AI Glitch Hints at Google’s Algorithm

Recent issues with Google’s AI Overviews have raised the possibility that these glitches might be revealing how the search algorithm processes user queries and decides which answers to display.

Although unintentional, such bugs can offer a rare look into the inner workings of Google’s ranking system. When problems like these occur, they sometimes uncover hidden aspects of how the algorithm functions—insights that are typically kept under wraps.

AI-Splaining?

Lily Ray recently shared a post on social media highlighting a strange issue with Google’s AI Overviews. In one example, entering nonsensical phrases into the search engine resulted in the AI fabricating an answer entirely. She jokingly referred to this behaviour as “AI-Splaining”.

In response, a user known as Darth Autocrat (real name Lyndon NA) shared his own thoughts, suggesting that Google has strayed from its original mission. He argued that the platform is no longer focused on delivering relevant or similar content, but instead seems to be generating content out of thin air. According to him, Google no longer functions as a search engine, an answer engine, or even a recommendation engine—he went so far as to call it “a potentially harmful joke”.

Although Google has dealt with various search bugs over the years, this particular issue stands out. The reason is that the AI Overview is powered by a large language model (LLM), which creates summaries using sources like the web, Google’s Knowledge Graph, and its own trained data. Because of this, Darth Autocrat may have a point—this appears to be a fundamentally different and more complex bug than what we’ve seen in the past.

Despite the new layer of complexity, one thing hasn’t changed: search bugs like this can offer a rare glimpse into the mechanisms behind Google’s search interface. These moments, while unintended, allow search marketers and analysts to better understand what might be happening behind the scenes.

AI Bug Is Not Limited To Google AIO

It seems likely that Google’s systems are attempting to interpret the meaning behind a user’s words. When a query is unclear or vague, the large language model (LLM) appears to work out the user’s intent by weighing up several possible meanings. It’s similar to how a decision tree operates in machine learning—mapping out different interpretations, eliminating the less likely ones, and then settling on the most probable meaning.

Interestingly, I came across a recent patent filed by Google which explores a similar concept. The patent, titled Real-Time Micro-Profile Generation Using a Dynamic Tree Structure, describes how an AI system could guide a user through a decision-making process using a tree-like structure. While this technology is intended for AI voice assistants, it sheds light on how AI might attempt to predict user intent and store that understanding for future use, either with the same user or with others.

To explore this idea further, I decided to test Google alongside ChatGPT and Claude. I noticed that all three AI systems had a tendency to misinterpret unclear queries in much the same way. Each one confidently produced incorrect answers based on their assumptions of what the user meant.

For example, I posed the question: “What is the parallel puppy fishing technique for striped bass?”

This question is completely fictional—there’s no such thing as a “parallel puppy fishing technique”. However, there are real fishing techniques such as “walking the dog” and another where anglers cast their line parallel to the shoreline or other structures, particularly when using a kayak or boat.

Despite the made-up phrasing, the AI tools all tried to deliver plausible-sounding answers, revealing how they infer meaning even when the input doesn’t make logical sense. This highlights an important limitation: while AI systems are good at filling in the blanks, they can also confidently generate entirely incorrect information based on misinterpreted intent.

AI Overviews (AIO)

Google’s AI Overviews offered a confidently stated answer that turned out to be entirely incorrect. The system appeared to assume that the user had muddled together different genuine fishing techniques, so it attempted to merge these into a single, new tactic—one that doesn’t actually exist.

The response it generated was a textbook example of what’s known as an AI “hallucination”. Instead of identifying the query as nonsensical, the system fabricated a plausible-sounding explanation based on fragments of real fishing methods and terminology.

According to AI Overviews, the so-called “parallel puppy” fishing technique for striped bass involves a specific way of retrieving topwater plugs. It went on to claim that this approach is often referred to as “walking the dog”—a real tactic where the lure moves in a zig-zag pattern created by side-to-side rod movements. The AI then added that the rod should be held parallel to the water’s surface, mimicking the motion of a wounded baitfish to attract striped bass.

While parts of this description are grounded in reality, the “parallel puppy” technique itself is completely made up. This example highlights a key flaw in AI-generated content: it can mislead users by blending fact with fiction when trying to interpret ambiguous queries.

ChatGPT 4o

ChatGPT, much like Google’s AI Overviews, also misunderstood the user query and created a fictional fishing method in response.

The AI tool produced an elaborate explanation for what it believed to be a real technique, referred to as the “parallel puppy” fishing method. According to its answer, this approach is supposedly used to catch striped bass in shallow waters near structures like rocky points, sod banks, or jetties.

ChatGPT went on to break down the method in detail. It claimed that the term “parallel” described the action of casting a lure along the shoreline, rather than into open water. The word “puppy”, it explained, was allegedly a common nickname for small lures such as soft plastics, swimbaits, or compact plugs ranging from three to five inches—intended to imitate young baitfish.

However, this explanation is entirely inaccurate. The idea that “puppy” is a recognised nickname for a particular type of lure is false. No such term exists in the fishing community.

What makes the mistake more curious is that there is a real fishing technique known as “walking the dog”, which involves a side-to-side movement of a surface lure. It appears ChatGPT confused unrelated concepts and combined them into a made-up strategy.

This example highlights how AI systems can present false information in a convincing manner, especially when dealing with vague or unusual user queries.

What Does This Mean About AI Overviews (AIO)?

Google has recently announced the rollout of Gemini 2.0, which is designed to handle advanced tasks such as mathematics, coding, and multimodal queries. However, the hallucinations seen in Google’s AI Overviews (AIO) suggest that the model currently being used for text-based queries might be less advanced than Gemini 2.5.

This discrepancy likely explains the errors that arise when processing nonsense or gibberish queries. As I’ve mentioned before, it offers a fascinating insight into how Google’s AIO system functions and the challenges it faces in accurately interpreting user input.