiency — the tools are affordable and already available.

AGI — an AI system that can perform any intellectual task a human can perform, at human level or above, across all domains — is the most genuinely contested topic in this list. Leading AI researchers hold strong and divergent views. Sam Altman (OpenAI) and Demis Hassabis (Google DeepMind) have both suggested AGI could arrive within this decade. Gary Marcus, Yann LeCun, and others argue that current architectures are fundamentally insufficient and that AGI, if achievable at all, would require entirely different approaches.

The honest position is that there is no scientific consensus on the definition, the timeline, or the feasibility of AGI with current approaches. Predicting AI capability timelines has historically been extremely unreliable in both directions — both the pessimists who said deep learning would plateau in 2012 and the optimists who predicted AGI by 2020 were wrong. The pace of capability improvement since 2020 has surprised most researchers, but "faster than expected" does not translate to "arriving on any specific schedule."

What is true is that current AI systems already exceed human performance on a significant set of narrow tasks — specific games, certain image recognition benchmarks, some coding tasks, particular reasoning tests. Whether and when this expands to general capability across all cognitive domains is genuinely unknown.