Eighty-eight percent.
That is the failure rate Saturn's researchers recorded when they asked the leading large language models a hard question about UK pension tax. Not a riddle. Not a jailbreak. A pension question, of the sort a 58-year-old with a decimated retirement pot types into a chatbot at 11 p.m. because the free version is the only advisor they can reach.
The same test, run across 121 questions and repeated five times each to produce more than 10,000 discrete answers, found an aggregate failure rate of 57 percent. The best-performing model โ a paid, reasoning-tuned system the report labels "Claude Opus 5" โ still failed 39 percent of the time. The free tier failed 63 percent.
Here is the part that should stop you cold. According to FCA data cited inside the same report, 44 percent of UK consumers believe the financial information these systems give them is regulated. It is not. Fifty-six percent say they trust AI more than they trust television and radio. So we have a population that trusts a system it does not understand, over an object it does, while that system gets the wrong answer to nearly six in ten questions about money.
Read the code, ignore the roadmap. The roadmap says these are helpful assistants. The code says they are stochastic text generators that have been handed fiduciary-adjacent tasks with no compensating control.
This is not a story about hallucination. Hallucination is the symptom. The story is a structural mismatch between three curves that were never designed to intersect: adoption rising fast, capability rising slowly, and accountability not rising at all.
Context: What Was Actually Tested
Saturn is a UK fintech. That fact matters more than the report's authors would like, and I will return to it. First, the construct.
The methodology was a behavioral evaluation, not an architectural one. Nobody fine-tuned anything. Nobody measured perplexity or examined attention maps. The researchers assembled 121 questions drawn from UK personal finance territory โ savings, investment allocation, and a heavy weighting toward the two domains where errors are most expensive: tax and pensions. Each question was submitted to a set of unnamed frontier models, five times, yielding a corpus north of 10,000 answers. Each answer was then graded against a rubric.
Failure was defined in three primary categories: arithmetic error, failure to account for recent changes to tax rules, and fabrication of rules that do not exist. A fourth category โ omission of adequate risk warnings โ was also counted against a model.
That last one is where the methodology and the messaging start to diverge, and I will flag it now rather than bury it.
The test's headline finding is a 57 percent aggregate failure rate. But aggregates are the enemy of diligence. A single scalar tells you a problem exists; it does not tell you where, or how fixable it is. The report offers the scalar. It withholds the distribution. I want the per-model breakdown, the per-category breakdown, and โ critically โ the confidence intervals. None are public in the coverage I can verify.
The domain composition is likewise opaque. If a disproportionate share of the 121 questions concerned UK-specific tax and pension rules, then the evaluation is partly a measurement of geographic knowledge asymmetry rather than a measurement of financial reasoning. A model trained on a corpus weighted toward US English will bleed accuracy on HMRC thresholds for reasons that have nothing to do with its ability to reason about compounding interest. That is a confounder, not a finding.
The repetition design โ five passes per question โ is the one methodological choice I will credit without reservation. It signals that the researchers understood that these systems are non-deterministic. Same prompt, different output, every time. For a financial decision tool, that is not a quirk. It is a defect, and the report's design implicitly acknowledges it even as it fails to report the variance.
Which is the missing number I want most. Not the failure rate. The inconsistency rate: how often the same question produced materially different answers across the five passes. For a user, an answer that is wrong and stable is dangerous. An answer that is unstable โ wrong 40 percent of the time, right 60 percent, never predictable โ is worse, because it destroys the user's ability to calibrate trust at all. The report does not give me that number. It gives me the average and asks me to move on.
I have been dissecting crypto whitepapers since I was a high school junior in Chicago, dismantling 42 of them during the 2017 ICO boom. The lesson that survived from that year is that the most important data is almost always the data that is absent. When a project publishes a headline and buries the distribution, the headline is the marketing and the distribution is the truth. This report is not a scam โ I do not believe that for a second โ but it is also not a peer-reviewed artifact. It is a marketing-adjacent document from a company with a commercial stake in its conclusion, and it should be read the way I read every whitepaper: as a claim to be verified, not a fact to be absorbed.
Core: A Systematic Teardown of Three Failure Modes
Let me reverse-engineer the machine, because that is the only way to understand why the number is what it is, and why the number will not improve on the timeline the market is pricing.
The report identifies three failure modes. Each maps, one-to-one, onto a known structural weakness of the transformer paradigm. None of them are fixable by scaling.
Failure Mode One: Arithmetic.
Language models do not calculate. They predict tokens. When a system tells you that the tax-free lump sum on a ยฃ400,000 pension pot is ยฃ104,000, it has not solved a multiplication problem โ it has produced the most plausible token sequence following the phrase "25 percent of ยฃ400,000 is." Most of the time the plausible sequence is also the correct one, because the training corpus contains many worked examples. Some of the time it is not, and there is no internal mechanism that breaks character to catch the error, because there is no calculator inside the model. There is only the probability distribution over the next token.
This is why I keep telling people that the failure is architectural, not experiential. You cannot fine-tune a transformer into a reliable arithmetic engine by feeding it more arithmetic. You can only bolt an external calculator onto it and teach it to call that calculator โ a tool-use architecture โ at which point you have not improved the model, you have built a wrapper around its weakness.
According to my own audit work, this distinction is the one that institutional due-diligence teams most consistently miss. In 2025, I led the technical review of an AI-generated content platform backed by a major ETF sponsor. The pitch deck described a proprietary reasoning engine. The reality, once I pulled the API logs, was a thin wrapper around a deprecated base model with a prompt template and a rate limiter. The "engine" was a marketing layer. The latency profile exposed it: genuine multi-step reasoning shows up as variable, elevated response times, because compute is being spent. This platform returned answers in flat, suspiciously uniform time windows โ the signature of a single forward pass with no reasoning scaffold.
That same diagnostic applies here. When the report notes that the paid reasoning models outperform the free tiers by fourteen percentage points โ 49 percent failure versus 63 percent โ it is documenting the fact that more inference compute buys you better token prediction. It does not buy you a calculator. A fourteen-point improvement for what is almost certainly an order-of-magnitude increase in inference cost is a terrible exchange rate, and it is the single most important economic fact in the entire report.
Failure Mode Two: Knowledge Cutoff.
The second failure category is failure to account for changes to tax rules. This is the most damning and the least discussed.
A model is a frozen snapshot. Its weights encode the world as of the moment training stopped. In the UK, the personal allowance, the pension lifetime allowance (a moving target that has been abolished, restored in rhetoric, and reconstructed in practice), the dividend allowance, and the capital gains thresholds all change on an annual or sub-annual cadence. A model trained in early 2024 does not know what happened in the Autumn Budget of late 2024. It will not merely be uncertain โ it will be confidently wrong, because the token sequence that was correct under the old rules remains just as plausible under the new ones, and the model has no way to know that a fact it learned has expired.
This is not a knowledge gap that more data solves. It is a structural property of any system whose knowledge is baked into parameters rather than retrieved from a live source. The fix is retrieval-augmented generation โ RAG โ where the model queries an authoritative, continuously updated knowledge base, such as HMRC's own published guidance, before answering. Every serious enterprise deployment of these systems in a regulated domain uses RAG. The report does not mention it once, which is either an oversight or a deliberate omission, and I will explain below why the second is more likely.
Failure Mode Three: Rule Hallucination.
The third category is the fabrication of rules that do not exist. A model invents a tax relief, cites a nonexistent allowance, describes a pension wrapper that was never legislated. This is the canonical hallucination failure, and it is the most dangerous of the three, because a fabricated rule is indistinguishable, to a lay user, from a real one. It comes with the same confident register, the same apparent specificity, the same absence of hedging.
When a system gets arithmetic wrong, a careful user can catch it with a calculator. When it fails to update for a tax change, a careful user can catch it by checking the current rate. When it invents a rule, the user has no reference point from which to detect the invention, because the failure is the creation of a fact that does not exist in the world to be checked against. Fabrication is the only failure mode that is unverifiable by the person most likely to be harmed by it.
My 2022 investigation into TerraUSD taught me to look for exactly this pattern in systems that present as self-regulating. The dual-token mechanism of Terra and Luna was, on its surface, an elegant algorithmic stabilizer. It was, in reality, a self-referential loop: the peg held because the market believed it would hold, and the belief was the collateral. When belief wavered, there was nothing underneath. I published a 40-page teardown a year before the collapse arguing that the model was mathematically unstable under stress, citing the specific code dependencies and the incentive misalignments that made the death spiral inevitable.
The LLM failure here rhymes with that structure. The model's apparent authority is the collateral. There is nothing underneath it โ no calculator, no live database, no grounded rulebook. The confidence is real. The substance is inferred. And when enough users act on the inference, the aggregate of their individual errors becomes a systemic event, because HMRC does not offer a do-over.
The report floats a specific figure: a single bad pension answer could expose a user to up to ยฃ17,500 in HMRC penalties. I cannot independently verify that number from the coverage, and I treat it as directionally illustrative rather than precise. But the mechanism is sound. When a system advises a user to make a withdrawal that breaches an allowance, the penalty is real, the user pays it, and the system that gave the advice has no legal obligation to compensate anyone.
The Compensation Void
Here is the load-bearing wall of the entire argument, and it is a legal fact rather than a technical one. Saturn's CEO states plainly that AI financial advice is unregulated, and that a consumer who follows it has none of the recourse that a human advisor's client would have. If a regulated financial advisor gives you negligent advice in the UK, you have a complaint path, an ombudsman, and a compensation scheme. If a chatbot gives you the same advice, you have a chat log and a shrug.
This is not a gap in the market. It is a gap in the law. The advice is happening. The harm is possible. The liability is assigned to the person with the least information and the least power in the transaction โ the user. A system that is defined as "not advice" while being consumed as advice is the purest form of regulatory arbitrage ever constructed, and it is running at scale right now.
The Trust-Cognition Mismatch
Now to the number that actually terrifies me, and it is not 57.
It is 44 and 56.
Forty-four percent of UK consumers believe AI-generated financial information is regulated. Fifty-six percent trust AI more than they trust broadcast media. Read those two together and the failure rate becomes almost irrelevant to the risk calculation, because the risk was never contingent on the model being wrong. The risk is that the model is wrong at a rate of 57 percent and is being trusted at a rate of 56 percent by a population that misidentifies its regulatory status at a rate of 44 percent.
Entropy is not the danger. False certainty is.
I want to be rigorous about the comparison, though, because the report uses it for drama and I distrust drama as a data source. "Trust AI more than TV and radio" is a constructed contrast. Baseline trust in UK broadcast media is historically mediocre, and it is deteriorating independent of anything AI does. Framing AI's trust as superior to a low and falling benchmark inflates the apparent endorsement. If the comparison had been AI versus a trusted institution โ say, the NHS, or a chartered accountant โ the numbers would look very different. So I discount the comparison's rhetorical weight.
But I do not discount the direction. The signal underneath the framing is real: consumers are extending trust to these systems faster than they are updating their mental model of what the systems are. And that lag โ trust running ahead of comprehension, adoption running ahead of capability โ is the exact condition under which every historical financial technology bubble has transferred wealth from the uninformed to the informed.
The sociological read is that trust here is not a rational assessment of accuracy. It is a transfer of a prior habit. People trusted search engines to find facts; they now trust chatbots to generate them, not realizing that "finding" and "generating" are opposite operations. One retrieves what exists. The other produces what is plausible. The interface looks identical. The epistemics are inverted.
I saw the same inversion in the 2021 NFT market. I ran a statistical analysis of 15,000 OpenSea transactions and found that 85 percent of the reported volume was wash trading by coordinated wallets โ a mechanical fact that the community narrative of "organic demand" completely obscured. The data said one thing. The culture said another. And the culture won, because the culture is what people act on. I was harassed for "ruining the fun" when I published it. I have stopped expecting the release of accurate data to change behavior on contact. It changes behavior only after the harm, and often not even then.
Contrarian: Saturn Is the Judge, the Jury, and an Interested Party
Now I have to be fair to the bulls, because the failing of a cold dissector is the failure to notice when the critique itself is captured.
Saturn ran this test. Saturn reported this test. Saturn's CEO used the test to call for FCA action. Saturn did not disclose its own product's failure rate. Sit with that sequence for a moment.
This is the forensic pattern I look for in every incentive analysis: who benefits from the conclusion, and what did they decline to measure? Saturn benefits, unambiguously, from a world in which general-purpose AI is declared unsafe for financial advice. That declaration does not harm a specialized, licensed fintech โ it clears the field. If the FCA responds by requiring a license to give AI financial advice, OpenAI and Google are suddenly facing a compliance burden they are not built for, while Saturn already holds the license, the compliance apparatus, and the position to sell the safe alternative. The report is a warning to consumers that doubles, with remarkable precision, as a competitive moat being poured in real time.
None of which means the test is wrong. This is the part the cynical reader misses. An interested party can also be a correct one. The 2017 project whose whitepaper I dissected was run by people with every incentive to lie, and they were lying. But the forensic conclusion was not "the messenger is compromised." It was "here is the centralized database this project calls a blockchain." The claim stood or fell on the code, not on the messenger's motive.
Apply the same standard here. Saturn's motive does not invalidate the 57 percent failure rate. The failure rate has to be assessed on the methodology, and the methodology, as I said, is under-disclosed. So the correct verdict is neither "the report is exposed as marketing" nor "the report is a neutral scientific finding." The correct verdict is: the report establishes that a problem exists with high confidence, and does not establish the size of the problem with any confidence at all. Those are two different epistemic objects and the report's framing collapses them.
The bulls, for their part, are right about something the report buries. General-purpose models fail, yes โ but the failures are engineering-addressable. Every one of the three failure modes has a known mitigation. Arithmetic fails? Call a calculator. Knowledge expires? Retrieve from a live source. Rules hallucinate? Ground the generation in an authoritative rulebook and refuse to answer outside it. None of these are research problems. They are deployment problems. The reason the models in this test scored terribly is not that they are incapable of better โ it is that they were deployed as bare chat interfaces with no scaffolding, because scaffolding costs money and a bare chat interface is what the user-facing product ships.
So the report is correct that AI finance advice is dangerous today, and misleading if it implies the danger is intrinsic to the technology rather than to the deployment decisions that deliberately omitted the mitigations. The bulls are wrong that the technology is ready. The bears are wrong that it cannot be made ready. The actual finding is that it is being sold before it is ready, to people who cannot tell.
The one comparison the report should have made and did not: how do retrieval-augmented, tool-calling, human-reviewed systems perform against bare chat interfaces? If the answer is 3 percent failure versus 57 percent, then the entire consumer-protection debate is really a debate about deployment standards, and the fix is a regulatory one โ mandate the scaffolding โ rather than a technological one.
What the Report Gets Right That Its Critics Will Miss
I have spent a great deal of this piece on what is missing. Let me be precise about what is present, because it is not nothing.
First, the recognition that failure must be evaluated at the level of the consumer decision, not the level of the model benchmark. Most AI evaluations measure trivia, reasoning puzzles, and coding tasks. This one measured whether a person can use the output to make a financial decision that survives contact with the tax authority. That is a materially different and more useful frame, and it is the frame that institutional risk teams should be adopting.
Second, the inclusion of "missing risk warning" as a failure category. This is subtle and correct. A model that gives a technically accurate answer without noting that it is not a licensed advisor, that tax situations vary, and that the user should seek qualified help is still failing the user โ it is failing at the level of the system's social contract rather than its factual output. The report treats that as a failure, and it should. A correct answer delivered with false confidence is not the same product as a correct answer delivered with appropriate epistemic humility, and the industry has been allowed to conflate the two for years.
Third, and most importantly, the report refuses the comfortable conclusion. It does not say "AI is a useful tool that should be used carefully." It says the tool is being used for something it cannot yet reliably do, at scale, by people who cannot evaluate it. That is a hard claim and it is supported. The bones are good. I am arguing about the flesh.
The Missing Numbers
For the record, and for anyone building a diligence model on this, here is what a complete study would have to publish, and what this one does not:
One โ the per-model failure rates, disaggregated, with confidence intervals. A 57 percent aggregate across an unnamed set of models is close to uninterpretable.
Two โ the per-category split between arithmetic errors, staleness errors, and fabrications. These have entirely different mitigations. A study that lumps them cannot tell a builder where to spend money.
Three โ the consistency rate across the five repetitions. As argued, likely the most decision-relevant number in the dataset, and absent.
Four โ the geographic composition of the question set, so the reader can separate "this model cannot reason about finance" from "this model was never trained on UK tax law."
Five โ a comparison of bare chat versus retrieval-augmented, tool-calling, human-reviewed deployments. Without this, the study cannot distinguish a technology ceiling from a deployment choice, and those two conclusions lead to opposite policy responses.
Six โ third-party audit or peer review. A study run by a party with a commercial stake in its conclusion, whose methodology is disclosed only partially, is a press release with charts until someone independent reproduces it.
None of these omissions make the report wrong. They make it incomplete, and incomplete is more dangerous than wrong in a market that is about to regulate on the strength of it.
The Model Name Problem
The report refers to "Claude Opus 5." I cannot verify that this model exists as named. It may be a typographical error, a future-dated snapshot, or a naming convention I have not tracked. This matters more than it sounds.
If the name is wrong, it is a small signal that the report's technical details were not carefully checked, which lowers my confidence in the parts of the methodology I cannot independently inspect. If the name is right and refers to a released model, then the report's core claim โ that the leading system still fails 39 percent of financial questions โ is a claim about the state of the art, which is a much heavier claim than a claim about a mid-2024 model. A single label shifts the entire weight of the finding, and the report does not resolve it. That is a diligence failure, and on a document that is being positioned as a basis for regulatory action, diligence failures are not stylistic complaints.
I flag this without hedging: I am not certain what model was tested, and neither, apparently, is the document.
Takeaway: Price the Absence of a Control, Not the Presence of a Flaw
The comfortable reading of this report is that AI chatbots are unreliable for financial advice and should be used with caution. That reading is true and it is useless, because the people who most need the warning are the least equipped to act on it, and the entities most able to fix the problem have no incentive to until they are forced.
The load-bearing insight is elsewhere. The 57 percent failure rate is a technical fact, and technical facts get engineered away. The 44 percent regulatory misperception is a legal fact, and legal facts do not fix themselves. What this report actually documents is a market running a financial-advice product with no compensating control: no liability, no required disclosure, no live data grounding, no human review, no compensation path. Every other high-stakes advice market in the developed world carries at least one of those controls. This one carries none, and it is growing at consumer scale.
My judgment: the technology will improve on a multi-year horizon, the failures will shrink, and none of that will matter if the accountability vacuum is left open, because the harm is generated by the gap between what users believe they are getting and what they are actually getting, and that gap survives every model upgrade. Volatility is just unpriced risk. Here the volatility is consumer trust, the risk is the eventual correction, and the market has not priced either one.
I have been watching this pattern since 2017: a system presents as one thing, functions as another, and the delta between the two is where the money moves and the damage lands. The only question that matters now is whether the FCA prices it before the users do.
Logic doesn't lie. The average is 57 percent. The trust is 56 percent. The overlap is the story, and nobody is selling the hedge.",,