If you have noticed newer AI models hallucinating more in your own testing, you are not imagining it, but the full picture is more mixed than that. The evidence grade here is mixed: some newer reasoning models post higher hallucination rates on specific, difficult tests, while those same models and others improve on many other factuality benchmarks at the same time. There is no clean, one-directional ai hallucination rate trend running across the industry. The honest answer to whether ai models are getting less accurate is: it depends on the model, the task, how the test is scored, and whether the model is being rewarded for answering or for admitting it does not know.
That distinction matters more than it sounds like it should, and it is worth sitting with before you look at any numbers.
What the claim actually means
People use "hallucination" loosely, but the studies behind this claim do not, and neither should you if you want to read them correctly.
A hallucination is an answer that sounds confident and plausible but is wrong, invented, or unsupported. Accuracy is a different measurement: whether the answer matches the correct one. Those two numbers are not opposites of each other, and that is the part most casual reads of this topic get wrong. A model can answer more questions correctly overall while also producing more false answers, because it is attempting more questions than it used to.
Abstention is the piece that ties this together. When a model says "I don't know" instead of guessing, that counts against its raw accuracy score, even though it kept a false claim out of the world. Attempted accuracy strips out the abstentions and measures only how reliable the model is on the questions it chose to answer. Calibration asks something related but distinct: whether the model's confidence matches how often it is actually right. A model can know a lot and still be poorly calibrated if it sounds certain when it is wrong.
One more distinction worth keeping in your head: short-answer benchmarks test something different than long-form writing does. A test with one verifiable fact per question is not the same task as a 2,000-word article carrying dozens of claims, where the odds of at least one slip go up simply because there are more chances for one.
The evidence that newer models can hallucinate more
The clearest example driving this whole conversation is OpenAI's own system card for o3 and o4-mini, published in April 2025. OpenAI reported accuracy and hallucination rates on two evaluations, SimpleQA and PersonQA, for o3, o4-mini, and the earlier o1 model side by side.
On SimpleQA, o3 scored 49% accuracy with a 51% hallucination rate, compared with o1's 47% accuracy and 44% hallucination rate. On PersonQA, o3 scored 59% accuracy against o1's 47%, but its hallucination rate was 33% against o1's 16%. o4-mini fared worse on both: a 79% hallucination rate on SimpleQA and a 48% hallucination rate on PersonQA, well above o1's numbers on both. This is the clearest documented case of ai models hallucinating more within the very same evaluation family, not across some vague industry-wide comparison.
Read that o3 result carefully, because it is the crux of the whole hallucination-rate trend argument. o3 was more accurate than o1 on both tests, and it also hallucinated more. OpenAI's explanation was that o3 tends to make more claims overall, which produces more correct claims and more incorrect ones together. The system card says plainly that more research is needed to explain why. This is solid evidence for a specific, narrow point: a newer model can get better at answering while also getting worse at not making things up. It is not evidence that every newer model follows that pattern.
A 2025 preprint titled "Reasoning Models Hallucinate More" adds a possible mechanism. It found that reinforcement learning aimed at getting the right final answer can increase fabricated or unsupported statements inside the model's reasoning steps along the way, even when the model lands on the correct answer at the end. The paper tested this using question-answering and math benchmarks and proposed a training method to reduce the effect. Treat this as one study pointing at a plausible training tradeoff, not as proof that reasoning training always causes this.
A newer benchmark called AA-Omniscience makes the same point from a different angle. It asks models 6,000 no-context questions across six domains, and it scores models in a way that punishes wrong guesses rather than rewarding raw attempts. GPT-5 high answered 39% of questions correctly, but its hallucination rate on the rest was 81%. Grok 4 had a similar accuracy figure with a 64% hallucination rate. Claude 4.5 Opus, by contrast, had one of the lowest hallucination rates in the same test. The lesson is not that frontier models are simply worse now. It is that on hard, closed-book questions, some models keep guessing when they should stop, and that habit shows up as a high hallucination rate even when the model is otherwise capable.
The evidence against a universal decline
Now look at the other side, because it is just as real and it is why "newer ai models factual errors" is not a story you can tell in one direction.
A 2024 study in JMIR looked at hallucinated references generated in an academic research context and found GPT-4 fabricated references at a lower rate than GPT-3.5: 28.6% versus 39.6%. Bard, for comparison, fabricated references 91.4% of the time in the same study. That is a clean improvement from one model generation to the next, in a narrow but real task.
SimpleQA itself, the same benchmark behind the o3 and o4-mini numbers above, tells a more layered story once you look past the headline pair. Claude 3 Opus scored 23.5% correct with 36.9% incorrect, and Claude 3.5 Sonnet improved that to 28.9% correct with a similar incorrect rate but far fewer unanswered questions. GPT-4o answered 38.2% correctly with only 1% left unattempted, well ahead of GPT-4o-mini's 8.6%. o1-preview reached 42.7% correct, ahead of o1-mini's 8.1%. Stronger, larger versions inside the same family generally outperformed their smaller siblings, and reasoning models tended to abstain more often than standard chat models, which pulled their raw correct-answer rate down without necessarily making them less trustworthy.
The FACTS Leaderboard, introduced in December 2025, tests factuality across four different settings: whether an answer is grounded in supplied documents, whether it reasons correctly across images and text, whether it answers from internal knowledge alone, and whether it stays accurate while using a search tool. The best model overall scored 69%, which says something on its own about how much room even the newest systems have left. But the ranking shifts depending on the task. On the search-assisted setting, Gemini 3 Pro reached 85.6% accuracy with a low hedging rate, GPT-5 reached 81.4%, and Claude 4.5 Opus reached 80%. On the closed-book knowledge setting, the ranking and the gaps between models both changed. No single model or model family wins everything, and no generation of models is simply worse across the board than the one before it.
Vectara's hallucination leaderboard, running since 2023 and covering more than 160 models on summarization tasks, is often cited as longitudinal proof of an industry trend. It is useful, but it has real limits worth knowing: it tests summarization specifically, not open-ended writing or research, its automated detector is imperfect by its own published numbers, and the detector itself has changed versions over the years it has been running. It is one data point among several, not a verdict on its own.
Why the results seem to contradict each other
Once you hold all of that at once, the apparent contradiction mostly resolves. A few things are doing the work here.
Accuracy and hallucination rate are measuring different behaviors, not the same thing from two angles. A model that attempts almost every question can post both a higher accuracy score and a higher hallucination rate than a cautious model that leaves hard questions unanswered.
The scoring rule matters more than most people assume. A newer paper published in Nature in 2026 makes this point directly: when a test scores abstention as a failure, guessing becomes the mathematically better strategy for a model trying to maximize its number, even when guessing produces more false claims. The paper walks through a SimpleQA example where o4-mini's raw accuracy looked stronger than GPT-5-mini's, even though o4-mini made errors on more than three-quarters of its answers, because it answered almost everything while GPT-5-mini abstained more and made fewer mistakes as a result. That is not proof that models are secretly worse. It is proof that the scoreboard can reward the wrong behavior.
Closed-book knowledge and search-assisted answers are also not the same skill, and a model can be strong at one and weak at the other. And a "newer" model is not one clean variable. It might be larger or smaller than what came before it, a reasoning model or a standard chat model, trained to answer more readily or to hedge more, given tool access or not, or simply evaluated with a different prompt. Release date alone does not explain a hallucination-rate difference, and treating it as if it does is where most of the "AI is getting worse" framing goes wrong.
What this means for teams drafting with AI
None of this is an academic argument if your job involves publishing what a model writes. A February 2026 release from NP Digital, based on an analysis of 600 prompts across six large language models and a survey of 565 U.S. marketers, put some numbers on the day-to-day version of newer ai models factual errors. Treat it as directional agency research rather than an independent industry census, since it does not name all six models tested, but the pattern it reports lines up with what the benchmarks above already show, and it is the clearest picture yet of the ai model reliability risk marketers actually deal with week to week.
Nearly half the marketers surveyed, 47.1%, said they run into AI errors several times a week. More than 70% said they spend one to five hours a week fact-checking AI output. And 36.5% said a hallucinated or incorrect AI claim had already made it into something published. Only 23% said they felt comfortable using AI output without a human review step, which tells you something about how the people closest to this work actually trust it.
The risk is not that AI got dumber. It is that higher capability does not automatically buy you more reliability, and a more fluent, more confident draft can make an error easier to miss rather than harder to produce. The conditions where this bites hardest are predictable: niche topics where a model's knowledge is thin, current events and anything that changes month to month, prompts that ask for many facts at once, and any claim about a competitor, a price, or a specific statistic. A single wrong number buried in an otherwise well-written paragraph is exactly the kind of error a fast read misses.
The practical response is not to pick your model by release date or by a leaderboard rank and call it done. It is to keep a verification step in the loop regardless of which model produced the draft, because that step is what actually catches an unsupported claim before a reader does. Managing ai model reliability risk marketers face is a process decision, not a one-time model choice. DeepSmith's own writing pipeline grounds every draft in stored product, persona, and brand context rather than a from-scratch prompt each time, which narrows one common source of drift, but grounding a draft in real context and fact-checking its specific claims are two different jobs, and neither one replaces the other.
Verdict and what would change it
The defensible conclusion is narrower than either "AI is getting less accurate" or "don't worry about it." Are ai models getting less accurate as a category? The evidence collected here does not support that framing, and there is no single ai hallucination rate trend you can point to and call settled. Progress in reasoning and general capability does not guarantee progress in factual reliability. Some newer models answer more confidently and more broadly, and that combination can raise hallucination exposure specifically in the moments where the model does not actually know the answer. Other newer models measurably reduce hallucination on the same kinds of tasks. Both things are true, and a single release date tells you which one you are looking at.
What would strengthen this verdict in either direction is a series of independent, apples-to-apples evaluations run across successive model generations, using the same prompts, the same tools, the same scoring method, and an explicit accounting for abstention and for long-form writing rather than only short factual answers. Until that exists, the safest reading of any single hallucination statistic is to ask what task it measured, how it scored a refusal to answer, and whether it tested short factual recall or the kind of long-form drafting your team actually does.
