Skip to content
Multigrid

AI myths, hype & media literacy

How to read a claim about AI — from a launch post, a demo or a headline — by finding the step where the evidence stops supporting the conclusion.

There are two ways to be wrong about this subject and they are equally easy. One is to believe the announcements. The other is to decide that all of it is marketing, which feels like sophistication and is a forecast in its own right — one that has already been wrong several times about capabilities that arrived. Neither position requires knowing anything, which is why both are popular.

These pages take a different job: for each widely repeated claim, find the exact step where the evidence stops carrying the conclusion. That step is usually specific and usually boring. A benchmark measures something narrower than its name suggests. A demo was cut from many attempts. A number that was calculated per datacenter per year gets quoted per person per question. A capability appears in one prompt format and vanishes in another. Once you can see the step, you no longer need anyone to tell you what to think about the claim.

Where a question is genuinely open — and several of the biggest ones are — these pages say so and describe what would settle it, rather than picking the side that reads better.

Does an LLM “Understand” Anything?

The question splits into three different questions with three different answers, and most arguments about it are two people answering different ones.

5 min read

“AI Is Just Autocomplete” — What It Gets Right

The dismissal is mechanically accurate and its conclusion does not follow, and the gap between those two facts is the interesting part.

4 min read

The Stochastic Parrot Argument, Fairly Presented

The phrase became a slogan for a position the paper it comes from did not quite take — and several of its actual arguments have aged extremely well.

4 min read

Does AI “Learn” From Your Conversations?

Three completely different mechanisms get called learning, they have different privacy consequences, and the honest answer depends on which one you meant.

4 min read

Can AI Be Creative? Define It First

Under the operational definitions creativity researchers actually use, the answer is already yes for some and clearly no for others — the disagreement is about which definition counts.

4 min read

Why AI Can’t Count the Letters in a Word

The most-shared AI failure has a completely mundane cause, and understanding it tells you which other tasks will fail the same way.

4 min read

The “AI Will Replace Programmers” Claim, Unbundled

Three separate claims travel under one headline, they need different evidence, and they are not all going the same way.

5 min read

Benchmark Marketing: How to Read a Launch Post

The presentation choices that make a launch chart look decisive, and the one question that defeats each of them.

4 min read

Demo Videos and What They Hide

Six things a demo can be doing that are not lying, and the questions that distinguish a capability from a performance.

5 min read

“Trained on the Entire Internet” and Other Loose Claims

The phrase bundles four separate claims, three of which are false as stated, and the true one has consequences people rarely draw.

4 min read

Parameter Count as a Marketing Number

Bigger genuinely did mean better for several years, then three separate developments broke the relationship — and the number mostly stopped being published.

5 min read

Context Window Claims vs Usable Context

The advertised number is an architectural limit, not a promise about behaviour, and the gap between them is measurable and large.

4 min read

AGI Announcements and Goalpost Movement

The target moves in both directions, for reasons that are mostly not dishonest, and the fix is to argue about capabilities rather than about the term.

5 min read

AI Doom and AI Boosterism: Reading Both Critically

Incentive analysis applied evenly to both camps, plus the reason incentive analysis cannot settle who is right.

4 min read

Misleading AI Statistics: Six Recurring Errors

Six error forms that account for most bad AI numbers, each with the arithmetic that exposes it — and no blame attached to anyone.

4 min read

“AI Uses X Litres of Water per Question”

How a per-query water figure is constructed, the four choices that move it by orders of magnitude, and what a number must state before it can be compared to another one.

4 min read

Emergent Abilities: The Mirage Debate

A concrete argument that sharp capability jumps are an artefact of how the task was scored — and what it does and does not overturn.

4 min read

Sentience Claims: Why They Keep Recurring

Four mechanisms make a language model an unusually strong trigger for attributing inner life, and none of them is evidence either way.

4 min read

AI Detection in Schools: The Harm of False Positives

Why a detector with an impressive accuracy figure still accuses large numbers of innocent students, and why the errors are not evenly distributed.

4 min read

A Bullshit Detector for AI Claims

Ten questions to ask of any AI announcement, ordered so that the cheapest ones eliminate the most claims.

4 min read

Other topics