Back to blog

Hyperfence Blog

Money doesn't buy a better answer. It buys a second opinion.

By Oli Cundale

Money doesn't buy a better answer. It buys a second opinion.

Two developers are handed the same brief. One pays for a leading AI assistant; the other uses a free one. Two students are given the same essay task, one with paid AI access and one with free access.

The intuition is obvious. The one who pays wins. And if money buys better work, then a fairly small subscription buys a large and arguably unfair advantage - over classmates, competitors and anyone who cannot afford it.

The evidence complicates that intuition. Large studies of AI at work have found it narrowing gaps rather than widening them. Paying can certainly improve an answer, but it cannot guarantee one worth trusting. It may, however, give more room to question what comes back: more attempts, access to different models, and more opportunities to catch a mistake before it matters.

The best known study followed 5,172 customer support staff at a software company - real employees doing their real jobs, as their employer rolled out an AI assistant that suggested replies for them. Productivity was measured as ‘problems resolved per hour’, and it rose 15%. But the gain was not evenly spread: the least experienced staff got both faster and better, while the most experienced got slightly faster but slightly worse.

In a planned experiment with 453 professionals, each was set a writing task drawn from their own occupation. Half chosen randomly were provided with ChatGPT. They finished 40% faster and their work scored 18% higher - graded by experienced people in the same professions who did not know which submissions had been written with help, and who were asked to judge each one as though they had met it at work. “Inequality between workers decreased.”

The clearest result comes from an experiment published in August 2026 with 1,174 adults. Given a workplace problem to solve, the participants with more education substantially outperformed those with less. Given AI, about three-quarters of that gap disappeared. So if the worry is that AI makes the strong stronger, the measurements do not currently support it. They point the other way, and consistently. These studies compared access to AI with no access; they did not compare paid and free tiers.

Two questions remain.

The first is what happens when you take the AI away again. The 2026 study tested that too, and found that “a sizable gap re-emerges”. AI levelled how well people performed, and some of that gain stuck. But most of it did not: the levelling lasted largely as long as the access did.

The second is more serious. These studies measured how good the work looked, not whether it was right - and the writing study’s authors say so themselves: they did not think it feasible to hire fact-checkers to assess the accuracy of what their participants produced. Quality was judged by professionals reading it the way a colleague would. Which is exactly how most work gets judged, and exactly how a confident error survives.

One study looked underneath. A Stanford security team gave developers a set of programming tasks, some with an AI assistant and some without. The developers using the assistant “wrote significantly less secure code” - and were “more likely to believe they wrote secure code” than those without it. Confidence and quality moved in opposite directions.

But one group did better than the rest: those who “trusted the AI less and engaged more with the language and format of their prompts”. Many of the others simply accepted what it produced and moved on.

What money actually buys

Paying for a stronger model does not reliably buy safer code. A 2023 benchmark called SALLM put five models through security-sensitive Python tasks. It found GPT-4 the best of them at producing code that actually worked while, in its authors’ words, “not generating the most secure code” - and its own older sibling, GPT-3.5, struck a better balance between the two. Nor has that been solved since: across eight current models tested in 2026, every one produced insecure code in more than half of the scenarios designed to invite it.

What the frontier has become dramatically better at is finding the problem. One team ran the same evaluation in October 2024 and again in September 2025 and watched detection and repair rates climb from roughly a half to around three-quarters. The same team notes that AI-written code is about as likely to contain flaws as human-written code, and that models “may confidently provide incorrect information, posing risks for less experienced developers”.

These are separate studies with different tests, but the direction is clear enough. The advantage money buys is less about authorship than about review.

A paying user may well get a better first draft. What they reliably get is more of everything that follows it: more attempts, a model strong enough to criticise the previous model’s work, and the headroom to run the job twice and compare. A free tier can do some of that too, but against tighter limits and usually a weaker second opinion.

So the divide is less about what you can produce than about how thoroughly you can find out whether it is any good.

The student and the essay

The same problem appears in education. The student who can pay asks for an essay, then asks a second model to attack it, then checks that the sources it cited actually exist. The student who cannot pay can do the same, but fewer times and against a weaker critic, which makes it likelier they stop at the first confident answer - including, sometimes, a confidently wrong one, with citations to papers that don’t exist.

Meanwhile, the systems intended to catch AI use remain unreliable. OpenAI withdrew its own on 20 July 2023 “due to its low rate of accuracy”: it had correctly identified 26% of AI-written text while wrongly flagging 9% of human writing as machine-made. Worse, they fail unevenly. A Stanford study found that detectors “consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified” - apparently because they read a smaller vocabulary as a machine signature.

The same paper found that simple prompting both defeated the detectors and removed the bias. The student who knows how to prompt escapes. The student who writes plainly gets accused - whether or not they used anything at all.

Watermarking offers a different approach. Instead of guessing whether text looks AI-written, a model can embed a statistical pattern as it generates it. Google uses this technique in Gemini, and Anthropic began rolling it out across Claude in August 2026.

But a watermark is not proof of authorship. Anthropic says it can show only that Claude was probably involved somewhere in the text. Short passages are harder to detect, substantial rewriting or translation can remove the signal, and there may be little to detect when AI merely edits human-written text.

Most importantly, the absence of a watermark proves nothing. Not every model uses one, and even when one is present it may not survive editing. For universities, that makes watermarking useful provenance information, but not a reliable way to decide whether a student cheated.

What a subscription doesn’t buy

A more serious concern is that a well-resourced company could turn a powerful AI on a competitor. The capability is real and growing - Britain’s AI Security Institute now puts open-weight models four to seven months behind the frontier on offensive cyber work, down from six to ten months a year earlier.

The subscription is not how it would be done though, because the expensive models are the ones that refuse. In a recent independent evaluation, an open-weight model refused none of the offensive security tasks it was set, while Claude Opus 4.7 refused so consistently that the evaluators could not complete the test. Anthropic does grant vetted organisations - critical infrastructure operators, mostly - access to a version with looser safeguards, on the stated grounds that nobody has yet worked out how to release cyber capability safely to everyone. But that is granted, not purchased, and no amount of money moves it.

The way around that is well known. Downloadable models can have their refusals removed by editing the weights directly - a technique called “abliteration” - and such versions are published openly, advertised as uncensored. That does cost money, but not to a vendor. It costs hardware, because you have to run the resulting model yourself.

A study published in June 2026 took a family of coding models, stripped out their refusals, and measured whether they would then write a specific well-known security flaw into otherwise safe code. The largest model tested had refused every such request beforehand. Afterwards it complied between 88% and 97% of the time. It is small enough to run on a single consumer graphics card.

The researchers’ more interesting finding was that willingness and ability are separate things. Their smallest model also stopped refusing - and then simply wasn’t good enough, succeeding only a quarter to half the time. Stripping out the safety training does not make a model capable. It uncovers whatever capability was already there.

So there is a threshold, and it is low. The subscription that refuses costs $100 to $200 a month. Running something that doesn’t refuse needs one graphics card (or a rented one from about $1.50 an hour).

The refusals in expensive AI products are real, and the evidence suggests they work. But they only constrain people using those products. They are not a financial barrier to anyone running an unrestricted model elsewhere.

So who is paying?

It is tempting to picture all this as rich countries against poor ones. Microsoft’s diffusion research found that in early 2026, 27.5% of working-age people in developed countries had used a generative AI tool, against 15.4% in the developing world, and that the gap had widened over the previous six months. That measures use rather than paid use, but the paid tier sits on top of it: up to $200 a month, every month, which across much of the world is not a subscription decision at all.

The more uncomfortable figure is domestic. Bank of America, looking at its own customers’ accounts, found only around 3% of its household customers paying for any AI service - in the country that builds the models. This is not a line between nations. It is a line between a very small number of people and everybody else, drawn in roughly the same place everywhere.

What is actually being rationed

Look at the three protections in this article together and they have similarities. Refusals hold for the people using the expensive products. Watermarks mark the people using the marked products. Detectors are trusted on the people who never had a way to argue back. Each one works on whoever stayed inside the system, and on nobody who stepped outside it.

The advantage that holds up under measurement is not a better first draft. It is the means to find out whether the first draft is any good - and in the one case where somebody checked properly, that was exactly what separated a secure result from an insecure one.

So the question is less who can generate than who can check. And checking is the capability that has quietly become a premium product: the better the reviewer you want, the more you pay. The people producing the least examined work are the ones who can least afford to examine it.

This runs further up than individuals. On 1 September Anthropic released Claude Mythos 5.1 to vetted partners without giving Britain’s AI Security Institute the pre-release access it had been given for earlier models - the first such exclusion, according to the Financial Times, and one UK officials reportedly read as part of a wider tightening of American control over who may use frontier systems. Anthropic has not publicly explained the decision. The body whose entire job is checking could not get the thing it exists to check.

That seems like the wrong thing to ration. Verifying what an AI produced should not itself require renting the most expensive AI available - it is the kind of check that can be made small, fast and local, run on your own machine at a cost that does not climb with how much you use it.

Two developers, two students, the same brief. The one who pays is not guaranteed the better answer. They are guaranteed the second opinion - and on the evidence so far, the second opinion is the whole game.

Sources

Does AI widen or narrow gaps

Code, security and checking

Detection

Watermarking

Refusals, and who is allowed to check

Access and pricing