'Put glue on your pizza' embodies everything wrong with AI search — is SearchGPT ready to change that?

The role of artificial intelligence (AI) in influencing how we use the web looks set to increase inexorably, especially with OpenAI — the company behind ChatGPT — teasing SearchGPT . This is an AI-powered search tool designed to serve up direct answers to your queries rather than pages of ‘optimized’ results.

If you’re experiencing a sudden burst of déjà vu, that’s because Google has already tried something similar. Using its Gemini AI model, Google trialed its "AI Overviews" tool which, like SearchGPT, is designed to scour the web and provide summarized answers to search queries. The simple idea was that this tool would give you a summary of the core information you wanted without needing you to pursue a load of search results.

Only it didn’t really work — at least at first. In some egregious examples, Google's AI told users to add glue to their pizza sauce to give it "more tackiness," suggested washing clothes with the toxic gas chlorine, and even noted that a solution to feeling depressed would be jumping off the Golden Gate Bridge . The issue here was that while AI Overviews could pull information from a mass of sources, it appeared to be no good at separating satirical, incorrect or malicious information from useful and correct information.

SearchGPT is underpinned by ChatGPT, which is arguably a more mature AI model than Gemini, and so could yield better results with less heinous answers. However, the tool is at a prototype stage so nobody knows how it will perform when released to the public.

Related: Researchers gave AI an 'inner monologue' and it massively improved its performance

But it does raise the question of how effective the role of AI will be in the future — if finessed, is there potential for AI to kill off traditional search engines, or will the accuracy of AI search remain a let down?

Robust not rampant

"Current AI has a lot of inconsistencies because it isn't very cohesive. The thinking patterns can go in strange, ditsy directions. However, research shows that it's possible to design models to think much more effectively,” Nell Watson , an AI researcher at the Institute of Electrical and Electronics Engineers (IEEE), told Live Science. Some models can be married with logical programming languages such as Prolog to greatly increase their reasoning capabilities, she said, meaning that mathematical processes can be trustworthy.

"It also helps models to be a lot more agentic — to understand a situation and to form plans and take independent action in response. However, without such scaffolding in place, AI systems will be extremely limited in their ability to provide accurate and trustworthy information, and to retain sufficient focus on a desired context," said Watson.

Therein lies the rub of AI and search — the potential lack of any robust framework behind these systems to ensure accuracy and trustworthiness. And it would appear that the desire to strike quickly, while AI interest is blooming, could be the crux of the shoddy results they spit out. Watson said: "It is clear that some AI features were rolled out far too early without adequate testing." Beyond this, they lack user context.

But such issues don’t just start and stop with AI models. Blame can also be attributed to the state of web searching via some of the biggest search engines, notably Google Search.

"Beyond the AI system not being helpful, are the broader issues with Search itself these days, with changes made to facilitate paid search results making it far more difficult to find content," said Watson. "That's aside from the issue of AI infrastructure bias to prevent 'undesirable' content from rising to the top, which again does not fundamentally respect the desires of users. It is important to remember this is a feature designed for customer use, and resolving these issues will only further their search engine optimisation experience."

Agents of accuracy and trust

In that case, what does the future hold? Watson noted that the current state of AI search is hinged on agentic models — autonomous AI models that are designed to carry out defined actions and solve problems without constant human oversight in a goal-oriented manner. This is different from generative AI models that create content. Plus, these agentic models will only grow in sophistication.

"Agentic AI systems will be used to go off on a mission to perform a deep search and analysis of far greater sophistication than a simple keyword search. They can find answers for questions that users didn't even know how to ask," explained Watson, although she added that such AIs will need to understand human values, boundaries and essential context.

“We are putting responsibility for aligning these models in the hands of everyday citizens, which is a disaster waiting to happen. A great deal of public education is needed to ensure that we get the best out of this next wave of AI, instead of AI running circles around us."

— New supercomputing network could lead to AGI, scientists hope, with 1st node coming online within weeks

— GPT-4 has passed the Turing test, researchers claim

— Claude 3 Opus has stunned AI researchers with its intellect and 'self-awareness' — does this mean it can think for itself?

While concerns over the effectiveness and accuracy of AI systems in search raise questions, there’s a lot of potential to shake-up how we find information on the web — or at least offer an alternative to classic search engines.

"As the ‘agenticness’ of AI systems increases, AI Agents will likely one day act as our ambassadors, actively seeking out products, services and experiences which may surprise and delight us, and dovetail with our existing plans," said Watson. "This will transcend the primitive search-optimized market, by users not even needing to search for the product in order to sell it to them. It also means that marketing to bots may be of more value than marketing to humans. Moreover, there is evidence that AI systems find content written by other AI systems more stimulating," added Watson.

With this more positive outlook, however, comes a caveat — and it’s one of trust, as Watson concluded: "Pushing too many products to AI consumers runs the risk of diminishing trust and causing frustration. Future successors must seek to maintain the trust of their customers."