Journalist Challenge: Stop Saying "AI" for Machine Learning and LLMs

What the Challenge Actually Says — and Who Issued It

The original post, published on November 22, 2025, reads:

"Journalist challenge: Use 'Machine Learning' when you mean machine learning and 'LLM' when you mean LLM. Ditch 'AI' as a catch-all term, it's not useful for readers and it helps companies trying to confuse the public by obscuring the roles played by different technologies."
— Robert McNees (@mcnees.bsky.social), Bluesky, November 22, 2025 — primary source

A pointed journalist challenge posted to Bluesky on November 22, 2025 is cutting through the noise of tech coverage with a precise, three-part demand: use "machine learning" when you mean machine learning, use "LLM" when you mean a large language model, and retire the blanket term "AI" as a catch-all. Robert McNees — a theoretical physicist and professor at Loyola University Chicago who studies black holes and quantum gravity — issued the call, and its implications for how the public understands technology run deep.

McNees is not a media critic by trade. He is a theoretical physicist. That is precisely what gives the observation its sting. A scientist trained in precision notices immediately and viscerally when public language around technology becomes dangerously imprecise. The challenge is not about pedantry — it is about accountability. If readers cannot distinguish between a recommendation algorithm, a fraud-detection classifier, and a generative text model, they cannot meaningfully evaluate claims made about any of them, and they certainly cannot hold the companies deploying those systems responsible when something goes wrong.

The framing matters because it is also, implicitly, a challenge to editors, PR teams, and product managers. The problem does not start in newsrooms. It begins upstream, in the press releases and earnings calls that reporters quote, and it compounds when journalists accept imprecise source language without interrogating it.

Why "AI" Has Become a Liability for Readers

The phrase "artificial intelligence" has a legitimate history. It traces back to John McCarthy's 1956 Dartmouth Conference, where it described the broad scientific ambition of making machines exhibit intelligent behavior. For decades it stayed a research umbrella — useful in academic framing, but never a precise engineering term. What happened over the last decade is that the word migrated from research papers into marketing decks, and then from marketing decks into news coverage, where it now does the work of four or five distinct technical concepts at once.

Consider what a single news cycle might bundle under "AI" without differentiation:

• Supervised machine learning classifiers used to detect spam, flag fraudulent transactions, or sort medical images by pathology — trained on labeled datasets and evaluated against known ground truth
• Recommendation engines built on collaborative filtering, matrix factorization, or gradient-boosted decision trees, serving content on streaming platforms and social media feeds
• Large language models (LLMs) — transformer-based neural networks trained on massive text corpora to predict and generate human-like text (e.g., GPT-4, Claude 3, Gemini 1.5, Llama 3)
• Diffusion models and GANs (Generative Adversarial Networks) used to generate images, audio, or video — diffusion models work by iteratively denoising random noise into structured output; GANs pit a generator network against a discriminator in an adversarial training loop
• Reinforcement learning systems that learn strategy through simulated trial and error by maximizing a reward signal, as in robotics, game-playing agents, or supply-chain optimization
• Rules-based "expert systems" that predate modern ML entirely — hard-coded conditional logic dressed up in press releases as cutting-edge AI

Each of these has a different architecture, a different failure mode, a different regulatory surface, and a different set of societal implications. Calling them all "AI" does not simplify — it obscures. And as McNees's challenge makes explicit, that obscurity is not an accident. It serves a purpose.

The Corporate Fog Machine: How Vague Language Serves Big Tech

McNees's sharpest observation — the one most likely to be quietly skipped over — is that the imprecision of "AI" actively helps companies trying to confuse the public by obscuring the roles played by different technologies. This describes a mechanism, not just a side effect, and it deserves careful unpacking.

When a company says "our AI detected that pattern," it is making a claim that sounds impressive but carries no falsifiable content. Which AI? A logistic regression trained on labeled data? A fine-tuned LLM? A heuristic rule set wrapped in a neural network scoring layer? The vagueness is useful precisely because it prevents scrutiny. If a journalist had instead written "the company's machine learning classifier, trained on user behavior data, flagged the account," readers would immediately ask reasonable follow-up questions: What was the training data? How large was it? Was it representative of the user population? Who labeled it, and under what guidelines? How is demographic bias monitored? What is the false-positive rate?

The same dynamic plays out across product announcements, regulatory filings, and earnings calls. "AI" absorbs all of those questions into a single aspirational word and redirects them toward science fiction imagery — HAL 9000, Skynet, sentient machines — rather than toward the mundane but consequential engineering choices being made right now. This is not a conspiracy theory; it is straightforward brand management. And tech coverage has been enabling it for years.

When technology reporting uses "AI" as a catch-all, it doesn't democratize understanding — it forfeits it. Readers are left with a sense of awe and no tools for judgment.

There is also a liability dimension. Companies that obscure whether a consequential decision — a denied loan, a flagged job application, a moderated post — was made by a rules-based system, a trained ML model, or a human reviewing model output are in some cases actively managing regulatory exposure behind the fog of "AI." Precise language, in that context, is not a style preference. It is an accountability instrument.

Machine Learning vs. LLM vs. AI: A Terminology Reference

Part of the journalist challenge is simply knowing which term fits which technology. The reference table below is a working guide for reporters, editors, and technically literate readers who want to apply the discipline McNees is calling for. It includes reinforcement learning as a distinct and commonly mislabeled category.

Term	What it actually means	When to use it	What it does NOT mean
Machine Learning (ML)	Algorithms that improve at a task through exposure to data, without being explicitly programmed for each case. Includes linear models, decision trees, random forests, support vector machines, and neural networks.	When a system was trained on labeled or unlabeled data to perform classification, regression, clustering, ranking, or anomaly detection.	Does not imply language ability, reasoning, or generative capability. A spam filter is ML; it is not an LLM.
Large Language Model (LLM)	A specific class of deep neural network — transformer-based — trained on vast text corpora to model and generate language. Examples: GPT-4, Claude 3, Gemini 1.5, Llama 3, Mistral.	When a product or feature involves text generation, summarization, question answering, code completion, or chat interfaces powered by a transformer model.	Not synonymous with all ML. Not synonymous with "AI." Not the same as a search engine or a rules-based chatbot.
Generative AI	A broader category covering models that generate new content — text, image, audio, or video. LLMs are a subset; diffusion models (DALL-E 3, Midjourney, Stable Diffusion, Sora) are another major subset.	When the key feature is content generation across modalities and you need a term broader than "LLM" but more specific than "AI."	Does not cover most ML applications such as fraud detection, recommendations, or medical imaging classifiers.
Reinforcement Learning (RL)	A training paradigm in which an agent learns to maximize a cumulative reward signal through trial and error in an environment. Used in robotics, game-playing agents (AlphaGo, AlphaZero), and increasingly in fine-tuning LLMs (RLHF — Reinforcement Learning from Human Feedback).	When a system learns a policy through interaction and reward, rather than from a static labeled dataset.	Not the same as supervised ML. Not inherently generative. RLHF is a training technique applied to LLMs, not a synonym for LLMs themselves.
AI (Artificial Intelligence)	A research field and broad philosophical concept covering the goal of building machines that exhibit intelligent behavior. The term dates to the 1956 Dartmouth Conference.	In historical context, academic or policy framing, or when genuinely referring to the broad field rather than a specific product or technique.	Should not be used as a substitute for ML, LLM, generative AI, or any specific technique in news reporting about a concrete product or system.

What Precision Actually Looks Like in Practice

The journalist challenge to use "machine learning" when you mean machine learning is not an abstract style guide note — it changes what a story can actually say and which questions it forces into the open. Here are concrete examples of how swapping in the right terminology unlocks more precise, more useful sentences.

Before and After: Rewriting the Vague

• Vague: "The platform uses AI to moderate content."
  Precise: "The platform uses a machine learning classifier, trained on previously removed posts, to flag content for human review — with humans making the final removal decision."
• Vague: "The startup's AI writes personalized emails."
  Precise: "The startup's product uses an LLM — specifically a fine-tuned open-weight model — to generate personalized cold outreach emails from user-provided context."
• Vague: "The hospital adopted AI for diagnostics."
  Precise: "The hospital deployed a supervised machine learning model trained on radiology scans to assist radiologists in flagging potential anomalies, with radiologists retaining diagnostic authority."
• Vague: "The company's AI assistant was jailbroken."
  Precise: "Researchers found a prompt injection technique that caused the company's LLM-based assistant to ignore its system-level safety instructions and produce prohibited outputs."
• Vague: "The hiring platform uses AI to screen candidates."
  Precise: "The hiring platform uses a supervised ML model trained on historical hiring outcomes to rank applicants — a practice that regulators have flagged for perpetuating historical bias."

Each revised sentence opens a door the vague version closes. The reader now knows what kind of failure is possible, what data was involved, what the human role is, and where regulatory scrutiny might apply. These are not stylistic preferences — they are the difference between coverage that informs and coverage that merely generates impressions.

The Developer Perspective: When Imprecision Has a Price Tag

For engineers, the stakes of imprecise language are immediate and practical. Consider a concrete scenario: a product manager tells the development team "we need to add AI to the onboarding flow." The engineering team has received zero actionable information. Does that mean:

• Integrating an LLM API (OpenAI, Anthropic, Google) to generate personalized welcome copy — a matter of days, but with ongoing per-token inference costs and data privacy considerations?
• Training a custom supervised classifier on existing user behavior data to surface relevant features to new users — weeks to months, requiring a labeled dataset and MLOps infrastructure?
• Adding a collaborative filtering recommendation layer to suggest next steps based on similar users — a distinct architecture with cold-start problems for new accounts?
• A rules-based decision tree disguised as "AI" in the product brief?

Each option carries wildly different timelines, infrastructure costs, regulatory considerations under laws like the EU AI Act, and long-term maintenance burdens. The journalist challenge, applied internally to product communications, would force these conversations to happen with real vocabulary rather than marketing language — producing better software, more honest roadmaps, and fewer blown deadlines. Developers increasingly serve as translators between what leadership says ("AI") and what engineering actually builds. Precise public language would reduce that friction — and the whiplash tech workers feel when grand AI promises collide with engineering reality would diminish with it.

The Broader Problem: Why Journalism Drifted Here

Adopting "AI" as a catch-all was not entirely a failure of rigor — it was also a failure of incentives. Headlines containing "AI" tend to outperform headlines containing "machine learning" in engagement metrics. "AI" carries narrative weight: it evokes agency, consciousness, and drama. "Machine learning model updated its fraud-detection feature weights" does not produce the same engagement, even when that is precisely what happened.

There is also a genuine expertise gap. Most newsrooms lack dedicated ML reporters who can make these distinctions fluently on deadline. General technology reporters are expected to cover everything from semiconductor supply chains to social media policy, and "AI" has functioned as survivable shorthand — a way to signal that a story involves algorithmic systems without requiring the journalist to understand which ones. That is understandable as a coping mechanism in under-resourced newsrooms. It is no longer adequate.

The technologies now being collapsed into "AI" are affecting hiring decisions, loan approvals, medical diagnoses, legal judgments, child welfare assessments, and electoral systems. The vocabulary must keep pace with the consequences. McNees's challenge is elegant because it asks for something genuinely minimal: not a full technical education, but simply the discipline to use the correct word.

A journalist covering a court case uses the correct legal terms — mens rea, voir dire, habeas corpus — because imprecision would be professionally embarrassing and substantively misleading. A journalist covering a drug trial uses the correct pharmacological and statistical terms. Technology coverage deserves the same standard. The argument that ML terminology is too arcane for readers falls flat when outlets routinely expect readers to absorb far more specialized vocabulary from finance, medicine, and law without complaint. As the gap between corporate promises and user reality continues to widen, the need for accountability journalism — journalism with a precise vocabulary — has never been greater.

How to Apply This in Your Newsroom

The McNees journalist challenge does not require a newsroom to hire a machine learning engineer. It requires editorial will and a small number of concrete process changes. Here is a practical starting point.

For Reporters

• Ask the source to be specific. When a company representative says "our AI does X," ask on the record: "Is this a machine learning model trained on data, an LLM, a rules-based system, or a combination?" Most will answer — and the answer is almost always newsworthy.
• Use the terminology table as a self-check. Before filing, scan the draft for every instance of "AI" and ask whether "machine learning," "LLM," "generative AI," or a more specific term fits better.
• Note the human role explicitly. Precision about technology includes precision about human oversight: does a human review the model's output before a consequential decision is made? That distinction belongs in the copy.

For Editors

• Add a terminology section to the style guide. A one-page internal reference — defining ML, LLM, generative AI, and RL with worked examples — takes an hour to draft and pays dividends indefinitely.
• Push back on "AI" in headlines as a matter of course. Ask whether a more specific term fits within the character limit. It often does: "LLM" is shorter than "AI system."
• Require sourcing for technical claims. If a story states that a company uses machine learning, the story should be able to state — at minimum — what kind of task the model performs and whether the company disclosed training data provenance.

For Developers and Product Teams

• Enforce precise language in product briefs and announcements. Require that any external communication specifying "AI" also specifies the category of system, its training basis, and the human oversight mechanism — even in summary form.
• Use internal terminology glossaries. A shared Notion page or wiki entry defining your organization's specific use of "ML model," "LLM integration," and "rules-based logic" prevents cross-functional miscommunication and accelerates engineering scoping.

Key Takeaways

• The challenge is simple but consequential: Robert McNees's November 22, 2025 Bluesky post challenges journalists to use "machine learning" and "LLM" precisely, rather than defaulting to the semantically empty catch-all "AI."
• "AI" as a label actively serves corporate interests by making it harder to scrutinize which specific technology was used, how and on what data it was trained, what its failure modes are, and who bears accountability when it errs.
• Machine learning, LLMs, generative AI, reinforcement learning, and "AI" as a research field are distinct terms with different architectures, training paradigms, failure modes, and regulatory surfaces — the reference table above should be used.
• Precise terminology changes what stories can actually say: rewriting "AI" as "machine learning classifier" or "LLM" immediately forces questions about data, training, oversight, and risk that vague language permanently forecloses.
• For developers, the stakes are practical and financial: imprecise language in product briefs creates misaligned requirements, inflated expectations, incorrect cost estimates, and extended translation cycles between leadership and engineering teams.
• The fix requires minimal expertise but real editorial discipline: this is not about requiring journalists to write code, but about insisting on the same terminological rigor that any other specialized beat — law, medicine, finance — already demands.
• Newsrooms can act now with three concrete steps: update style guides, require sources to specify technology type on the record, and use the reference table as a pre-publication self-check.

What Comes Next: Can Newsrooms Actually Change?

McNees's challenge is unlikely to restructure tech journalism overnight — but challenges of this kind have a documented history of seeding style guide changes, editorial debates, and individual reporter decisions that compound over time. The more technically fluent corners of the internet, where developers and researchers read and share coverage, are already applying informal pressure on outlets that conflate these terms. Reader corrections, social media callouts, and letters from researchers form a slow but real disciplining mechanism.

The regulatory environment is creating a harder external forcing function. As LLMs and other ML systems become the explicit subject of legal proceedings — under the EU AI Act's provisions for general-purpose AI (GPAI) models, which impose additional systemic-risk obligations on models trained using more than 10^25 FLOP of cumulative compute (a threshold above which a model is presumed to pose systemic risk under the Act); under ongoing FTC scrutiny in the United States; and in emerging intellectual property litigation around training data — the language of coverage must become precise enough to track specific regulatory categories. When a regulation explicitly applies to GPAI models above a defined compute threshold, a journalist who writes "new AI rules" and moves on has failed to communicate anything actionable to affected companies, investors, or the public.

The path forward is not complicated. Editors should add terminological guidance to style guides. Reporter training should include at minimum a working distinction between ML categories and between ML and LLMs. Sources who use "AI" as a catch-all should be asked — on the record — which specific technology they mean. And when they refuse to answer, that refusal is itself a story.

McNees's challenge is a prod, not a curriculum. It asks for one thing: the right word, used at the right moment. In a landscape where the stakes of algorithmic systems touch nearly every domain of public life, that one thing turns out to matter enormously.