AI Is Literally a Scam? Separating the Hype From the Capability

The Scam Argument, And Why It Collapses Under Its Own Weight

Somewhere in the last two years, "AI is a scam" became a load-bearing opinion for a certain kind of commentator. The claim usually goes like this: large language models do not think, they only predict the next token; prediction is not understanding; therefore anything built on top of prediction is theater, and anyone charging money for that theater is running a con. It sounds airtight until you notice it is actually three separate arguments stapled together and presented as one.

Nick Saraev's response to this critique, in a video released under the blunt title "AI is literally a scam," takes the staple out. He argues that the accusation conflates a claim about consciousness (does the system have an inner experience, does it "really" understand) with a claim about output (can the system solve a problem it has never seen before) and a claim about the market around the technology (are vendors lying about what their products do). These are not the same question, and treating them as one lets a weak argument about consciousness smuggle in a strong-sounding but unrelated conclusion about fraud.

The consciousness question is genuinely unresolved, and it probably stays that way for a while, because nobody has a working test for subjective experience in a machine, or frankly in another person beyond inference. Saraev's point, as he frames it, is that you don't need to settle it to evaluate whether a tool is useful or whether its output is novel. A calculator does not understand arithmetic in any phenomenological sense either, and nobody calls long division a scam because of that.

Capability, Incentives, Evidence, Utility: Four Questions, Not One

The most portable idea in Saraev's argument is a kind of unbundling. Instead of asking "is AI real intelligence, yes or no," he suggests you get further by asking four narrower questions separately: What can the system actually do (capability)? Who benefits from you believing a particular story about that capability (incentives)? What would count as evidence either way (evidence)? And does the output solve an actual problem for an actual user, regardless of how it was produced (utility)?

Run the "scam" accusation through that filter and it mostly survives the capability and utility checks while collapsing on the incentive and evidence checks. People who dismiss AI outright often have an incentive too: credentialed professionals whose work is partially automatable have a reason to prefer a world where the automation is fake. People who boost AI uncritically have the obvious opposite incentive. Neither incentive, on its own, tells you anything about whether a given model can do a given task. You have to actually look.

This is the frame worth stealing from the video even if you ignore everything else in it: stop asking whether AI is "real," and start asking what, specifically, a given system can do, who is telling you that, and what it would cost them if they were wrong. Applied honestly, that filter is uncomfortable for hype merchants and skeptics alike, which is probably a sign it's doing something right.

What "It Just Copies Patterns" Leaves Out

The strongest plank in the dismissal argument is the copying claim: a language model is trained on human text, so whatever it produces is just a remix of that text, dressed up to look like reasoning. Saraev pushes back on this by pointing to performance on problems that, by construction, were not in the training data — contamination-free math and coding benchmarks, and novel logic puzzles generated after a model's training cutoff. If a system that has supposedly only ever learned to imitate surface patterns can nonetheless produce correct answers to questions nobody has asked before, "it's just copying" stops being a sufficient explanation, even if it remains a partially true one.

It's worth being precise about what this does and doesn't prove. It does not prove the system "understands" in whatever sense would satisfy a philosopher. It does show that the pure-mimicry story is incomplete, because mimicry alone does not generalize to unseen problem structures the way these systems sometimes do. Saraev gestures at a specific example — he raises the long-unsolved Navier-Stokes equations in fluid dynamics as a case where AI-assisted approaches have made unexpected headway. That specific claim is the kind of thing that circulates in AI-optimist commentary without a clean citation attached, and it could not be independently verified against a primary mathematical source for this piece, so treat it here as Saraev's claim rather than an established result. The caution matters more than the example: when a creator's evidence for "AI reasons novelly" rests on an anecdote that doesn't trace back to a paper, the honest move is to flag it, not repeat it as settled.

What is independently verifiable is narrower and less flashy, but it still cuts against a flat "it only copies" story. In July 2025, Google DeepMind announced — and the International Mathematical Olympiad's own coordinators confirmed — that an advanced version of Gemini achieved gold-medal-level scoring on that year's IMO problem set, graded under the same rubric used for the actual student competitors. IMO problems are written fresh every year specifically so they can't be solved by memorization. That's a real, checkable data point in favor of "novel problem-solving," even though it says nothing about consciousness and shouldn't be over-generalized into a universal claim about all models on all tasks.

The Scam Is Real — It's Just Not the Technology

Where the "AI is a scam" framing earns some credit is in the economy that has grown up around these models, and Saraev does not pretend otherwise. Course sellers promising six-figure "AI agency" income in thirty days, SaaS wrappers charging monthly fees for a thin prompt template anyone could write in ten minutes, and dashboards claiming proprietary intelligence that is in fact an unmodified call to a commercial API — none of that is imaginary, and all of it deserves exactly the skepticism critics direct at the underlying technology. The distinction Saraev draws, and it's a fair one, is between fraud riding on top of a real capability and the capability itself being fraudulent. A patent-medicine salesman on a real railway doesn't mean the railway isn't real.

This is also where incentive-tracking earns its keep as a habit rather than a one-time check. OpenAI has reported, through its own usage research, that ChatGPT reached roughly 700 million weekly active users by September 2025 — a figure that, whatever it says or doesn't say about "intelligence," is at minimum a measure of sustained voluntary usage at a scale that is hard to fake for long. People do not return weekly, for free or for a subscription, to a tool that produces nothing useful. That doesn't settle the consciousness debate. It does make "nobody gets anything real out of this" a harder sentence to defend than "some people are overselling what this is."

FAQ

Does Saraev claim AI models are conscious?

No — his argument explicitly sets the consciousness question aside as unresolved and separate from questions about output and usefulness.

Is every AI business legitimate, according to this argument?

No. Saraev distinguishes between the underlying technology and the marketing layer built around it, and treats course-seller and wrapper-app hype as a genuine problem distinct from whether the models themselves work.

What's the strongest reusable idea here?

Splitting any AI claim into four separate questions — capability, incentives, evidence, and utility — rather than accepting a single verdict of "it's real" or "it's a scam."

Is the Navier-Stokes claim verified?

No. It appears in the source video as an example offered by the creator and could not be independently confirmed against a primary mathematical source for this piece, so it is presented here as his claim, not as fact.