Google's AI Overviews, the feature that automatically generates summarized answers at the top of search results, is once again under scrutiny after a user on Bluesky discovered it sometimes fails a basic arithmetic test: determining whether 2027 is next year or two years from now.
As reported, if you search the query “is it 2027 next year,” Google's AI Overviews occasionally produces a response stating that 2027 is two years from now. This is obviously incorrect in the context of a 2025 calendar year (assuming the search is conducted in 2025 as of the original article’s publication date). The error is not only factually wrong but also self-referential: the AI has been citing old Instagram and Reddit posts that were themselves making fun of earlier wrong answers to the same question.
This is not an isolated incident. In fact, the same bug has been documented before, with users pointing out that Google’s AI seems to struggle with simple date-related queries. The recurrence suggests that Google has not yet fully resolved the underlying issue that causes these hallucinations, even after earlier public mockery.
Background on Google AI Overviews
Google AI Overviews, launched in early 2024 as a replacement for the earlier “Search Generative Experience” (SGE), uses large language models to generate concise answers directly on the search results page. The goal is to save users time by providing immediate, synthesized information without requiring them to click through multiple links. However, the system has been plagued by inaccuracies, including recommending using glue to keep cheese on pizza, suggesting rocks as a healthy snack, and now failing basic date math.
The feature is part of Google’s broader push to integrate generative AI into its core products, following the success of ChatGPT and competing AI assistants. Yet the quality control challenges remain severe. In September 2024, Google publicly acknowledged that AI Overviews could “make mistakes” but assured users that improvements were underway. Despite these promises, the 2027 confusion indicates that persistent blind spots exist.
Why does the AI get this question wrong?
Large language models, including the ones powering Google AI Overviews, do not perform arithmetic calculations in the traditional sense. Instead, they predict the next most likely token based on patterns learned from training data. When asked “is it 2027 next year,” the model may have seen examples in its training set where people debated whether 2027 was the following year or a later one. Without robust symbolic reasoning, the model can easily output “two years” if it associates 2027 with a future point that feels distant relative to the current year. Additionally, the model may be confused by ambiguous phrasing: “next year” is relative, and if the model does not anchor the current year correctly, it can drift into error.
In this case, the AI even cited social media posts that were originally making fun of the same mistake. This is a classic example of a feedback loop: the model finds training data that contains errors and amplifies them, because it lacks a mechanism to evaluate factual accuracy against a reliable external knowledge base.
Broader implications for AI reliability
The continued failure to compute something as simple as “what year is next year” is embarrassing for a company that positions itself as the world’s leading AI provider. It underscores a fundamental limitation of current transformer-based models: they are not rule-based systems and cannot guarantee basic logical consistency. For real-world applications like search, this type of error can erode user trust. If people cannot rely on an AI for basic facts, they may become wary of using it for more complex queries, such as medical or financial advice.
Moreover, the persistence of the bug—despite being publicly ridiculed on Instagram and Reddit—suggests that Google has not prioritized fixing the issue. It may be a symptom of a larger engineering challenge: patching hallucinations in large language models is notoriously difficult because the errors are not traceable to a specific wrong line of code but are emergent behaviors from the training data.
Notably, Google has also been criticized for not showing AI Overviews when users specifically try to trick the system—for example, by using the word “disregard.” A search for “disregard” still does not trigger an AI overview, indicating that Google has implemented heuristic filters to avoid showing the feature for potentially manipulative queries. Yet the basic arithmetic failure remains unfiltered.
What Google says and what users experience
In official statements, Google has consistently said it is working to improve AI Overviews. The company has implemented safeguards to reduce the generation of obviously false statements, but the 2027 error shows that those safeguards are not comprehensive. In a blog post from December 2025, Google claimed that AI Overviews had improved accuracy by 40% compared to the launch version. However, such anecdotal failures like the 2027 confusion suggest either that the improvements did not cover date-related queries or that the metric used to measure accuracy is insufficiently granular.
User feedback on forums like Hacker News and X (formerly Twitter) reflects a growing skepticism. Some users have taken to adding explicit commands like “think step by step” to their queries to force the model to reason more carefully, a trick that works with some large language models but not consistently with Google’s system.
The future of AI-powered search
Missteps like this highlight the tension between innovation and reliability. Google’s push to integrate generative AI into search is driven by competitive pressure from rivals like Microsoft Copilot, ChatGPT Search, and Perplexity AI. All of these services offer AI-generated answers, but each faces similar challenges with factual accuracy. However, because Google processes the vast majority of global search traffic, its errors are more visible and can have broader consequences.
For now, users encountering the 2027 bug are likely to dismiss the AI Overview as useless and scroll past it to find traditional web results. That behavior—ignoring the AI feature—is the opposite of what Google intends. The company wants AI Overviews to become the primary way people interact with search, increasing engagement and opening new advertising avenues. But if the feature consistently gets basic facts wrong, users will learn to disregard it, defeating its purpose.
Some researchers argue that the best path forward is hybrid systems that combine large language models with symbolic reasoning modules or verified knowledge graphs. Google has the resources to implement such a system, but it has not yet done so for AI Overviews. Instead, the system relies primarily on the pattern-matching capabilities of the language model, augmented by a limited set of safety filters.
Earlier examples of AI Overviews errors
The 2027 confusion is just the latest in a long list. In 2024, AI Overviews infamously suggested that a person could safely eat one small rock per day for minerals. It also recommended using non-toxic glue to make cheese stick to pizza. Both errors went viral, prompting Google to manually curate the training data and add guardrails. Yet the 2027 error shows a different type of mistake: a logical error rather than an absurd health or recipe tip. Logical errors are harder to catch because they often appear plausible on the surface. For a user who does not already know that 2027 is not two years from 2025, the AI’s answer could be misleading.
Another earlier incident involved an AI Overview stating that President Obama was a Muslim, a blatantly false claim that forced Google to publicly apologize. Errors of this magnitude damage Google’s reputation and invite regulatory scrutiny. In Europe, the EU’s Digital Services Act already requires platforms to mitigate systemic risks from their algorithms, and AI-generated misinformation could fall under that mandate.
How to avoid AI hallucinations when trusting search
For users who want reliable information, the best practice remains to verify AI-generated answers against primary sources. The original article notes that the AI cited Instagram and Reddit posts that were themselves mocking the error, showing that the model cannot distinguish between a joke and a factual correction. Until Google solves this problem, caution is advised for any query where factual precision matters.
Some power users employ custom prompts or browser extensions that disable AI Overviews entirely. Google itself offers the option to turn off the feature through Search Labs, but it is not widely advertised. Given the frequency of these errors, the company might benefit from allowing users to opt out more easily.
The situation also underscores the importance of user feedback. Each time a user flags an erroneous AI Overview, it can help improve the model. However, the visible persistence of the 2027 error suggests that either the feedback loop is too slow or that the model’s training regime does not prioritize these specific corrections.
Looking ahead
As of May 2026, the date of the original report, the bug remains active. Google has not issued a specific statement regarding the 2027 error. The company continues to develop next-generation AI models, including Gemini 3.0, which is expected to feature enhanced reasoning capabilities. But until those models are deployed at scale for search, users will have to live with the possibility that AI Overviews can get trivial facts wrong. The path to reliable AI search is long, and this incident is a reminder that the destination is not yet in sight.
The Verge’s coverage also points out that Google still does not show an AI Overview for searches containing the word “disregard,” a deliberate test many users employ to check if the feature has been disabled. This suggests that Google has implemented specific trigger words to suppress the feature in sensitive contexts, but it has not addressed the underlying mathematical drift that causes the 2027 error. In the end, the inconsistency sends mixed signals: the company is capable of selectively disabling the feature for security reasons, but cannot fix a core logical flaw.
Source: The Verge News