Physicist's 'Journalist Challenge' Takes Aim at Vague 'AI' Label

A physicist's short, seven-word dare posted to Bluesky in late November 2025 crystallized a complaint that developers, researchers, and technically literate readers have nursed for years: the word "AI" has been stretched so thin it now obscures more than it reveals. The journalist challenge — use "machine learning" when you mean machine learning, use "LLM" when you mean LLM, and retire "AI" as a catch-all — sounds disarmingly simple. But following it rigorously would force a discipline on tech coverage that is long overdue. For developers and engineers who live inside these distinctions every day, getting the language wrong carries surprisingly high stakes. For the general public navigating a world shaped by these systems, the stakes are even higher.

The journalist challenge, applied: Before you write "AI," stop. Ask what the system actually does, how it was trained, and what it produces. Then use the most specific accurate term available. If you cannot find out, say so — that opacity is the story.

The Post That Started It: Robert McNees on Bluesky

The challenge originated with a post by Robert McNees, a physicist active on Bluesky, published in late November 2025. As McNees framed it, the challenge to journalists is straightforward: use "machine learning" when you mean machine learning and "LLM" when you mean LLM, and ditch "AI" as a catch-all term — because, in his view, the catch-all is not useful for readers and it helps companies confuse the public by obscuring the roles played by different technologies.

Readers who want the exact wording can consult McNees's Bluesky account directly; the substance is what matters here, and it is unambiguous: swap the vague umbrella term for the specific one whenever you can.

McNees is not a media critic by trade — he is a scientist. That a physicist, rather than a journalism professor or a newsroom editor, felt compelled to issue a journalist challenge to use "machine learning" when you mean machine learning says something important about where the frustration actually lives. It lives in the technical community: the people who build, train, evaluate, and deploy these systems, then read mainstream coverage and wince at how thoroughly the vocabulary has collapsed.

The framing echoes a basic scientific instinct: science demands operational definitions. A chemist does not say "stuff" when she means "a covalent compound." A physicist does not say "energy" when she means "electromagnetic radiation." Yet the technology press, and by extension much of public discourse, routinely says "AI" when it means something far more specific. That slippage has measurable consequences for accountability, policy, and public trust — all of which this article examines in turn.

Why "AI" Became a Catch-All — and Why That's a Problem

The term "artificial intelligence" is genuinely old, tracing back to the Dartmouth Summer Research Project on Artificial Intelligence held in 1956, which was outlined in a 1955 proposal by John McCarthy, Marvin Minsky, and colleagues asserting that "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." That definition was contested from the start. But the specific crisis McNees is describing is a product of the last three to four years, as large language models became commercially dominant and every product — from a spam filter to a voice assistant to a code autocomplete tool — began appearing under the single banner of "AI."

For editors and journalists under deadline pressure, "AI" is attractive precisely because it demands nothing. Writing "the company's AI system" sidesteps any need to understand whether the underlying technology is a rule-based classifier, a random forest, a convolutional neural network, a transformer-based LLM, or a retrieval-augmented pipeline. That editorial convenience comes at a direct cost to every reader who trusts the coverage to tell them what is actually happening.

The Corporate Incentive to Blur the Lines

McNees's post identifies something beyond mere laziness. He argues that imprecise language helps companies that are trying to confuse the public by obscuring the roles played by different technologies. This is the sharper edge of the critique, and it deserves careful attention.

When a company announces an "AI-powered" product, the claim is functionally unfalsifiable. If the product is a lookup table with a chatbot skin, calling it "AI" is technically defensible but practically misleading. When it later underperforms, the company can retreat to "AI is still maturing" — a narrative that simultaneously protects both the product and the broader hype cycle. Precision in language would require companies to say what they actually built, which would then allow journalists, regulators, and users to hold them accountable to what it can actually do.

This dynamic can be amplified where a publisher and the companies it covers share a commercial relationship. As newsrooms sign content and licensing arrangements with technology companies — a trend that has grown as generative-AI products became commercially central — there is at least a theoretical pressure to accept a vendor's preferred framing rather than probe it. That pressure, where it exists, is structural rather than merely editorial, and it is worth naming as a reason precise terminology matters most in exactly the coverage where vagueness is most convenient.

The Reader Deserves Better

Readers who consume tech journalism are not a homogeneous audience. A significant portion are developers, data scientists, product managers, and technically literate professionals who immediately notice when "AI" is doing heavy lifting it cannot support. For these readers, vague language is not just an annoyance — it is a trust signal. A piece that conflates machine learning with LLMs, or treats a fine-tuned binary classification model the same as a frontier generative model with hundreds of billions of parameters, signals that the writer does not understand what they are covering. And if the writer does not understand the technology, their conclusions about its implications, risks, and benefits may not be reliable either.

For general-audience readers, the problem is different but equally serious: when everything is "AI," the term simultaneously overpromises and underdescribes. A recommendation algorithm and a system that generates synthetic medical images are both called "AI," yet they raise entirely different ethical, legal, and practical questions. Collapsing them into a single word makes it impossible for a non-specialist reader to develop any accurate intuition about what these systems can and cannot do.

Machine Learning vs. LLMs vs. "AI": A Practical Glossary

The journalist challenge to use "machine learning" when you mean machine learning implicitly requires writers to know what each term actually means. The working distinctions below are grounded in how practitioners use these terms — not in marketing copy.

Term	What It Actually Describes	Example Use Cases	When to Use It
Machine Learning (ML)	A family of statistical methods where systems learn patterns from data without being explicitly programmed for each rule. Includes supervised, unsupervised, and reinforcement learning paradigms. Encompasses decision trees, random forests, support-vector machines, gradient-boosted ensembles (e.g., XGBoost, LightGBM), and neural networks of all sizes.	Fraud detection, recommendation engines, image classification, predictive maintenance, credit scoring, spam filtering, churn prediction	When the system learns from labeled or unlabeled data to make predictions, classifications, or decisions — and does not primarily generate open-ended natural language as its output
Large Language Model (LLM)	A specific class of deep neural network — almost exclusively transformer architectures, as introduced by Vaswani et al. in the 2017 paper "Attention Is All You Need" — trained on massive text corpora to model the probability distribution over token sequences. Capable of generating, summarizing, translating, classifying, and reasoning over natural language. A subset of deep learning, which is itself a subset of ML.	Chatbots, code assistants, document summarization, retrieval-augmented question answering, structured data extraction from unstructured text, long-form content generation	When the system is specifically a large transformer-based language model (GPT-series, Claude, Gemini, Llama, Mistral, Command R, and similar) or a product built directly on top of one via API, fine-tuning, or prompt engineering
"AI" (Artificial Intelligence)	A broad, historically contested umbrella term covering any computational system designed to perform tasks that would require intelligence if performed by a human. Encompasses 1960s rule-based expert systems, 1980s symbolic reasoning engines, and modern deep learning equally. In 2025, largely uninformative as a standalone technical descriptor; in most product coverage it functions as a marketing or rhetorical choice rather than a technical one.	Too broad to anchor a specific use case; any system from a chess engine to a frontier generative model fits the definition	Rarely, if ever, in technical or investigative journalism. Potentially appropriate in explicitly philosophical, regulatory-framework, or historical contexts where the broadest framing is deliberate and acknowledged. In all other cases: identify the specific technology and use that term instead.

The table above illustrates why the challenge matters so acutely. "Machine learning" and "LLM" are not synonyms — an LLM is a specific type of machine learning model, but the vast majority of machine learning systems are not LLMs. Treating them interchangeably is analogous to writing "vehicle" when the story is specifically about a hydrogen fuel-cell semi-truck: technically defensible, practically useless, and likely to mislead any reader trying to form a view about infrastructure, safety, or environmental impact.

What Precision Actually Looks Like in Practice

Adopting the journalist challenge is not merely about swapping one word for another. It demands a small but genuine reporting effort at every stage of production. Here is how that discipline plays out in concrete editorial situations.

Before Publishing: Questions Every Writer Should Ask

1. What is the model architecture? Is it a transformer? A convolutional network? A gradient-boosted ensemble? A recurrent neural network? The answer determines which term is correct — and if a company spokesperson cannot or will not answer, that refusal is itself informative.
2. What is the primary training data? If the model was trained primarily on large-scale text corpora scraped from the web, "LLM" is likely the right term. If it was trained on tabular business records, medical imaging data, or time-series sensor readings, "machine learning model" with a modifier (image classifier, anomaly-detection model, etc.) is more accurate.
3. What is the primary output type? Open-ended natural language generation strongly points toward LLM territory. Probability scores, class labels, rankings, or anomaly flags point toward general ML.
4. Is the company using "AI" to avoid disclosing specifics? If a spokesperson defaults to "our AI" and cannot or will not elaborate on the underlying technology, that evasion is worth naming in the article. Readers deserve to know when a company is being deliberately opaque about what powers its products.
5. Is this actually a rules-based or heuristic system dressed up as ML? Not every "intelligent" feature involves a trained model at all. Some products use deterministic logic, lookup tables, or hard-coded rules. Calling these "AI" is not vague — it is wrong.
6. Would a data scientist or ML engineer reading this wince? This is the practical gut-check. If the answer is yes, revise before publication.

Rewriting Common Headlines

The practical effect of the journalist challenge becomes clear when you apply it to examples. The headlines below are illustrative composites, not descriptions of real companies or deployments; they show how a generic, vague headline sharpens once precision enters the picture:

• "Company X Deploys AI to Detect Fraud" → "Company X Deploys Gradient-Boosted Machine Learning Models to Flag Fraudulent Transactions in Real Time"
• "Startup Uses AI to Write Marketing Copy" → "Startup Builds Marketing Automation Tool on a Commercial LLM Accessed via API"
• "AI Predicts Patient Readmission Risk" → "Hospital System Uses a Supervised Machine Learning Model to Predict 30-Day Readmission Risk from Electronic Health Records"
• "AI Chatbot Answers Customer Questions" → "Retailer Deploys LLM-Powered Chatbot for Tier-One Customer Service Inquiries"
• "AI Screens Job Applications" → "Company Uses a Machine Learning Classifier Trained on Historical Hiring Data to Score Résumés"

Each revision is longer. Every one of them is more informative. The added words do genuine work: they tell a technically literate reader what class of system is involved, what its plausible capabilities and failure modes are, and — in several cases — whose infrastructure or intellectual property sits underneath the hood. That last point matters for questions of liability, data privacy, and vendor dependency that the original vague headline wipes out entirely.

Why This Challenge Is Especially Urgent Right Now

The timing of McNees's late-2025 post is not coincidental. It arrives during a period of extraordinary regulatory, legal, and public-policy activity around these technologies. Precision in language is not a journalistic nicety at this moment — it is a structural precondition for coherent governance.

The European Union's AI Act, which entered into force on 1 August 2024, creates a tiered risk framework that assigns obligations based on use case and system type, with its provisions phasing in over subsequent months and years. That framework only functions correctly if practitioners, regulators, legal counsel, and press can agree on what technology they are actually discussing. Imprecise coverage can feed imprecise public comment, which in turn can shape imprecise interpretation and enforcement. The argument here is not that a single sloppy article rewrites the law, but that persistent vagueness in how technologies are described makes accurate public and regulatory reasoning harder than it needs to be.

At the infrastructure level, the distinction between ML and LLMs is simultaneously a distinction in computational cost, energy consumption, cooling requirements, and data center demand — topics that are live policy debates in multiple jurisdictions. A lightweight image classification model running inference on a low-power edge device and a frontier LLM inference cluster handling continuous traffic from millions of users are not equivalent objects by any relevant engineering or environmental measure. Treating them as a single category called "AI" makes it nearly impossible to reason accurately about grid load, water usage, or carbon footprint. These concerns are far from hypothetical: several U.S. states and localities have moved to scrutinize or restrict new large data centers, and the logic of any such policy depends entirely on understanding which specific workloads are driving demand, and at what scale.

There is also the cumulative question of what persistent imprecision does to public understanding over time. When every product is "AI," the term loses all signal value. Readers become simultaneously over-alarmed (because "AI" conjures science-fiction superintelligence) and under-informed (because they cannot evaluate actual capabilities or specific risks). Both failure modes can serve corporate interests and undermine the kind of genuine public scrutiny that a mature technology ecosystem requires to function democratically. A society that cannot distinguish between a spam filter and a generative model cannot make informed decisions about which deserves regulatory attention, public investment, or collective wariness.

Objections — and Why They Do Not Hold

There are predictable counterarguments to this kind of precision campaign. They deserve direct answers rather than dismissal.

"Readers don't know what 'LLM' means"

This is the most common objection, and it is weaker than it sounds. Readers also did not know what "broadband," "algorithm," "API," or "cryptocurrency" meant until journalists started using those terms consistently, defining them in context, and trusting readers to keep up. The answer to unfamiliar terminology is never to retreat to meaningless vagueness; it is to explain the term once, briefly, and then use it precisely. "A large language model, or LLM — the type of software that powers ChatGPT" is a single parenthetical clause. It is not a burden on the reader; it is a service to them.

"The distinctions are too technical for mainstream coverage"

The distinction between "machine learning" and "large language model" is not appreciably more technical than the distinction between "a vaccine" and "an mRNA vaccine." Health journalists successfully navigated that transition in real time, under enormous pressure, during a global pandemic, in publications ranging from The New York Times to regional newspapers. Technology journalists can do the same — and the technology beat has, if anything, a more technically literate core readership than the general health beat does.

"'AI' is the term companies and regulators use, so we have to follow their language"

This is precisely backwards. When companies use vague terms strategically, following that usage is not neutral journalism — it is amplification of a marketing frame. The press's job is to translate corporate language into terms that accurately inform readers, not to reproduce it uncritically. If a regulator uses "AI" loosely, that imprecision in the regulation is itself worth noting and interrogating.

"What about multimodal models and AI agents? The lines are already blurring"

This is the most technically serious objection. Multimodal models such as OpenAI's GPT-4o and Google's Gemini family process images, audio, and text together — are they LLMs? Systems described as "AI agents" chain LLM calls with tool use, retrieval, and code execution; does "LLM" still apply? The honest answer is that the vocabulary is genuinely under construction at the frontier. But that complexity is an argument for more precision, not less: "a multimodal foundation model" or "an LLM-based autonomous agent" are more accurate and more informative than "AI." When the precise term does not yet exist or is contested, say so explicitly. Naming the uncertainty is far better than hiding it behind a catch-all.

Applying the Journalist Challenge: A Practical Summary

The following checklist distills the journalist challenge into a repeatable editorial standard that any writer, editor, or fact-checker can apply before a piece goes to publication.

Step	Question to Ask	Action If Unclear
1	Is "AI" appearing as a noun or adjective in this sentence?	Flag for review. Assume it needs replacing.
2	What does the system actually do — classify, generate text, rank, detect anomalies?	Ask the source directly. Check technical documentation if available.
3	What type of model underlies the product?	If the company will not say, report that explicitly.
4	Is the output natural language at scale? → Use "LLM." Is it predictions, scores, or classifications? → Use "machine learning model."	If genuinely ambiguous (multimodal, agentic), use the most specific available descriptor and acknowledge the ambiguity.
5	Have you defined the technical term for general readers on first use?	Add a brief parenthetical definition. One sentence is sufficient.
6	Would a working ML engineer or data scientist accept this characterization as accurate?	If uncertain, run it by one before publishing. Most are willing to give thirty seconds of feedback on a single sentence.

Key Takeaways

The following points summarize the case for the journalist challenge and its practical implications for anyone writing about technology today.

• Robert McNees's late-November 2025 Bluesky post issued a direct journalist challenge: use "machine learning" when you mean machine learning and "LLM" when you mean LLM — and retire "AI" as a catch-all term.
• The challenge identifies a dual failure: imprecision withholds information from readers, and, in McNees's framing, it actively assists companies in obscuring how their products work and what they cannot do.
• "Machine learning" and "LLM" are related but categorically distinct; an LLM is a specific kind of ML model, but most ML systems are not LLMs. Using the terms interchangeably is a factual error, not a stylistic choice.
• Precision in tech vocabulary is a prerequisite for coherent regulation, enforceable policy, and meaningful public accountability — not a pedantic nicety confined to specialist publications.
• The main objections to adopting precise language (readers won't understand; it's too technical; everyone says "AI") all fail on examination; journalism regularly introduces and normalizes new technical terms when doing so serves readers.
• The frontier of multimodal models and LLM-based agents complicates the vocabulary — but that complexity demands more precision, not the convenient shortcut of a catch-all term.
• Every journalist and editor can apply the challenge immediately: before publishing, identify the actual model type, ask what it does and on what data, and use the most specific accurate term available.
• If a company refuses to clarify what type of model powers their "AI" product, that opacity is itself a story worth reporting — and the journalist challenge gives you the precise language to frame it.

---

What comes next depends almost entirely on whether enough journalists and editors take the challenge seriously enough to act on it. A single viral Bluesky post will not rewrite style guides overnight. But challenges like this one have a history of gaining traction precisely because they give a name — and a concrete, actionable rule — to something the technically literate segment of a readership has long felt but struggled to articulate. If newsrooms begin treating "machine learning" and "LLM" as the default terms of precision, and treating "AI" with the same skepticism they would apply to any other piece of corporate euphemism, the coverage will improve, the accountability will sharpen, and readers — technical and otherwise — will be better equipped to evaluate the most consequential technology of their lifetimes.

The challenge is not hard. It requires understanding the difference between a family of statistical methods and a specific class of large neural network — a distinction that can be learned in an afternoon and applied in a sentence. More than anything, it requires caring enough to be specific. That, as any good journalist will tell you, is the job.