Ask ten people what AI actually is, and you'll get ten different answers. Some will describe something close to a digital mind. Others will call it autocomplete with a marketing budget. Both camps are usually working from a few beliefs that don't survive contact with the research.
Here are six of the most common ones, and what the evidence actually shows.
Myth 1: It's Conscious, or Close to It
Chatbots that say "I find that interesting" or "I'm not sure" sound like they're reporting an inner state. Mechanically, they're doing something narrower: predicting which words are statistically likely to follow, given the patterns in their training data.
There's no established scientific method for detecting subjective experience in an AI system, and no consensus that one currently exists
What looks like personality or feeling is a property of well-trained language modeling, not confirmed evidence of an inner life
Some AI researchers, including at companies building these systems, treat the deeper question of what's actually happening inside these models as a genuine open scientific question, not a settled "no"
The honest position sits between "obviously a mind" and "obviously nothing at all": current evidence doesn't support the first claim, and serious researchers are increasingly cautious about asserting the second with total certainty either.
Myth 2: It's Neutral and Objective
A well-organized, confident-sounding answer can feel like it's coming from an impartial source. It isn't.
AI models learn from data generated by people living in societies with real, documented inequities. A hiring model trained on past hiring decisions inherits whatever bias shaped those decisions. A lending model trained on historical credit data reflects decades of lending practice, good and bad. The model doesn't invent the bias; it finds it, formalizes it, and can scale it.
Myth 3: Hallucinations Are Basically Solved
Rates have genuinely improved. Independent benchmarks that showed frontier models fabricating answers 15-45% of the time in 2024 now show meaningfully lower rates on many tasks in 2026, and reasoning techniques like extended thinking cut error rates further in testing.
"Solved" is still the wrong word. No frontier model tested by independent evaluators reaches zero, and the consequences are showing up in places that matter. Researchers tracking biomedical journals found the rate of papers containing at least one fabricated, non-existent reference has grown more than twelvefold over three years, from about 1 in 2,828 papers in 2023 to roughly 1 in 277 in the first seven weeks of 2026.
"I'm thinking this is just the tip of the iceberg," one of the researchers tracking the trend said.
| Myth | Closer to Reality |
|---|---|
| AI is conscious | No confirmed evidence of subjective experience; treated by researchers as a genuinely open question |
| AI is neutral | Reflects and can amplify biases already present in its training data |
| Hallucinations are solved | Meaningfully reduced, not eliminated; still causing real-world errors |
| Bigger model = smarter model | Parameter count is one factor among many; largest isn't always best |
| Chatbots remember you | Only if a memory feature is explicitly enabled; off by default in most products |
| AI will replace most jobs soon | Current impact is concentrated task-level automation, not wholesale job elimination |
Myth 4: Bigger Always Means Smarter
Parameter count makes for an easy headline number, but it isn't the whole story. Training data quality, architecture choices, and what a model is actually optimized for all matter as much or more.
Some AI labs have made this explicit. When Thinking Machines Lab released its first open-weight model, Inkling, this month, the company said outright in its own launch notes that it wasn't the strongest model available, open or closed. It was built to be broadly capable and easy to fine-tune, not to top a leaderboard. Bigger, in that case, was a deliberate trade for something else.
Myth 5: It Remembers Everything You Tell It
Most consumer AI chat products start each new conversation with no memory of previous ones, unless a specific memory feature has been turned on. When that feature exists and is enabled, it typically stores summarized information the person can review, edit, or delete, rather than a running transcript of every past chat. Without it switched on, the assistant genuinely doesn't know what you talked about yesterday.
Myth 6: It's About to Replace Most Jobs
The most rigorous recent attempt to measure this came from MIT and Oak Ridge National Laboratory's Iceberg Index, released in late 2025. Researchers modeled 151 million U.S. workers as individual agents across more than 32,000 skills and 923 occupations, then measured where current AI systems can already perform those skills at a competitive cost.
The finding: AI could technically already handle tasks equivalent to 11.7% of U.S. wage value, about $1.2 trillion, concentrated in finance, healthcare, and professional services. But the visible layoffs and role shifts people usually point to, in tech and IT specifically, only account for 2.2% of that total wage exposure. Most of the real capability sits in less-discussed areas like routine HR, logistics, and office administration tasks.
Crucially, the researchers were careful to separate technical capability from actual job loss. "External factors, state investment, infrastructure, regulation, mediate how capability translates to impact," the report notes, and its authors stress the index isn't a prediction engine for exactly when or where jobs disappear.
AI Myths: FAQ
There's no scientific evidence that current AI systems have subjective experience. Technically, they work by predicting likely sequences of words based on patterns in training data. That said, some AI researchers, including at labs building these systems, note genuine scientific uncertainty about what internal processes in these models represent, and treat the question as unresolved rather than settled in either direction.
No. AI models are trained on data generated by people in societies with existing inequities, and they reflect the patterns in that data, including biases in hiring records, lending decisions, or historical text. The model doesn't introduce bias out of nowhere; it can reproduce and scale biases already present in what it learned from.
No. Rates have dropped substantially from 2023-2024 levels on many benchmarks, and techniques like extended reasoning reduce errors further, but no frontier model has reached zero. Real-world consequences are still showing up, including a documented rise in AI-fabricated references appearing in published biomedical papers.
Not necessarily. Parameter count is one factor among many, including training data quality, architecture, and what the model is optimized for. Some labs have explicitly released large models that they themselves describe as not the strongest available, prioritizing traits like customizability over raw benchmark performance.
Not by default, in most consumer AI products. Each conversation typically starts fresh unless a product specifically has a memory feature turned on, in which case it stores summarized information the person can usually view, edit, or delete. Without that feature enabled, the AI has no persistent memory of you between sessions.
The evidence points to something narrower: task-level automation within jobs, more than wholesale replacement of jobs. MIT and Oak Ridge National Laboratory's Iceberg Index estimated current AI could technically perform tasks equivalent to 11.7% of U.S. wage value, concentrated in finance, healthcare, and professional services, while stressing that technical capability doesn't automatically mean jobs will disappear on any set timeline.
Jans Bock-Schroeder
Publisher & Founder of AI Angst
Coming from the world of art, photography, and the luxury market, Jans launched AI Angst in 2025 to explore the cultural, ethical, and psychological impacts of artificial intelligence. His work bridges creative vision with critical technology analysis, offering clarity in an era of rapid technological change.
Sources and Citations
This article draws on the following sources:
-
Fortune — "AI hallucinations are slipping past experts into papers and books to enter the permanent record" (May 24, 2026)
Source for the biomedical fabricated-reference growth statistics.
https://fortune.com/2026/05/24/ai-hallucinations-scientific-research-authors-medical-journal-treatment/ -
CNBC — "MIT study finds AI can already replace 11.7% of U.S. workforce" (November 26, 2025)
Primary source for the MIT/Oak Ridge National Laboratory Iceberg Index findings.
https://www.cnbc.com/2025/11/26/mit-study-finds-ai-can-already-replace-11point7percent-of-us-workforce.html -
Fast Company — "MIT study finds AI is already capable of replacing 11.7% of U.S. workers" (November 27, 2025)
Additional source for the Iceberg Index methodology and the technology-sector wage-exposure breakdown.
https://www.fastcompany.com/91450119/mit-study-finds-ai-is-already-capable-of-replacing-11-7-of-u-s-workers -
Gizmodo — "MIT Report Claims 11.7% of U.S. Labor Can Be Replaced with Existing AI" (November 29, 2025)
Source for the study's own caveats about correlation versus causation.
https://gizmodo.com/replacement-study-mit-2000692601 -
AI Angst — "Mira Murati's AI Startup Just Released Its First Model. It Admits It's Not Even the Best One." (July 16, 2026)
Internal reference for the Inkling "bigger isn't always smarter" example.
/discover/inkling-thinking-machines-open-weight-model
Published: July 21, 2026. Sources verified at time of publication. All external links open in a new tab. This is an evergreen explainer piece; figures cited reflect the most recent available data as of publication and may shift as new research is published.
