By Hamza Faisal

Every week brings a new AI "breakthrough," a record funding round, and a benchmark that somebody's model just shattered. Most of it is real progress. Some of it is marketing. The difference matters, because the same announcements move stock prices, shape policy, and decide where billions get invested. This is a reader's guide to separating signal from hype — five questions to ask before believing any AI claim.

1. What does the benchmark actually measure?

Benchmarks are tests, and like all tests, they measure what they were designed to measure — no more. A model that tops a coding benchmark may still write buggy code in a real codebase; a model that aces a multiple-choice exam may still fail at open-ended work. When a company announces a record score, ask: was the test public or private? Was the model trained on similar test data? And does the benchmark resemble the job you actually want the model to do? A record on a narrow test is a clue, not a verdict. The most honest announcements report several benchmarks, including ones where the model does not win.

2. Is the funding round a valuation or a business?

A startup "raising $1 billion at a $16 billion valuation" sounds like success. It is, strictly, a bet: investors buying a share of future profits that may never arrive. Valuation is not revenue, revenue is not profit, and a funding round is not a product launch. The questions that matter: how much revenue does the company actually make, who pays for it, and what does it cost to serve each customer? AI inference — running models for users — is expensive, and several celebrated companies lose money on every query while hoping scale fixes the arithmetic. A big round proves investors are optimistic. It does not prove the business works.

3. Who is making the claim, and what do they sell?

Every AI announcement comes from someone with an interest. A chipmaker benefits from stories about insatiable demand. A model lab benefits from stories about imminent breakthroughs. An investor benefits from stories about the company they just funded. That does not make the claims false — but it means the claim should be weighed against independent evidence: peer-reviewed work, third-party evaluations, customer testimony, or audited numbers. One company's blog post is a press release. Two independent sources saying the same thing is a story.

4. Is this new, or is it a new name for something old?

AI marketing loves new vocabulary: agents, copilots, reasoning, world models, superintelligence. Sometimes a new term marks a genuine capability — models that can use tools and complete multi-step tasks really are different from chatbots that only answer questions. Often, though, the term is packaging: a familiar product with a new label, or a demo that works in a controlled setting but not in the wild. The test is always the same: can an ordinary user, without the company's help, reproduce the result? Demos are promises. Products are proof.

5. What would change your mind?

The strongest habit a reader can build is asking what evidence would falsify a claim. If a company says its model is "the best," what would show it is not? If a pundit says AI will replace a profession "within two years," what happened to the same prediction two years ago? Claims that cannot be checked — vague timelines, secret data, results "coming soon" — deserve the most scepticism, not the least. And claims about the distant future deserve less weight than claims about the present, because the future has a habit of embarrassing forecasters.

Red-flag phrases to watch for

Certain phrases should raise your guard immediately, because they are doing persuasion work rather than information work. "Revolutionary" and "game-changing" describe nothing measurable. "Industry-leading" is true for whoever defines the industry narrowly enough. "Up to" before a speed, accuracy or efficiency number means the best case in the best conditions, not the typical case. "AI-powered" on a product that previously worked fine without AI usually means a feature was renamed, not improved. And any claim followed by "experts say" without naming the experts is an appeal to authority with the authority removed. None of these phrases prove a claim is false — but each one is a reason to look for the number behind the adjective.

A worked example

Try the five questions on a typical announcement: "Startup raises $500 million at a $10 billion valuation for its AI agent that beats humans at office work." Question one: beats humans on which benchmark, under what conditions, judged by whom? Question two: what is its revenue, and what does each customer interaction cost to serve? Question three: who announced it — the startup, its investors, or an independent evaluator? Question four: is an "agent" here a genuinely autonomous system, or a chatbot with a few connected tools and a new name? Question five: what evidence would show the claim is wrong — and is the company sharing anything that could show it? If most answers are "not disclosed," you do not have a breakthrough. You have a fundraising announcement wearing a breakthrough's clothes.

The bottom line

AI is genuinely transforming technology, and the pace of real progress is remarkable. That is precisely why hype discipline matters: when everything is called a breakthrough, nothing is. Read the benchmark, follow the revenue, check the source, ignore the label, and demand checkable evidence. The readers who do that will understand AI better than the people writing the press releases.