The Promise of AI Overviews

The premise was simple: a lot of users who log onto Google are looking for specific answers to specific questions. They aren’t looking to get bogged down searching through various blogs and homepages to get that answer.

AI Overviews hoped to provide those answers by combining their Google Gemini language-learning model (previously called “Bard”) with a technology called “Retrieval-Augmented Generation” (RAG). Traditional LLMs are limited to the training data they were provided; a RAG system combines the existing data pool with some quick pings of available online data to (hopefully) provide better answers.

We’ve all known people who ask you for information they could Google for themselves; the idea behind RAG is that the AI bot does the Google search for you and summarizes the answer. Rather than a pure search engine, AI Overviews would bring the technology closer to a true digital assistant.

So, what went wrong?

What Went Wrong With AI Overviews

Soon after the launch, there were hiccups — many of which were connected to issues with the combination of RAG and a language-learning model.

Some examples:

• Searching for pizza recipes prompted the AI Overview to suggest adding glue to the pizza. Adding glue to cheese is an old trick to make for a more-dramatic “cheese pull” in TV commercials, which is likely how Overviews got confused, but we wouldn’t suggest eating it.
• AI Overviews also said experts recommend eating at least one per day, which Overviews pulled from the famous satirical website The Onion. In this case, Overviews likely confused The Onion’s popularity with trustworthiness, and also failed to recognize the joke inherent to the article. Overviews had a particular weak spot for The Onion, presenting several other gags as facts.
• Overviews also extolled the “health benefits of tobacco for tweens,” which should go without saying is not something a doctor would tell you.

There’s a few overlapping issues at play here. Some of it is that the RAG process depends on Google’s algorithm while simultaneously operating at cross-purposes with it. Google searches are often trying to find the most-popular, most-entertaining, most-linked-to sites on the Web, with a particular emphasis on things like the site’s technical safety and adherence to Core Web Vitals. AI Overviews, on the other hand, needs to provide safe, accurate information — which requires an entirely different type of trust and vetting.

Additionally, LLMs don’t really possess a genuine capacity for understanding. They operate based on statistical probabilities, predicting the next word in a sequence based on what’s most likely to come next, the same way the autocomplete on your phone works. Depending on the breadth and depth of the training data, this can lead to “hallucinations” or flat-out lies. Adding in the RAG abilities seems to exacerbate this issue in some cases.

The Danger of Rushing AI Innovation

Generative AI and Language-Learning Models were held up as the “next big thing” in tech, so there was a lot of incentive for companies operating in the space to be the first-to-market or, failing that, to at least avoid being left behind in the ensuing gold rush. But the rollout of AI Overviews highlights the dangers inherent in rushing AI technologies into widespread use before the kinks have been worked out.

After all, the issues encountered with AI Overviews aren’t just about minor inaccuracies — the wrong information at the wrong time could hurt or kill someone. AI apps designed to identify plants, for example, had a worrying tendency to label fatally-poisonous mushrooms as delicious and harmless. On a longer timeline, AI Overviews could advance a lot of misinformation about basic facts, current events, and history, which will damage user faith in the technology and potentially damage society itself.

Google has acknowledged the issues and has been gradually introducing technical fixes to improve the results, including blocking some of the obvious satire from RAG results and creating newer detection mechanisms.

But the biggest problem may ultimately be that, while AI is a powerful tool, it’s important to leverage that power towards the sort of tasks that AI is good for.

Tasks AI Is Well-Suited For (vs. Tasks That Are Too Much of a Stretch)

At Insticator, we use Artificial Intelligence to enhance productivity, streamline processes, and generate valuable insights for our clients. But the effectiveness of AI varies wildly depending on what you’re having it do. Even as we like to stay on the forefront of adtech, we pride ourselves at understanding the places where AI excels versus where it falls short.

Tasks AI Is Well-Suited For

• Data Analysis and Pattern Recognition: AI shines in analyzing large datasets and identifying patterns that might elude human observers. Algorithms can quickly process and analyze complex data sets, making them invaluable in areas like market analysis, and customer behavior insights. Adtech is a data-intensive industry, and AI helps us move quickly in response to changing market conditions.
• Automating Routine Tasks: The ultimate goal of technology is to help us do things that we would have previously had to do by-hand. Whether it’s launching thousands of personalized creatives using Dynamic Creative Optimization technology, or combing through thousands of comments on COOL Comments every day to help flag problematic content for our human moderation team, repetitive and labor-intensive tasks are where AI shines.
• Personalization and Recommendations: Things like your Spotify Discover Weekly, ads run via Dynamic Creative Optimization, and the Amazon Suggested Products are all great examples of AI being able to curate personalized recommendations based on past user data and behaviors. It’s just pattern recognition, but when done well it’s something users really enjoy.
• Image and Speech Recognition: All those years clicking on which pictures contain a crosswalk are starting to pay off: AI has achieved significant success in recognizing and interpreting visual and auditory inputs.
• Predictive Analytics: Adtech runs on analytics, and so do many other industries. By analyzing historical data and identifying trends, AI can make accurate predictions about future events.

Tasks That Are (Currently) Too Much of a Stretch

• Understanding Nuance: Google has successfully used context to solve a lot of problems of nuance: if you search for “hot dogs” you don’t get a bunch of results about overheated canines. But AI can’t recognize satire, and tasks that involve sarcasm, humor, or complex emotional states remain challenging for AI. Famously, Commander Data on Star Trek didn’t understand jokes, and that remains largely accurate.
• Making Ethical Decisions: AI lacks the capacity for ethical reasoning and moral judgment. Decisions that involve weighing ethical considerations, like in healthcare or legal scenarios, often require a level of human insight and empathy that AI currently cannot replicate.
• Creative Problem-Solving: Because AI hinges on scanning existing knowledge and providing a statistically-likely answer based off of it, it’s not really capable of what we’d call “innovation.” New ideas, artistic expression, and unique solutions to complex problems still require human ingenuity and creativity.
• Unstructured or Ambiguous Data: Though we may be sluggish before we’ve had our coffee, human beings are still better at reasoning out unstructured, ambiguous data and situations. For instance, interpreting the meaning of open-ended survey responses, or understanding diverse and unstructured customer reviews can be tough for AI systems.
• Handling Unpredictable Situations: People like surprises, but computers don’t. AI systems are typically designed to work within specific parameters and may struggle in highly unpredictable or dynamic environments.

Understanding these distinctions helps in setting realistic expectations for AI applications and ensures that the technology is applied where it can deliver the most value. Some things you should absolutely let AI do for you; others you should definitely still be doing yourself.

The Way Forward