Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
The evidence points to uneven progress, not a clear industry-wide plateau. Some models have reported lower factual-error rates on specific tests, but performance varies with the question, the scoring method and the benchmark. A benchmark can also stop distinguishing systems without proving that real-world reliability has stopped improving. There is no single, comparable long-term measure here that settles the trend across AI providers and everyday uses.
What does “improving reliability” actually mean?
Reliability is broader than whether a model invents a fact. It can include whether an answer is accurate, whether its claims are supported by evidence, whether the model admits uncertainty, and whether it behaves consistently across different phrasings and levels of difficulty. Those properties do not always move together.
Measurement matters, too. A claim-level error rate, the share of answers containing a major error, and overall answer accuracy have different denominators. A model may answer more questions and still make more unsupported claims; a model that declines difficult questions may be safer but score worse on a test that rewards only correct answers.
What evidence shows improvement?
Some reported model comparisons show fewer factual errors
In its 2025 GPT-5 system card, OpenAI reported that GPT-5-main had a 26% lower factual-claim hallucination rate than GPT-4o, and that GPT-5-thinking had a 65% lower rate than OpenAI o3. These are relative comparisons from the developer’s evaluations, not independent proof of a broad industry trend. They apply to the card’s tested prompts, models and grading methods; they should not be read as a percentage-point reduction or as a guarantee about other tasks. OpenAI’s GPT-5 system card also reports different measures, including claim-level errors and responses with one or more major errors, which should not be treated as interchangeable.
#1 Best Overall
Results vary sharply by task and framing
Stanford HAI’s 2026 AI Index reports hallucination rates ranging from 22% to 94% across 26 models on a new accuracy benchmark. That range describes results on that benchmark, not the probability that an AI answer in general is wrong. The report also finds that performance can change substantially depending on whether false information is framed as another person’s belief or the user’s belief. Even a small change in how a question is posed can therefore affect what a score says about reliability.
Difficulty is another dividing line. The authors of FactBench, an ACL 2025 evaluation built from 1,000 prompts across 150 topics and tiered by difficulty, found that factual precision declined from easy to hard prompts. On that evaluation, a larger Llama 3.1 model performed comparably to or worse than its 70B variant. This is evidence that scale alone did not ensure better factuality in that test—not a rule that larger models are generally less reliable.
Rank #2
- brand: Pearson
- ARTIFICIAL INTELLIGENCE: A MODERN APPROACH, 4TH EDITION
Does benchmark saturation mean AI reliability has plateaued?
Not by itself. A benchmark is a measuring instrument: if models begin scoring similarly near its ceiling, it may no longer separate newer systems well. That is a problem with the test’s ability to discriminate, not direct evidence that performance on useful real-world tasks has stopped advancing.
A 2026 analysis of 60 language-model benchmarks found that nearly half exhibited saturation, with saturation rates increasing with benchmark age. The study also associates expert curation with greater resilience. Its finding concerns benchmark longevity; it does not establish a flat trajectory for everyday AI reliability. The PMLR study on benchmark saturation is therefore a reason to scrutinize old scores, not to declare a universal plateau.
Why can a reliable model lose to a guesser?
Because a scoring rule can reward the number of correct answers while ignoring the risk of answering when uncertain. A 2026 Nature paper on accuracy incentives argues that common accuracy or pass-rate measures can favor a model that guesses over one that abstains on an unanswerable question. If wrong answers carry little extra penalty and abstentions count as failures, the test may encourage confident guessing rather than calibrated caution.
For a meaningful comparison, ask whether the evaluation gives credit for abstaining, penalizes unsupported answers, and reports both correctness and answer coverage. A system that answers a larger share of questions is not automatically more trustworthy.
What makes a model comparison fair?
Before treating two reported scores as evidence of progress or decline, check whether they measure the same thing. A useful comparison should make these details visible:
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware match- Task and difficulty: factual recall, open-ended writing, biography generation, document-grounded summaries and adversarial questions stress different abilities.
- Error definition: determine whether the score counts erroneous claims, answers containing a major error, or incorrect answers overall.
- Abstention and coverage: see whether the model can say it does not know, how that is scored, and how often it answers.
- Evidence and grading: check whether answers are grounded in sources, who or what grades them, and whether human judgments validate automated grading.
- Benchmark health: ask whether the test still distinguishes systems or has become saturated.
- Scope and independence: distinguish a developer’s comparison of its own systems from an independent, cross-provider evaluation.
Uncertainty around the scores matters as well. NIST’s 2026 report, based on three benchmarks and 22 API-access frontier language models, explains that some common benchmark metrics can produce invalid uncertainty estimates or rely on unrecognized assumptions. It demonstrates statistical modeling approaches for estimating generalized accuracy and uncertainty. The report’s model count describes its evaluation, not the entire market. NIST AI 800-3 makes the case for asking how certain a measured difference is, not just which system ranked higher.
Best Value
So, are we starting to plateau?
The available evidence supports three different conclusions, which should not be collapsed into one. First, some specific measured factual-error rates have fallen in developer-reported comparisons. Second, some benchmarks are saturating and may no longer reveal differences among systems. Third, a clear long-term rise or plateau in everyday reliability across providers remains unresolved: the available results do not form one harmonized, independent time series using the same real-world tasks and error definitions over multiple years.
The most defensible answer is uneven progress rather than a clean plateau. Hallucinations have declined on some measured tasks, while difficult prompts, varying evaluation rules and aging benchmarks make the overall direction hard to summarize with one number.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →
